<%BANNER%>

Comparing Poisson, Hurdle, and Zip Model Fit under Varying Degrees of Skew and Zero-Inflation

xml version 1.0 encoding UTF-8
REPORT xmlns http:www.fcla.edudlsmddaitss xmlns:xsi http:www.w3.org2001XMLSchema-instance xsi:schemaLocation http:www.fcla.edudlsmddaitssdaitssReport.xsd
INGEST IEID E20101117_AAAABO INGEST_TIME 2010-11-17T18:26:16Z PACKAGE UFE0018500_00001
AGREEMENT_INFO ACCOUNT UF PROJECT UFDC
FILES
FILE SIZE 8765 DFID F20101117_AABIRC ORIGIN DEPOSITOR PATH miller_j_Page_185thm.jpg GLOBAL false PRESERVATION BIT MESSAGE_DIGEST ALGORITHM MD5
295b24155a444b470d08ce0e763a614e
SHA-1
a01f38ea9386815f965a58df482a3c0f0ca10793
2550 F20101117_AABIQO miller_j_Page_171thm.jpg
7f3a8e0ea85f0a4fbffdb71dc9423ebb
3fdd833fa0c6f3fc7528d0c8ab2aef7752538227
30082 F20101117_AABIPZ miller_j_Page_151.QC.jpg
9fbee4a5d658f04467a25ac36f71a3bb
c1aa7ea71f502cadc35e48fe59ffc465b4ce6daf
1051982 F20101117_AABHOA miller_j_Page_179.jp2
5bee1f915ae3c9cf1033c4e91af7a880
582973aa8afc583788bbb685b7abc10a12fe5af6
192337 F20101117_AABHNM miller_j_Page_161.jp2
d00eb6df3a12b3f93f3e3cee943480d3
d16a0e772a4ca61656c513ee3dbaf750a0dd6feb
813277 F20101117_AABHMX miller_j_Page_143.jp2
ef76f8d3cd55c42b2c7918a5bbd46443
d6bd1db3af7d8d2521f5b869f5dcfcaaa9878e0f
31576 F20101117_AABIRD miller_j_Page_186.QC.jpg
e6e30b0f4ac74b19f35122a4fdbb342c
095ed71b456ba5ba173fbd29f6bc09a266f82b81
8021 F20101117_AABIQP miller_j_Page_172.QC.jpg
f87145a0c75f8c02ba382fa3b78883f6
bc84cf154929f1c983eccb4277ba5efe6a21c188
1051959 F20101117_AABHOB miller_j_Page_180.jp2
48f117c043e1f0afe72b51b0fd43f554
f3c24774d84ac7ef2619096b7c5d025e3c789d2c
182210 F20101117_AABHNN miller_j_Page_162.jp2
a1290a626c00aa3b56d54bd2308921e3
8fbbd006a132e65f88d9ca3b0d4f5eee67a0e568
831036 F20101117_AABHMY miller_j_Page_144.jp2
ed10428805a419693bf5d4683a947546
ddebde20a94c499d2d3fcd8114a10ad9786767bb
34524 F20101117_AABIRE miller_j_Page_187.QC.jpg
46236d3e8ce504100f0f49b27066651e
9a9b99410a3b851e6c641deacbdddb703b836a6f
12518 F20101117_AABIQQ miller_j_Page_173.QC.jpg
2809cb4145e130409b2ac4ced1e9cfc9
54d448023923e54c673fada21ce22e2020279975
1051956 F20101117_AABHOC miller_j_Page_182.jp2
84287b0c12925d1b1607c94ced14b740
27fd245bd6d9ec1d4494dc740b50889a00cb51fa
178998 F20101117_AABHNO miller_j_Page_163.jp2
de4aa8443c98ad981ccc027375b67a94
30ea3d187d58f362d9b1bfe6160974ce40450394
799350 F20101117_AABHMZ miller_j_Page_145.jp2
a1502edeed78b4a0d84947e44901b1cf
14e23e3f7527bfcdfae9c2aba88b22d6331350a0
37082 F20101117_AABIRF miller_j_Page_188.QC.jpg
ae5cd932ed49631d9d0742fd016d0cf5
763d86d6a08e36be7318fe4588587653fa8732b8
3843 F20101117_AABIQR miller_j_Page_173thm.jpg
69445e31aa1bd211572fbf84b76a33cf
9d848077e96602d60136513ab4b6d169ce106f1a
1051967 F20101117_AABHOD miller_j_Page_183.jp2
a6e6d473277addf007f468484b40457f
3f58985bfc78cf3eede9b2fa30d240b3c1363728
164429 F20101117_AABHNP miller_j_Page_165.jp2
c40e1262d692f5203e4d3abf09a920b9
2f6d3a1a72234146d7e51f15d1f8f73a3ee37c1b
9024 F20101117_AABIRG miller_j_Page_188thm.jpg
3b06b5fe944e026b250ce22d78b72436
7ceaad8c87bd2acf49525e5d3cb6f80d599c3477
10634 F20101117_AABIQS miller_j_Page_174.QC.jpg
3c077c696fc3876018788cc6394eee04
a600b922d66d6da26850412c7f62f5e28d0b029b
F20101117_AABHOE miller_j_Page_184.jp2
85867f73813fa05e279a14ed24153bdc
cae832c34448df433a63bb774268e54b0fd298a8
160812 F20101117_AABHNQ miller_j_Page_167.jp2
877757301d2cfed4a4e60b986e3f2515
a38aa5fd93d330aaec95a9654563a6bb2909c03f
34037 F20101117_AABIRH miller_j_Page_189.QC.jpg
521a35b1a6939c842fd783518d28c659
d415f73acdc684c505a2a80b564c8c8357a8d008
1051963 F20101117_AABHOF miller_j_Page_185.jp2
4106b459314c71151df3cb88835fe69e
0d1acc21a537daf717400fff1515b927393cf872
36592 F20101117_AABIRI miller_j_Page_190.QC.jpg
e82ff847d2bfcb47f8e4ead9ed787c0f
9db9e65ff8050e637d62efca2c5d8c30beae028d
12263 F20101117_AABIQT miller_j_Page_175.QC.jpg
1038993a75befce59eda17e26390e6de
c9b4a16a5ebb05f1f60e6229b63216f917faa891
F20101117_AABHOG miller_j_Page_187.jp2
d1580f8ee6e985dfd63cef74568276e8
6329340829c2fdc5d7de2bd9a0261dbaea9a38c1
153131 F20101117_AABHNR miller_j_Page_168.jp2
7827236bac046909e40c6ea5f122689b
2c0b7bd38a2ec43d1a40d8d479ce98c42bdd7b60
8815 F20101117_AABIRJ miller_j_Page_190thm.jpg
8c03572d72b32afab5de4535a9562df4
f6b565e27a1d431fb773dfb978eafa4618df240c
4057 F20101117_AABIQU miller_j_Page_175thm.jpg
42fbba9271e33276a38397f882c0f2fa
4860d30c0d78eea121a7186312053bbe263d6726
1051950 F20101117_AABHOH miller_j_Page_188.jp2
73bf682f183e55106d8290eb21e8930a
fd05cb276c15da90c0deb8aa715f0a1496dd5102
151971 F20101117_AABHNS miller_j_Page_169.jp2
bc495c76adf9a26ef5a1416e99c08a46
c3343f2ce9680d1b7a0dce1322cc0f9fa4e1e2a1
27139 F20101117_AABIRK miller_j_Page_191.QC.jpg
1fcb6e16ad2576da37e4276695f6f5c0
09592ce279d6845b4d9641fe30b6ff9b6af0ddcb
31467 F20101117_AABIQV miller_j_Page_176.QC.jpg
86222b0c0a540303918d700a8161e2a5
9c5ce85a3d6511629ffc3e66ef291aec4014eb30
1051965 F20101117_AABHOI miller_j_Page_189.jp2
26a36db9b080542fdd9813547dc3abad
987c3d039493543e7479d38e6baca56c501f55ae
151620 F20101117_AABHNT miller_j_Page_170.jp2
d63011bbed4fc52e1751f6c26ff1aeb7
2207c319935a952eb23aeb5c7813cec444f82298
8557 F20101117_AABIRL miller_j_Page_192.QC.jpg
cdc9f679659ec31b77c29c28864ddf4d
49f1eaf116a3fe142bb6ce54ef5b79d212705dfb
8927 F20101117_AABIQW miller_j_Page_177thm.jpg
dae7fb8b33082493437f9ff9be6bf8c7
7ce05e5458de50987ca70fc8ecf8082ab6f061c3
1051914 F20101117_AABHOJ miller_j_Page_190.jp2
0cc63dc91a0881c0298366716ff25ec2
a9008ad00fe8be52a6de9407c91ed8288b18a413
155618 F20101117_AABHNU miller_j_Page_171.jp2
89dccbae70bfbcb1ecc06b145786dd83
a0d6ac1fd349d063d737a3ae01272e712c4d8cfb
9061 F20101117_AABIRM miller_j_Page_193.QC.jpg
1e3f086e4bdb50cfbf67efa396fb9ae8
b919b0c256cb15016bdba33f7030035d20bda202
35728 F20101117_AABIQX miller_j_Page_178.QC.jpg
6861e80bb3e876cbbdb5c65d226bbf27
5e723bcf3630e61c379cd1d86b259872f43fa3f2
905143 F20101117_AABHOK miller_j_Page_191.jp2
d81e36b6dbdab82429a09c9ff9a4f663
d73a607bc3c06090a411985614c02c881509cc85
257497 F20101117_AABHNV miller_j_Page_173.jp2
9444eec89ff07f789e7c6be4c5768d67
a2a3523115ce6a377c37848a139bb2650d7ca3d6
3159 F20101117_AABIRN miller_j_Page_193thm.jpg
f2b7fe560b9884e8ab57b5aa43ca87fd
b57ae8da284b9aa015b44c5bcf809fddb46cd4e6
8798 F20101117_AABIQY miller_j_Page_180thm.jpg
54285c79b3e4a1774487cde43c4d1787
575d48f8b6bbe607012afcd31fdf0c6839ff55e1
21459 F20101117_AABHOL miller_j_Page_193.jp2
d66d10d83b90eeaaa76d3fea9dc64cd0
c6f37cb11cc74cfdd454c25632c5d915cbff42dd
207363 F20101117_AABHNW miller_j_Page_174.jp2
96fe12ab8e5c575d189af0378d1ed81f
8f2d6d8ed30ee63da71eeeac37a9f1c368a8249a
3162 F20101117_AABIRO miller_j_Page_194thm.jpg
0d338925d99e7d7bab3142f4612bc6df
994587c163bae2fb9f04ec71607b7115df30ef6b
9011 F20101117_AABIQZ miller_j_Page_181thm.jpg
086d198ae7d576afc3ea80fcc012cd56
0ce062565f1c86ef746e616b1ff620ef8d3c447b
1051915 F20101117_AABHOM miller_j_Page_195.jp2
657651fc75687a488572c91f425aade7
945d8cda4a1705804e7b769541fddb8c00d61584
248687 F20101117_AABHNX miller_j_Page_175.jp2
3ebc254d269adb7d2a864b7bae9c8bf3
63365644b9a13b955cdd16f02ffa837779e36795
25271604 F20101117_AABHPA miller_j_Page_011.tif
4e38f1f396f8aed375b069523993d888
6594882332b512473686278cfd0eb6d388b66381
32351 F20101117_AABIRP miller_j_Page_195.QC.jpg
5a4462e8cf00668753fb729511466890
9df06d47d20028a485ae0ba8adf5935d51d56855
F20101117_AABHON miller_j_Page_196.jp2
18400246d19f100129c86f1b2cf30928
39ec43ddd2b9d534944b3cea9fb5eadb5f19034e
1051892 F20101117_AABHNY miller_j_Page_177.jp2
c263c49b0e1b4118bc56feefbf0dccf7
12f1e86b5db6711d2bac70d260578bbe7193fc78
F20101117_AABHPB miller_j_Page_013.tif
a639fdb0325d67ef327c7efe46578302
19ddddc3c6079497c4a068e73b4816b9b6360f48
8445 F20101117_AABIRQ miller_j_Page_195thm.jpg
c6cfb50073217412bda3d1d5931cd6ca
5a43c3cbd7ace4984439f40367688782354902ef
1051980 F20101117_AABHOO miller_j_Page_197.jp2
6e07043eaf99c52722ef7a629a48ad0d
f083064fe3516a8fffef17f0a578fb6535927f02
1051986 F20101117_AABHNZ miller_j_Page_178.jp2
c225512730bf8ece231cd211597db5b9
8e4d40a13f51f3c5b01dbddc4d124eeb07c94316
F20101117_AABHPC miller_j_Page_016.tif
89b056fd82db8c6cd362401b60b53605
5e882e947b679c570de2442a90a82aa9bfe13a77
36580 F20101117_AABIRR miller_j_Page_196.QC.jpg
bb6ed4f58b02a7de324f97b01f6b294d
fc097d5af471f3894ccb9bf0909f45862483c98c
1051983 F20101117_AABHOP miller_j_Page_198.jp2
730f9da6384ed0b673b55c433633cd49
1c10b862716efa555dd538c62a5bd56388a02967
F20101117_AABHPD miller_j_Page_017.tif
5f0a7bb985319ce882a3cbf5e0f8a33d
724a6a7c9805bc472d0779c080d6c2c48079bf4b
9070 F20101117_AABIRS miller_j_Page_196thm.jpg
980dc08bf53648ad237e9ee05b9bf486
fd582296dd312a09b48594bae7b3c0cda47b8bbc
F20101117_AABHPE miller_j_Page_018.tif
bd4464e47ccb507a37777c2482da56a1
10e9747a35594f9f27b9bd727f69688c96b07ce5
197427 F20101117_AABHOQ miller_j_Page_200.jp2
bc15de503de6bc42938ae58a5c6ed8f9
5aa7505d3115c3e1273ba541fe6947167d7ae031
31975 F20101117_AABIRT miller_j_Page_197.QC.jpg
247c44f2da0f69a2ff8271863fac1f4b
999736df147c266d7c4170693349e9a730493bf9
F20101117_AABHPF miller_j_Page_019.tif
fa1d61c4c6e8309e0e26247d8d695c17
346739ac144f92dd2e0911bc44ca96177c5c8e1c
452719 F20101117_AABHOR miller_j_Page_201.jp2
4474fc88a050f3b8a6dfe1b14beab459
b49b30b9aa5ed1314ae5c7949e4af6e71e2e5ef5
F20101117_AABHPG miller_j_Page_020.tif
8f2313091e86ad09391cec180d433f82
bbafb29e2eadb29bb19e321a37ea31f5c3a5123f
8388 F20101117_AABIRU miller_j_Page_197thm.jpg
1f8bfe0354a62e4a9cfe89c63f992693
6e97e84968affb09c9346054e7e4dd64047e9b40
F20101117_AABHPH miller_j_Page_021.tif
b8dd3f43eb0e4febc743b500ca70c4df
ea3e2fe6c341b912e3f207d77406fc2108526c64
F20101117_AABHOS miller_j_Page_001.tif
de87f83f59c87aa610dcfad37f463980
7d9462cbb71d5b43d764d19a0db222118eb1dacf
34131 F20101117_AABIRV miller_j_Page_198.QC.jpg
48d85593b1785413498ff31c1676a3af
fd462619dfd6d30d7153f4e4db10761be11dac38
F20101117_AABHPI miller_j_Page_023.tif
1cc66302165991149c3a532e6ab7a83f
ffcb9a4e617c95f9a46d99aa71cf5a1a54896c5c
F20101117_AABHOT miller_j_Page_002.tif
b9b32aa1b24b3f05116e1649b65377dc
a2830a01f00d40adc08827603da1c400d527c6a4
9074 F20101117_AABIRW miller_j_Page_198thm.jpg
8ae7f0ff0b94c303944406e9917ed832
47e7e62c896137b5fe4eab97a27426ef74218bc8
F20101117_AABHPJ miller_j_Page_024.tif
52fba4c50b1e0a90c958642043bc8787
41d6596450ea9640539cf3b4928e8373324f3a56
F20101117_AABHOU miller_j_Page_004.tif
fabf174f842d82c5af954da6a66464fb
74c2b49e574b68903f465457ef17754ba6f7f2dd
34265 F20101117_AABIRX miller_j_Page_199.QC.jpg
aca51d4e192ef3890c4c51287f8fb05b
842bc62cba04163e666feff423b630d72e8ee224
F20101117_AABHPK miller_j_Page_025.tif
d23a8c63f86a6e02cc7f3f061a5cb316
15ecd49dc1eee15f77f04f6dbb6f275a34f47ea5
F20101117_AABHOV miller_j_Page_005.tif
7d5309587ae87477b5a983e02e7ece3e
8cdcfb726e399eb50c12125e612a3583b800eca2
1540 F20101117_AABIRY miller_j_Page_200thm.jpg
cffa3d485ce78240f9e3cbf68ba5612a
cbe4a10e4d47d5db612b276414b91e1a96d7908f
F20101117_AABHPL miller_j_Page_026.tif
f2497165e678fdd048348578e6b6cf79
5f1360285323c7998120b4008da92ae2bc6aae36
F20101117_AABHOW miller_j_Page_007.tif
eb83a3924c33ea3826371956419428a9
720557e94d431c3ed5e5bdaaa9706f831d31bf5e
13990 F20101117_AABIRZ miller_j_Page_201.QC.jpg
928013d03f085df662d7286cdaec78c8
23bfff9d9e132e00c66115552783bba95b997a8c
F20101117_AABHQA miller_j_Page_047.tif
ee036fdd49a2942cf32363f91a6e2ddc
574eebc63b73e26b43bb7aa0f7d166a7508b039e
F20101117_AABHPM miller_j_Page_027.tif
2fc006f25e21d7f80820d0340e4b5214
9dfa2b72457bf1ccdbcd49c3712f336207232e95
F20101117_AABHOX miller_j_Page_008.tif
ed23ac99101dec7a4121cf8ea8b0ac9c
3c36475d1c77290798ff382c52824fbd7ffe1824
F20101117_AABHQB miller_j_Page_048.tif
7ff8427c5c2f75e66a891e2591f76326
1b53392c7c816a8cb24fd5271f06123e10692ba9
F20101117_AABHPN miller_j_Page_028.tif
dbbd97aadc43d132015064d79854b62a
bab188d62c319e0231251062a353fdb9cc7cbb10
F20101117_AABHOY miller_j_Page_009.tif
391426236e583823f09243cd7546633f
a8eebe4d95cf34befc1cd81052855ffee10ee78d
F20101117_AABHQC miller_j_Page_049.tif
e868e538174c99db9727818399e7fb60
dbb6c8a0d56088e7eb7c1e7937ced16f26510612
F20101117_AABHPO miller_j_Page_029.tif
f710913aabed9fa96dd6a736747e9794
75401c120f049eb6ea8595e8c36bd69bf124e0d5
F20101117_AABHOZ miller_j_Page_010.tif
46d4c51b82afb0fe4cb8581a80d61be7
659de88ca97513c9f4f7893b43f5f47b6d22a9b4
F20101117_AABHQD miller_j_Page_051.tif
0b70c0e28a1aec37067300aa3f71989b
23867364a6f644e600821b214b6623db2e362321
F20101117_AABHPP miller_j_Page_031.tif
cff64b4bcf90c2aab33eabd8ced002c6
cfa31225ce39072fdc8841c62a5ccfd4a21e2dd1
F20101117_AABHQE miller_j_Page_053.tif
6024bab24e0c97a0b07037fe9d4100ab
0d73ee8320979f1b2bf15cb502defdeb4bcf21da
F20101117_AABHPQ miller_j_Page_033.tif
6b602c1ad80e324d8886358e945978d0
8c1492ca7dbc591a949d93bbd6ab90e1af51952c
F20101117_AABHQF miller_j_Page_054.tif
da4b285df8c915b17b8e8bf80bfc424b
bc53e93a36d54195c64b0f88a7a8a771d4a38a9e
F20101117_AABHPR miller_j_Page_034.tif
d937dd7ff4ad3d96ea283a693f2d5d6a
435d430ddeaf2acee89712b7e8edebdce91f9040
F20101117_AABHQG miller_j_Page_055.tif
723163247e80838b59e497f77277324c
1e88f2696ac159674a8ac369e8ea72aafafca995
F20101117_AABHPS miller_j_Page_036.tif
b11e4824c23882c47b14159aba51f79f
23cb092dfe8f9783e6f80dc3ee341d3f70c5d5f7
F20101117_AABHQH miller_j_Page_056.tif
c8996583b83b47b509706509749a247c
07a4958abb810b101f3b52faab6c8b9541ceffd1
F20101117_AABHQI miller_j_Page_057.tif
05c33643693d64ce1a913bd3bbc50d79
336be3ef4eb993e42495e52e4eb5268462151e49
F20101117_AABHPT miller_j_Page_038.tif
96e8b977de47cd3e11f0d4dc55f1dc98
8a1616464840d5edb524eead9911650a7687f80e
F20101117_AABHQJ miller_j_Page_058.tif
a79c1c45bfac80ade5520e2b7853b655
166e7a15b18e5bb37821a7085fe39504d5ba06b5
F20101117_AABHPU miller_j_Page_039.tif
0c68a7da44e0e1f82250b88924994f4d
13c879e7b5b6becef6a66b183aba5ba85567b090
F20101117_AABHQK miller_j_Page_059.tif
fea0732531a96fed7b9ed982f64ad915
60840bf1e67b9fd2f20dd4f9ea247cc30e99af58
F20101117_AABHPV miller_j_Page_041.tif
fef6bba459aa07f41c39ceebac416b8a
cd1c2f67263c1d2ad8bd7f8a58c5fb5cb69c92d1
F20101117_AABHQL miller_j_Page_061.tif
3e2816bb85251617c3da58736e6d3e0e
72917f92243e92a3456f5f1eb259339656326a74
F20101117_AABHPW miller_j_Page_042.tif
a6c68bd12502dc558ca52dacc1c0b96d
3b06c689a59224ac842f7699961136187f29a7e4
F20101117_AABHQM miller_j_Page_062.tif
f25d37ff5ff1742b04da0940632d5211
ea98bf5eb9cc01143738beb57ec67864345d3ecb
F20101117_AABHPX miller_j_Page_043.tif
d8f7f3e062afcd508c53f1c11fa1f0d3
2b4b581474fe4e2f62e4d8c009202b90a9ce5142
F20101117_AABHRA miller_j_Page_079.tif
44daf03889bf99fe0ab0680e999318a8
d7a206701c05b1f855cbedcf51d80c947a0e09ce
F20101117_AABHQN miller_j_Page_065.tif
6fd56493ffd4e77fd9d236f863b28c3b
e7e403846e5ae8a8de3918aa4370b76877089092
F20101117_AABHPY miller_j_Page_045.tif
31a0e541511b6d3942a75a685bfd2326
a92812fbc675084d18f4632f8ad340832e8b4a98
F20101117_AABHRB miller_j_Page_081.tif
42fb5f02329e0e89cbebb27379b38ae0
930979fb418463e089474d8349a0dabaf483e534
F20101117_AABHQO miller_j_Page_066.tif
76fd037fbe98cc7a968b9c6e1a2bea87
5e06b793dd2ab077eceb44abac99d8105c18e069
F20101117_AABHPZ miller_j_Page_046.tif
a210e7c7c3595f92301eb3160dc2b0e1
dfca79e4d7a91df6261269440292193c3e633c27
F20101117_AABHRC miller_j_Page_083.tif
7e23e54c8df475650d7abae39bceae48
10b3975463a3b75875ac4d2b0edd5417e234199f
F20101117_AABHQP miller_j_Page_068.tif
7de9203c40ecf43c499f8ab1a17e89ce
90deb3b534e8ae2e3c191f18cbc8aaf0de7d7194
F20101117_AABHRD miller_j_Page_084.tif
386058255379491f052190169b30236c
3d2298273dd4de3d35ef973bdc9c9fcd9d09f42f
F20101117_AABHQQ miller_j_Page_069.tif
3b6c116c9725d9579d9da2f49bec3af2
c29fce6364a4f577d96d802f62d447aa32e4d20c
F20101117_AABHRE miller_j_Page_086.tif
7a98ecc56dd6692e5b15dda2c599b1a0
dbf5be5a03e45edb5137bc0c7f4fc0716ff1e79f
F20101117_AABHQR miller_j_Page_070.tif
b00bcf03b2c2480cc1662500d084f1fe
c00f12f2459cddd32334b4ea9b402a05a4b5b825
F20101117_AABHRF miller_j_Page_087.tif
afc56480f68bc2660bba3fe9825da2e2
797f1436a81ff0e605a7b0444955a685c6d699e3
F20101117_AABHQS miller_j_Page_071.tif
586e291bd03873a6505847eb4154a01c
e0c7c8594b10e0f4aec782b0076900a41e114d87
F20101117_AABHRG miller_j_Page_088.tif
3e47b727def9d5b2abd5199538dba0ed
0a4b97170efbe930ec64fe75d0961c2145cd32e3
F20101117_AABHQT miller_j_Page_072.tif
f63c390880247eaf7c7b7ed792bd0925
47bebafb6ef14a4047fad338b0cc1db1ff44b985
F20101117_AABHRH miller_j_Page_089.tif
b0e9e492d4553cb95fe3d722524e3b7b
f7968583632475965ef8b0f2b2cbb76b19e2b95d
F20101117_AABHRI miller_j_Page_091.tif
4821a2c2b0dac88402f7b24e47bf668b
d34dd2600c9b1c14010a31c781204734d4160355
F20101117_AABHQU miller_j_Page_073.tif
8da0df341fefab3aaa206766e480d515
686e886d9faf21e51deda4d1ff5d947a70426aed
F20101117_AABHRJ miller_j_Page_092.tif
e51343dd623062fe6c394ffba8b51ba7
9d23f66fa209df47a287dd06c6e8a6dc9a925ef9
F20101117_AABHQV miller_j_Page_074.tif
2ab8d9a10feae62df92a1c633469740b
75bb23a6367e6562bd2172d30da4437e14780b4d
F20101117_AABHRK miller_j_Page_094.tif
d96387f63826776572b76c176c56bbf9
beb9f3813f0e940333a4a6e147fda76d86f58f75
F20101117_AABHQW miller_j_Page_075.tif
c261b45f8add8617a6fe8a7c8826ba6e
c7b9b7117c5aaaf1a792ccd3c9b052f4ea88794b
F20101117_AABHRL miller_j_Page_095.tif
823c462225dc72c38164b99b6c6e2c9d
0401c960c420dd348092e6b3fa746bb15c192d73
F20101117_AABHQX miller_j_Page_076.tif
56d0b2410d244f7663873fc73bda4041
d9626f7b48e5edc30082e4df8d1e26b6d4b74a6d
F20101117_AABHSA miller_j_Page_121.tif
5ca99f243e23769ab13562c0c00cf1ac
01410ef39306d3f10398220c7b7667d2758cb174
F20101117_AABHRM miller_j_Page_096.tif
56846661414a08eb15d38bb45d0022b3
0ef1ea9b62a1a45bd801b1e8a313138cfeecc352
F20101117_AABHQY miller_j_Page_077.tif
f79a88a1126279cf37a65c71338dd47d
f6beb14293327de2202f15050b5e02158c7a23bb
F20101117_AABHSB miller_j_Page_122.tif
b8a56d50df81dc22f654f5de755e654c
12415334cdd44639006d35807976a9668f12031f
F20101117_AABHRN miller_j_Page_097.tif
f6db42a450663e3e9d88698c6af206af
2d96e9fcdfb1b5bbe098d0f6f123b42f219810ac
F20101117_AABHQZ miller_j_Page_078.tif
f17669a1d4b7a11112c118fc49d230db
8abbcd20a9ab05e430c48c4009ec8dc873a70efe
F20101117_AABHSC miller_j_Page_124.tif
d87d4af449000a8cb1cad1585e7e057f
3371bc324b28ca3d5bd018e2bfd2764ba6d66265
F20101117_AABHRO miller_j_Page_099.tif
6920091b2e8892be6bcc312cf331d7bf
43b57ced7453de6b3a97aaaad94923eb5a8a884c
F20101117_AABHSD miller_j_Page_125.tif
d63bc1cdcd2262049b9f29cf75c1c880
ae580cb6df6ca897a885837db1f3b95fc0768408
F20101117_AABHRP miller_j_Page_101.tif
84ba9f5b33509c6e57db24f24c1bd2aa
a59b891a8d5e1b52210356efeaa9c52ad4d63a8d
F20101117_AABHSE miller_j_Page_129.tif
5322edf82914b994c5e45fdd8dd9568c
af359b5b7047e6ab3a8cc002387a7b0a1d74a80a
F20101117_AABHRQ miller_j_Page_102.tif
9ad8e8e5afa19d1541147e9e5734ebf6
e07f1576503d5bae36141c7019ebc20236503fac
F20101117_AABHSF miller_j_Page_130.tif
d9ad77376bdf3622a632f77f167a4707
76adc29c0992f5ba98435f14cbedb2497191f936
F20101117_AABHRR miller_j_Page_103.tif
4aef98e979cbe5db7bea7547ced81935
e9d58bfb057b4234a6dedc6ef2260b1a784f4b10
F20101117_AABHSG miller_j_Page_131.tif
d18d0ae578bf99f0459c5d30b3acc5ff
3677bfd54ce4fdf13d65d37c74d0368f947edb63
F20101117_AABHRS miller_j_Page_104.tif
b2db1150648e965894bd4d668800fea7
e829ddf03d878f29aea6bc9b65e696aa0a0c83e7
F20101117_AABHSH miller_j_Page_132.tif
1d8a9ccfe1647a6ba0906801ea86501d
81c608b6e5c023ca40596957c0dd90425b9f2d8b
F20101117_AABHRT miller_j_Page_105.tif
df9b6cb58f3d24f0f8d86cdd0999049a
7778bf524a17aa6b3c0dc4793919b2254a5da9ee
F20101117_AABHSI miller_j_Page_133.tif
3ef47cfc479e9578e184d61b23a870c3
8615bd7643fa8f4f4267bd71a3e1352f46895d68
F20101117_AABHRU miller_j_Page_107.tif
eeca3cc329cc0e7c7ae69ea58702f204
afa4af12122058250cb431776dea768b90630861
F20101117_AABHSJ miller_j_Page_135.tif
547063f052a15e32277395b2fe0e6221
33d0ba6bcb1f5ce8b8b2136cd2c8d1c6ce480d6a
F20101117_AABHSK miller_j_Page_137.tif
88c487697c0cffe4fa65bd90f9c73b47
21912a56103510484a7fe20d908fa2c66a12eb08
F20101117_AABHRV miller_j_Page_108.tif
fbecdc4adbf0c3271044b0d849c3a68b
549718b371a241faa5de38184262682c1bc56c84
F20101117_AABHSL miller_j_Page_138.tif
f6d9959c10745206138a02fed6fb7eb1
1d8ad9d0090831d3a6e9d3f6d78673545048709d
F20101117_AABHRW miller_j_Page_109.tif
15fa05cfe2f7f9f2c665f5203e7ccdab
b2aa16237a7e1d07c967799e954eb4e8d7751e98
F20101117_AABHTA miller_j_Page_161.tif
562874de5dfd54bb674dd8d46fee49cb
edd77798372e8dc656938a4b7bfe53e9f437ed55
F20101117_AABHSM miller_j_Page_139.tif
d31920b2ab470206b255a87a9a466367
74cb4e890e54a57d42326cd4fd5b6652cd24f044
F20101117_AABHRX miller_j_Page_111.tif
34c7097b666591ff293d692696412cfb
d332550263f6a008cc93eff66641261883234b7f
F20101117_AABHTB miller_j_Page_162.tif
6bd9151dd18ba7e4ec26606953b1e355
2daa3d48b55471749d9caf1bad81edeedb8f89b1
F20101117_AABHSN miller_j_Page_141.tif
2bdbbad30b3ff4843272e44993cc5024
67fbc765eb63d067fc2898565235371c76389664
F20101117_AABHRY miller_j_Page_114.tif
886dbb5c37c67223883ed721650cccf4
9b026affbedfd48d752f83bbf7eca04491ee5d85
120173 F20101117_AABGQA miller_j_Page_199.jpg
3a279c2ab6b1e7d58f5f6312e2cf9397
62729aaf29cced9e9fac6a666918240ba464ae41
F20101117_AABHTC miller_j_Page_163.tif
0896f7316572f4a67505c59d5b89c99b
512bae4c65dbc05d33823fff79c08e8dbcfdeda5
F20101117_AABHSO miller_j_Page_142.tif
722703c150058f3664ef458fce327c9a
fdafec1670a07e7e17a93750f1b33329dc0bd61b
F20101117_AABHRZ miller_j_Page_115.tif
47b9022fd273a4f65168ec838824d7cd
51c4e7366804028dd6dba575e816cca13c7c8bee
33887 F20101117_AABGQB miller_j_Page_142.pro
3e407a135594b0a44440273f58c4579f
67397c6c30a1ac433531145cb94832836e4d973d
F20101117_AABHTD miller_j_Page_164.tif
8c3a0b68b21cbfcf72b6a76d71ee2e1d
90b6e84884de07ea9c4847404bd8501c76a320f0
F20101117_AABHSP miller_j_Page_145.tif
6a061928759588d0b070bbca30fd4d42
16f3366d51ab3e8713c3c214fb12d5343fae0dba
F20101117_AABHTE miller_j_Page_165.tif
63474d68c7c6646082aed9e9feebbf7e
038379ae40f0c1c55ee017358153096af6855641
F20101117_AABHSQ miller_j_Page_147.tif
d050a9137180e3bd187132202979c55e
4b10634c231ce7f684b00a160b8a0b867c546bca
34182 F20101117_AABGQC miller_j_Page_127.QC.jpg
97ea7deb45c3b8eefe176c1c16e9df98
3d015c6c32080bef352ae9e27d96212c02d51173
F20101117_AABHTF miller_j_Page_166.tif
7130aa3f82a36afe8a57f93ad8a8953a
0b1e4d9f02e41979b08e60f549074143070ff135
F20101117_AABHSR miller_j_Page_148.tif
800f652efbc38aca635ba12e11cfacef
286b8ca598ae90dbb451b7caaf2a5ad4b63876a2
8345 F20101117_AABGQD miller_j_Page_085thm.jpg
3313e486e2eb71d4e9a652655b00772c
6561cc6b8c16bcde825aaa6834827036af09ac6a
F20101117_AABHTG miller_j_Page_167.tif
2986f33dd1fc392d65aea3932d92d868
d50f4c025f655e162ece7283f7d84717db8c7761
F20101117_AABHSS miller_j_Page_150.tif
ac24cc9020b8254717db9e16a169520b
659eb645103b0703d3ec6fd70fc18e049e5c8dd4
5880 F20101117_AABGQE miller_j_Page_003.jpg
1851deb9095a2340ded14fa263af0612
f8abf4cd5cb73b040902df7fc6860ba56bed60cc
F20101117_AABHTH miller_j_Page_168.tif
5e0575980f01fac1815f58787cf0f4ad
82e18833ea292f439a52801a274f5f705b939a07
F20101117_AABHST miller_j_Page_151.tif
9e5e7a48930647f7363c840dd8e31460
01dbc8519848602563b39ab34e1f2820c62764dc
24511 F20101117_AABGQF miller_j_Page_161.jpg
7d59db6ec6b0e86db15ef9847ec51b26
39a62663d0754de382b35ed2753c3fb2dbb97d63
F20101117_AABHTI miller_j_Page_169.tif
f2f98d997c7d129b95cbdc149534c888
0b0557ffcdc1ba7a45056b0d929d5c68d6ee47f8
8659 F20101117_AABGPS miller_j_Page_075thm.jpg
cea5dd1a27b88ecfaa12dbc0259fa8ad
59a158829fd100112ac88ad075508da3a7ee6481
F20101117_AABHSU miller_j_Page_153.tif
729baaaf60a92e5b5a2eb941c732b4a5
db1b6d3c9c58645dbd88204eb04932406702e8a2
1578 F20101117_AABGQG miller_j_Page_026.txt
c87233f3777219e09096f7d706de2c48
a14936166258950ff2c8cdf2bcf35dda388a2264
F20101117_AABHTJ miller_j_Page_170.tif
db6dbe5daaa72f8aa8b6c3be0fd6036e
9c2d788db314bf28abc13b5e2f4cf967596e9005
F20101117_AABGPT miller_j_Page_110.tif
99a461bdadbf0140b1590da2a5a78242
23c69f139a72db298f91baef3a1e392d3cbd5a31
F20101117_AABHSV miller_j_Page_155.tif
520f6d94ce2f9aabfded638c4811ac64
081f39bbcd30d8d979e078d5cbb74d6d477b5474
19185 F20101117_AABGQH miller_j_Page_168.jpg
4c5ecf80a4ba882787b03c1278aff46b
40f3216b25e5ff56bbd60c0ac3c4b04624268238
F20101117_AABHTK miller_j_Page_171.tif
9ae5bb5e91540ba79a6055be2c354bc0
060f648207b8102e98335f17741d91d89132eb4b
29032 F20101117_AABGQI miller_j_Page_148.QC.jpg
20e398d531ba689c1d43084f51bae489
02ece74323ffb68a11d2d1c1f87d7ebdf4d8f9b8
8423998 F20101117_AABHTL miller_j_Page_174.tif
bfa9d196c7f133414c216ea964224908
c898f2b1eaebba9cc03267bacdca4c76b5f859d8
33989 F20101117_AABGPU miller_j_Page_051.QC.jpg
8f1906eb0badec710cb94bbf27aecaaf
01a1477f0626d6a64a07b8247eeb79deecde0171
F20101117_AABHSW miller_j_Page_156.tif
03c9338e58c8eed6ebf7c7c6f83d9562
0161bce21160f7309d793675b5a94ed31f0f3f9b
109194 F20101117_AABGQJ miller_j_Page_136.jpg
8a1592fb4760e39c0fcdc57f2dfb6956
7e05d95f56b4290e28f25ec7f4e3d50e4e5add81
F20101117_AABHUA miller_j_Page_196.tif
1684346032c894fb6c0876bfaf1a4026
4e2c3a696a1e2f49d01133d1dccc35f5d8543cfa
F20101117_AABHTM miller_j_Page_175.tif
4e0305121b64cefcd7bc23db97d246f5
70d78e7006925c5f608bf972394057a478c5988a
6590 F20101117_AABGPV miller_j_Page_026thm.jpg
028871f24a2ac4fa193baf0c06a4d71a
df2303d1c6225b4b4e553f1f200b166867655f09
F20101117_AABHSX miller_j_Page_158.tif
2961f6e074160bb97fd8b4484c0e5e2f
4d9594bcaeccbe864f4115438c0837d8a921c1c6
36291 F20101117_AABGQK miller_j_Page_052.QC.jpg
a574c431987c028fecfd45cd6f08fa3d
98c4caae63c746670dcddea2ceb0cf512823fe6e
F20101117_AABHUB miller_j_Page_198.tif
6101ac905e7cdc1d54812f3c3f8cd609
2dae2af76e9a8e05151cdc3ac5d442a54040b9bd
F20101117_AABGPW miller_j_Page_063.tif
773022a9bd6e86d86c5408bda3973f2a
28e3e29ca7f033601afbef2ec61a82a18ed50937
F20101117_AABHSY miller_j_Page_159.tif
48e2644c7427edad142fcd673c632b17
77348128d92430954871a19caa5c08e9127603b4
54292 F20101117_AABGQL miller_j_Page_052.pro
112eca06f46460cc8d43ce2b313187bd
632488d509878036d51d351fd69b009448196795
F20101117_AABHTN miller_j_Page_177.tif
25650aabb80324083eede78df4035632
f4572f6ed73354958a0a91398a7f46b05d4e7e3e
F20101117_AABHUC miller_j_Page_199.tif
89744b0bb73ee76dd301d388cc794120
4c7dd49d53bff79a64a68d2e45c70c4f9f1b35d2
F20101117_AABGPX miller_j_Page_003.tif
96422dd49175f242a4a55ccdc81f0736
74f30be587780e8e385ac29c90323d8ddc00b91b
F20101117_AABHSZ miller_j_Page_160.tif
872271d62095c14cb8b6ad59f06ef6d6
ceb5c4908ba5c940c55815cc0c1ccdc804fb4259
F20101117_AABGRA miller_j_Page_049.jp2
5ddedc24b11e64f92e8e74bb4d367c60
e1f81f13c56d2963b36711b441b94e00439e365b
F20101117_AABGQM miller_j_Page_006.tif
91b816dd9dd3ec3afeeee7ffc3f559de
ba81ea10b351202c79d72a9d176843310159c03b
F20101117_AABHTO miller_j_Page_179.tif
2c0736ba45b7fce9300582340a9348b3
7db3ad84d2596b0eb8a0aec964c6b1524ad547eb
F20101117_AABHUD miller_j_Page_201.tif
5df1363fa6db668d73236d1ea25a34cb
c28252a61aa5103874729869acd133b43534f480
6713 F20101117_AABGPY miller_j_Page_191thm.jpg
731e03aaef13c8f1531990cd42084958
1a7e8f123faa9bd9dd18e1897daeab77064f8d6e
F20101117_AABGRB miller_j_Page_037.tif
78896bf9212e5fc741ab20668b4a9f15
8e4e07571f835d4b18191021d90940e9fc0a27b1
5315 F20101117_AABGQN miller_j_Page_192.pro
14997a7baf24d2becdee560c84b0dbd1
3600830e0408046e9ac61b1c21b9164cf8ad0fef
F20101117_AABHTP miller_j_Page_180.tif
44a634387928c2405b0d06609889b079
3eb80f9b81a814eb394436f72270ba5359d0c264
1662 F20101117_AABHUE miller_j_Page_003.pro
9873398a4b048b2a9c07308eddee6941
86b947038ed4213fe844343b36dba1f5fc76405b
36679 F20101117_AABGPZ miller_j_Page_088.pro
1783bd51453a57f56a88a0cd783c642d
9dd5688b4f3e76dfb040075ba4ad5c8a5a57e989
F20101117_AABGRC miller_j_Page_093.tif
0e0bd100a680bd6b90e83e4a4de2551c
df9b34ff4044f0964bc362e1123c563fb1fefad8
2589 F20101117_AABGQO miller_j_Page_168thm.jpg
1bac2fdbbcfb8f821eab1ad8635ff3cc
7bace551c544db40fe7df2e949174b2acfee66b7
F20101117_AABHTQ miller_j_Page_181.tif
a3c6032706a2d01924947b9df84eb496
eb8c24aeddd0f815a97e79a0ba8cca515e84db57
122 F20101117_AABIAA miller_j_Page_003.txt
741db761dec7eae9ee1e55a2e83785f5
d534cc2b47b192d1f847c4283c619a7baa6e4851
77310 F20101117_AABHUF miller_j_Page_005.pro
163f093d5417101e5792a994d8b34da8
c7735a1593f896fcac0410ef5cc81d292b3a5864
8087 F20101117_AABGRD miller_j_Page_077thm.jpg
693cd8976c5f0b56673265f9f0dca78b
10cffbe17effe5b65d5597cd74f21bc532c8137a
48278 F20101117_AABGQP miller_j_Page_111.pro
d73d6939167f07dd0dd51a0ed7784853
e006ee9ffb9a4d8e2c02eb6e1b3fe6a9b9e88ac1
F20101117_AABHTR miller_j_Page_182.tif
977d6fd18c0e2ed091b217328afcdf79
157aceb28cd2e1b77181593c0c65b1677768ac05
1322 F20101117_AABIAB miller_j_Page_004.txt
04e6c70beaf7d975b44a511d4c09f068
14cd86b1d7006680c31c8c9d9c45fbb3eb591b8e
35699 F20101117_AABHUG miller_j_Page_007.pro
d573b53bb6accac14aa06d1549aa35f7
0d4b73e2eb32f93cbc272474ba645b1258c9431c
F20101117_AABGRE miller_j_Page_178.tif
075c54f63b1480609922bfadbe3d0482
340584462bb9d665dfb00c1dec1de142951710d8
31031 F20101117_AABGQQ miller_j_Page_022.QC.jpg
931d960e22136b135a1852b949866157
2bbac6e36c9d7e76e66c81b309539c1a53c786bb
F20101117_AABHTS miller_j_Page_185.tif
be99b4d5dc6de7490b3ba03ca56e4000
ab584ae45eb4ad19440f27161f6e417286d4b7a5
3482 F20101117_AABIAC miller_j_Page_005.txt
386edec28417aff5e40169232c890088
87e06a8dfac047beb36cb653ab47418431d5dfb9
62486 F20101117_AABHUH miller_j_Page_008.pro
1a0820a074f0f37691b782d8a58693e2
e18207504f596221993490b82cfd0e391a0d3243
7860 F20101117_AABGRF miller_j_Page_029thm.jpg
45601608942fa9b0bd1adf65ee8550d5
17480bd01236f8dd2032110d5bddc3823c0b9cb8
36353 F20101117_AABGQR miller_j_Page_044.QC.jpg
c091d6e21fb68ad42f6039699ff81cef
43a5389f7d648b36148a0cd77da399810491a5f1
F20101117_AABHTT miller_j_Page_186.tif
df9d6237424d72aba7fbcabe630ee655
df423364d8cec8e1f2b0d9d64e1013b449397281
68743 F20101117_AABHUI miller_j_Page_009.pro
ddb4067bad18c03b8981fccd12f33f16
cb00db591513878b638d4f66e5f7df9f3a015e2e
F20101117_AABGRG miller_j_Page_052.tif
e36b9f321b91e0be8b44e9868b81eaff
52b650e93bded208790f743b40824187750c17e9
36437 F20101117_AABGQS miller_j_Page_107.QC.jpg
81e8f57a97be67f2d5da34ed5d8652fc
da77a88775632bbe79bbc707f3e7dd1cf39246f2
F20101117_AABHTU miller_j_Page_187.tif
994327344efab572630e0a62bcb5710b
ff1225479352da35d297e42db5983fadf86a5da2
3901 F20101117_AABIAD miller_j_Page_006.txt
f3acf0f615631bd5c46dc23701b7d3a7
c764318de072bdf807ccf8cbb8fee80eab16306c
57582 F20101117_AABHUJ miller_j_Page_011.pro
9f726cd2e7b7e68eccdbaeec8eaad78c
c25a04e5d5b7363862c3f7e85dce90e8b3926bcd
9142 F20101117_AABGRH miller_j_Page_100thm.jpg
2aa987eb621760a9691b96acc5ed622c
e5169d0b2073e8a30efbfcda4998a19d1e9d85b1
2055 F20101117_AABGQT miller_j_Page_077.txt
fb7e26ffd385a05a1dae58549374305c
d0c5d02abf59ea57eee857534158ae389035bd39
F20101117_AABHTV miller_j_Page_188.tif
8b6cd8e190150fc2d694e393fc254587
04a803ea26744b4ca8182b2d7a99dfc44498bedc
2455 F20101117_AABIAE miller_j_Page_008.txt
7f260c6eeed3cfb291deb6135ca77fc6
8f1ed9ebd41aabfc3a49a18483e430002034e8bd
44262 F20101117_AABHUK miller_j_Page_012.pro
b7b779b77a68527621023600a42568f4
4547c6217a24dd0239bbe033462ddd0072d50e62
53461 F20101117_AABGRI miller_j_Page_131.pro
faefe1f46f2ceffb3b6100da1c899fda
52177fbdd2e5be5d9b6f14146bede53949bfc3b6
91038 F20101117_AABGQU miller_j_Page_006.pro
20435589a37b77a0e4ed62cdbaa2feaf
32f4ded27e4d186b6c82214beccd1ccace97f5db
F20101117_AABHTW miller_j_Page_189.tif
c623272d446054d02d1fa25e37048b16
caf0a00b4e43b23bf9a7cbfbaeebff733d60870a
2675 F20101117_AABIAF miller_j_Page_009.txt
ea88900ebb7731a7b8e9b2fdf63c1e5c
01c3f3286ddb42c033aa4d2c2ce6770dc65e77d0
17081 F20101117_AABHUL miller_j_Page_013.pro
90d063b231021ffba130a58e12381495
04b7949f01207c8fa3081f232876b4988b1fea98
67814 F20101117_AABGRJ miller_j_Page_138.jpg
bf959b22707170103b34ba43bb55cfcf
79d27847cccc83b0fcb3f41b6eab54bfa5840b73
1167 F20101117_AABIAG miller_j_Page_010.txt
bf1176c642366877e723a43587963036
590c40a2643fcb3751ec7b38706ea70bdfc0d25a
52164 F20101117_AABHVA miller_j_Page_034.pro
91360e2630adcf560c1361f63e6f06f8
b74c364d5988f638fffc3296f269671aec8d0de0
51179 F20101117_AABHUM miller_j_Page_014.pro
b6b67954e3c380682cdf1605d4505452
0f01bd7a115884d785caf2e953029dfd20d9643a
27640 F20101117_AABGRK miller_j_Page_154.QC.jpg
0c06cbd332f9bfc0db17dfb64ce4bea4
f89efe03dfc49d300b32c97bb3cded95669059d7
F20101117_AABGQV miller_j_Page_126.tif
c5912c5f0b807977c7cba462615cfdb3
639df16d93ba2eeb82ffa0b4b29ad8cad588fd84
1053954 F20101117_AABHTX miller_j_Page_192.tif
90f8d17eba8fe27a2a9d938a6f509829
41c74175f7e2c03569ce069daa07747b5d3ef43b
689 F20101117_AABIAH miller_j_Page_013.txt
b3c052331926022a1ba29b0c4b0b89bd
71b531ee9201e306f96efaba85d27d7fda291278
50773 F20101117_AABHVB miller_j_Page_035.pro
b542f0593293266457400cadc91c7594
fafd6fe707a9469907f057c78da299a0c2be429e
52122 F20101117_AABHUN miller_j_Page_015.pro
38ab4fb72169e591dcb9f6ffdb0d8f19
bbfe8978da84b8245c94f21188f2f23c07ceda4d
F20101117_AABHTY miller_j_Page_194.tif
dacb048c1bf1a461f4a0176cd3bed8d9
994e170c7210db2ca40d2c5a5e78db0400ff3193
18857 F20101117_AABGRL miller_j_Page_086.QC.jpg
4fc7f9db3d61bed772eb61b8d5687af0
ce6e4c597146035f5d55fcccd2e4ff514bb86ead
1051937 F20101117_AABGQW miller_j_Page_007.jp2
e2512fefa44804e632f62d62b75b57e5
b54469cc21ed5647ea1299dd142f3fe782269d0a
2103 F20101117_AABIAI miller_j_Page_014.txt
7201d1ab3cba65bf4e302cd694df6304
b95dca890219eb69235da44bf9862af187c4001b
47812 F20101117_AABHVC miller_j_Page_036.pro
301942f06570920c7206a698fcd859f3
f704df2a043063132b3438ffe08c26939f845e75
49420 F20101117_AABHUO miller_j_Page_016.pro
84fb2388da9cd389547f947ec703cb0a
37a5ff6b63628412def0cd24b4202d575c30ce20
F20101117_AABHTZ miller_j_Page_195.tif
2f85de306edb07e333ac4e610d6412ef
80ec82bee1cfa3a6814f830d7bba9298b07d6c9c
2012 F20101117_AABGSA miller_j_Page_062.txt
4b92206af96bdf8d0539e818ee037f16
10500019770493b056ccd7501bd354455b1348a6
7161 F20101117_AABGRM miller_j_Page_171.QC.jpg
9656c5274e44ce192b8520f584bfab96
52c100a7b104f59beed1459d8d4cadf80187dd15
2062 F20101117_AABGQX miller_j_Page_110.txt
9b5e180384942163b532d71bf0882c7e
d347c4199dffc5b1da1cf1317590025533993686
2128 F20101117_AABIAJ miller_j_Page_015.txt
c6066824071e562236a74a9a13e0c785
43c97b5f6d816b91520baf97bd2029d0dc4e15bb
39175 F20101117_AABHVD miller_j_Page_037.pro
83f04a6bef9ed9de46dbf02891730b44
9094f5afcef58350c97978d9a3e2e6812c6a6789
16793 F20101117_AABHUP miller_j_Page_017.pro
8b6df9fa114d7ef0db3059e2447070e1
e62cda8cee2cd53b9c42adfd7d0db4c321a0fd4f
115246 F20101117_AABGSB miller_j_Page_067.jpg
1640691c02eff8295676484fceceb3e6
d44d99cca7133628dda2441b5fb46921ac03b53e
F20101117_AABGRN miller_j_Page_067.tif
1e12cac2daf610439f3833daac1b2742
66c1db4f517ce6552c2c12161aa63870d5203a97
F20101117_AABGQY miller_j_Page_143.tif
73f9e89c161290231eb13dcf16341b5e
020ccddc9bc5c5bdd99c325e8420d1a1ace059e9
677 F20101117_AABIAK miller_j_Page_017.txt
5fcfae44e2f454942988394ac634b296
087e95c0c4a6a9e982f4d2294dc9e3068449c48b
52378 F20101117_AABHVE miller_j_Page_038.pro
362751646347653411ae6f0f8256f180
cd601e451491a0db77e3ab8d855ce7dc6eb4971e
46520 F20101117_AABHUQ miller_j_Page_019.pro
f49f7ed4be1f954ac6587c41661a157b
6838fecb5743c4545a0fc78911f759b7458354fe
55532 F20101117_AABGSC miller_j_Page_181.pro
8f2f6596c33b6aff7e8e3862cc251a2f
1e3d3ac8a6a93ea0c4a3f487c00ae9d16b7c7519
3554 F20101117_AABGRO miller_j_Page_161thm.jpg
c2c1940e887d49fbd7e404479ae9fb79
47e38f9f9b7d4204289a16e862669676ecfb7081
7777 F20101117_AABGQZ miller_j_Page_165.QC.jpg
137e44ce0ac49c3b834026d257552c69
024db81703220f70115455878e9b35135f3142aa
2045 F20101117_AABIBA miller_j_Page_039.txt
4abee85c55dea920dd861be4283891e1
f032227926a50835de46a4d75b3f72ee7c38b14f
1869 F20101117_AABIAL miller_j_Page_018.txt
596e45e4ae6048f201e62ad220725ee3
d9260b3fb2b91241dc6dce87eacc33ca679e1873
52095 F20101117_AABHVF miller_j_Page_039.pro
df5aa5ede47336f1226ba2e47dc32c96
fb5d1b4b1e75253571e95ffe22c6fc01d7b04c44
57080 F20101117_AABHUR miller_j_Page_021.pro
b3ade0706819daf65dcc830564b60d6f
89d026b3834bc6b927fa65e66737b0959306b50c
34501 F20101117_AABGSD miller_j_Page_094.QC.jpg
da2f3340bd24107f5f954837918fe375
219fbae56611dca34851a4d4ee1d9484d34b8017
8311 F20101117_AABGRP miller_j_Page_076thm.jpg
76816bc89317de5921a83c48d4450cfe
c9fbd28b1e4fabd9cc6410842a1571fd250e5ed2
2283 F20101117_AABIBB miller_j_Page_040.txt
6c516acebdf47892daf1c19e4156974c
3a60a9abb85ca5f1ec4e0a00dd373590dbd21574
1879 F20101117_AABIAM miller_j_Page_019.txt
766bba8a1b46252274020a028b06f4b7
fa63601b58085bb8d3a9a78b5168ad17ff1c0b44
58414 F20101117_AABHVG miller_j_Page_040.pro
825225a61369c981f794fe18bc28d214
4acc717135be3b7a1e7376a793cb75371e39c5ee
33229 F20101117_AABHUS miller_j_Page_025.pro
08e6fbe3b83195e28a408dc83f688e40
2cd80008d41217a100503fed629d89aeb8d2bbf6
87440 F20101117_AABGSE miller_j_Page_041.jpg
832e575731a14a1ba73768d39f963c81
9f88922e075653e7c8a4b292231b8a8a69f4d230
F20101117_AABGRQ miller_j_Page_172.tif
642d4b25f9eb140482042e6ef44e8f66
2184003c85845b64d8de569a0a013815e12442eb
1693 F20101117_AABIBC miller_j_Page_042.txt
74a347fe1dc42d61061895967f9adf9c
a7aebe57df4489540f6b6946f07666cba4b8eb91
2319 F20101117_AABIAN miller_j_Page_020.txt
9205863bb3781bf5aad1ebe35220dde2
b4b1ae33e01a40f066a0ec3db68ab3353eb05590
42126 F20101117_AABHVH miller_j_Page_041.pro
f31b54e604ebdfdcecaf05bafae64286
4d9080e9c8a50e2d7962b3689936fb8b647658f2
34744 F20101117_AABHUT miller_j_Page_026.pro
068359b55eaf0309403c5d67288ed0e2
6b3f45a62779bbbca79ac2dc2f2fb94e98433a44
8964 F20101117_AABGSF miller_j_Page_078thm.jpg
47e14c6114def095bc0a8fd77e46b31a
d84d5dc58d4ff9b88b060ada2d5d1054e7417873
275 F20101117_AABGRR miller_j_Page_158.txt
89a710c11968a00997da0bde525ec881
cdefb70352f22dcce55b9007c22a763e08e142ac
2082 F20101117_AABIBD miller_j_Page_043.txt
46ab66096a68c5663b03385c8a590765
bf56e0c8a529bb04d8aac87bd25c218ed47e30fb
2243 F20101117_AABIAO miller_j_Page_021.txt
d11c1df128f63df96606b29e719a53f8
bf546cb53a1577f1017cc764738e15283328ef23
42130 F20101117_AABHVI miller_j_Page_042.pro
af0b2584d31533ef9afdb91c32a3881f
64b6528d85e13f9e3b4dc750dcae2dc3fd2407b0
52665 F20101117_AABHUU miller_j_Page_027.pro
ab2320f92d8f4a63f300be90c507c09a
d4beacfb126bf02aeac4ebb9e27320af198e23c9
1051961 F20101117_AABGSG miller_j_Page_066.jp2
1d97aeb799f32e1b9b445a19ed165cfd
0c5b10cbe703ed5e6c82dde144877bba757f47ad
1051931 F20101117_AABGRS miller_j_Page_073.jp2
4f3129f6a3b2bbd5fc56ff0d9459f527
45e82303bd51a7e79c2bd717677d3a5b9f207b8e
1856 F20101117_AABIAP miller_j_Page_022.txt
e94fc104374623eadd3e229397e1c5b9
f5b4c2894c26b32c37963b5784811b8a73280c5f
56866 F20101117_AABHVJ miller_j_Page_044.pro
faca9c16a7dac93b51726979b81071f1
3dd565454e6cb15c5e7cf495cbd8d8620eb1032c
47784 F20101117_AABHUV miller_j_Page_028.pro
b3b167aabf3cb06e75b4f36486c03ed1
3b8c68599b5b3177d493f723cbd0b686e74cf527
5093 F20101117_AABGSH miller_j_Page_158.pro
cac695a01a0fb1bbaac62f805477744b
b10bad86eae96dfaf69026f024ef6ee0a0ab662d
60114 F20101117_AABGRT miller_j_Page_049.pro
5baa6e30a87a685b6fb225c083864547
7a83d9c2fb29c768bab76a3a8436ddc8767b432f
2228 F20101117_AABIBE miller_j_Page_044.txt
d12e73e36a996d5dad5e1fe7b4ba20ec
6391fec06d2339c492bc13f53d5c4a7275ff07f0
F20101117_AABIAQ miller_j_Page_023.txt
24002da59ea4d5351ef299915b6ef893
3e5fe3f0bf4821453ffd5d087a9192d5fb415ac0
55300 F20101117_AABHVK miller_j_Page_045.pro
95e82f2bfb9aa5382c6aee6a6275455f
29a0ac6ff80c5395da976de7f1954c6290de6992
43773 F20101117_AABHUW miller_j_Page_029.pro
806cad1215f408c94c99a2948ff49c5e
86c5b2f562dab9e079348b24f44f1ce05bb687f3
57146 F20101117_AABGSI miller_j_Page_067.pro
ec4bc3ec79237635cda5f436a4412732
9dff482a150c35e7e7f86dce51ed897fb88019fd
732794 F20101117_AABGRU miller_j_Page_090.jp2
7b0086c3eb81eb71fb70f771532792f4
9391389bcb15bfd7fb66284879fa1cc8be03963f
2183 F20101117_AABIBF miller_j_Page_045.txt
6b2d0d1874fec85bab4728e47bcc8538
53a28f3e1369fe77df7e4bcab8ac99f4c75cdb5b
2208 F20101117_AABIAR miller_j_Page_024.txt
9767a87327727acbf371ef3fad67dbf3
f40f5ea024bd72499df2eab029c7bc32cbb59c93
45900 F20101117_AABHVL miller_j_Page_046.pro
ceba528c534509100138e1183b4888ac
bcf2161a5616b1f80581250b3ceccb7d518614a2
47551 F20101117_AABHUX miller_j_Page_031.pro
0805cb54f2b1c9cbadc979317c287655
0b41151aa9b7b0455ae052335b365f8c13f21236
113668 F20101117_AABGSJ miller_j_Page_054.jpg
5d422461c6274a778ddc047474713917
313c2c8f0119f927564056307183c57f90c98453
F20101117_AABGRV miller_j_Page_144.tif
2dd58b4f81233c189b97d793f5a8329a
f5dc01557f94988b19076295725bb42d4a26b0f2
1813 F20101117_AABIBG miller_j_Page_047.txt
838e1594f1be0ac3d513915926741d68
60a1298060116e3c2f081c638574b6ea386084f8
1497 F20101117_AABIAS miller_j_Page_025.txt
b059114f8f849b9f987eb0440efde8bd
ba3874761bec92d70d5247499fa288d0a4cc06c3
58344 F20101117_AABHWA miller_j_Page_069.pro
e77092890a7fe706bf29145a034ddf9b
229ce35f669deef2d08c2d2036fb2c8003568f28
47333 F20101117_AABHVM miller_j_Page_050.pro
6888adcf7e9036b798c8024a70913a2c
be30ffb596ddd6f99aeb4e10f1d0efe05fe31ee7
F20101117_AABGSK miller_j_Page_040.tif
3823338dc3bbba4a7ac44d51706e08f5
ee04efbe436536b17560862d2fab6166a3c29558
1792 F20101117_AABIBH miller_j_Page_048.txt
3ed1c01df80c2546c2ee1facfe38cc71
3c02ca34487d4244fa7232555a6091c7ec24998f
1925 F20101117_AABIAT miller_j_Page_028.txt
4f5ed72ba0eec57540d13745b1f432a9
63ebe07cd6a073fe4ff8231da7a33b7a6c5b5956
19447 F20101117_AABHWB miller_j_Page_070.pro
32c3fce64e77d39fc947d7d3854402fe
22d93895baa5aa6c89ecfbef2529686c87f31ae1
54092 F20101117_AABHVN miller_j_Page_051.pro
a9e6d70d0ea3c0d967706e494a0569df
055055c2dfd800618b866d6d7d312442f9b6fd12
56284 F20101117_AABHUY miller_j_Page_032.pro
63d5dd47c377c139eab02ba88e7fbd21
c0c97f2335956ff8d1353afa0bd85bc71680f554
1051899 F20101117_AABGSL miller_j_Page_043.jp2
ace02422af409ce9dc0820e10c368a5d
48c61787fe99c8af1abf12cd71ddd7db7f70caf5
1605 F20101117_AABGRW miller_j_Page_037.txt
8e9e95f5c80bca8c2b6e56f88ef13d89
2e186a5d134dcb498bae17035d822e56378c2eb1
2393 F20101117_AABIBI miller_j_Page_049.txt
d670ff0db6cc97ab2514de0ac7da24df
a07444faab9dd9be0a23c7dbf7a19c1b084bc786
1997 F20101117_AABIAU miller_j_Page_030.txt
e55cc60a1c170c88366acff4e0c4d59d
5a805c2b16d5636135ba829ee59ae494a6fd4d92
38050 F20101117_AABHWC miller_j_Page_072.pro
6702b017c6889d350fca3da9ad5f2387
c4e32ff99b76e51df1a1f346ea0eb871a36ef8ee
56805 F20101117_AABHVO miller_j_Page_054.pro
c1d15c26a5e236c38172244aa6eca6d3
382858945e85d4cb2c9fb13d2b73d05bbfebd7cc
55746 F20101117_AABHUZ miller_j_Page_033.pro
2d28869a9a5d3657e1d2b61c1fa5007b
277d94bc1c2fa919c310b5f1e7425cd2ee3df9c5
2085 F20101117_AABGSM miller_j_Page_001thm.jpg
7a5e3b0622d0ee38eb32a0771d847313
ea5508e8158675f174887906e80f119b3d0d60c8
52905 F20101117_AABGRX miller_j_Page_079.pro
116cd6b8435e184df4d6c5ec2ca1fb36
5b5532cf9d8bf726a7b25ca94c56bf030999306f
7450 F20101117_AABGTA miller_j_Page_149thm.jpg
89392b9e926469cfc234e0107064d196
4551d418b8a475e57d3ea8972e12059abf07b4d2
1919 F20101117_AABIBJ miller_j_Page_050.txt
309852f8d8ab3a730f1c5c6dcf6208af
8ea0f4f11d6b3db365619881eebad4f3fded756a
2203 F20101117_AABIAV miller_j_Page_032.txt
df1b8621054db9674dbb18502864f161
e0453774b262a421c34b9a3c710f4ce4f5856f13
55867 F20101117_AABHWD miller_j_Page_075.pro
ec46ef6b3f32e246be2e3c338d9afd52
6faddf0c45579803170354ad50bc4ec52911e682
54037 F20101117_AABHVP miller_j_Page_055.pro
58e3d016dfdfdf50179d46045b148fa3
2c3bc0b8e5d6ff43c2e079443187dad2d8478058
F20101117_AABGSN miller_j_Page_157.tif
41d9714756f48777d7e1f954a88cf65a
109db8edb3cdaa271b7700939534ebcc5a325a77
112433 F20101117_AABGRY miller_j_Page_100.jpg
9c97e3e94e6e5e8cf474265184c85b6e
0915a506334c7c8fa1871f5196ab15f370daa6dd
1937 F20101117_AABGTB miller_j_Page_031.txt
f47df2fd34661c53c3064e7802b60198
1a4ec3bb52d7cb2cce35972dc76a712888ac364a
2149 F20101117_AABIBK miller_j_Page_051.txt
522b2ddd7e42731ae354440e05d52885
f2b14b4306c1a31e1ea92412d49145211c0100bb
2185 F20101117_AABIAW miller_j_Page_033.txt
4cc685f3a908a3c435c028bd7c7e3c61
4aedcb928a942fb6b54920560f5540b7cf75e39c
53567 F20101117_AABHWE miller_j_Page_076.pro
6264c4ca5a5fe65e4b364e9d8ca25fdc
870ccf607f7ff92b21fe62fadd36d69f784c4c2a
51442 F20101117_AABHVQ miller_j_Page_056.pro
6e39c3e11fe79437b0d4eb5e0b1614e5
7d4d39dec38e5b8cee06810a47ad25d27a4e4ba1
2242 F20101117_AABGSO miller_j_Page_067.txt
ca9184159e5ed0d639e1e303cf36c945
5418a2893a0a7967c09c2f84fd3a67722e184621
56086 F20101117_AABGRZ miller_j_Page_060.pro
076d7066e1e88bad30743ea4dbf11aa4
591500c4335f3e6a0d3b17e37adfe7590ecfb2e7
2058 F20101117_AABGTC miller_j_Page_187.txt
b08a61977e953af564c96ef48c3224b2
2cabb79340e63f426990127ede389a80507e1d34
2303 F20101117_AABICA miller_j_Page_069.txt
f5af8406f9024fd47ee8f2b6ea8156b2
74133d50aea8244ee60c46a2a9e11ae0f9186be3
2136 F20101117_AABIBL miller_j_Page_052.txt
2d5836bebf590e663bf7c1f9e21ddd59
cc2956a0e8bfe0cd4ba57442fcd4b42bb5a2f803
2065 F20101117_AABIAX miller_j_Page_034.txt
5bf3144338d178113f87c8b4283ec841
bdf88c41d4f1eee98cce9ee0266932dcf95ee1b2
51979 F20101117_AABHWF miller_j_Page_077.pro
271b8beddc334e033974577e4b70a4ca
4e69de0c5415f154d5aa6cd22ef4749bead91e63
54027 F20101117_AABHVR miller_j_Page_057.pro
f08e44e63a9757220f49d576d4383a3a
50eb92beeb84db879fcc63b60225eca4270fe9d7
31963 F20101117_AABGSP miller_j_Page_036.QC.jpg
bc53f54f73054043b7c693d17181fff1
9513ac161bd1b1e3bb0e99bd340ff8e303bf8cd6
F20101117_AABGTD miller_j_Page_123.tif
0bfabf1ffe4f89a3f66d24b68fda92f3
6b7127135ddb578217d9c44742bd765bbdddffd7
292 F20101117_AABICB miller_j_Page_071.txt
72f024db8821fc3f9a977be8ac304cc3
0c2249fa0f3419404b9993a098aa80d4b0d6ac3c
2117 F20101117_AABIBM miller_j_Page_053.txt
2b8887a07a199b6eec266ab5f9c6404c
c1d81b081be33f8c28778f6bfcaebcf082ec4134
1911 F20101117_AABIAY miller_j_Page_036.txt
34c529baba1b27e1621bd2dee42df21d
5b668818924785afba3b06186fa32f4688631947
54316 F20101117_AABHWG miller_j_Page_078.pro
d75d9ce3f86e10829343dcaa791709af
15550e1c02ab1c96448c40dfec70b618900fb45b
49361 F20101117_AABHVS miller_j_Page_059.pro
42a27796dc21f2ef8d696c85b8998811
d312d46829e0be34e2b8422c738605ad2eb32d9a
F20101117_AABGSQ miller_j_Page_117.tif
cb5f7b1666065f308387eff569e5c122
8823b87df82f620e253e6603ff5eb28a0540367d
313 F20101117_AABGTE miller_j_Page_164.txt
5dc4e9396d2a27aab51f9b279c310ca9
1ea36f5eb850d60112da4adabf8f780b48b13258
1703 F20101117_AABICC miller_j_Page_072.txt
80deb059f0c882cdf1a2bc1ce8cd5664
c4a702cd614975c3542f7b565ae68c866904ebd7
2231 F20101117_AABIBN miller_j_Page_054.txt
ac60175a6993c5c2da703ddf10e5a36f
18f3029ff5000e2dca3c0298341c5a62b2207d85
F20101117_AABIAZ miller_j_Page_038.txt
24056792b12cdc31e24bf268e824372c
b56df87bbaae61e7dae98e679efa3d8dfd8b6d60
45862 F20101117_AABHWH miller_j_Page_080.pro
6838482fe8cf9a98876b36c96b7fe445
3a9931d1c3ef1ca70da501a4aed7d0d08c4ef49c
54340 F20101117_AABHVT miller_j_Page_061.pro
302f0a408575ca26f737bfcb3d5966ab
e758109d7c6f0346bebe96267f65eb61f32e6242
19073 F20101117_AABGSR miller_j_Page_167.jpg
52518a6e256bfdcbf940953fe1880e31
3c30346fa4a03716ca77ba14f15907b676baa373
459 F20101117_AABGTF miller_j_Page_161.txt
47b38969e90591957dece19a062d5ae4
d307b52175f4d54a9e98a68ea16235bc34153e89
2180 F20101117_AABICD miller_j_Page_073.txt
e54f40ac1adb71112bbb6403877808e4
b206ba62f4823d72c05330b662c067fcd4b0838a
2124 F20101117_AABIBO miller_j_Page_055.txt
31e131135fb2587bcfc6a4fa051d568d
9b3b258650cff085f3e075ddb08accf16d97cd70
48375 F20101117_AABHWI miller_j_Page_081.pro
fe1cbfc51148f82c988ce292f1e13802
6969aeb11fabfd9c57613ecd81f7ce056890dc77
51022 F20101117_AABHVU miller_j_Page_062.pro
ec3db3abc9f272ae15ad9231698c50af
647fff4084fa51ce255f7ad2ade460d3ee0077d0
23623 F20101117_AABGSS miller_j_Page_090.QC.jpg
eaed05942cc68cbf4fbce31cd3db6b8a
196de525715be7b21f11b710489d600d57d77b96
107917 F20101117_AABGTG miller_j_Page_057.jpg
209e9d0c8d7c946906e152a055d8c2c0
6a042ec4a86f2632ee36bdb70992cab25a5bf529
2438 F20101117_AABICE miller_j_Page_074.txt
0ecc9e8cbb918b20039763c523d27edb
16790881590c208d23b2b2c84e50fd08e6834438
2031 F20101117_AABIBP miller_j_Page_056.txt
e0d1579cdc594383cf6dba362544a7a9
0bafc9b5ef766b806b476bdec822a7af5651cf3d
47232 F20101117_AABHWJ miller_j_Page_082.pro
bd5735aaaa9ea4413e6397adfca35a45
989791338b02e55b849ab95b0d1d7ee74fc1866d
55792 F20101117_AABHVV miller_j_Page_063.pro
b7b3d55ab43b7c9cb0459086cd3d2465
290dee0c1c5096e726fd0ec0f7534b25142afcd9
81552 F20101117_AABGST miller_j_Page_153.jpg
7ed1c978099f1a40ba3550b298b8d434
97ebc6b993749d59cb90b190af5630b2627e7158
1956 F20101117_AABGTH miller_j_Page_012.txt
d34991f8a62611cde3a0d7928b6c8af4
25d85e3c14960a8bbefb2154b9d22d656a37a3fd
2129 F20101117_AABIBQ miller_j_Page_057.txt
94b70c66cdfd5314c2fbff4170953873
7f4f27e46fa5a6136eea60f9a2ab4567d2f29a0d
67587 F20101117_AABHWK miller_j_Page_084.pro
e506929f783a689943f90f6e9c50484b
075a297f2e98792110626192fc787238b6116380
52222 F20101117_AABHVW miller_j_Page_064.pro
b282c7474e78cb1abc56270e1aa42e48
0c394e82bb4b2bd9309333de389d589d31ad7544
6561 F20101117_AABGSU miller_j_Page_071.pro
9c5f072acf18177a454876ec35fc1e25
e95cb7f4ef01f3ad654b0a49f1a4285b2f4b85f3
705855 F20101117_AABGTI miller_j_Page_089.jp2
beb45192ea5ca40d044ed98639b7e1c1
e19b5bffa9ba0ded00389cfc2b9488f8b9b73945
F20101117_AABICF miller_j_Page_075.txt
443ae24e1203dc40d370c81d509fee40
59b0364fc499f6e5c819203e42bb3f7b660f996e
2170 F20101117_AABIBR miller_j_Page_058.txt
22e7860cb577ef5050a77c3c84be8630
7fd65c01564e65477cd383305c1df91ec60ade0a
51332 F20101117_AABHWL miller_j_Page_085.pro
9cc7cba3ccc4bb661f5e8f5f5924a708
e22f6a38cb8ac441cc218b21ce2e5a194f242607
55368 F20101117_AABHVX miller_j_Page_065.pro
5868c74ca2a5a83b3d3eee0bcb2c9ccf
233721edd3ab4928d075898779a4e374a967393b
6355 F20101117_AABGSV miller_j_Page_139thm.jpg
35e6d8c08cc651b6ec69040e3ca21b64
8abfb7589c7b9f94bc4375b0d3902c21949102f8
2258 F20101117_AABGTJ miller_j_Page_011.txt
2163fd1ea003299048dd32e7cc04b14f
54ca556083ac4a6d540564fd891b0178cf71147c
F20101117_AABICG miller_j_Page_079.txt
066ea6fa58b59d98ad568ff2913230a1
dfbb3c8dfcd02ab2e6cf3686f9f43cc3280444fc
1959 F20101117_AABIBS miller_j_Page_059.txt
8b5c57bd0ecb29f3da6382e5865d8e78
c79be40bdcb2ac529736e4f7b93c1a9a9c6a5a73
54129 F20101117_AABHXA miller_j_Page_104.pro
dc7030fda7bc5a820613819ed0ba0176
930747ddcdca7c4919005c1caf045a4e258fade6
26498 F20101117_AABHWM miller_j_Page_086.pro
476b3821ad5c16f1f7bc2fe2b5c787f8
1a38d8696204328c8feedf0437ecdb60a43a7778
56777 F20101117_AABHVY miller_j_Page_066.pro
267afa29ae5bd3612d549a0cf064b690
b444a1858aa6158a73c1a2974dfaa0fedfaa46b6
7723 F20101117_AABGSW miller_j_Page_158.QC.jpg
55bfc6f981a08f9856bb51d1dd2dc5c0
25014008aa0f5f0f639f4b4b352007506a25c3e6
2150 F20101117_AABGTK miller_j_Page_078.txt
959d6737e6bed4af4608a6613c8c61e4
36f1aac6db816124892672632fe6853fa343f104
1918 F20101117_AABICH miller_j_Page_080.txt
c56ff9d79d72282667bc1403c4206dfc
1299725be55e1f79e0d15d3259ec300ddf60fc83
2201 F20101117_AABIBT miller_j_Page_060.txt
d3423cfec67b2aa139482cb882a6755f
4a349e6719dd6fded46d5ec41e460214ad6cb0a7
53107 F20101117_AABHXB miller_j_Page_105.pro
dbe69b076bed4c3073b338733d4893d9
68214cee8151a8cedddf3022b4478ca16e20edfc
33171 F20101117_AABHWN miller_j_Page_087.pro
1f2d695aab49d4267d856461daf03e95
10e833c58c105a4c32f11b151a8ec726479f22ad
1889 F20101117_AABGTL miller_j_Page_029.txt
864ad4c569b661e69c6ce28ba39883c1
faee2cc16be84f2c06691b9c3bb2823b81015d8b
1960 F20101117_AABICI miller_j_Page_081.txt
c67d210ef09e7aea7fd68f2ca7c4960d
0e0dba37a1312bdedfa2bca2dea9a7b87675f9b1
2148 F20101117_AABIBU miller_j_Page_061.txt
dd4819f3dd17fbd49bc36b90b9795ecd
f2549448c9292b5847bfd6a64a116a3d1e17ee44
53194 F20101117_AABHXC miller_j_Page_106.pro
e40d9d7862cf29af9ec38fd821377919
08d32afc2c738591fb5dcd8246c71b5608acc46b
22688 F20101117_AABHWO miller_j_Page_089.pro
3d99fca70fe9c2006acc5d10c2e6343e
24caee92824c6fbbb7be1e395d8012098ee46eaf
51750 F20101117_AABHVZ miller_j_Page_068.pro
8368ac5dfcd9212a09a2aa5084004d47
578299e7a2b3ec52eb9532a02792901a7b9dc121
1984 F20101117_AABGSX miller_j_Page_071thm.jpg
9e70b23871c2668e8353f95880c1b124
d3d20c364019c4bccb75da93049f3a95f31d0035
8651 F20101117_AABGUA miller_j_Page_015thm.jpg
5a888d160cf5ba51db4581bbcadd8839
fbec0382a842f1039b046f7db072bd8cb1c527a0
1051907 F20101117_AABGTM miller_j_Page_100.jp2
f85c8cc2b9087d8593590a591f3544e1
2c9b918d41b1ad586c539672b3b9f81dfc9047e9
1962 F20101117_AABICJ miller_j_Page_082.txt
1ccf17362438b14ba1a539539b39e3f7
2ba23d3cc0f4547f9853a8a297e4fa7a20f52ea0
2197 F20101117_AABIBV miller_j_Page_063.txt
b6b0315f54250ee78463825f5c8894ae
d1e9bd2cd9addebb0822f43b4c33f34013c05786
55013 F20101117_AABHXD miller_j_Page_107.pro
356c02fe1cbd99f11f77217556c20fcb
16d2200a8a75afa95649a630753325492c1fc9e1
28382 F20101117_AABHWP miller_j_Page_090.pro
97b104618719c1281bad9139975e5364
a6cc87037f89997d10c632d1bb13fddf4fd44631
F20101117_AABGSY miller_j_Page_078.jp2
446a6dd5f225872e02aa4ca5e69b00cc
01006f6160993deef287583c357e81361f6925cd
6348 F20101117_AABGUB miller_j_Page_193.pro
75c9824facab7c3e5af9f5a1b1557313
622d4f9db60e287aa5f7d5ac5ba7c1d5f89d8293
36465 F20101117_AABGTN miller_j_Page_002.jp2
8441fe69aeeac0e187814e08cae216b8
1e06ab03089125b25f07221ed218fe93e5770290
F20101117_AABICK miller_j_Page_083.txt
e764bbfc3e4490dea3a11256eb6959c2
5257c415cf4d038f911b06b1610244ade82d1cc7
2051 F20101117_AABIBW miller_j_Page_064.txt
f680571cb07a756225ef3ab2df76f571
49636ae91d53e8f3f74c58f0885dc7e72240cda8
53341 F20101117_AABHXE miller_j_Page_109.pro
4f1483314a4c91905c46a2b53dc06fd4
70209104127faa854dc1c136cc548c927e6eb31e
52874 F20101117_AABHWQ miller_j_Page_091.pro
29a109ac3b4aaffab9427685ce887cad
b282391f4fb9bb45e227f27ee400f750b8a1cf88
2068 F20101117_AABGSZ miller_j_Page_179.txt
c145bd7d02028d7e5003c1b15c0feac4
f918000110f6dcb5422aae796bb7baf1226bde0a
F20101117_AABGUC miller_j_Page_014.tif
d8a2a4c2778ab94a29d2faed5bbae9f2
34bd8c4a906489f14c5ace6ecaef0b500e609149
82668 F20101117_AABGTO miller_j_Page_148.jpg
1788ca436243df37bd4075ddeb8d44be
f58ee90b4d2c2d5845af656cee896c185e46cbcb
2143 F20101117_AABIDA miller_j_Page_104.txt
42b0abc741257858955db4d1bda0589c
e09f1529c662d0b4853394ac5a72d4332bb4cb8e
2024 F20101117_AABICL miller_j_Page_085.txt
401780b6361ae3c42db4411a3f2b8cae
85cf40628c1ed4c12378a7303cc9cc08dfb656d4
2182 F20101117_AABIBX miller_j_Page_065.txt
3d1262c56f8ed27cf8712903d5504261
46db92b9264135bbe726cd356b60daa6d878725e
51983 F20101117_AABHXF miller_j_Page_110.pro
0d851084dbebfb880dad40c7187cbe64
59bc18e3a0f4d37f504b28f25860a90bff846510
39780 F20101117_AABHWR miller_j_Page_092.pro
5f42d43ae15b8b99b53aea5598c4d2fb
637a4c5d435de2d7b12331facef1860577e3572a
F20101117_AABGUD miller_j_Page_176.tif
09b389eb92f05f784c0df562e195bbb9
3e90badedc86e481b458d35f358bfc900f0a7c0a
2041 F20101117_AABGTP miller_j_Page_126.txt
5e95bfbd2673fdacaca879c54a596a74
dc29d7f45b2972a9ece38542f7ae17fef44847b8
2099 F20101117_AABIDB miller_j_Page_105.txt
fcf8ca9fc144e24ae494fafb5078abae
eb12aaa6607427f4e5b7974aff83c46efb9faad6
1065 F20101117_AABICM miller_j_Page_086.txt
31e6523bd37a48a4a0872e3eee182d30
deac6c0c1e4975dcf092d19c9748dd7105f4a72a
2261 F20101117_AABIBY miller_j_Page_066.txt
406096e61628f1d69c3c879ac556870e
3b396aceb38a58422802e321fc27ba4334903ede
50234 F20101117_AABHXG miller_j_Page_112.pro
888b87f872d1c92f26486716df405613
bb26140e795ad9b45b5e398fe2e361cdcf6bf330
22899 F20101117_AABHWS miller_j_Page_093.pro
2bf125d7b3e129bd73025c6580cc1bd1
9ef89400c61c9d31327155e45ae306e4cf9771bc
F20101117_AABGUE miller_j_Page_015.tif
dcb0d82677f19c7d1497592eb8972fb2
d5b89666bc75aa45f5413cb15a378254dc8f69c9
8415 F20101117_AABGTQ miller_j_Page_189thm.jpg
3711da5cf8ff13e746b66505c65d8f1a
5ecaa3e00e28f99dcccbd6c5ba29010b284ea039
2104 F20101117_AABIDC miller_j_Page_106.txt
afc33330276899090573626d7bf7e149
9734c435e1ea5d642507e8100ea0254238b33558
1409 F20101117_AABICN miller_j_Page_087.txt
0c8f6e88010044a6df4ade851b3ee9b1
e9f440acad018ea76653d5348e9c1fb848e70c1c
F20101117_AABIBZ miller_j_Page_068.txt
40ee89768854cfc33759a70248240615
cc40ccb30da7fa188af6f0522768e1f66abd681e
49500 F20101117_AABHXH miller_j_Page_113.pro
2e3fe21d685209819a6c47981746ad42
26f9774716dcbc28d6b9b6379fa42ed662145dad
56736 F20101117_AABHWT miller_j_Page_094.pro
8620cfa421429d21e65ead98669c6bcb
ff998640d723be74e819081a3758eb71e5448b8f
F20101117_AABGUF miller_j_Page_140.tif
300c3a3167d7ac7291f1280bedd7886a
efe69074bd009cfc3bbbccb1bf43cea0d42811e7
51262 F20101117_AABGTR miller_j_Page_127.pro
f19b3a7772be5d35dd0d0dc7fea32aea
356c8dc39b09576bee8feceaa0acd6d30c848ec4
7882 F20101117_AABHAA miller_j_Page_150thm.jpg
db758931b68ceedbadd94eb9ab02bb9a
5885ddf18bd59de942718bf4e811b680df5b8fcd
2175 F20101117_AABIDD miller_j_Page_107.txt
929d2abce8c98b3f7c69b28a1822a61d
ea4997b1ccfbcf63085bfda4c6e5c1c5c5e03e87
1225 F20101117_AABICO miller_j_Page_089.txt
668d27d9f4f1504f53bde93b4ef5eeee
54fd237c02ef051d472a15902466a499056af041
51920 F20101117_AABHXI miller_j_Page_116.pro
ebfe3ea88fb42dadeb0747981762fc49
e57f9b3edb0768ae3d3ecec40cc5cfefba238af2
51064 F20101117_AABHWU miller_j_Page_095.pro
03f36375e46350bc23bc5747e19c4e8e
dc8467488c4105aa0b7dcd654dd602df3da763dc
F20101117_AABGUG miller_j_Page_154.tif
6eb6c24d6699b1d2e85b99759a607f1b
a97dd9a6384443356ab47f5b8f03f9f982c32953
1051977 F20101117_AABGTS miller_j_Page_199.jp2
1b6c7179eab9f60031373c18cf5d1f44
ebe621357499447eae6209736f035b9268f90040
1051984 F20101117_AABHAB miller_j_Page_053.jp2
67a4b0ffba49379c56ec466695421159
293f07a992bd164e0856c3cd4e579a1cea60a2a7
2050 F20101117_AABIDE miller_j_Page_108.txt
6b6a14fa366467c2f0b2dec1400f0b62
cb90294215cdefdca8ce75c34dc15d753781ea5d
1637 F20101117_AABICP miller_j_Page_090.txt
ddf04be0b00a5e8a118f78eebcebdc72
165f71009ada8faef60310bb343c26b657b74157
51357 F20101117_AABHXJ miller_j_Page_117.pro
bf24af8f73e168a090681d66a2e97d03
fb278a2102007b6e056c6fd8966092d5da8f48bb
51248 F20101117_AABHWV miller_j_Page_096.pro
8d7f79f65cc74653f2d64fd4ab1ab63d
bc0871ff9a548c0181b6b52dc04460675d6f3f26
34991 F20101117_AABGUH miller_j_Page_133.QC.jpg
554425a2df4751601a13e948d50e0924
fed2424f11d9bf6d14c505a15050852d7f3728c6
50479 F20101117_AABGTT miller_j_Page_128.pro
c37790e29b2caf3153a5161929d4884d
45112e6951b2f26ce72c92dc7fbc4094345f7bf4
F20101117_AABHAC miller_j_Page_133.jp2
0e067c1d030ab557314fdcf6e07e8a5d
1e935b56571a89d1c474f4b618174ddb9d19a9c2
2108 F20101117_AABIDF miller_j_Page_109.txt
8761be513546ccd38bbcab3c4f098a26
eb6760bb727c495da94844c119ac7121c359033f
2938 F20101117_AABICQ miller_j_Page_091.txt
bdec3e833378b039b745750b1ad5772c
09cb8fcdc10c4a8703d1242ff204fca6f9d02cf0
50857 F20101117_AABHXK miller_j_Page_118.pro
8c1458db468ebdfd5be2711a3dc4ba2d
ea5b7d57587809713c54deeff620c240da1a7f1b
54957 F20101117_AABHWW miller_j_Page_097.pro
4f0f3722b5e272fccdaa838f96d4f221
b9bc513b499f23e02e0e7952f381c9ea0714f58c
8988 F20101117_AABGUI miller_j_Page_107thm.jpg
36f376866d3d0c06cf5e2b3db84a9c37
39fba7d8dc53dd5522624406ea9227fd4c766dd3
9631 F20101117_AABGTU miller_j_Page_009thm.jpg
c33b2a398634c510440e1216f4564c16
1c91511b2e7621bfae2720ed8127b18cef19ee7b
105344 F20101117_AABHAD miller_j_Page_096.jpg
ff1c206a7754a9e39aa36a52be5fa974
c3f973785875283dd7b7b35a017e579e229a945c
2328 F20101117_AABICR miller_j_Page_092.txt
098d680b18410e5417c31833b650a154
b139ddcfbecf5eb612391fa1c6a2b9d03ba20322
54588 F20101117_AABHXL miller_j_Page_119.pro
d15c54a10e3ef29b0cba9843b7a31de8
6768bbbf02c98844d9ba4cf9b4ee57caa6f6b15d
54244 F20101117_AABHWX miller_j_Page_099.pro
1907e5876cb9565eb5429942be8bb257
2635793c0d5924c30687ce28a198de9f14e688d6
70095 F20101117_AABGUJ miller_j_Page_025.jpg
061c4bfb100ca5ab34ba14cda35950cf
8b4f19d6549c85bed8614407dee1f946df1d420d
2026 F20101117_AABGTV miller_j_Page_127.txt
0349e235c68f8e7429bb834a78e5290b
b7cd1b0c145b6aa26a093867855dacb78a911a7f
F20101117_AABIDG miller_j_Page_111.txt
f22917c7e617f4f32329b495cc82ad11
363cf8a925a51729307a8a95762a4b0ca73f5138
1202 F20101117_AABICS miller_j_Page_093.txt
9725362b369f7c085fbe37c7c5c9553d
fbfed685407e3589a842d8ee4db244cbd4e2be7c
26692 F20101117_AABHYA miller_j_Page_138.pro
d7e53762f5fe30263d0790c96b92c5da
55593f61e7566d5f45d35ba8da3da97171d484fe
52491 F20101117_AABHXM miller_j_Page_120.pro
a77e30fc826e9b63ebb40ef926596de3
51e72f4afbb8b02fb3a2811eb6da80423572ef24
50892 F20101117_AABHWY miller_j_Page_101.pro
3d0efce93267a5d8047c6517e161cbd6
9a6b51b9b5e6c2c29031b1247ba2fc6dbb6bd732
F20101117_AABGUK miller_j_Page_197.tif
b6ba53c4a54e2e905a8b8e069b90c886
4571fea017963c288824989f9fddb2fc21680954
2028 F20101117_AABGTW miller_j_Page_096.txt
5ed198befeba0d674c99d896691f8d75
006faf3e877327a56b3ec46187246770b8e0b041
2115 F20101117_AABHAE miller_j_Page_125.txt
84d7bc03ab8cd742acc317e1f89be645
d8ae12710180e25003f6d5c4000ef0bb01668cd0
1993 F20101117_AABIDH miller_j_Page_112.txt
e0a89aa2411224366e0ec89cc7f2d47e
4719be9bbede80ab50a0533ab724494ec6e76a01
2356 F20101117_AABICT miller_j_Page_094.txt
a47337a17c50a648e3fe5684bd87dae1
3e09416414382965b7f4f8e6d83a376cdea4e690
28211 F20101117_AABHYB miller_j_Page_139.pro
517bf88980dc72d9b419a9ec2499d3fc
0cbdcabd1961868d8e5f7778da1a880310f22e35
50401 F20101117_AABHXN miller_j_Page_121.pro
565898eecf682f1ba94946e9c96a446f
2ee9e5498a4515b9cfe769010eff323384fb24b4
49828 F20101117_AABHWZ miller_j_Page_103.pro
024c272162220b1ef2b067b9e28f0159
2f21572ecea4eacdea6948fa22129b723331baaa
154795 F20101117_AABGUL miller_j_Page_164.jp2
86bee34d2c8119a3ba655ea54bc92431
4568c6fd95de41fd68173fca0c09bd2cc7e3c7ad
36073 F20101117_AABGTX miller_j_Page_184.QC.jpg
d367f79283e879e3a59332d32b938cac
e7a0ef135186bdb8e2b343e74c7f60c20dc9bfd8
1051934 F20101117_AABHAF miller_j_Page_084.jp2
1ec03dbb90b6bb65462835fb57938f5a
19c6206cf2691df62eaeeedbc66b3579592cfd98
1957 F20101117_AABIDI miller_j_Page_113.txt
96c609e65a655d2f7230539f0ab6dc4d
68c6c99993c95b5a4af049a969ed77ed1037c220
2018 F20101117_AABICU miller_j_Page_095.txt
40074ca7177280f488895e67a203885e
c51ef4eb0c68f329611623811246e527c81a6832
31715 F20101117_AABHYC miller_j_Page_140.pro
9ac539084bde45eb86491646d60de3dd
477fe1f7996dc77eb6d3ae12a5d56eb8dd0db0ca
52225 F20101117_AABHXO miller_j_Page_122.pro
1f60854c92cd560baa7aba053b1a2bb1
b6fd6037b2bb0c881dd5ec63db0bf42825254196
28084 F20101117_AABGUM miller_j_Page_080.QC.jpg
ad7a05531887928adefa62013125b7d8
e623d948f3d6c106cf99ed8d0a046e1949e17742
104228 F20101117_AABHAG miller_j_Page_038.jpg
553fa25f8b19bf823cf4119cc3aa3b97
198ee2b99c140c2f0301c7beef669d5015f5cd84
34858 F20101117_AABGVA miller_j_Page_015.QC.jpg
215243921b190a58ae1cb5b527687d57
d7988db7e4c463f1142c05d192d912ffe59f0b70
1931 F20101117_AABIDJ miller_j_Page_114.txt
e5be359778ed03f84eeec1a5ea526ca3
9b7166ef29552ddc4a1fa3c839575392a5f445cb
2174 F20101117_AABICV miller_j_Page_097.txt
b8ed9bff083ab2a2a8b55be89dbd820b
ff757014c748b9115f958d0720fb8ff2f273b09e
34126 F20101117_AABHYD miller_j_Page_141.pro
1fc5bd8e9fab2d1af039e82435efe966
8c6692fb29e179b1b77574124b4be8620456da2f
46949 F20101117_AABHXP miller_j_Page_123.pro
1f9ab309d3d4c581808fb63a1ae8bdb4
049280c9d395ac4bfc401c2fef5eaeb4e2d56226
F20101117_AABGUN miller_j_Page_127.tif
ed3b2f670206799ee3303ec580f844f2
de7e072337e429ddc677180e5fce8b4c047a8794
F20101117_AABGTY miller_j_Page_015.jp2
7c544ff541f4489dd369a66a2c78cbba
594c09853fa7dc3c27da905586feccb75dad0a5e
21848 F20101117_AABHAH miller_j_Page_194.jp2
dd2ad3b43a02b57b7c684486d2e90406
78d9f038209e7a288b043bff430d054964030e79
8431 F20101117_AABGVB miller_j_Page_043thm.jpg
82652a56eb03877ceb853e860fc7aad4
678c3b4aced6035f84ca50ac1caf88c21de7efd4
F20101117_AABIDK miller_j_Page_116.txt
0d50533dc71a203b635726c4606a70cc
a730b1362abbce311c7f4ba120ff527d3e90a988
1986 F20101117_AABICW miller_j_Page_098.txt
469af4f4ae71ea46a7d0ffe4495c36ee
df308b71e73f6858e2a28aca9b5c7f7b4dadb1c6
32399 F20101117_AABHYE miller_j_Page_143.pro
678b750035b24c53e5ddf4c922f2f9ce
6134f0617992248841560c24413794724e575dd3
51474 F20101117_AABHXQ miller_j_Page_124.pro
eae4bace3aad8d2de5a1ade7e16e4821
e912eddff7c23a237d8afae5b1ab9ff4796d19dd
31386 F20101117_AABGUO miller_j_Page_081.QC.jpg
6127017cffd88737b70b0d1feb43c4d8
65bad6dce80e71d5b3ff4ef2a2c8bc322906a664
74143 F20101117_AABGTZ miller_j_Page_089.jpg
e04eff29c8da63a683b2bb116afd4e52
6eca7f5151c6531d5e7714ec7c85187dbd4e2013
179997 F20101117_AABHAI miller_j_Page_160.jp2
b10d24746580f5944fc31e27fe64a7c4
367a4361f542ef50cf932b66f7ea0ce115ec9544
7065 F20101117_AABGVC miller_j_Page_072thm.jpg
d818285ec5afbfdcfd294f039634a99e
6b1fd46253a674a31c0baefba3f979a0677d5566
2127 F20101117_AABIEA miller_j_Page_136.txt
4e3af76f68113bdd94f520f463099ae6
361051819d3470f0449c46800f0627954831047b
2034 F20101117_AABIDL miller_j_Page_117.txt
de48458c893d6962db65a93e57e08540
47df046d7f961c8ae80907875cb4984b2e925d6a
2166 F20101117_AABICX miller_j_Page_100.txt
461176a2232f776673c39c8a77199c94
ba834eda6c3d4f4b488ea4e583876550ee876fa4
33019 F20101117_AABHYF miller_j_Page_144.pro
f67fcbabc7dc4763ae6d7ac025a39b76
beb58eaaafde5d015ca600b1fec1358e9e6c198e
53556 F20101117_AABHXR miller_j_Page_125.pro
068858ea48d11096f697c34e595cd434
25430599ff987ac2b28349ad733a6bd31eec5ef3
1334 F20101117_AABGUP miller_j_Page_143.txt
40daa20b7ff0b9130c86292c2f69c1d9
13fd875ceb0db5ff06f5c5167356b7aafc5639a8
8650 F20101117_AABHAJ miller_j_Page_159.QC.jpg
15920fcc0e37f9fadf0445eaea3347ba
fcac842998a4a37a2c27272bcaeefd76ac79a11b
8389 F20101117_AABGVD miller_j_Page_055thm.jpg
5cb09137b579d413478992f6d7f3d591
9e634f4ec9f668faa331ee8443b8401a57a9e4c8
813 F20101117_AABIEB miller_j_Page_137.txt
fbd673ca3cf9d1f652cdaa5fa9329a66
b1b0b0eeb0b228a89efd7caccf92d201292c157c
2014 F20101117_AABIDM miller_j_Page_118.txt
0a023cd7872336a94613a52af68e61d5
177cea6320e86e64a82d0dc89b30eb96b294135e
F20101117_AABICY miller_j_Page_101.txt
4d9a5cea98c46292fd706487f8c01dea
c858247cf5b917405234652c6626b1f3eaeeb915
13778 F20101117_AABHYG miller_j_Page_146.pro
680e67111192d0af3a4833a37917b782
f525a9d99f58c4a1446319d119c2996a0b031b46
51509 F20101117_AABHXS miller_j_Page_126.pro
979d3fd8276ff62a2645cfef44a9e587
2528e2527ad4fd2af998adbad03a68ea46d9ab13
F20101117_AABGUQ miller_j_Page_012.tif
e9d226587de623acb4d844cf0ae72f5c
fab5dd2b6cc49dce3ffaa8a8979cf4f5f589dbaf
19180 F20101117_AABHAK miller_j_Page_164.jpg
df37753a20d44cb5424b1c557ab5fc65
645fee6832dde640262053c5d50914657c09e35b
2722 F20101117_AABGVE miller_j_Page_084.txt
2339394fc785191748f5d0a47ffde59a
6bad43c8f1b78e901b048f9d4e47bbab6a1f1790
1331 F20101117_AABIEC miller_j_Page_138.txt
7aad4ad4c544fdb75691f983274da955
6d0470ddcb044e8db4f41b85d5eab65b9c6e3849
2155 F20101117_AABIDN miller_j_Page_119.txt
2a6bca8203f8ea590ce90710fa938a6f
256f5dcefaef0c4bd0d415572ac119712dfab4ef
1907 F20101117_AABICZ miller_j_Page_102.txt
1069cc2a5d3e080368c90376173649f2
534a3d373f716a1a0f05beb9430ea19357877450
33816 F20101117_AABHYH miller_j_Page_147.pro
945a5dc3a82956407dacfb8f79573884
335fb85777a6a554892cfdbb13db994c74afd3c8
52466 F20101117_AABHXT miller_j_Page_129.pro
8501ded2eb8229fcbfecc05f4d3d5030
291248d9b3bdaa5d3cf999106c752499bc409b2b
F20101117_AABGUR miller_j_Page_149.tif
dbe4852e3ad1997fed08aa1912798c85
b57be58d15982ce28337d7476a6eb0155f726d31
25664 F20101117_AABHBA miller_j_Page_072.QC.jpg
632e5a5e3ec32b503b69313a5da8758a
62f575ca00c0e2a18170e1a5e3f103877dd29197
F20101117_AABHAL miller_j_Page_120.tif
2d4d27a4564984bc18c244bf0493a8a8
780287f0ce1b277f0c9b612333429de1aa526e68
33606 F20101117_AABGVF miller_j_Page_077.QC.jpg
8d4e9670ea9919037608c9918a14b53c
ef81d39969a6a9c5915defb1a37361f76162f1c4
1481 F20101117_AABIED miller_j_Page_139.txt
7b6ed781fd9d25c3030beffb607983ff
e97ab906decea0b5d7803f00a873da7efe5a65ad
2077 F20101117_AABIDO miller_j_Page_120.txt
ac13a3626cac1eaf1372653c91e2f719
a8e557ba48ddf2ee301fc6963cd83a9feb9b1967
34782 F20101117_AABHYI miller_j_Page_148.pro
cd6ba9b65a1811b5a1d47e2c04cc5423
a0fb21252216130aa1ef4a56c6ac8c202c00d264
52398 F20101117_AABHXU miller_j_Page_130.pro
98f58096efd56e5cab57e8e638661888
53bf8bb0592318d47835a9984ba437d513863169
8641 F20101117_AABGUS miller_j_Page_136thm.jpg
33675d52196a0abcdd38ecd812c60b57
974af2657efe2c98bbbe0d3ca06235773cb9d914
85376 F20101117_AABHBB miller_j_Page_141.jpg
284e6f3e54fbdfec9bdf0daea9462658
a99b2f398e8fb032c9384aab1e15dca740e65a8e
34994 F20101117_AABHAM miller_j_Page_059.QC.jpg
80f4db58a3a358b9886a25deb3ddb357
2f68c1262fdcb6fe2863ba2f4a3a56bf5e9a6bed
188223 F20101117_AABGVG miller_j_Page_172.jp2
6033f03305c45e9c18176b656d8b2741
41e5b57f003d05291979b037374cdbbdbb2d5182
1325 F20101117_AABIEE miller_j_Page_140.txt
153118c2231b68de3e190044b21b3339
5b88080c1ff2fa06305027521bfe2236d7ebaec4
1998 F20101117_AABIDP miller_j_Page_121.txt
3c489941136ccf4341ac0f7ced230c60
2dc0610a1214506c624280edbec28c59401f2bcb
32546 F20101117_AABHYJ miller_j_Page_149.pro
cede280d1ade038a1cc882351d6bb58e
098b664af58cc92093ebf1e21a66d3b77f7dca81
51214 F20101117_AABHXV miller_j_Page_132.pro
be51243968a2901f93c17cd03e151fd8
6a148b5cffac1c50e542cb216f93555f09322b3a
34708 F20101117_AABGUT miller_j_Page_131.QC.jpg
dba1cecaa3111fc4c9267293e17fb23a
956ab2d3c7edb6d48445ab1e9085349fcbf6cdc2
156904 F20101117_AABHBC miller_j_Page_166.jp2
b254d01ed6aa9b3a03a8bf67d1833b03
c88d3b623367604e4f31b66aca7eeec49a7c6e21
F20101117_AABHAN miller_j_Page_050.tif
71834f622dee806217d14a4513554dc3
d0cb053812d114473d5624a933c5c2329715dd84
109609 F20101117_AABGVH miller_j_Page_061.jpg
f1ad232f6e4ae12a14bc9134c1e4c38c
4efb950ed6ed66172ba02c78265616f03e372aaa
1500 F20101117_AABIEF miller_j_Page_141.txt
a47d345294d9fca3993eaa13fa1efd1a
355cd32c631bd7ce2147f319ed47cb5e11486439
1864 F20101117_AABIDQ miller_j_Page_123.txt
4be8215117431e852671dc572eb39f0a
02eb67396f86fb2227c7fa08131e9bfb1ffd7d15
32383 F20101117_AABHYK miller_j_Page_150.pro
d0ae3dc11686b87f3c9246f75eed3d1d
a435830ad020255e1c704d97e4f9006f737c969f
54817 F20101117_AABHXW miller_j_Page_134.pro
23732bebb799a64b2e1553742bba8066
e73eaba892ad7964e8f755659a6240baeab6c3a8
F20101117_AABGUU miller_j_Page_116.tif
710c2e2cb05551c6a7d883bce0e8f935
938091953554e3d4a86ce49a81059f88d2eace89
F20101117_AABHBD miller_j_Page_146.tif
bfbe9324b4de3ed399f9cb8c6ea3e713
58db47b74fa4b7b95495e19a28ce5d92ea25b9f9
1051976 F20101117_AABHAO miller_j_Page_010.jp2
57f74082147bec9d295ff741e4af939c
ccfc9bb0a4d032dd2b536cfcf7302ec4ab9fcc25
44625 F20101117_AABGVI miller_j_Page_018.pro
1c453645f1226404113971f113a7f190
0b6962205617d8e28ff3bf29ae756ecafff21dcb
1425 F20101117_AABIEG miller_j_Page_142.txt
12bda30c100c34654e009594fa23da81
e71fe5e87f34158a5c87fa7506420ec252759f8f
2039 F20101117_AABIDR miller_j_Page_124.txt
e405a91d9f28a02e68d65769146f32e7
8e38b231ed8fe7dd7fb254fad3a676c5574bdb42
32638 F20101117_AABHYL miller_j_Page_151.pro
09598e525af68be21b1758383fccc8f5
8dfb7ae4b10443cf54ab0547dce0c8d95f7f031e
53066 F20101117_AABHXX miller_j_Page_135.pro
174bad3495eba3d67f1949e365fe69f3
764ba3f3a7f1f561bf3ef28ede2cc37d199d5cd7
6743 F20101117_AABGUV miller_j_Page_166.QC.jpg
940ac44f4eb6b4d2930b52132e6a788f
1838b11e9522c9a405fa489838195e8590b563c4
F20101117_AABHBE miller_j_Page_183.tif
5c37d4117877cc06514d173837d18f0b
cf9d062344aaee755237be90c8ceb947f5f8cdbe
76192 F20101117_AABHAP miller_j_Page_047.jpg
d8c3caf71d04e13cbd7812dc6f5c6cb7
bbd6c526c9e2d501c6c18ac8f7d5027537ff0f50
49932 F20101117_AABGVJ miller_j_Page_098.pro
7d5f07198383469a7fe2685101acff6b
37b1f801b888d6432a08b64c421a38e13e74614b
2006 F20101117_AABIDS miller_j_Page_128.txt
b94bbbcca608abb8f779c0d62dfa1087
6af0da4b7faf1c9bb56e4d4ae9e428a007df7a53
5555 F20101117_AABHZA miller_j_Page_168.pro
aad3a84d1e4970a7ccf1c21092fa320d
e46e8dd46b581e0d16e21defc671ce0e21d08229
33350 F20101117_AABHYM miller_j_Page_152.pro
949adf01cab243dff5672acd670e289f
56d1068ac685bdb683930e0629031d06de3f1a37
53427 F20101117_AABHXY miller_j_Page_136.pro
73d0b0fb791e69d9c343679ad1a09009
df6633acbbac5e8a96e4aa3394c8cccbdec50303
80922 F20101117_AABGUW miller_j_Page_150.jpg
f199fa7b40498ced7c091c84a68a956c
3b73346a1d9c35a29ac2ab68f08ebba481028283
114414 F20101117_AABHAQ miller_j_Page_020.jpg
eb2bd7050be77494bb049327508f36e3
455c83926bc7be286f3037273206b93977115089
39187 F20101117_AABGVK miller_j_Page_047.pro
062bd91bbdaa0de2c02bd8d2fbd1826a
ea00e87fea13a23d5816f4916f841824119d1496
1360 F20101117_AABIEH miller_j_Page_144.txt
7ca73968ea3a1a3d73cf9f7d6f516674
4580797db3c9f03d18ab7bacaad70cab230f73c2
F20101117_AABIDT miller_j_Page_129.txt
322a7aa7ef9435a15a4a89682540210a
2c890261fb233f561281a59d63680d780667d5f6
6196 F20101117_AABHZB miller_j_Page_169.pro
a630225623bca33218bb432e1e7ba5e4
8564d6dd113040d4ee2ec318ae7b35fbc38d68a6
33852 F20101117_AABHYN miller_j_Page_153.pro
94e4bf5ba0d335ada56a621c4fa9202f
8f75974d81e75c113d7b1cefe7a1a6d7d333afd8
20319 F20101117_AABHXZ miller_j_Page_137.pro
28afc6ff99e994b1f4ac9550d3f1c6f0
0e7b04b29158305231ce13bd47a6004abb028b81
107813 F20101117_AABGUX miller_j_Page_126.jpg
0308ed3519760091e0eca79c92a1f48e
babf55466d3c8b86b8140011bfcae9812dcf10cf
1043726 F20101117_AABHBF miller_j_Page_186.jp2
24f5d004aeaf85b029bbd910467da052
6839cf4c4bbfd9ca23742dc17a7248d8d86842ce
52202 F20101117_AABHAR miller_j_Page_023.pro
8c04552510d995b1f5c444be62592593
251cb1861e91e4426c1cec6c37dd0952c85727dd
F20101117_AABGVL miller_j_Page_113.tif
d746abebfebaaa6eb340281347733453
db7221e6744d50221babf5a472d05df0d5a5f53a
1817 F20101117_AABIEI miller_j_Page_145.txt
6439f92edeb2e6562ff7b296c2b44a11
b43dcbcbff8be18d6acc937aae86b9800b88fa0d
F20101117_AABIDU miller_j_Page_130.txt
f52c4b7d72f8b822c5da2edad55efcde
d6efe211bddbc6000cf0b990b9fc376736d4882a
5921 F20101117_AABHZC miller_j_Page_170.pro
4d4bb0b46a5ca75278103985a767fb31
2c1d7721771ac2633c5c76a08f16a60d56393580
33061 F20101117_AABHYO miller_j_Page_154.pro
a43ab7f7ba9e1dd21a5aca1a1789791f
636f8d2bc78cf82cb7fb78e03732215c776ddc41
183133 F20101117_AABGUY miller_j_Page_159.jp2
6a3f9c871a3c82f7924ac6a7ab82ef33
c4b57151f4a72daeec45bef4b58800db5846231f
35022 F20101117_AABHBG miller_j_Page_145.pro
4aafe380963c8b59a35aaa9ccc16bc77
7bb8f63a9a365ad980e910f6f3e59b13b3833acd
21403 F20101117_AABGWA miller_j_Page_163.jpg
6cd2f2867865e6a353ef243c887c2206
c80dca10d2471c78d60f5803cddc819c38e1d795
100207 F20101117_AABHAS miller_j_Page_102.jpg
6db89b111ee8bdaef8bcf34f88689c61
33a6f7d4ca18556fe0e336d283ee7b7a90126680
104771 F20101117_AABGVM miller_j_Page_043.jpg
d10d82e4b912be827f902079509ebebd
669e01653e04149cb88e22f5a2541a397d507942
551 F20101117_AABIEJ miller_j_Page_146.txt
640c62d58be872184483c9b4285a77cd
148ea92148537d3e9aeb8ce2d78576cc053a6a3d
F20101117_AABIDV miller_j_Page_131.txt
918239ecb507a660c1fb06ae85fbde3d
f0ff96eae7bf837fdb2c499d6e3ccce94087c891
5721 F20101117_AABHZD miller_j_Page_171.pro
50c48aa8d7828f8563030c714978af5c
6068bb2c7169ecfd61f5b928097d7639e7405bb9
39218 F20101117_AABHYP miller_j_Page_155.pro
49ee41def5b00597090cc3a922ebf7a5
c5389673305354aa5e42ec7adb53a90ebbfb1483
8710 F20101117_AABHBH miller_j_Page_001.pro
c9e520cea9d0e4207178713212b7a162
f637d46f2ebf28bd795a040f3cd2524cb6bb8e6e
8795 F20101117_AABGWB miller_j_Page_032thm.jpg
0b288c84d7511ff3224a1812714da66b
78756ff2905c8a4ecc0ecb76f20d548a7050058d
2121 F20101117_AABHAT miller_j_Page_076.txt
6b9f8d646cd69da0565c845247c9d84c
635a64940d5f7641c4946236cd27eee7d939188e
F20101117_AABGVN miller_j_Page_152.tif
68b64a4a5790608e2080a641820b8f1c
56fa436ec6c43db7064403a793d556ef4b77f583
1639 F20101117_AABIEK miller_j_Page_147.txt
3fb8b9e3cda99eded7fcb57873ad982c
93bc213175a9c0faaad674bfebf0293c31bc5931
2027 F20101117_AABIDW miller_j_Page_132.txt
ffaf8ee510e20a5db51bb3f5c5f3243e
13dde9019fcdde0745d4a203d0abc7d89110f8a7
7328 F20101117_AABHZE miller_j_Page_172.pro
bd7621e46b70e2d342643989410bd420
de31b63477194fcc7cd37bd56e4f5979dbff3b82
36849 F20101117_AABHYQ miller_j_Page_156.pro
fad66425e0c384c5c8e50a22a8a8c7c4
6cd95b0533187ff3d6cf3fbf33a5a0d0b7a9302a
1051985 F20101117_AABGUZ miller_j_Page_109.jp2
a48f6dce273fa3705f934d4292710225
6f66474a47020b3bf4f554546de539d033803480
F20101117_AABHBI miller_j_Page_185.txt
2cfe3e34cf9ad959ee650793545bb3ff
c9776c034240ccdfdd6d99952844d368f4e7c836
F20101117_AABGWC miller_j_Page_173.tif
aedde8ffa1239dedb8f537decb40f05f
f5418e058fc5027e1cac6d666c0dbbe7c7a4557c
F20101117_AABHAU miller_j_Page_073.pro
edc074bcd5d1a78ed561d733d61c225c
5ec565aabaa36b9a33272626c7160f58739a598b
2074 F20101117_AABGVO miller_j_Page_027.txt
1c72795c90f54e1fb6ff039254aee104
2036a4ea68ec2c43af00d4d78877c6d4864eabb2
685 F20101117_AABIFA miller_j_Page_173.txt
7454a225595501ea79067a049c8c9419
942bf14f33b7329f96d1db517f00f41ef1ee1257
1329 F20101117_AABIEL miller_j_Page_149.txt
9e2d9e82a6ff707a9144978f6aac05d1
2d589f1f334a4db0efcdbc3f2ca8f8c8d17c27f6
2097 F20101117_AABIDX miller_j_Page_133.txt
04be37e6fc8d3e62eaa1da8e8093221e
448df2896cf4f2b452098d7fcb3aae2a921252cf
8790 F20101117_AABHZF miller_j_Page_173.pro
01ab9397f53b758ab00094a6751dbdaa
b6e77cec165c4cf09d421b8c343ad006ab1f767c
8594 F20101117_AABHYR miller_j_Page_157.pro
470aab700d0b88f96e0ccaef7a2376cc
652b07896e5891c6a478068ef325362efc18983d
F20101117_AABHBJ miller_j_Page_011.jp2
3ac0a3b9cb086da90157e3cf5d4610ce
72bbe37241347c7f3a3f5c7f9797257b53cf431c
1051947 F20101117_AABGWD miller_j_Page_115.jp2
f3428cda188a3fa33f283b30f0450f28
9f8c09265b1abb5ffe8d0dd7f710b965bea041ef
83761 F20101117_AABHAV miller_j_Page_145.jpg
4b7deae3dee0339633dc6a21bf33e955
4489c4492db436281e52a9d3fe4707e4f8ec44ad
82247 F20101117_AABGVP miller_j_Page_191.jpg
439fb782b1865459e7e6861924e7b175
ab568189ecccb4c604b0dd44b05b2daffa3e148c
714 F20101117_AABIFB miller_j_Page_175.txt
191dde07d22728771b28eba3f9539bf9
5e5065151b1d81dbca623941e5af8d9b5a71d720
1658 F20101117_AABIEM miller_j_Page_150.txt
463222d42de0d369b7194272b10c31f3
32f69f4d0c8e1663f961fca20b05dcf52d05c0e2
2158 F20101117_AABIDY miller_j_Page_134.txt
6194e1c8431fe5199db103a88933c961
29fabe708f7799c6e668c6f5ebeeb71a23074b54
9040 F20101117_AABHZG miller_j_Page_175.pro
82dac98e3464e33911aefa24c9775a6f
c156e97b8906c2d1d28b63ad75e72937fa0b225c
5319 F20101117_AABHYS miller_j_Page_159.pro
23c9b83e8cf0923147d76fd55a847a09
e179a2d2996e10eb6e4be95cea0ca40641f406b7
55119 F20101117_AABHBK miller_j_Page_100.pro
0287402292a76b74f675a55d2ece4430
84b2016f90278a59eb2cfe231713c4afb1690081
8985 F20101117_AABGWE miller_j_Page_060thm.jpg
a816a072acd045835352617bd49d0ec6
e0faa71910cf6d7231737f3d284bfedfea9fb786
1328 F20101117_AABHAW miller_j_Page_002.pro
feeeb790225945f7e9b5f8edb51a1b32
22d355100acd2bc8c7b3350fbafef33c8b432d53
9006 F20101117_AABGVQ miller_j_Page_058thm.jpg
e86028d10f5cfdc316d562a0cb017ff3
a026d913e500a020e890ff2a8d0a0c577220ea7f
2288 F20101117_AABIFC miller_j_Page_176.txt
f65782e062d75b425eb00e268cacc2c2
391668cc3e20edb7aa6125ff05b870318e4671c5
1616 F20101117_AABIEN miller_j_Page_152.txt
9c7924488eb5fa885d07b2c1ebb68276
a99b8d548ed8825cbee607fc4c2b42e76bf6b668
2114 F20101117_AABIDZ miller_j_Page_135.txt
ba6916d33eb4177c52070552996861ea
6a14809b9528d626cce98b33b9b6aa2390fc70e9
54963 F20101117_AABHZH miller_j_Page_176.pro
3efea33c66c93dcb6f363677a2a4cfd3
df399ccbf10babf17df664e64391942c2b2b3dfc
4854 F20101117_AABHYT miller_j_Page_160.pro
0c5948058045563c7db9665340012b35
419de158180d71c2ef3db2cc0c419415c33159c1
55025 F20101117_AABHBL miller_j_Page_058.pro
2eff93532737f8bd7f0f67fec724135d
89f70b06e603d017f281956db267d4491f0e311a
56300 F20101117_AABGWF miller_j_Page_024.pro
4b2506177c9cbf7bdaad03322dcfbb71
41eaec4ddd684fbc7405fed60a62e56320f3d8cd
784620 F20101117_AABHAX miller_j_Page_150.jp2
a20c166072d4ab0de786549858e62f15
6567a5a47a604ee980b89a6a186f649f5f4f4e07
7834 F20101117_AABGVR miller_j_Page_147thm.jpg
b91536859efdb6c0394e1f5e6c085c15
0f5204101fb039e06f23bfe34736ec474f6b1006
8223 F20101117_AABHCA miller_j_Page_062thm.jpg
e7d619ea5d37f8eb1708a54a6d175a88
bbc921dd9c002679a103f9b29bfa73d588bdffb5
2233 F20101117_AABIFD miller_j_Page_177.txt
78f3c4190724a5c9e7a9bf1c0642b468
475f4c652cab6699180c5d75648c8df62265f78f
1683 F20101117_AABIEO miller_j_Page_155.txt
cd7ea94bb5b4c641bf8120e0561c32a6
3f6ee07bb5fde969725828277e88630f5add0210
56534 F20101117_AABHZI miller_j_Page_177.pro
25fbe79daecededa700e786074c2b14f
2c23ff165fc7c62e62def91873fc863bb29e6e53
6487 F20101117_AABHYU miller_j_Page_161.pro
37ee6dcdfbd9cad856c925bc12047876
7f406cd43606561a2afbc6323996237fd05bf77f
1051957 F20101117_AABHBM miller_j_Page_099.jp2
913637206c5495ea92444c42a7c959b3
0f13355875e681f33302f6751c917bc7c0b545a1
1421 F20101117_AABGWG miller_j_Page_151.txt
ca6688642eb0d4bd9dbbcef0f3510bd9
0d9b19ee98796e8868f6330c13c40505451c6190
36663 F20101117_AABHAY miller_j_Page_177.QC.jpg
c2e13fdfa1a66cf8f367b7489f27d225
eaa0cc58ffe602394dcf0db8666db670ba8c38be
1051972 F20101117_AABGVS miller_j_Page_176.jp2
2d802cd558dadda81c9dda75a4923729
771eb718b77c838fd54b92da90b1d3c80b4771c4
8978 F20101117_AABHCB miller_j_Page_027thm.jpg
bfa281e9a9c86707ae2edf63b636e36d
1e77670a4fb3cff0cf5b7f45c0bbf8040a3ea9a7
2137 F20101117_AABIFE miller_j_Page_178.txt
640a51e74584051e17b41114d6475211
450576e1639e7445648de9a8acb69708096da4d1
1568 F20101117_AABIEP miller_j_Page_156.txt
1296f16d776d126c75daadc26700dda4
b7fc5bb1f7c6019777c32cf79225d3aa55090716
52269 F20101117_AABHZJ miller_j_Page_179.pro
d06933598640b095ca0e765738badeee
0d60c55aee8489b541f591fe22894ce47bdfd588
6452 F20101117_AABHYV miller_j_Page_162.pro
52007f1f3162bb67ad0981d566963d5c
11dd0daeb8cda9d6eeaa485308a0a8065c7318d8
51948 F20101117_AABHBN miller_j_Page_108.pro
dc1dacf2a8149de95f88419dc718b111
ccc61a07bbc424f07e40d46cfb1816927fec2a07
112054 F20101117_AABGWH miller_j_Page_063.jpg
03fdb9fd06191a20755992ac047c4a39
dae42b5e0bed3a9dbee4869bc8e2db9419d7a33e
F20101117_AABHAZ miller_j_Page_122.txt
90d61ad64221ca231bd8cb6aaf2ba67d
85650a8222b6eae24cca0189d2d0e12061a2b317
F20101117_AABGVT miller_j_Page_035.tif
ed94ffb69cfc5ca145a6843ffd77f3c4
8b872aa7191bc61356f082e49e159659b0ec8d07
7910 F20101117_AABHCC miller_j_Page_028thm.jpg
14c413fc84963fff16fe7827ba5bfd6e
b81f13547061f7f811dcfbb28e9e3593ec542e33
F20101117_AABIFF miller_j_Page_180.txt
fe909900c0e696ee1a1bdc67ca1908d4
567de7b84b3f4d82924c3687b2ee7fd96ceacd32
362 F20101117_AABIEQ miller_j_Page_157.txt
7fe155c4373818bb9f25a3798904e408
80e5a85fe76a481b8814ef53d73f982a1f028592
50591 F20101117_AABHZK miller_j_Page_182.pro
89b361b7221cb0671afe6468bd5dddc3
55f1ae8be9d6ef18f7ebc87e7031f0f4b91dff7e
4829 F20101117_AABHYW miller_j_Page_164.pro
9f34531d089e6d70bdb935ebc9e6a663
9e978f06cf8bdb03486a0100d75dc887a40e1610
F20101117_AABHBO miller_j_Page_030.tif
c00a2bf7021c2c3d111ccf554e71522b
00fbaa2d9d86326ba1e7457fdc64d780efd7089a
F20101117_AABGWI miller_j_Page_034.jp2
da9f2db2cb2769271fdf97e5b8104565
40c82f79d85ff8a0311ba2f7d4abca1376b2621f
89944 F20101117_AABGVU miller_j_Page_029.jpg
58685cba5b2dd7e90eb293fcf47db31c
20292e7cd9d8286518fd22f9501d6ccc786dc64e
8300 F20101117_AABHCD miller_j_Page_022thm.jpg
f723266723ce059c4b349aefc46eff44
d8e06f5421292c7993c9f39839a38d3abf3433a3
2186 F20101117_AABIFG miller_j_Page_181.txt
0a5727044a90c9776875e3d9d2807ab8
4b0b01f643f593666dcd9dfdee8caefc606c3f1d
411 F20101117_AABIER miller_j_Page_159.txt
a086edffefdb28e1d2a1457383dbd0fe
8e985a63ebcc5bcfb79134dd3b4f05da25e4e637
54641 F20101117_AABHZL miller_j_Page_183.pro
abe155930fcc88daeaa0af9b44deaef2
c52254d1cb96b7e1db6ae6ec8c49a44a1021edbf
5992 F20101117_AABHYX miller_j_Page_165.pro
60235ccef69b21a7316f30037b01637a
2cdd2616fdaae2ea79962fac9f374ea5b7275b16
48486 F20101117_AABHBP miller_j_Page_114.pro
02f21801e0b4fe51e3a07714fd5f7641
f7e6cc3b6ee3e0c3c16d0d3edda13b5fb0b544a1
1676 F20101117_AABGWJ miller_j_Page_148.txt
444a62002be2c3aa3993b4677433913c
9dac247c7d9f22278e9ce3a84ebea5e8e4018bfc
6866 F20101117_AABGVV miller_j_Page_091thm.jpg
53c772a456cc526337d7bd9e189677b9
e44b1173139c57523a463772313a6d47fcadb130
24129 F20101117_AABHCE miller_j_Page_192.jpg
bd3a1f690c0864234fbe3ddc4aaedc29
32b3a2170583dc6c19efbda5c8eb5c210db5a7f5
2007 F20101117_AABIFH miller_j_Page_182.txt
6c0dcf367e5d2d54923e637795460976
a81ecc099d47387cde98e1487752366a41cd4ee6
245 F20101117_AABIES miller_j_Page_160.txt
750a5a712944f38ac8bfc1270375c404
db22acd95c9ae2293be2bfdacc60821eb5bd266e
55103 F20101117_AABHZM miller_j_Page_184.pro
b9696f4341e63f9b9e0c4ba12b91dd5a
ecf68206c42cae51d9640e222805e774a0bae21a
6508 F20101117_AABHYY miller_j_Page_166.pro
0b580786760ed895393b51c4bef3b2bc
bbd0303701442fe2d18452e48636cab27141b410
8669 F20101117_AABHBQ miller_j_Page_131thm.jpg
d13c7f8c3f36e814e7b14b9593514e0b
dc3e6d45fee103fb916a5ea91b1c770ab364ba92
7298 F20101117_AABGWK miller_j_Page_174.pro
1c60ebf898e4e38e7148087bdadb1f3a
1c80a46a5ff7e1094ec6e41b5de1aa43febd98c7
33638 F20101117_AABGVW miller_j_Page_043.QC.jpg
17c173c22e03c7cbe8e052ae292e8655
50368281523c54d242d0d36741b0b201896e4efd
29248 F20101117_AABHCF miller_j_Page_010.pro
355eeafee3a0cec3f32a0b0221f1fa17
c33e1afb7f60a0f266c35aafae3e8ff47f30d27d
349 F20101117_AABIET miller_j_Page_165.txt
5abe4ca4d9621eb1b071e99740110e6a
826ee55ec811a7fd3d6fabe49968b70850ceee50
55142 F20101117_AABHZN miller_j_Page_185.pro
bb0cb052266f878dfd41b4ee9cdbd949
d8f47c5c56b49ceda431b59725ac4e3ee1965bf4
6125 F20101117_AABHYZ miller_j_Page_167.pro
422aa7b85aa0f2ca30f52b8bb25801fb
cad9dec30926b3f32f86add2b7c2a664838409f5
F20101117_AABHBR miller_j_Page_106.tif
e66e1a5ba753a0b8f0c95e274d1f2509
d52306e92d559c36961208c2da0a1f28d9882d92
104489 F20101117_AABGWL miller_j_Page_062.jpg
e3ecdf984e3d9c3974a5a4c18dc45ad8
5d9e9dd83fe26d75372a9250c782cafdb8ed826c
58612 F20101117_AABGVX miller_j_Page_199.pro
fe842177744966d894b19fa29a980eab
57c4886220b0d458047d44feb2a84099b7133c86
2156 F20101117_AABIFI miller_j_Page_183.txt
5a37c0311a74f4f65b032d871fed3320
b6360802fcbb9ad8c2f9ee33177803f514b840b6
430 F20101117_AABIEU miller_j_Page_166.txt
467caaad43e3b948dcdf256038a6aa37
de31c24969087981bcde8f3bcd2dde8c3c36bef1
52045 F20101117_AABHZO miller_j_Page_187.pro
a87c21f388b7a7e000d8efd1c7d423cd
dad74a23c77e115e0379e9d70b8a061ba265b784
F20101117_AABGXA miller_j_Page_128.tif
4973e9bc89fd9f87f83fe14f1432eab6
c25e1c1e901d0dbeb5b2d3a41c745f3292f1df77
51318 F20101117_AABHBS miller_j_Page_083.pro
c70595c973923e5def86f2ce12eca438
3043a459d33ba3327a30f4f62e8b3a7075628cdd
1574 F20101117_AABGWM miller_j_Page_154.txt
2948cbdc8040c27126d3d1fb04868ae6
2e494ab94395591d5ef77f0023efe01a47ee572b
7816 F20101117_AABGVY miller_j_Page_082thm.jpg
7352b4a278fec7267b769d7fae560318
46ce63261d69ab00f965b0bbc56ec6160c97ce2a
F20101117_AABHCG miller_j_Page_060.tif
5cacc7610ede6239cdfc0c72298f7239
cac3a974a3810705273bc281e25f0af867703010
2163 F20101117_AABIFJ miller_j_Page_184.txt
3bda3e015370c9797763a404e5fe7c51
dfb443d5dad49dcb74349c659539325ac33cc592
290 F20101117_AABIEV miller_j_Page_168.txt
3c736f46439659ee63f17b6d26fec22e
a9f06f7a8047c7fa9dda1cc9c131b18c7b07f88e
53756 F20101117_AABHZP miller_j_Page_188.pro
3a730c08461192336d1fe1fef1d4e552
d260947ed2e26ce39032162f8b1cab53f4b29fbf
114387 F20101117_AABGXB miller_j_Page_021.jpg
12eadd5659ceade2b4df906bebcd5227
388da9d272dc91b9a93b6b47b0f08aba88b21d2d
110937 F20101117_AABHBT miller_j_Page_058.jpg
10ebaa1aeb26e6c321b522a79170c5cd
3590824953913ba94930516c8a9465ddb7e0bf68
1051979 F20101117_AABGWN miller_j_Page_021.jp2
fc31c835d897873f53426f1b08286429
f77a4a8c70229bb329bec8f03200b9851c422e42
F20101117_AABGVZ miller_j_Page_112.tif
c6cb0e315e94c26571d2d9fd947e5877
a19850d511fe73eac1a6d6076ed3f3d49200fe09
F20101117_AABHCH miller_j_Page_118.tif
ddbcb6bbcfbe27b63654838401317790
561a16c9229280deb5014ca73f7724647c041ce5
F20101117_AABIFK miller_j_Page_186.txt
2413c693fd5a9b2cf19db945cf27b898
935530c65250f308a64c41d34a519f93790b37f8
501 F20101117_AABIEW miller_j_Page_169.txt
182e93ae1a8561bd1e92f373e7b0fcb3
b191dba540fbd1ee867828fea019e825a8e9f3dc
51216 F20101117_AABHZQ miller_j_Page_189.pro
1e641748ff75ce44331df65992541621
f761a25f8e56ab92079ceab6171d831f403ecd94
1977 F20101117_AABGXC miller_j_Page_103.txt
71cba98dddfef485c73e8753f8e1891a
09abdcd6963155d0aea5c8fb8a6c0e0df88f4aab
48139 F20101117_AABHBU miller_j_Page_102.pro
eedcd0843638f8eb1051ae6989bed838
d9cebdf524e9f6c1551f8a58f7b3e453447dbd80
6838 F20101117_AABGWO miller_j_Page_037thm.jpg
e61475a247b3fcc9ead1f4125e926362
14bd28c8d4252b1aff51fc125d8a9778fb006456
53737 F20101117_AABHCI miller_j_Page_115.pro
f27176fafdec790b5bc7443a77583e43
a864cb0dd2bf2b15691a536f3d1a6966083661c2
28518 F20101117_AABIGA miller_j_Page_155.QC.jpg
0e44a29c247a6c0c8d0b93d43122cc32
a7d8ad9e44c82a5ccbea3ad2a736329e6d0bd35b
F20101117_AABIFL miller_j_Page_188.txt
fd4e8d3a3c8b37c55388da40a9609161
1fa59ecdcf3047fecd67a5eb8e8f9fe0e8f059aa
395 F20101117_AABIEX miller_j_Page_170.txt
41aaec7ed6c4ef5f989f477f11d7f5eb
8312a0f0e3f4276f839ff2373d5ecd4010eb8192
56078 F20101117_AABHZR miller_j_Page_190.pro
211b3aae265afac539fba177295e013f
ca026620044bf6dddc544bf9faa334294469b681
F20101117_AABGXD miller_j_Page_123.jp2
fb436e99efaf57741d94f4c0e7d7e0b7
923ba62f114c3e0090eedef67af470eaa6998835
1767 F20101117_AABHBV miller_j_Page_088.txt
f1fa4ade3324fad2bda5f33cc906cbab
16b86f98bcaeb20c1f34d22e7a33039bc54db934
46822 F20101117_AABGWP miller_j_Page_022.pro
6d8d288c7d159e345df7ef183f1fa62d
7af14e11e57f4db6cf76c3c614a607512b70636a
1556 F20101117_AABHCJ miller_j_Page_007.txt
e81bb02006a2155bb442929b8bc0ffd1
6f663b050c717bf504b16959374be7891530a1e8
35694 F20101117_AABIGB miller_j_Page_040.QC.jpg
d273467d52d9e19ee6b8655d992090e6
48f9e777ceaea1088363e2052c114198ad7d30cc
F20101117_AABIFM miller_j_Page_189.txt
412388119fd1d0fd37fec9fced0f25f1
5a16e404e3be5f661cd46a379927be694582e633
487 F20101117_AABIEY miller_j_Page_171.txt
10b69b5222c62410feed060ae7c378a5
dcb93645b8c1a39378d896319d1220cf0975327b
39943 F20101117_AABHZS miller_j_Page_191.pro
cf97a5615a739d0c959c75d491e16eb8
04d47480fdf8386e27a3f0b26a703b6826bd4be3
F20101117_AABGXE miller_j_Page_098.tif
735a47d8686fdf19bb1d2345dea4c8ad
f8a0139d481614f17635b3d3bdd830d375fc19fc
143439 F20101117_AABHBW miller_j_Page_009.jpg
7e992b91031ea05587d06b77b545f5b6
065291c24768ccf6876407d0aa3baf8dc9b201ab
58279 F20101117_AABGWQ miller_j_Page_020.pro
4792da51aa96c8aeb8d73c7fe0bfefed
4f6535989e4318d9ce196ba3c9f9e5384cb3a699
482 F20101117_AABHCK miller_j_Page_163.txt
1f40f2a3f6c245a94c2d14e0398e01fd
aa644f28a4f858b101e456a38be7c5a389cdaed0
33979 F20101117_AABIGC miller_j_Page_115.QC.jpg
6cc794d3c64363ee8faf5b0b099f1fde
ab014a16e3ad39eb8c5b7bc5b40c3b59829f6896
1628 F20101117_AABIFN miller_j_Page_191.txt
862796044806ccb50993ea9ba29cf020
39ab08f1d9fc9b5d4a6b725c5f62af7d4394a427
578 F20101117_AABIEZ miller_j_Page_172.txt
088925e882be28084db5e302b334c139
dfe64214d1e6aa903893c1e6c5abc5183d322719
5382 F20101117_AABHZT miller_j_Page_194.pro
29bdd2f72f7eccd3e208f834ca14e53e
9541a436d6eb16046b679b5544558f49a8c21926
F20101117_AABGXF miller_j_Page_191.tif
0befe4036af7f936da182ba6d796ee11
5c65e0623a209c09e2795046819ec4cd32055051
35484 F20101117_AABHBX miller_j_Page_179.QC.jpg
b077870b54cd09a7c623bde787f2cc9a
65a4d187c4f9dff0df52e9a5a60cbc8e481c80f5
F20101117_AABGWR miller_j_Page_082.tif
5856c123e3bdaf8339ecd38520df9f58
43fb59622b8b7a95c1a3930cd9aca704bb84248a
62862 F20101117_AABHDA miller_j_Page_010.jpg
7448eed9cf5fdb7b35098c182342d3ad
423fbdbe8bc1ce54034737459f2fb40c2ddeb744
6991 F20101117_AABHCL miller_j_Page_155thm.jpg
21fb830cd648cecb5a1ccd5b434c820b
17332295c49d3786aa4102f21d27e8b604d30fcf
31148 F20101117_AABIGD miller_j_Page_142.QC.jpg
a2df5ca95b19161ba1a561e77bcb875f
f7e214ae23d879e16b28e66b5a879a4d312eecba
432 F20101117_AABIFO miller_j_Page_193.txt
a4d2883ab22fa0bb0f1b5d4035a29665
3cda37835bbee670047653a2fe343945802793ad
56155 F20101117_AABHZU miller_j_Page_195.pro
3a061d08a97710af36af92893dd7a223
0438afcfb6e5be6ab07e8cde827f6be95ac9f431
7080 F20101117_AABGXG miller_j_Page_140thm.jpg
adbe6acebcceda284a9f12321fa9b47c
3bfec64be01ddd4bce13d29f202b21b0b7e01848
79198 F20101117_AABHBY miller_j_Page_152.jpg
51eefaba049917eca5998e525e90afd2
4acde87e7c83af5a060195b0d0d96db41d774a6a
106928 F20101117_AABGWS miller_j_Page_055.jpg
e34ab53a7affdab0dc359d6591dcfdb2
a5f24f30c63507b44a0fd90590c9f2a9dc8fd37a
115313 F20101117_AABHDB miller_j_Page_011.jpg
e14652f81ea0fe4cfb4e7c16e9f6631b
03fbcbcb80986c2c8668989c95f3f8dd182ff6c3
9084 F20101117_AABHCM miller_j_Page_084thm.jpg
dcec9a399b6a1fc59a0bede4cab3f55e
6e1cfb59e86ceff7975cb42134c8f4cbe2be5408
32484 F20101117_AABIGE miller_j_Page_113.QC.jpg
2f6b46a869ef7d4ea7445d314b3152f4
763cb10ebae4bbda74b71033bf87fd6f6954ff1c
357 F20101117_AABIFP miller_j_Page_194.txt
51ce9822cea84a0e0b3b08678f5aa51b
e01126fc0d0b61ed608a172229e8f9ae7eeaa97c
63347 F20101117_AABHZV miller_j_Page_196.pro
46183b48a61ba1187a79d3530c60ded4
72473178771566600d662d66a36953b0730e72cb
48388 F20101117_AABGXH miller_j_Page_030.pro
80d17329f93a54366ed885678fa45294
33066e090f4ab16f020d9bd5e805e481b92747a6
F20101117_AABHBZ miller_j_Page_100.tif
ab70e8a9d9d2f3f57c671346d51928b4
bfa5761396435af0e42cefd07d103d8e4cf4d9f0
598 F20101117_AABGWT miller_j_Page_174.txt
abcb4a9e94b40d451157f13da999f560
5acd50fd8ec206be7a747bfbabace2ded1ef0b09
93309 F20101117_AABHDC miller_j_Page_012.jpg
d784b2d4cb7b48452da688375f18ec0e
c1995ad1179630cacb8ae45fa172de764c148a80
F20101117_AABHCN miller_j_Page_193.tif
6bb203931f312eaec43333d63b19206e
7fc1990c1e01567f6706386eac6fc0e605e72176
8997 F20101117_AABIGF miller_j_Page_099thm.jpg
c5fe6cb2ac2cd95798af2a00af043d16
f23e217537be3fcf29e79dfae0644a348190a85e
2269 F20101117_AABIFQ miller_j_Page_195.txt
3636515257f9e7d545217bb8492a62d1
eedc705fed2857e54aecf539155f9578def1e638
53497 F20101117_AABHZW miller_j_Page_197.pro
a4b81af92e2d98aaf114b56c15e2a1d3
f870ec670565bdd950a39b7b30103b855d17d594
8666 F20101117_AABGXI miller_j_Page_051thm.jpg
6f212459c2c735b11f3f7f9e05e4cfcb
7901f71adc1a778baa0a4ad7b65fdbc3878797ae
337 F20101117_AABGWU miller_j_Page_167.txt
f80602c4ecca77738e091a9dcf55c713
dab051b6daf6fa37aec53f0df026c045c0e9e77b
36708 F20101117_AABHDD miller_j_Page_013.jpg
9ff0ff8ba25c4fb8b37c8994136f9959
10ffe6537c911b201d4568a56c79238c67b55268
1840 F20101117_AABHCO miller_j_Page_046.txt
d2aa4f86e9709ce87fc95baea425a3b9
caf6ebb3a5042f6ac334f5dd35ce1cb320be2b75
2502 F20101117_AABIGG miller_j_Page_017thm.jpg
020b9d3f96dfdca6ffec8add4ae9b264
e3c67da72c68409d0d64b5ef3c5f31037beb0abb
2553 F20101117_AABIFR miller_j_Page_196.txt
abf4c3406bb81eccd3da631fb531143d
c2702cdf355115f67f0c8b4e16fb2a771660751f
57692 F20101117_AABHZX miller_j_Page_198.pro
89a4476d2e6f70ae8d7984e70a5eec6a
1006e0ba05d4f3a1b8cf71e921bea175759d7c7a
77317 F20101117_AABGXJ miller_j_Page_090.jpg
9c44b68cd09dbc62cc6597504a1c96e8
9399eeda875d0c4e49ee8929af99472e7bf670a7
7899 F20101117_AABGWV miller_j_Page_073thm.jpg
135fe73ffae191e61fc0fca453a2adff
bc622eb6840f764530fac8db35a52ea7ee00ef36
102332 F20101117_AABHDE miller_j_Page_014.jpg
907922b272d14aa618c76991caa773dd
fc5b3e8c919e8d0d9b20d218f158c2da29834f7d
F20101117_AABHCP miller_j_Page_119.tif
26934b2bfd789b909976d111145427ee
ef16f77af26b7e4ebc99c20d3fc3aa3d2a0941de
33861 F20101117_AABIGH miller_j_Page_038.QC.jpg
2f5e0afa846fb344994e2d89e7987643
b2ef0a7100c3240d8dfec54d0a59af362e674f3e
F20101117_AABIFS miller_j_Page_197.txt
df1e8ea1a6bc258e918d9d1a31207669
956db4971ff5b24e97267811ca3bc1957af424ce
19280 F20101117_AABHZY miller_j_Page_201.pro
09f07b56734e2f9648af7b95df34e74f
bdd39fcc67a9e3b5c81709cc0baf3ef1d96ae86f
34350 F20101117_AABGXK miller_j_Page_062.QC.jpg
b6ae7ab3561bb9a04c2c2a2d06178d79
0ec6c520b218dae7ac38c38403a0e82dd8d424ac
F20101117_AABGWW miller_j_Page_190.tif
222e19a7113acce391597a2c07ba5afb
1156a0911ac4fe1fff9dd49ab4c85258a419adc0
105976 F20101117_AABHDF miller_j_Page_015.jpg
2fbff93057cb0a74a2bd0a4c5eb7210d
9363731db28271f5c85a55083f464a3d274e39d1
2033 F20101117_AABHCQ miller_j_Page_035.txt
186e53ac648b3dec53125adf1ac7e1b3
41e84c291877ace72646bbe5776d455596f42310
8628 F20101117_AABIGI miller_j_Page_128thm.jpg
61336e72e839f4fffaae3751b1809570
9e3b58aee46f3849d368fc202973de4f15d668c8
2325 F20101117_AABIFT miller_j_Page_198.txt
0730ecca3fd8e24d6f7b8fee35389bfd
47ab3341f6527d6eaa3cc396b6dbeec10a627c04
475 F20101117_AABHZZ miller_j_Page_001.txt
b1ace4e6fcdf6a013ed2edd8a7fed55c
49c59ad920ea7d3b8926e83280d6fa860246d0e2
364 F20101117_AABGXL miller_j_Page_192.txt
c5af9b6f646ab8cae75ebdb045ac33bf
b055eb46a31d1e5923c372abae3a574d39427295
3354 F20101117_AABGWX miller_j_Page_162thm.jpg
c13df25ed74f7ec9e5b85dcefd1a1c4d
76b226cf92fb318d78ca9abd70196cd1bbbd2a92
100449 F20101117_AABHDG miller_j_Page_016.jpg
5872c959ac3b82ceb68fc4aafd7a7d62
005ed0bfbb1f37ac64c8169e98d5ce4d8b16662c
231315 F20101117_AABHCR UFE0018500_00001.mets FULL
b5b9cffdc7f2620c7dfa42d54a9b3e42
b37a2cf186401435addeed089127ebc4088b1c14
2374 F20101117_AABIFU miller_j_Page_199.txt
638fc0855010bb7a6f6d26c1a4600279
8ce4d7c37b5b2dd71813c11f0866d6b37251d0cc
9167 F20101117_AABGXM miller_j_Page_065thm.jpg
62b2da1bfbb6af74eec8152b736d773a
9f167e5e6ef8aef196c6f72825fbca6d758bef35
3621 F20101117_AABGWY miller_j_Page_201thm.jpg
285d3c869b2b4d4007c4a87b08e2f30a
572be5f7554793b009ba459c17676cae94f44473
31931 F20101117_AABGYA miller_j_Page_028.QC.jpg
6077351ad3f2ed88d8f0aecfcad7f90e
b5917a76f244cb00a392a98965b07a424cd95e53
7070 F20101117_AABIGJ miller_j_Page_156thm.jpg
9146967320a00055819ca4c0524201da
265e3e86a19b4caf604dbb8d6f84e5c4edf09399
354 F20101117_AABIFV miller_j_Page_200.txt
09f012da295b47ed65171fe230affbf3
8f028357c09fc3e9de7bcdadbdd00753493650e3
F20101117_AABGXN miller_j_Page_181.jp2
e6d491a2c26b4bb23f06b29396bcd048
4dfb85a416612fb53032bc3c320be99a59f476ad
101955 F20101117_AABGWZ miller_j_Page_128.jpg
3b47ae828a3af6c9ae397c47be80339a
23310f70eaced2069702a4381ea8fcadde5943cd
F20101117_AABHDH miller_j_Page_017.jpg
95a07525a305c5c8054d5c689b8cf954
5d2e9233c09b74f7054f23a8aa2d87fb7baeab2e
108103 F20101117_AABGYB miller_j_Page_053.jpg
21e9049efd05ddb166527fc2d975b88c
d6b632b6c99ed17f5507967374e696bcf5da8dfe
8183 F20101117_AABIGK miller_j_Page_019thm.jpg
bd2628a4f947ece3154f2b34d49943ef
bd331a95d73052e8152eb52d0f5a6bb5b10d43ae
804 F20101117_AABIFW miller_j_Page_201.txt
174c4fa1f6e716b8ad6281fe9a007d73
5521ee0e124d22df9ccb642cdcbb0a0087c1c9d4
28264 F20101117_AABGXO miller_j_Page_152.QC.jpg
f51911d85d6f08be96d7d96e7a51564c
c208262793629ecaac9571384aa0e59f2b5333ec
91853 F20101117_AABHDI miller_j_Page_018.jpg
734ddff5678b3644a87a482cacf11a3d
bbb119bfce631e435171155bdbe74c87342ccf6d
8437 F20101117_AABGYC miller_j_Page_112thm.jpg
69eeaec3ada505d4a16771b53b128f27
edd2a02a25e2d8860792723752a69ed294b62b20
5103 F20101117_AABHCU miller_j_Page_002.jpg
4bde9ca83f7ca3702e3e46324949cd25
eeb1bc49e2871dadbdb46b6b6dc75d2eded7adb9
7465 F20101117_AABIHA miller_j_Page_148thm.jpg
c8239650b5564f5eaaa3f6cfba159194
ad06a36f57539f3517ded35cd8f148cf729fffea
7324 F20101117_AABIGL miller_j_Page_152thm.jpg
166da19cb46d080497fb86c71d2bade7
3c748311bc7632f3d60fbcc266a1af2051710534
29135 F20101117_AABIFX miller_j_Page_140.QC.jpg
8cb3431fb822e41121a6220ceda8ae81
15700cab807cbf7e39002293bfb8fc8184f34a4c
8206 F20101117_AABGXP miller_j_Page_030thm.jpg
2f530cc3567dcaa7b67e5dbca9150127
dadc6fe0c540a8f2265b6bdbc12679772e439dfe
95213 F20101117_AABHDJ miller_j_Page_019.jpg
f3d8a89306e03b72dc396fb110afca16
6e73329115cea62f744f30f3f756aef26cb939d4
60183 F20101117_AABGYD miller_j_Page_074.pro
fb9f71d230a6c59f47595dc33faaf87d
ab43bf7897f3e5af8feb5651ebc8f7aefb3a1fab
69026 F20101117_AABHCV miller_j_Page_004.jpg
b4a5ef6e25bbfc35478eb336b42a21f8
92b22245118638afd74d3b00b077246afa8fe0c4
34079 F20101117_AABIHB miller_j_Page_056.QC.jpg
9d8ce7c02180538c20d3afd7d6d71e86
6e74c6d52a63e053df703633237a1a22e1c23957
35563 F20101117_AABIGM miller_j_Page_136.QC.jpg
38c865c7705270847678a8f4a11c72b2
11a77069bef3b14aa4907ef7f8589cec524e50ce
8811 F20101117_AABIFY miller_j_Page_045thm.jpg
487f8e6b500523c23ae8a4d931e561c3
0c892d76844554f508759ca050ca27bcb05cc378
F20101117_AABGXQ miller_j_Page_099.txt
e7ae0cbccc809556dd4ebab6fd34a394
409387b411027223de9c59031c8b88ab0c716159
94298 F20101117_AABHDK miller_j_Page_022.jpg
996e1dba6241a825776db7b9d6a955f6
03b533d1c2a1795e8cedad7f4d3b559472d187e3
2019 F20101117_AABGYE miller_j_Page_016.txt
4640e30c38d7b960491beb117ff86aad
14996e86162826fc84838de2b3f5c5baa0f865ac
126864 F20101117_AABHCW miller_j_Page_005.jpg
372b2a95896aebc9e908b6d7b8b87e2f
be0106a0476a6082eeac331305eb18ca74c2497c
8585 F20101117_AABIHC miller_j_Page_116thm.jpg
69d287913504690b1ccf4fea53587de4
6cd2deb9e7072a5b33e7bbe14328330a910df930
3156 F20101117_AABIGN miller_j_Page_013thm.jpg
df3a7b4dad85316136c3bea95d0aae8e
8b59e239a81942a65825452b0fd000d342a8e2ed
29680 F20101117_AABIFZ miller_j_Page_041.QC.jpg
a7e5c2bc0d7eee153201848be0140f85
71de14901aa845f8201b6b294a2c96d97529debb
F20101117_AABGXR miller_j_Page_064.tif
6c59390364d5e5eca9712039b7821195
98704821b606b3320c1e687d55e5f8e81304eae2
112292 F20101117_AABHEA miller_j_Page_044.jpg
faba8e45687c2c4e9d877594906a478c
98c7e0cfa1148c89e087148aad03d952a31a4c8e
102903 F20101117_AABHDL miller_j_Page_023.jpg
77b90fdd7b325642c86ae68684d05c5b
99e8aea02430e2c8e1013cf31ff199be6650c55b
25120 F20101117_AABGYF miller_j_Page_087.QC.jpg
e5c55c2e55dd70b8d36e9cbe6cca49fa
7a897ed8e00f2ff97d67fe2abcc9eca7a69b5239
138084 F20101117_AABHCX miller_j_Page_006.jpg
020d5598e071c3870649bc953f821950
4c75e9fa8418617319cb0e2994032543788def7d
32860 F20101117_AABIHD miller_j_Page_114.QC.jpg
78185c5d44c54fb767b538be3adc0d55
7d70d6b61b835b0b7dc2596a6f0212030f946c06
8721 F20101117_AABIGO miller_j_Page_104thm.jpg
a723e96c745215dcbdc8897d8bbfd02e
ea4ccd35fa9ef018d0c9935d7013d2815de06b2d
2122 F20101117_AABGXS miller_j_Page_115.txt
f74ed33390c5a836e6f2af0854e29c5a
6471e5396cccef313560343522a8205014faa5e0
109727 F20101117_AABHEB miller_j_Page_045.jpg
a6a51b9692fff68e4f454b1c8a36e425
52da1c4f0287674253ebdbc4fc1e3fe2c215f537
111586 F20101117_AABHDM miller_j_Page_024.jpg
59e85fe7d993cca06fe872fdb48ac1ef
43712eb4ce6e64c79154a1e7e034d0a80f22a1cc
110564 F20101117_AABGYG miller_j_Page_185.jpg
006f85834b1c92736c327b0a0c3fc9e3
977b3649c45be6e886ec808dcaa16b9006fafef7
62813 F20101117_AABHCY miller_j_Page_007.jpg
adcaeaf95e2be2815e5e1c9b6055d5dc
79cc1efbc52d7d540732a206aef140931199f291
2809 F20101117_AABIHE miller_j_Page_166thm.jpg
e919bc625440cb9b5f073722d731e759
bea91583884039f95da9be3ed4cfa2abfd61b40d
7402 F20101117_AABIGP miller_j_Page_154thm.jpg
528426d97142641360f32ebd1023ccaf
f3e67638a00d459e85db77204db8f8f8d8ae76d7
1051941 F20101117_AABGXT miller_j_Page_132.jp2
744cf194a2fc2188dc04fed04fbd0fa2
06ebfc449425c4bb1b4e4c8fea02622a6b2db234
96684 F20101117_AABHEC miller_j_Page_048.jpg
a9ac3b2037438642d45be40edc3ceef6
dc9bca03be60254fd79c375c3ff46abf6eda30d5
73960 F20101117_AABHDN miller_j_Page_026.jpg
fe211ee7a41e4244131f7fc8a783599c
e311d078e060e2b50b5837aee9f622c67df17944
F20101117_AABGYH miller_j_Page_044.tif
3b0e89e5e4700c71fe0e3d44da2c60fe
d35872c7786e05da4ba94fb794976eb66c833d08
126049 F20101117_AABHCZ miller_j_Page_008.jpg
c498ac7722b47b2d22d9deae56478b33
3479e1842d38f5059e4b30193c19d5261ec2eaf0
6519 F20101117_AABIHF miller_j_Page_005thm.jpg
962b8d1459be622c843650b3b670773b
17b3844801d3b2d753898b02995508adef69c757
3053 F20101117_AABIGQ miller_j_Page_163thm.jpg
b9e98999d28e4b79256e0d03e804f071
3640ed4c8e7ddcb447fad36211392c4dd159b75f
394 F20101117_AABGXU miller_j_Page_162.txt
5e954bad4ad248b24f422a1d1faab07a
3897a68f18829e1c5a4287144fff9b9364876d8a
118156 F20101117_AABHED miller_j_Page_049.jpg
3eb342ffaf0ba6518adeb808bc98bb7d
b2d16d809cba5feb6dbe4ea075085976a7bd6c88
104261 F20101117_AABHDO miller_j_Page_027.jpg
3e96ff64ac21b5b256804e730a1f57b0
a84c412ba61af95b7fbcb457e87230bdd5521e2f
102260 F20101117_AABGYI miller_j_Page_098.jpg
052f69d3b0263a1ed3274613066638f6
5fbb7337653afdf5ebc509f84ffc4a54e8efa63c
29069 F20101117_AABIHG miller_j_Page_147.QC.jpg
9346db7ab4a69849fae69ec2d2ad7f84
b47b9e9221b83fdc272015ae4b1b62218d40e1b7
36907 F20101117_AABIGR miller_j_Page_097.QC.jpg
65e2b8a02eca3794fede81ba1be9ed57
603771237768f3884d9f55f54cac883f9d154e67
777 F20101117_AABGXV miller_j_Page_070.txt
c51d66b61f8a9845b41f963fd43f31e9
e82726e47bd1619cc332c51ec821157dbda9b39d
96304 F20101117_AABHEE miller_j_Page_050.jpg
ca501706fc222bf44970425fecbe3083
ec2c3be57d1b794b4bb114d2720417bc7e3bcbb7
95197 F20101117_AABHDP miller_j_Page_030.jpg
4d76a9b69eccc21e95976783360d0d45
8e7fa1fde2b5d37d46e02792b0f6b9251a3f71d9
53492 F20101117_AABGYJ miller_j_Page_180.pro
f87b0917a5195e18638386b3b2ec11f8
395677db68e6767db6d95cd676041dcfbbcd90e5
37034 F20101117_AABIHH miller_j_Page_084.QC.jpg
73ac6cd1b3b4cd486ef6edadc3efda5e
632d563179d7b7c2af9c30f64fba5c8032e1ce14
9256 F20101117_AABIGS miller_j_Page_194.QC.jpg
505583b9d846d857bbdfe3fb15941e03
ca2ba5bdbcd28ac999206be395a0c1972d0931a6
8482 F20101117_AABGXW miller_j_Page_200.pro
521a71426a0a8a5ecc6ce143ed162d6f
5e2fa5e290fc1eb8c1a1361c69bbc57896cce6f2
110639 F20101117_AABHEF miller_j_Page_051.jpg
5acd01e506290a29e8a055a91a533cfb
6741a949a5ce1b132d2e4ee4b61731fdc1ac0f56
95579 F20101117_AABHDQ miller_j_Page_031.jpg
2f095cd6c20f8f3c9a437823decc1871
cfe35b42549940b8edb5eee15c5c212a07f7b192
1051960 F20101117_AABGYK miller_j_Page_106.jp2
c712e17f7bc172f3790ca7e2f417f69f
dda80663f5656dc617775a072810085e09d35933
5598 F20101117_AABIHI miller_j_Page_004thm.jpg
37b04eadcf4ff84d6c8d0cae5c3acbd9
63a4010b96162d1315d2810a1929bd495cae844e
8703 F20101117_AABIGT miller_j_Page_178thm.jpg
7a921db99349892dabadb4ea1ea28b61
b4b16b57926f0da1fc03326f1cdf970c2cab1015
35906 F20101117_AABGXX miller_j_Page_069.QC.jpg
bbfe3572652100fb4e00abfa43ee1935
fd839716dd9146ab4d83c76328cd4a9202346ef6
108751 F20101117_AABHEG miller_j_Page_052.jpg
147fe79e771123d20106937f7c11bfbd
08b96d835fe087aae4ebdc444e830f7e4c904ed5
107008 F20101117_AABHDR miller_j_Page_032.jpg
47d4999122b5a59ddde1cd1409212cde
9a4fe23a59daced74d02df712507f3f63d160224
96328 F20101117_AABGYL miller_j_Page_028.jpg
192d4ce196c8419080c178d21aa7985a
592eaf5af85875211214fe6f137c2b27019d8bce
37566 F20101117_AABIHJ miller_j_Page_067.QC.jpg
1c57cf6587a15bb6d519b070e5271394
bfe0a44d142a6afae567cd58bc045b7b05d0077d
24711 F20101117_AABIGU miller_j_Page_139.QC.jpg
6d4ff4d44cc3c17117f875bcf0517e04
73494fb65e7764d6e9cb3abf076f8408026ff260
105024 F20101117_AABHEH miller_j_Page_056.jpg
f7e33f48faceeef0d007f45298bb0e5e
89553a589c51b6abde4425ac0efd71ab7a6d54b9
F20101117_AABGZA miller_j_Page_032.tif
bf505096c900673699f456beccf38cfd
d73436353a0ea787cf9f5be4cae53b600d599e22
111668 F20101117_AABHDS miller_j_Page_033.jpg
f5ce66498df531e3efc5bcd27e0467ff
627aa5d3779418841a4527d9420cba5d76951034
1764 F20101117_AABGYM miller_j_Page_153.txt
a311060e5374b19565238a7157a3a589
a517fd3ccbc1407b886b33328bce544324f06571
20548 F20101117_AABGXY miller_j_Page_192.jp2
f7a940aee1c9a69cab9cc31b5a7d96ed
1c25566c7cfbf8b82bb6fd8f9c71538d0423bfae
8020 F20101117_AABIGV miller_j_Page_011thm.jpg
1b8bcb122c627d3a00854dd95d6baaf8
2a2e3502bd75b7d68fba048a1630fe8ae899bbd4
8162 F20101117_AABGZB miller_j_Page_048thm.jpg
5270a17ba2eb006bfe62cfacf0c51a98
3f1b882be3f26b6e0c0f1999e81dface1c87d6a9
106107 F20101117_AABHDT miller_j_Page_034.jpg
b997a7ba589ce4f4f97e24fd1da666b4
a0bdf8de57fb052f6ca4bc14c878d428fce2df95
32160 F20101117_AABGYN miller_j_Page_004.pro
abdad7e810c8fea535b872b9bed0f2ed
5dbb2172c1f55a47a9e9db2e7df7997e6966fed8
F20101117_AABGXZ miller_j_Page_080.tif
52f9604b77bceef4abc192e6d7d957d5
b9ad6f28daea75afe2f7709cf10eb55638973310
38185 F20101117_AABIHK miller_j_Page_021.QC.jpg
cf02f601ad3094e4c4d2c6d8e97dfd7d
c9f158b4f0009eef75a35f56b33fe546b9b2f0c2
4571 F20101117_AABIGW miller_j_Page_010thm.jpg
ccb088204ede9a646104bf102de0788e
71d2ce767b8f344e0a036c27daac4ad04f902e6d
100373 F20101117_AABHEI miller_j_Page_059.jpg
06097d85f98c8cfa75570397e4965b42
9cff9793df8340769415a3016f5fc591d79ddce7
52771 F20101117_AABGZC miller_j_Page_133.pro
490f8036a8331d1c6579e5436615658f
474ee252267e2cd8e346d9a2ef5f945861f68d97
99963 F20101117_AABHDU miller_j_Page_035.jpg
8483e6bb7cfd5199a709f5611a7b2344
234434ebc0cdad546492b8a6035b46ac3e6bcc50
405478 F20101117_AABGYO miller_j_Page_093.jp2
47670dc96f55e74686247a69400f42b9
23e17cc2fb8a45e54a897e92fb3d7a8a7ce3bbc0
33040 F20101117_AABIIA miller_j_Page_182.QC.jpg
f4fa2af3e84ed3a31b9ea244ae7641cd
a008f3859098f7df60ffbcdc40cf31dbb8a178c8
32558 F20101117_AABIHL miller_j_Page_035.QC.jpg
b5a05c8ec94e4c9ab8a76c3e226c8758
824a82e362061cfc4145f3d42425ec845cc11f81
25067 F20101117_AABIGX miller_j_Page_091.QC.jpg
82ca0d35f555038c64f78c47a1e4f2f5
5a0891dc69e170509e4d5ef4df65ae5b76f13011
112211 F20101117_AABHEJ miller_j_Page_060.jpg
259a9a3180ef9633f8708f9e310fce98
1db3e56dcf8da8f80f52b280ad27f6a89b48264e
8583 F20101117_AABGZD miller_j_Page_069thm.jpg
d8ea607cd523f4808015e10b86134f46
126dae2f6817f4720df7dbba6137b8246aea62b4
97062 F20101117_AABHDV miller_j_Page_036.jpg
ea3923ef08dcadbe205c850e3f56bb2b
e06c5fcf8672f5be4cbe303ab6c3817f5c63c818
18188 F20101117_AABGYP miller_j_Page_170.jpg
f4b55f6f2e7f4b18953cfc50b7781e87
4d414edbdade1e2da9705c20b0da10c7dd8c2934
34111 F20101117_AABIIB miller_j_Page_034.QC.jpg
074760e1973ff48ccd29ae3f67c75958
55e2c8691f3f43b6df5bb5885055e17532ca3ef3
9310 F20101117_AABIHM miller_j_Page_067thm.jpg
057aa6d10c72f0f3e41ac2dba497cb91
cfd0106e4157abc6aa452fcf79da28fbb6017fa1
8890 F20101117_AABIGY miller_j_Page_126thm.jpg
2bacc99b2bd36b5d864d1769fc032d5b
76b73c0e390dbed890801fd93b7867596d304507
103864 F20101117_AABHEK miller_j_Page_064.jpg
acc62ca5bab59eed297e490967a259f6
70c05587692667e591346a1d85f1c11f6d453ed1
F20101117_AABGZE miller_j_Page_200.tif
53992c822c9c40396e613e8e4275d424
aee8605209e0a1ab9062cfabf073f777a8acc2e8
79685 F20101117_AABHDW miller_j_Page_037.jpg
120ea5c63dc788c8b20b302e51c7ae45
4183de7b4358d74b234b0090e7a4bab97ef20941
F20101117_AABGYQ miller_j_Page_130.jp2
b2ad377c2632d9d2bb0b6727f3e64618
f87d14e4babd0484b18492ac91b98e0cba339664
7751 F20101117_AABIIC miller_j_Page_168.QC.jpg
2465c30e72aa1c8a06cb93156f4c8cd6
702866465081db4b5d5cdaac60c8784d3a3e4321
8766 F20101117_AABIHN miller_j_Page_199thm.jpg
50c3c7505157d7ed13afbd56a8193cbb
5d206d106b2c1a9ebdc31902d9f9d8a4c0a6014d
8739 F20101117_AABIGZ miller_j_Page_079thm.jpg
7f9288efb888a2bbf17498169696bc48
a5b4e3a774d13a4bd165b8a668fc5f8134aa8c5a
110852 F20101117_AABHEL miller_j_Page_065.jpg
f7a74a6dd66aa651ceb67e0cea2cb99b
9ca2478e54cf05878915ee54a5bf1bdc5280fcb9
90058 F20101117_AABGZF miller_j_Page_046.jpg
8d76d420c5aa8e719b37d0687c6914f3
fdaf65bced594d37f8f18dc69ef7bb0cbe564f67
97200 F20101117_AABHDX miller_j_Page_039.jpg
004d01f6c7026a6f8fd1cd53d6d64da8
dbd051c8ecac712230d48aec63bde27c4bda5ef1
7828 F20101117_AABGYR miller_j_Page_145thm.jpg
ab56a7d625a507c29ef1f617ffb4b428
131a9f6f6dfa306f097cffc5a5b818ceef2b1a51
103826 F20101117_AABHFA miller_j_Page_083.jpg
f65ffadcf7a9a664849702696d23a9a3
70eb71d47d3fe42ba77d6184b065923c14171b16
34711 F20101117_AABIID miller_j_Page_110.QC.jpg
de9c250466c12f7cb1d41c98aaf156a7
ef4dca89dec48f5d6f7dbaa3e02927f4403e060e
22875 F20101117_AABIHO miller_j_Page_025.QC.jpg
be84a4c72684ef8873912d03476e9aa6
6038d0ce19e2e0c328700f12a0e717515923a212
113023 F20101117_AABHEM miller_j_Page_066.jpg
7d15055a03959ccac2be592606029e31
bfd3df0464f9b0ea67876e9db15c266487eb59d6
120 F20101117_AABGZG miller_j_Page_002.txt
0d52a472efca5a4b4cb8b049d9396aec
36d6a2178449a9c1b20185ac765fe563c7864c85
114254 F20101117_AABHDY miller_j_Page_040.jpg
74295a0605cb5a02a387d496988f288f
92e3ba0717002229acf98144ad4f9ab5db5dc075
1051936 F20101117_AABGYS miller_j_Page_104.jp2
63bd3dbf75ea8fb65666847b5640ad00
5ac4b81171df05a1f3c47aeee407f7d97497679d
131765 F20101117_AABHFB miller_j_Page_084.jpg
4a9b85fb37be1011d0506d13447cbcf4
c23b0374247a464dac9d1d964b5d9baa9ec629eb
33164 F20101117_AABIIE miller_j_Page_112.QC.jpg
01b3427a1d717e0564069565da0567ca
c0c0143777b18025967a1e975e38f9860c3231e1
3795 F20101117_AABIHP miller_j_Page_137thm.jpg
d1a26571f2d9ad42b3cbc0f87ca1314d
7e0746877771998270481c18a3990d50c5c47cf2
115865 F20101117_AABHEN miller_j_Page_069.jpg
5a1ecfbb110d955921bf267593c663fd
9ebf3b49fde44acbb4c298e0c0923bb003d54739
1040015 F20101117_AABGZH miller_j_Page_022.jp2
785698768a37743dbae458059c50b44f
4240d85ab4e3b5f22e8052989d50e9c3c4a2a670
87494 F20101117_AABHDZ miller_j_Page_042.jpg
4933ea14cd3b0a952387b4a43f0b78f2
9bad61e27305a577f66f0adaf6f54663dac6b812
F20101117_AABGYT miller_j_Page_136.tif
d1c9cf99b3dd37fc64e6af078ddaa32b
a5ac9e62eb7bf68ecdf0aaa8bcf85c68d68e85c7
103724 F20101117_AABHFC miller_j_Page_085.jpg
d32360029e2e16f6d2a30057fb6beef6
c7fe436c2db93858f9ebe9e58b7bb20a3c1a8e77
24226 F20101117_AABIIF miller_j_Page_026.QC.jpg
719bca4361176bd2f65e134d641b0043
bde6e42684f5268555e65249b8d06f927eba2b6a
7923 F20101117_AABIHQ miller_j_Page_167.QC.jpg
ae292f1066a8372ec8c6c2d596e79f5b
33be8ac7da320cc6f6f3f3b00aba01eb0ca02ad3
41272 F20101117_AABHEO miller_j_Page_070.jpg
138e485f0f00ce8b3eecfb0f354eca3b
b6e386384fa9d3553e5847243f0f9323d82e16f8
36220 F20101117_AABGZI miller_j_Page_058.QC.jpg
ea706f8846124bea69f98fed4fed1067
78df9cf06da6f73bfc3e5952e0af6eaf6ffe576d
F20101117_AABGYU miller_j_Page_085.tif
33c1f5ec7e1a8e19b1de8de5ed7f2149
10bfb4b19c7d8ed3461f5c8a23139139ada85baa
56275 F20101117_AABHFD miller_j_Page_086.jpg
1a1598044d22631f20e8bb7e2063e956
f18ae5ef9361a2a5deccba146ddb2605791115a8
25347 F20101117_AABIIG miller_j_Page_088.QC.jpg
e7e3a9cb2df14858219fbe9dd03d749d
4169b9b82e115de8253309f4b132e014b460eb36
8093 F20101117_AABIHR miller_j_Page_001.QC.jpg
46b7b0cf6ca28cf214d6225538a4df2d
39ad6e13d7632eeb5022f26206427b780c6ddff6
20074 F20101117_AABHEP miller_j_Page_071.jpg
62708d7e4b9f81d7d251e997c218f070
590053b46ace6ef581430a6ebee683f944f97617
F20101117_AABGZJ miller_j_Page_184.tif
6d678dee7cde14904317b9612e55cd11
af66aaaa9c4e10a4f683937b340dbbd8ea47f260
44613 F20101117_AABGYV miller_j_Page_048.pro
93b2e20fe351793d5c42c7dc60290062
1fc9c0afc264f797a89621db2fa051e04edbaed2
81429 F20101117_AABHFE miller_j_Page_087.jpg
31f2b0f492a55bbb66253c283f9b7d38
97c5671aa3d68ab241684a8d1165c42602895e96
34825 F20101117_AABIIH miller_j_Page_024.QC.jpg
9f2b2ffead26f40c1285460221a1414d
fa41220c605553dd64e4f059ee745a0274557697
8535 F20101117_AABIHS miller_j_Page_056thm.jpg
7eb4dbd2d72ee63df22ab701184aa40f
4bb7bbcb9fd6e777b40a7503ec6fbe3133c157fe
82022 F20101117_AABHEQ miller_j_Page_072.jpg
998940d46a54ab75cf522f5e7f70a2a4
c888914b0de4f1541aac998031235678c8670b7e
103611 F20101117_AABGZK miller_j_Page_068.jpg
5d3ba9a00cc95adc1905ef7a7a7da497
c840b9e820e6dad434b4d98dd2ebae24ae2fd72e
47612 F20101117_AABGYW miller_j_Page_186.pro
69fd3f1f3d4e3db7e7e3703cf9bdcf16
7b4a83fede0a50ad03b0f1b8fd879aec5856b131
81831 F20101117_AABHFF miller_j_Page_088.jpg
7fc85634faedabe980eaa4a13c6baa7c
136cec03232691a264a99ea4db4af7d441ee22a0
6846 F20101117_AABIII miller_j_Page_087thm.jpg
8cd4672f099b62ad0f29bf7a74039fdc
da240ca37d6144c0267c3480d28b6a02be9f409e
35124 F20101117_AABIHT miller_j_Page_180.QC.jpg
dedaa2e03304eb465a0fd602ebbd374f
30f7d50d3e97852e997e1268e73f2b207f239c62
103915 F20101117_AABHER miller_j_Page_073.jpg
b37cbe03a6fc6610360004599a5220e3
cd19cf20cf0d941e44a5f08595cc8db8d5b9aed1
7503 F20101117_AABGZL miller_j_Page_151thm.jpg
179d4808d9ecd23cce86a7b75aa5c097
36e53437135db7d92cddededc07778e40d9ea52b
F20101117_AABGYX miller_j_Page_090.tif
fb58e2acf6ab5e1511b1155dc885edca
0d81a6669ff319ec3f62589ef70108703ac38044
82069 F20101117_AABHFG miller_j_Page_091.jpg
8790d61dcd1556d6bb77f2795fab0ce9
26b2e3bc99799b3ff572c043e85f0f6ad29c4ff5
34987 F20101117_AABIIJ miller_j_Page_122.QC.jpg
100f59f1e2263270700a2a742bff5f7d
80792e5980cd0687591ed8829db8d41cd8c9bfc1
29586 F20101117_AABIHU miller_j_Page_149.QC.jpg
107f83e414c69ab595614473599b36c2
c71057e13cfd71eec6b84c7685a1ebc713af079a
117291 F20101117_AABHES miller_j_Page_074.jpg
3a8f2785b4f2582de088ecfda0661d98
b0fea5b0a515cb27a522d944948f56bfce51f3e2
53694 F20101117_AABGZM miller_j_Page_053.pro
be2553ef00d3212876181afc19d313cf
85a7deaf436349dca5ce78809d2a291363530de3
1685 F20101117_AABGYY miller_j_Page_041.txt
0ea3ff59dd68d7d9f4f0e7d698fe3070
44d76529f3064f29bd5abbf930ed38994e3cc9b5
78312 F20101117_AABHFH miller_j_Page_092.jpg
b65dfd11bedc0c220ea1246177932703
e1462de77b86a8d43da0d93944136c5810882f11
2994 F20101117_AABIIK miller_j_Page_192thm.jpg
f53e30488f47ea248bb538740e424e66
c3a8d75d0642094dca4c541298cc0706b7bc8426
36478 F20101117_AABIHV miller_j_Page_181.QC.jpg
4df8f67d7e7450024fea8b71bfbe4edc
17afd9022743fd0b8c7ef31041f0ed1dd4c958d0
108573 F20101117_AABHET miller_j_Page_075.jpg
ea0479368f0b87f69ea460a12bb428d8
78210907452626ffdf063c1c982a6a40f838cba2
52587 F20101117_AABGZN miller_j_Page_043.pro
1b77b75a87e6f033addc3664018fc45f
6fbf9f6b58ea14d8fa99f5a00d577917345f2b93
9056 F20101117_AABGYZ miller_j_Page_063thm.jpg
2df5ae223d254b035abcea19b375d56b
8db1eaf88e43f9202ec225df43457b932802ffbc
40378 F20101117_AABHFI miller_j_Page_093.jpg
50eebe896b374dd3f25666c3779646eb
8b67d5967811d0d11092e62a8ddbb66bb46aa5a5
37489 F20101117_AABIHW miller_j_Page_066.QC.jpg
fce181be95a062c23e40168335409376
7bf4a3eeacb5b006e1242fa0225f449130bed424
103859 F20101117_AABHEU miller_j_Page_076.jpg
9f5c2e266083c1c3595eec2d3fb79141
4daa849122b04603967086ef11557e57506312fb
99403 F20101117_AABGZO miller_j_Page_081.jpg
4fc9335db16fabe3534da65c1e9aabeb
ab33a51c98aa4c6916676ade1a35ce1e77f05db9
7981 F20101117_AABIJA miller_j_Page_074thm.jpg
59e57dc45d3bf0fefbb2cf8e7f6b6cda
8986f48db153a5f486bd4035a00125de52f7e7ef
8354 F20101117_AABIIL miller_j_Page_123thm.jpg
99196d26e739555b559fc7565eb0d6dd
1a93ba62a2862e6a40427e58871ee108439e4a7c
36279 F20101117_AABIHX miller_j_Page_078.QC.jpg
6d13b15ac9a02d6f2221691d1e923a42
221f1c6d003c50a680fced2041104f453802149c
7698 F20101117_AABGZP miller_j_Page_164.QC.jpg
6e8b605cee5a35626850388691c4b348
5a3702961fc6bb5dad548cbf1e5f4bba4031ac29
113517 F20101117_AABHFJ miller_j_Page_094.jpg
a0ce6af562d5c99ced5ace4d3cacedf9
2d5bc4f0f8f700a2815bd7a020365c7be6f4eb55
101744 F20101117_AABHEV miller_j_Page_077.jpg
fa981b5e4cd9e614320596448e92aaad
93be6f19a6be056b3eb9dda42c5c32d360ed7174
8591 F20101117_AABIJB miller_j_Page_183thm.jpg
aa4b073cf78ca226f48f5ed68d7622f9
93a906f2596110b3fed18ce8ae2bd6e909978f63
31847 F20101117_AABIIM miller_j_Page_039.QC.jpg
7d05f640cb65b97c5b21d904d67c2189
2da5a3fc82c8d7b9d44b43161e0a759da907c002
7382 F20101117_AABIHY miller_j_Page_143thm.jpg
aaccb64becd4de2a02dbe70bf76568d6
a752ac5adbea4d8b3744c45c1348f7c0b8cc8dc3
F20101117_AABGZQ miller_j_Page_134.tif
495e38d0160f434d12c16eaf2758626e
cbea3d48cfa05fdfa25e5619ba4336ba3fef000d
103815 F20101117_AABHFK miller_j_Page_095.jpg
e0556ed46f6aea2951b6ce56bc958da5
11d69126588396f89d5262dfe91c7209543fe10a
107686 F20101117_AABHEW miller_j_Page_078.jpg
09a6abe76535e52c680816e82e5ca264
37a443261c58fa29fc010833baf73d489bde291b
8191 F20101117_AABIJC miller_j_Page_039thm.jpg
bbed98a5f98ddde5c26591e7eac253ce
167e49d4cd3c766156290a1182ed661c560273df
F20101117_AABIIN miller_j_Page_157thm.jpg
6178ccc80f80b994c5d2e31af92683e8
f043e13005f96ef5fafea8c948ea4d85fb76f17a
F20101117_AABIHZ miller_j_Page_108thm.jpg
b306619d6bd7a81c81d3389b28aa062f
195cd28514b3818e497f4feffd4675c46c4464f7
34128 F20101117_AABGZR miller_j_Page_032.QC.jpg
83d3f884ad4cc2f4e2be58e0f50e7508
8d295e1244b20be97757cfa43458439fd6fa0799
110973 F20101117_AABHGA miller_j_Page_115.jpg
deaa4e2358fe23424b106b64b4971be9
a23989e039b141aa6f54e861e80cb2cfbae6548a
110332 F20101117_AABHFL miller_j_Page_097.jpg
132c58b921667e143f59f37799d596c6
6c695d348771998ecba17525181dad8f045798af
107545 F20101117_AABHEX miller_j_Page_079.jpg
640555872e493fded4664d31af9dbd19
1499650a18a41021efef6fe651427ce6de3dde72
8781 F20101117_AABIJD miller_j_Page_061thm.jpg
62def64f2ca46988433bc54858e943bd
f3dd97a44b9aa8ce21660e32b8e8f5594316df40
F20101117_AABIIO miller_j_Page_052thm.jpg
32118d11fdbf27ea91ce560a63d4ecac
04b22b6a8a4b34ae62b1176705df6720806d25e5
1225132 F20101117_AABGZS miller_j.pdf
21b8dc364624095b832908257d498f7f
48e1cf0e3df4e8be28d31ffd8e902e4c7c46009f
104485 F20101117_AABHGB miller_j_Page_116.jpg
b11ee639c82bede95aa5f484af3cf1b4
e2d04c42fefb53d7d051efeb7cb67915f16c4fdf
111135 F20101117_AABHFM miller_j_Page_099.jpg
92c2f3daf01bbf5ddacdf1e0ee33943a
f4dde3ba105433fbdd13862f49918cdb29bdf98e
94250 F20101117_AABHEY miller_j_Page_080.jpg
dd624c2766b055a7a58357463a6a96ad
1be422c55a99f3f27178e0786c6fdfb759fc633b
28637 F20101117_AABIJE miller_j_Page_006.QC.jpg
1d599d5f4d2882abc8a9b875c29a4cb1
5e9f80187a3cd6913c6e5f2a5fd8b47bb2d9b4a7
3460 F20101117_AABIIP miller_j_Page_070thm.jpg
f11ca949c14904206848275d4f2a2a15
8db23721cb461e3efa00e18a811675f5bc12069a
18603 F20101117_AABGZT miller_j_Page_166.jpg
12a24ef7eb6597420db34d5d47226fd6
23e6ee8824aa66d94751ee29d113cbd434a341d0
102193 F20101117_AABHGC miller_j_Page_117.jpg
cbdfed39793ad2fcb1ecbc526eb30892
7f9a80560fe2296496489e2d690ca1dcbb68cfe9
103415 F20101117_AABHFN miller_j_Page_101.jpg
bea8084d11b2054c36e8bca43cbefefb
cffde4a132078040de8c1a4a8846c4474441d0e8
97370 F20101117_AABHEZ miller_j_Page_082.jpg
03982deab3aeeaf98d13db1b5ce64388
4bc3ffc301d5cff8c7596ff009cd8571a57d7bd6
36343 F20101117_AABIJF miller_j_Page_033.QC.jpg
c460dd3b64dd9526d0b313d974446f16
45283aebaf9e61d7ed768d41bfb47040337ee806
8733 F20101117_AABIIQ miller_j_Page_162.QC.jpg
0197d5f4b0ac093b5cd5a2304a98b52f
e120aa434f78ea2e446576452fb00477c8420c3e
54249 F20101117_AABGZU miller_j_Page_178.pro
44994e615533cd8a617cefc749da10c1
24bd7aa4d95e2936539dab21fbfe098292d19c8b
104312 F20101117_AABHGD miller_j_Page_118.jpg
01c8a5acf8b9b058cac63b0b2e4ee190
b12c9df14414c70e584a16bbd9fc4462643b99ca
103066 F20101117_AABHFO miller_j_Page_103.jpg
bc9e962e8a704aad9818ac3361cc650c
18ca26ae527493b1ed1a3039d9552ef696b0e287
34895 F20101117_AABIJG miller_j_Page_049.QC.jpg
dc0f5e0ee34d5beefa9a37c8575a35c2
f2272f781441cbe36ff9946adedef049abc9e480
6546 F20101117_AABIIR miller_j_Page_025thm.jpg
1f82738a71f931b8d312d2170d9d4d1f
3b32e3077140c8883ff162819ec547152d57d228
2272 F20101117_AABGZV miller_j_Page_190.txt
1a421eba02303ebb0b34633233e37e34
bb970fa3609e50b4afd6cfdc8534661bb8bdcc02
108444 F20101117_AABHGE miller_j_Page_119.jpg
965178552d7a43d2df044953e47dd357
f6184edaf980db85a15f74b22895ebee6ee0409a
110006 F20101117_AABHFP miller_j_Page_104.jpg
cea15f65bac88e3fa48e3c6268b4f091
6713ca46437be72c96af85d2e460f8385609c63c
3200 F20101117_AABIJH miller_j_Page_159thm.jpg
21f4440f2d781d33f64a187e9b99419e
f2171ec281b476e19d0658ae25a8137829b6fd7d
8605 F20101117_AABIIS miller_j_Page_187thm.jpg
3d0c72125b499adc7a0742d1fe8f6a9c
5682e0f6eb567371688594f9c2e6104d346eb84c
26792 F20101117_AABGZW miller_j_Page_001.jpg
ee1d13be3cdb6cc06082d53c1c91d9c9
e0d9cb422e74d6325c22f8efc8c513639da8c36d
107000 F20101117_AABHGF miller_j_Page_120.jpg
2454df7f7222de8f6dcb0d9909927617
391712e196420517cff787dbb88c3026a9ce838e
107117 F20101117_AABHFQ miller_j_Page_105.jpg
f42f94eda28c22fc3ec7b50842ea00b8
9aa5b0458f116aab5029733387530a0cbb714075
8608 F20101117_AABIJI miller_j_Page_023thm.jpg
461ab57d3ad72101a0795411dc96e957
2d8dda09fc6333e3799cce4b027971f8d3b23fbd
31022 F20101117_AABIIT miller_j_Page_019.QC.jpg
0e2d84b8a4b5dec2893b1087f86ed24f
8a1de9bc0061226a27eaf7f3425f2abba4c23b41
F20101117_AABGZX miller_j_Page_022.tif
0b89929ddf33b3f01c2a6e082d554f26
3920406854f21582c7fa99a084ff05dacbb4e67d
103207 F20101117_AABHGG miller_j_Page_121.jpg
ee5508aebb59adba2275f0b5f2b26514
ab32cc5c4be653cc49a6918e4d87c981608f8f6b
107603 F20101117_AABHFR miller_j_Page_106.jpg
45a11238ce2865df84d8444a198f8393
d672c040d472671ae5a59dca2168a46bf7d81294
8346 F20101117_AABIJJ miller_j_Page_103thm.jpg
6c98a9447345a655deb6a23dd556c39c
9136d2fc0c69f701584c5a75b5da36cf552072e7
35157 F20101117_AABIIU miller_j_Page_045.QC.jpg
005f38fc86dab52a715e7ad7347d5cd8
6a7a4d3f3f2cf7bf246580d29b71cc6ebbb625c7
6319 F20101117_AABGZY miller_j_Page_163.pro
85642a8aa14afb31c2cb66d825dd78e1
9059ac4b31cb72a8c46bcaccb9b5f7b76788af9d
F20101117_AABHGH miller_j_Page_122.jpg
851036b606da8af412f4b999509c9f80
1d25f196fd1b00027936c9b8ec6f46e8df410e61
110878 F20101117_AABHFS miller_j_Page_107.jpg
657e2a8808abb95712aa3aa9ca9632b1
ca2681e3da5b07eefc2f389638dbdf2aae06933b
36920 F20101117_AABIJK miller_j_Page_185.QC.jpg
533829f6c8e67fa809a68518ae9e56e8
b9bef5d33c4cf927bcf04abf05f009d894d6107a
8110 F20101117_AABIIV miller_j_Page_182thm.jpg
1a761c698b2056f00296505aef4df70e
4e3bc6cb42557951a5e36ad96280b261d549a39b
733475 F20101117_AABGZZ miller_j_Page_004.jp2
c5744da2009506dc9ed0b285d5d62b0e
635efeae79297168eac51411391f87d0de7d8fc0
97536 F20101117_AABHGI miller_j_Page_123.jpg
b0a742ea5b9d74e11c105144d7b14282
d65901f6e12b47b6570315c727900b789b47b7d7
105491 F20101117_AABHFT miller_j_Page_108.jpg
d575a2cf96412922e91d641936091a9f
bb9ced41db7c78f038a34873eea140f336c0ec60
6103 F20101117_AABIJL miller_j_Page_200.QC.jpg
625c60bfefabe7a8d897864546c20309
d1217694d547346a4f84e458e850cdcee879b1f0
36455 F20101117_AABIIW miller_j_Page_065.QC.jpg
887b69d234a7e4e99da727382923cad4
b946b26314cafd494f6ce35ab9463b7fefc25dd4
105191 F20101117_AABHGJ miller_j_Page_124.jpg
0e68c25a509e692fb0edfbc899c46c52
642aac0bea27d67aafcddd9d04e4cbd2dc1c8b37
108665 F20101117_AABHFU miller_j_Page_109.jpg
c985af48cb3e4e920a4940e000d1c6c3
61264a5f4f70ede9afb4bafb410631c0afa07e71
3565 F20101117_AABIKA miller_j_Page_174thm.jpg
be1225e1f871727d45f2c89e16bbe99e
044638098761047861862783e77a8cbb235aa18b
36194 F20101117_AABIIX miller_j_Page_100.QC.jpg
6cb85e66f722648cd83f07b956c83e1c
456a5f1fc12f323636449b5bb718d54691efd35d
105334 F20101117_AABHFV miller_j_Page_110.jpg
74091c50c1dda739087d08bef185706e
bdeddb4b672295a0863501bea69eabbefa8602c5
9220 F20101117_AABIKB miller_j_Page_066thm.jpg
9b0fe4bfdc60fa12a11481d06e9a8944
b278aec57c693fb9bcb1405b1cbbe3f7990b87c0
7912 F20101117_AABIJM miller_j_Page_186thm.jpg
5f5fa1c532f766543d19fbcca0fa9c14
29c24c22ce46c63789cbc0d236108d7b750bc4a5
34428 F20101117_AABIIY miller_j_Page_111.QC.jpg
ec1743764b7d1c8d085b9399a77c1a2e
ae4c36bee2f52adf337a82a420dbbb1a9b1cccce
108054 F20101117_AABHGK miller_j_Page_125.jpg
9539e0872fe59b9d531c0d52a00627ef
76731ed121b973938d886ff5aac41ffa8a335661
97704 F20101117_AABHFW miller_j_Page_111.jpg
46f7b05b6f870e4cf190e7880d232f0e
2a6ec60f558c8dc311594a4175fae3aaf70d0bdf
7388 F20101117_AABIKC miller_j_Page_144thm.jpg
d9c5220e73e3c30e34f51be434f772c0
cf4f67d93240c4f59015ed8ec923c9eff3034ef2
34897 F20101117_AABIJN miller_j_Page_121.QC.jpg
e0d65cdf474368e13cb5d1d4d24728aa
096bfe138af35c36d8606c84d579d511bd05d20c
8634 F20101117_AABIIZ miller_j_Page_064thm.jpg
99b62c3dc2f1fc5f67408c8c7a78477e
6a1b8f67be6ddbab1b760fb9408508a1c27fd4c5
105490 F20101117_AABHGL miller_j_Page_127.jpg
6084ad2315a9d37da1152aa36f6573af
290f255a1fe61276a4f8cacc22f75f4fcc98f0b5
102478 F20101117_AABHFX miller_j_Page_112.jpg
c402504c54b781d1403503643471b907
47bb54d1c08e82f911fc49f4824e1a4729d652dc
81069 F20101117_AABHHA miller_j_Page_147.jpg
07eef507d0c1227459b877704587786e
348ab7f0b341240f3983b216c6334f57e3ab2799
4636 F20101117_AABIKD miller_j_Page_086thm.jpg
bbb27a2aa3ce35cc76b00dd47892e0e9
da3b8256e1ff905ce4264038659b53450b48bf70
22759 F20101117_AABIJO miller_j_Page_089.QC.jpg
366c8f40fcca17d2b38fd4d5a62b33bd
46c7ef116839777b121492eeffd1d3052e1e4649
106588 F20101117_AABHGM miller_j_Page_129.jpg
ac089321f341b6ed8c2889c601ddc8ab
86b85d47d9a72fa30b3e2577750168cfa9e950cb
99864 F20101117_AABHFY miller_j_Page_113.jpg
83d8200a8247eb49b09f71b3aad493d0
4b04479eb684ecb9d4fe16a1f837c771bf269ca4
84102 F20101117_AABHHB miller_j_Page_149.jpg
ede3659c74e47b711ea5aa5834ddcfcc
0901f980d9326ff59eed8df1032f6f6329c126c8
35074 F20101117_AABIKE miller_j_Page_057.QC.jpg
f296819dcf95dd27d6fb15ce3ec3c8c0
68132e4424e172e10743c0d33b685983b6771722
34754 F20101117_AABIJP miller_j_Page_119.QC.jpg
76498dce669aa6be1bd41a238e9f1b5e
2b714e626047ad91d4000df076c6b83b5135bf44
107705 F20101117_AABHGN miller_j_Page_130.jpg
21ee52a6ea077dd7f98a4af5c8fac96f
598d31c39083e287b4a4f15f221afede31bb060b
98705 F20101117_AABHFZ miller_j_Page_114.jpg
1c0254f6fcd16cafb2dc8683283c4717
dbb567c3413c100f0419fcaf6d52b390242dd1ed
80697 F20101117_AABHHC miller_j_Page_151.jpg
87daef63c032fb2f641d63b81bc2d41b
2e242aff3f6290c5dce2075c2acfd28750b8986e
34458 F20101117_AABIKF miller_j_Page_129.QC.jpg
360dbe6ed7c3d26237a1ba45341a22be
04b635eefb46bd90211296e11889adf6de435d8a
33502 F20101117_AABIJQ miller_j_Page_076.QC.jpg
2c6a2e7f5492f5e55ba0882e33013ad6
45ca5ea7df9f2d1f67ed44ba5ea89f2140bbe39b
107737 F20101117_AABHGO miller_j_Page_131.jpg
aa00cf008283570a88584118028c63eb
2d8b2be76e6526569b174626f0fa84743ed46741
78686 F20101117_AABHHD miller_j_Page_154.jpg
5a6c3c63561aa886c7666b08d7910529
a274968d18d3862f52aeea8455f82dd6dbbede73
27787 F20101117_AABIKG miller_j_Page_156.QC.jpg
f79bd69af7cbb5c49382ace8b9fcc172
919ac7c39b9e71f6d11a67af6dff15b5d66f1a28
32961 F20101117_AABIJR miller_j_Page_014.QC.jpg
e31d1e248dd18a434bbf1a0391b25fe1
31aa4b73e22d3a27d60ed38a0c9f9852c2ee9d68
101525 F20101117_AABHGP miller_j_Page_132.jpg
784fdf41a3e0d42706562dcd2cda1055
288dfbd9067b463c44ba4561fd262c2750c01ec1
91745 F20101117_AABHHE miller_j_Page_155.jpg
08702780ab05aa183b5d89eda6dbe4da
9f4ae25e722154a148c5895a23fe89e8c44cd0bb
8996 F20101117_AABIKH miller_j_Page_020thm.jpg
8fd191068a038950c485b6dd6436f5c8
37525e58722de22531eb08092f1416cde2594588
2670 F20101117_AABIJS miller_j_Page_165thm.jpg
8e51379ecc6f4ceaa9cfb42b1f7e096e
51fccecab78d9211b1dc27b78671b0ed2a3cf929
107082 F20101117_AABHGQ miller_j_Page_133.jpg
cce7e3695beb54b8a611a387645139cf
2a0879a6ffb6e5a1f0e42e1c8f0b82e3d7c9a00e
92196 F20101117_AABHHF miller_j_Page_156.jpg
1eb71fa89df7c3d4e3002c06a325fd65
0d16d5593aa0434c32b0c09b1b816729bc2549b9
7995 F20101117_AABIKI miller_j_Page_176thm.jpg
6e3787ecaba5f7dff18342c8424b131f
0117c2edf7216438725ee846dd0915172225040c
34106 F20101117_AABIJT miller_j_Page_117.QC.jpg
87f145bb19c4785649779cd1894f4fd0
622167ead6462bc30e9d0b9b39071764e390b0e4
109942 F20101117_AABHGR miller_j_Page_134.jpg
baecf39384c504bef526a2fb6a3c10eb
0915256ad7766d0b9d6fcd2fb4d013716f131f75
25992 F20101117_AABHHG miller_j_Page_157.jpg
4c884ec3766edd134e2347527c582567
ce8baae41a8fc50a54836167b0d291081cf8f6e4
34741 F20101117_AABIKJ miller_j_Page_064.QC.jpg
fb3018f4825a8afa121228cca61f44df
2ef4d1276625d69ea31801d3f3d9c5bbf4f5c75c
8834 F20101117_AABIJU miller_j_Page_179thm.jpg
e9eff956bfa5fc18298f2874dd214341
d80cb20b7e63b2f68e7de93176f765f03fcb3494
107391 F20101117_AABHGS miller_j_Page_135.jpg
b93e25c0686ef5ee6533add0eddb96d7
94163211353321f9da12cf58c5a1c797cdcdb8e9
19585 F20101117_AABHHH miller_j_Page_158.jpg
f1612d4f457f1e664e3aab583ed50e72
5e3059f091a04f2078eb1ada923edd2d69321982
31138 F20101117_AABIKK miller_j_Page_141.QC.jpg
2d19d21726d6398ca9eed429b1cd51e8
5a0ec649b927df72a769c40f754dce23b5d8b618
8962 F20101117_AABIJV miller_j_Page_109thm.jpg
f230a5570ba36f4db98d4380f3f699f1
f9218c2d538a751a535e57871d63aba286861435
43962 F20101117_AABHGT miller_j_Page_137.jpg
73506faff73f8af86386ab7d3e214120
3779015cc99e184970cfcbf9ebaafcfb65c16988
23285 F20101117_AABHHI miller_j_Page_159.jpg
64f7cc1a50cd92b5d4248e69d38b78f6
07eba1a0643f77bd33a8e13eda0540ac579dbc68
8526 F20101117_AABIKL miller_j_Page_034thm.jpg
e90029eb610033900cccaac4d1bbe18a
e87474adf18bd2950c00afd238c5a123a6b407c7
34299 F20101117_AABIJW miller_j_Page_096.QC.jpg
d79e04a6186c8ffefe5d776b1dc56efe
6cbd9a167f9aa48395529da87895ae7a2ab61963
71011 F20101117_AABHGU miller_j_Page_139.jpg
3f704ef64428bcb5ced9ceaa86f75c78
2f6f26e3a3030abce062445c8b255077493ab6d7
23368 F20101117_AABHHJ miller_j_Page_160.jpg
9bf60288e4552b6a0bdf458485d42fed
46d0426a2a937b3316af4f7617f5b685d4074272
6800 F20101117_AABILA miller_j_Page_006thm.jpg
302ac4a5fa36763bbf277fa97123fdfe
e2051f7d9a5105b7b9f202c99bac07087663e602
34621 F20101117_AABIKM miller_j_Page_027.QC.jpg
7968246ef7cee3c6fb323b91225953b4
641e9884f9123212834482b4189b893e5d444c80
37479 F20101117_AABIJX miller_j_Page_130.QC.jpg
fe3df3001873e7a0dd2a40c279d6ce1d
ef848e4136004a6b2fcf9451e26e9b3731525769
84876 F20101117_AABHGV miller_j_Page_140.jpg
06098243e51931e76ba907d6018ae249
74b916359a9142158d09541764f021425d4ad073
22533 F20101117_AABHHK miller_j_Page_162.jpg
f5002fdfa8a5b42dd4a57ac4b426a9fa
945e63555dc62c0c71b2b79a2973129de5a2e254
13832 F20101117_AABILB miller_j_Page_007.QC.jpg
5855fef79bcdb90a522b199d5f787c52
2b4055ebcb112aa3b13bcf279f23cc50da88c1a0
33626 F20101117_AABIJY miller_j_Page_132.QC.jpg
490336f03f8ef41f27803be31c1b02b6
9c73c39a5a721b5b4dcec6fd19632948c74e0d49
86091 F20101117_AABHGW miller_j_Page_142.jpg
5197dd7b8d7ab4885fb1829614509ed8
8947ee52dbf6e5b6c3fb4e154babe04e06f93609
3337 F20101117_AABILC miller_j_Page_007thm.jpg
fb5de73bf84b358aece3267c91aa1ce9
d3944efc97917e61670c60176dc46d56cda33dce
33385 F20101117_AABIKN miller_j_Page_085.QC.jpg
6b04090aaf2c31fb253129060991dff7
eea2c361ca2cadb1fbf8d8adfc05ed411093dbae
7845 F20101117_AABIJZ miller_j_Page_036thm.jpg
b2a38e4844d9bf0b955478ee897c0c68
aa0a5a8e90cbb110daaf210ab7295151accc3592
85736 F20101117_AABHGX miller_j_Page_143.jpg
f412e627e23d70211be8123525b4b3ae
00d2a999c35188867a09c05a016d6ac67dcd39b3
111595 F20101117_AABHIA miller_j_Page_184.jpg
51c24b0b53771eb5e9512e280368acd6
578eae736086599a05ad400be6535d6453347043
18482 F20101117_AABHHL miller_j_Page_165.jpg
691288d4cc781abb4ff85bd68244c5af
0f61604fe9b7d51114a835b23eb6cf5faf26b3a7
36885 F20101117_AABILD miller_j_Page_008.QC.jpg
a25bc588cc38a2f1bbfc2a8dbcce1649
b19355be63c32556aea9c1e277f198e7d395dd84
34927 F20101117_AABIKO miller_j_Page_108.QC.jpg
863c20d98068f1a5c0a2992eece6b428
2684b90872b7fb4fe7e21396fbf745b974b78880
83383 F20101117_AABHGY miller_j_Page_144.jpg
f3906656d3d8621b95adee6eb993d8b9
6b30a6fc337bc3011eea2305f718e1dc0213e081
96186 F20101117_AABHIB miller_j_Page_186.jpg
8d5d232ce0404d1a1fb3ca8a2a2df4e7
7e0ffae0afe77f65c2f2ff85ef924f9eefa65726
18448 F20101117_AABHHM miller_j_Page_169.jpg
2c9e90b6bb2997f51de0371f16d18f55
c2dd09c09432cfec70888ebcc27278d2d3e01769
8519 F20101117_AABILE miller_j_Page_008thm.jpg
161a89447fa699095aead8b546faf231
5ab7e0f42304a6ad9de29f839ae7b319a6b6e84f
8849 F20101117_AABIKP miller_j_Page_101thm.jpg
abf82afd7fc1e80c047bb093c2f2f5e7
3da4c721f059fca13e087d46bf2cb36155c604f3
37574 F20101117_AABHGZ miller_j_Page_146.jpg
de9631dc279f408212c90ecffda5e690
ec2851636e4b788bc9838a1dd4fd3669df1d69f8
107464 F20101117_AABHIC miller_j_Page_187.jpg
f836c615fceb32bce9509eec7f904f4d
56b86abcfab19aaaa26a55a005d4d02ee6494afb
18501 F20101117_AABHHN miller_j_Page_171.jpg
e96b9953eb144abfea695c77314edcc0
575cf5488305f291fe2299f5f59aa15df15a792d
42200 F20101117_AABILF miller_j_Page_009.QC.jpg
bf457542f7d8e88b238aec2e7f889b5a
f683155fd7efdd9ee50effe171a36cc696c6730a
2822 F20101117_AABIKQ miller_j_Page_172thm.jpg
912aebfb84ad5bc54e981d3e5f15c89d
52d414a9611cdf50f81faecf8cd574f12c38e165
108907 F20101117_AABHID miller_j_Page_188.jpg
88cf718cf1e11833f63647f1f5e65bb9
dfee13dbb8e0be102ce2c487ed8360107b242c1f
21314 F20101117_AABHHO miller_j_Page_172.jpg
6f037a9ceb4ed593c389ecac992391da
cf50c50ef654fc5456b8aea8a7755f5097afebe5
18968 F20101117_AABILG miller_j_Page_010.QC.jpg
a7857a38af0d9d1518e217a33a1f9bce
4ee00bf494d64ddcfa80609585f83bd33ea86cf5
6651 F20101117_AABIKR miller_j_Page_071.QC.jpg
e157df6e70fc7e43b91b781efb684e86
33e000137423b2be4f2895f1a70f98de86fd6f01
103766 F20101117_AABHIE miller_j_Page_189.jpg
551324fe4cf98918032179832096c1f9
81043f079d5c75fea9ec67a87abf3fb68e742af4
33688 F20101117_AABHHP miller_j_Page_173.jpg
8b1ae9f5a59a3def5885eadc5d31c3bb
2af37ed025529f11f6fdbcd7f1f153d4694c4428
34762 F20101117_AABILH miller_j_Page_011.QC.jpg
8cc095fea0b30514d629fa7ae07f8e59
37e7b71d734801587e4881bdaced1bb1a9cd5b36
35399 F20101117_AABIKS miller_j_Page_106.QC.jpg
dee7edaede3fecc21bb5d9f31681311f
d24a29247280aba8578e299b5c0f02d1d7bd5c80
112272 F20101117_AABHIF miller_j_Page_190.jpg
867160ff99cadc7d40d215ae34ecdbda
cf2be22f551b016fbd2293f5e91d0d5c11c00b93
27820 F20101117_AABHHQ miller_j_Page_174.jpg
47f850773f39992a1282830070c821e6
8adc798ab80ee94dff6dac83f8b8a451100dcc58
29159 F20101117_AABILI miller_j_Page_012.QC.jpg
ecf1c45c763f7ff31ae5a25dc3bd68b4
c492ef43079290460a467c6b22b995f53522e6bd
299559 F20101117_AABIKT UFE0018500_00001.xml
d4c88d9cc4f2c2f79c8e059a844fe019
89882e674776de69ede78188e63d1568d98598cc
25399 F20101117_AABHIG miller_j_Page_193.jpg
64578f5969eec24c4b5ce959f4ec8146
810340ab3cbf31eb2a4efedd13ca99b57f7841cd
32886 F20101117_AABHHR miller_j_Page_175.jpg
fca22577954764c314570ca48306f6b7
462ccb5b6a9e78d4a31ff0d96cf57a6a1bc7d290
7478 F20101117_AABILJ miller_j_Page_012thm.jpg
0e9fc2c3d687800207a006c12cb6d0cd
1e7f4d0c06c3fa0368cc3bf0729d351f4b3f81e4
1774 F20101117_AABIKU miller_j_Page_002.QC.jpg
e3da433e9096c7fc7d9183faec0e169e
b41343be012e7d852f980fe4657eddd25d43cb21
26020 F20101117_AABHIH miller_j_Page_194.jpg
85a6d7c61cadbe6d0d271a708b0fbe70
e9e799de97a98f7bba2ddb2369914e16adcb5dd6
107672 F20101117_AABHHS miller_j_Page_176.jpg
33ac013efb556c9f8ffa4f1467d90d5c
a5c9b693fb84026bdae78dc28f32a551119eda94
12696 F20101117_AABILK miller_j_Page_013.QC.jpg
3c3f7431d4c4494514bc4c52b7fcd6b8
e033fb885181b12ea9f21d01196241d0e910097e
646 F20101117_AABIKV miller_j_Page_002thm.jpg
f4b4d0cfc2f4afa82d94ca5c945573fc
c7ed4ce81a24c7da32fefad7a49d5a0812f30434
111698 F20101117_AABHII miller_j_Page_195.jpg
d7fab798a1f1da08447f63342ba91d88
b9477e359829c203027720109d86705fd47cf76d
112350 F20101117_AABHHT miller_j_Page_177.jpg
b3ec8e1f2e3003cf6dc688adf3e5ddfb
5e025064cea6507a8dad7dfde79d95696231d3e7
8690 F20101117_AABILL miller_j_Page_014thm.jpg
6336e4b1a26d85d80f7835745d6d9e07
e7c95624f9da3886613de14ba8d3819f08be49f6
1600 F20101117_AABIKW miller_j_Page_003.QC.jpg
fefacd10acf56f90201d599833b6b1a3
d795388160f7ebb82e1aa854322fb916804e1b81
127269 F20101117_AABHIJ miller_j_Page_196.jpg
0be82169197d65de4499226e409672f4
a99dd190e6df167d1a49bbfa30f739b661ef7512
106638 F20101117_AABHHU miller_j_Page_178.jpg
09d14e1e4f0879247b93d9c87516baf4
31e2f765ceaaa96dc8e2a45827d80c73b413685f
8467 F20101117_AABIMA miller_j_Page_035thm.jpg
92f411589ceab0f0bb29f3d8800c4d97
391cd53bcd871e3792d33037bfe323a0c1e47f91
32973 F20101117_AABILM miller_j_Page_016.QC.jpg
d9289559a6ab2c16fd16ccb9ce40f7fd
1aaab862cef4f37f05feccc478f7c4d21775bc57
669 F20101117_AABIKX miller_j_Page_003thm.jpg
b480c6571d78996f9ee84976dae0ac2a
b61bb7a55773e56426bb2b4697dfa93560d85997
108769 F20101117_AABHIK miller_j_Page_197.jpg
cc297777ab1800462c21301997b6f4b8
4962cced01d9c51c019c8de48ac8c8a03ba1da11
104899 F20101117_AABHHV miller_j_Page_179.jpg
021a20fe9c16eca5ad367f5a8d8409c4
e9606fab2464dcc1ab5b06d5bd4ae9895251e85c
27713 F20101117_AABIMB miller_j_Page_037.QC.jpg
49d38102db637332c16629a103119250
31c70728e59550c03e2d61f0ba9f00c2cb7b7cbb
8330 F20101117_AABILN miller_j_Page_016thm.jpg
64bed46a27999cd80cb835d7787d7200
4f2756417d859893c116d5a4814c09d242ab9a9f
22983 F20101117_AABIKY miller_j_Page_004.QC.jpg
fae310d18683fba92fb077b51063b6a9
c769ef31b68e57562beaf76712c44d677203ce68
114384 F20101117_AABHIL miller_j_Page_198.jpg
e6320f052925b51f62ad83e27b4459a8
eba0ae2dfec96d512587a24be49d8b5c495dc7db
107475 F20101117_AABHHW miller_j_Page_180.jpg
9569082941dd9efec3c6a71cffe6e7c1
274c1358986bec1ee0e4014795c908dc4931a471
8143 F20101117_AABIMC miller_j_Page_038thm.jpg
3066222b70cd573572681ccea8ebce75
a9de381af69f37c18c6c65c3412ec7bea2028826
27553 F20101117_AABIKZ miller_j_Page_005.QC.jpg
03b45c57865e25f960f5bb1e12ef03e2
1e023cafb89706eff97343bbb503a994d37a24fd
F20101117_AABHJA miller_j_Page_019.jp2
ec2b7b50edd4486cdde70c2f9494d760
c38ca14ef26aab41084ed62db521389b7055e091
111753 F20101117_AABHHX miller_j_Page_181.jpg
04a748660f303c833da4726f2339c1b3
2f6685847efc5d633c44b2b4b90254697e7b9094
9001 F20101117_AABIMD miller_j_Page_040thm.jpg
a9ce5204895ef5f08c43412167c72b60
ce21a97d7fe505fd01809482e916299e865895bd
9049 F20101117_AABILO miller_j_Page_017.QC.jpg
4c77a2120ccd16f356c01c9cea519cf7
c839b1ffa1dff2be7e16e83c79e002ac988b373c
1051929 F20101117_AABHJB miller_j_Page_020.jp2
879b91d3a40d99b06ec5138b1fe61b7e
2b88bc922455cc50ebccc04154989cdb2195f960
20530 F20101117_AABHIM miller_j_Page_200.jpg
1f17ba4cc5da6410625ad947c27ad102
c9e7197180f003c84ab91994a177ef75e428e775
101938 F20101117_AABHHY miller_j_Page_182.jpg
f07beb4bd955140bb821f4a7c64fa110
b3e43dbffd07ff50d1c9c8c807f1c3a8ec0f6f98
7487 F20101117_AABIME miller_j_Page_041thm.jpg
3dc4985f4864b3bb988817863c1a44a3
1997cd15b021184c9c41fab64ef037c2376d8ef7
29444 F20101117_AABILP miller_j_Page_018.QC.jpg
a251474d1b56417a5cb873473b72c66a
7c0ce5a6d76a84b0f5346ebc22180356093072fd
1051923 F20101117_AABHJC miller_j_Page_023.jp2
ccd8afa942e54a197b85909e24d28ce2
4ac5448e2607955419fa42a585cf5c5dcc621cf2
43712 F20101117_AABHIN miller_j_Page_201.jpg
70cd79e0ed12e30a9f7c93770f2b38ec
fd765300c5e0be99b642d296245fa15ec604bec2
109514 F20101117_AABHHZ miller_j_Page_183.jpg
c45956d7a0d937e8c94af966a590f7ac
0bfb0e8291fb30046aac07ef231585c75d61adff
29107 F20101117_AABIMF miller_j_Page_042.QC.jpg
9826cfbedf358983bb2b1f85cc92cec7
ee9a77176b0486e1f79b8c66a9181d4ad2ace7e3
7170 F20101117_AABILQ miller_j_Page_018thm.jpg
f7fcbf5766609f04c450944a7f6fe361
37df16810556c8dbcc2f6181a59303220434cf9f
1051893 F20101117_AABHJD miller_j_Page_024.jp2
548a44c42c6bb1ba659f94e736e345a7
72a6caa0e0e33ec2d736cc1d6e43ccfb925ca85c
261892 F20101117_AABHIO miller_j_Page_001.jp2
6fb38950b7893ec45f5de1115af60c5e
eefac957a88bd3f0850d253167632e26a2e9c27d
7663 F20101117_AABIMG miller_j_Page_042thm.jpg
985f822b0c63c3117ddd73f07f316a29
1ba5b9f164d3fd4ae29d51615f9789fb7f222c5b
35240 F20101117_AABILR miller_j_Page_020.QC.jpg
900b56c0e717c5dcd692265e132d2692
ca2f01421866beee5eb8518be30e095e450e2541
751003 F20101117_AABHJE miller_j_Page_025.jp2
3e45c7e7e74ba4635899b1926b76c974
b8286a18b3559bb79a1bb38383b5b006162723a1
42829 F20101117_AABHIP miller_j_Page_003.jp2
46b6609d6d6e7bd800ac1c9b0cc572bf
86802f92641d1239a663a4418be8defa25f5e0c3
8812 F20101117_AABIMH miller_j_Page_044thm.jpg
e4d244f7b2f5d5eb6bc09fff7f6c2baf
297a584836d6ab60201d0f09d43ae9391244a169
9138 F20101117_AABILS miller_j_Page_021thm.jpg
5697ded71f7f2249670d9bcc9c513e90
ac067bf6680cb3f62c68c3360d18088f96c90e35
782119 F20101117_AABHJF miller_j_Page_026.jp2
26828e9f3e8ddc8fd03fcd08f43fca5e
d0ffcab5ab91f2da44c10a1e83a597a84ae94619
F20101117_AABHIQ miller_j_Page_005.jp2
495caf9f74e2205b84d425a2f65e2577
60841580582f750e1446819d8f8949159fd91197
30060 F20101117_AABIMI miller_j_Page_046.QC.jpg
d09b471604064fc08b4231ae8c7f6d67
d7ba145956417a572dee0338703c44beb597da0a
34064 F20101117_AABILT miller_j_Page_023.QC.jpg
4d664ba926bd9080d38ec1849c3af4bf
dac13d6efceb15bc1d0a178e342716cf14e33105
1051909 F20101117_AABHJG miller_j_Page_027.jp2
027000a012f21824969cd7323d3d9b89
064a69b9fea85a7dd3b9a90f55c3e4a9d9c92217
F20101117_AABHIR miller_j_Page_006.jp2
32a900e017d8c554ec9a3087c38fdab8
ce42f3655763a80106830d330c0f54beca436e4c
7635 F20101117_AABIMJ miller_j_Page_046thm.jpg
b13aece1ddec06cc2fccbf7ab80489ff
33c972e0f1da377b561bd5b365feecbe1fac7071
8929 F20101117_AABILU miller_j_Page_024thm.jpg
cce5d56d745490e680d42230a207ae93
1e7ec4c0ea74791fe0394826431a284390bd6150
1049302 F20101117_AABHJH miller_j_Page_028.jp2
65cbffad2a4c9f31a8bb10d5c1191d30
9323ee9a1af45837fa0f8adc12483fd2bd1aa2d3
1051964 F20101117_AABHIS miller_j_Page_008.jp2
f95d998095443efd3c1e597845983537
69edc8a08e8656e9844d716bb670c45baffe5203
25690 F20101117_AABIMK miller_j_Page_047.QC.jpg
706ca0d6670fed8369fe46cfe7f1295c
1d5d5fdb4badbbc4d39a2345e6cc968b51d5058c
31969 F20101117_AABILV miller_j_Page_029.QC.jpg
22db7da8c47913695041a12225c754f2
0dca21509a8fb454e2965416fc5d435123dcc4ef
967006 F20101117_AABHJI miller_j_Page_029.jp2
e913c2bc115010b0a4b2948a151653b2
2345df56511db58a03faa3f18e7c6d8a76ead2af
1051973 F20101117_AABHIT miller_j_Page_009.jp2
e14b0bf77a29c0570b65de9cb3238f94
378f30924ec80695d1f2247423478d8b2ffe83f1
6939 F20101117_AABIML miller_j_Page_047thm.jpg
1afc634de54d453c701e23cf493bf9ab
33d59f6f1ae3b0abfeb07248f00ef096cc5f6238
31088 F20101117_AABILW miller_j_Page_030.QC.jpg
112383af3f687120425d388841ce3359
9f1a48c45ecf828762425ad936480c1c7f8303bc
1021162 F20101117_AABHJJ miller_j_Page_030.jp2
f1aa3bc708c8a57514101ef1251c252a
803954f7cca7fe4f1d169d99eb2a4c95c4676ea4
1008248 F20101117_AABHIU miller_j_Page_012.jp2
6fb40e770da96765911cf3b6cc7951d2
4f4a6f60ee39f2ea808aaf1298bc858fff646f57
33966 F20101117_AABINA miller_j_Page_068.QC.jpg
5d1573cba5d82433c316591d0d6a8194
a16df83a3b3ec0d1a83ae5b97b1ba1e17d9b36cb
31375 F20101117_AABIMM miller_j_Page_048.QC.jpg
d61ac2ed2716bbd1a441eed671f32f52
b2c706fb0239539acc38863c26ad9ce3ff252a90
31893 F20101117_AABILX miller_j_Page_031.QC.jpg
b1badb91d1330a1be99e0f16d510ded4
9efc69cb17cd03dd9f6c3cf0edddec2d9ac81e87
1043947 F20101117_AABHJK miller_j_Page_031.jp2
757e8a2501d2e3d9ba51c508513616cd
5ad5a466afc0335bb5bd97f61e02c5499481453f
389479 F20101117_AABHIV miller_j_Page_013.jp2
4b9e40b465f09dc5427b537d4752ba60
59bf4b292296052d42586e76521c818057fa0f5d
8872 F20101117_AABINB miller_j_Page_068thm.jpg
17f5d41e2c907e3fac6af82db8ff3f84
3ba2f79e3bc32badb60e5ca8fe53c87866f973d6
8684 F20101117_AABIMN miller_j_Page_049thm.jpg
65c870d7bbc2540c0514057c84da0d28
f3b9c3a72876dc91cdf6db10a8e9ec1e5d476727
8035 F20101117_AABILY miller_j_Page_031thm.jpg
fe64a63abe914ab09aefba5f272a3197
d0fd455fada3432c88140ced09cc45010182d5b4
F20101117_AABHJL miller_j_Page_032.jp2
5d3a9f0fbcdfbf5da6f8faaa80983e07
7e53e420fd227fd241372b0a80bc4c20411ccf1d
F20101117_AABHIW miller_j_Page_014.jp2
2fc6c89d39ea2cf9ca791881d80cec3e
654f1e9d6bc37c9f8e5aeb5dad49d23bc3ab8624
13942 F20101117_AABINC miller_j_Page_070.QC.jpg
84a82657745255e326bf72ad2d2215ce
3acff516944309cb4df28ac23131c7713e9a28d0
30711 F20101117_AABIMO miller_j_Page_050.QC.jpg
be35fdf1bde40de5edc8628f999c9d75
a2f0d8b9d8444513b1e9ac5d2dcf3373078085eb
8946 F20101117_AABILZ miller_j_Page_033thm.jpg
99bf56936bde67d9cba9ea9fc3db5c70
f1288bd0d42e41f9983b3eff67e0966e3403f109
1051970 F20101117_AABHJM miller_j_Page_033.jp2
b8e3d6ea7950164e85c24844d717c7e9
dfb9e41ddb5ed6bd8b82d1d20cccfe61a2207116
F20101117_AABHIX miller_j_Page_016.jp2
4ab0e7b87619f186412125c5206513fb
47f7a611e33d28265c80ff42f99864da4025bfcb
1050555 F20101117_AABHKA miller_j_Page_050.jp2
d3f400d9dd14f2ca81d906296173accb
058a79615a8d4da7284ea64c376e7d3b4851ef87
31498 F20101117_AABIND miller_j_Page_073.QC.jpg
533527f15d76d100198f0b1a5c59ded6
cbf73d78bfc871e4e24b681d3d480cd8a056733f
377511 F20101117_AABHIY miller_j_Page_017.jp2
63024a48230bbc52a6203ae7d39d9244
f8cd0b026c321ef4cab8d3ef8209f3619e62042b
F20101117_AABHKB miller_j_Page_051.jp2
af0b32b1d470d20217c3c13c7340a23c
3ea2d8a252114aa338c26a1e9a78f73cbf0fe443
34586 F20101117_AABINE miller_j_Page_074.QC.jpg
1c4386268e1dfa07147b1f25b319ea47
9f4abb4336948092e95f50bf61dc9ce6c658edf8
7830 F20101117_AABIMP miller_j_Page_050thm.jpg
5334e204e65fac061d357b9b17a820d4
c592cb20f6c551b31322f694d79e864f734b47db
1051940 F20101117_AABHJN miller_j_Page_035.jp2
4443e881e9e5141e219a3c4a03a9dad4
b886803ea659f3f102b78d8f7ae9d5ae0982daa9
998159 F20101117_AABHIZ miller_j_Page_018.jp2
b12185603873e39795280b0b7e34d0af
e53b1a37f65687e6d366cea510700bc44381cdc1
F20101117_AABHKC miller_j_Page_052.jp2
107661269f834cfbb86a575d046da79d
6bcfe6b31c8ffa051f5e2eaeb3a1ae7bd31a1b28
34273 F20101117_AABINF miller_j_Page_075.QC.jpg
9e27eef4f46484339dfb5f817e4eb0f8
29d19763c5d3af2ad59d0431fe4994fe8539f15e
36416 F20101117_AABIMQ miller_j_Page_053.QC.jpg
d45228ab20585c8327a2fc5693c2ec92
aeb397695d9eb075fae8b977055606dc88bafe72
1038547 F20101117_AABHJO miller_j_Page_036.jp2
49e7e245ae47ea1ed00f93864972028f
b598faea4a9a012251f7203f4c24434d255c8607
F20101117_AABHKD miller_j_Page_054.jp2
62d0198eb527c9733f49fb7d83f52c26
f964c39b03606664951ec44d8989d357e2b12645
34437 F20101117_AABING miller_j_Page_079.QC.jpg
922e09ad9bc17a171152401612717883
13b1bc17f4dcfd3861377f886c5734eb72a1fb33
8785 F20101117_AABIMR miller_j_Page_053thm.jpg
eb3fb0f531356b336e513f5959ec13f8
d2ab0a5ac651569522fa221073b4f7bddf828218
877467 F20101117_AABHJP miller_j_Page_037.jp2
dace1c81147f60bd7adef33614ee7aa6
d7ac19ce6d194c178732a434ebc9b497dbf15218
F20101117_AABHKE miller_j_Page_055.jp2
57d21c2c6f470f463fca745d872fdc10
896acb416b1192c8541249e2c06c3ba2044c87f2
7359 F20101117_AABINH miller_j_Page_080thm.jpg
be143ef0545fcaf9efc3c045fea7d181
632b859ff2a1de8d1aa69f50d639e41fc02eba53
37061 F20101117_AABIMS miller_j_Page_054.QC.jpg
4b6a27542940cb1de47b849418397d60
fb86384695e7c15ba2b44f046dac5ea6a269564d
F20101117_AABHJQ miller_j_Page_038.jp2
d04e533dd530dc0c3463d44963814cad
c7c9df0ae0ebef25c092837f9903509fbe1ddca7
F20101117_AABHKF miller_j_Page_056.jp2
8da06ad00e1ad3f690b1b3ae8771a13a
6ddbfe8a2e217bd863bb7ac7c96ccfee236c31cb
7888 F20101117_AABINI miller_j_Page_081thm.jpg
9f07b517c147ead8ba5c3e884168755d
e0f7675fac9f1be3127bbdb77f4cafd16a0e0e55
8880 F20101117_AABIMT miller_j_Page_054thm.jpg
286033f5e81e6442c9669243b6d62080
a7c306692e3924551a3a5f8ff36f04e813cafbf4
1051969 F20101117_AABHJR miller_j_Page_039.jp2
3da02d3e3c3867c97996dfbe39380ad1
098324b1fea570ff6a8d896999fee427734c4e46
1051968 F20101117_AABHKG miller_j_Page_057.jp2
71a2ae3f7540826aa7f5615611f94f54
5ad517f9663bfa53a55f56f4018ce6fc441436b1
32977 F20101117_AABINJ miller_j_Page_082.QC.jpg
b620d50a95af0a20428fc809423d82f7
7e5c281f9a20519f8d0c32deafc3c011c15e84bf
35217 F20101117_AABIMU miller_j_Page_055.QC.jpg
d1d54bb888df1d31fec386b054df83d1
7cb989d3fa506cad751882cfe0fe4787fb2dec6a
1051954 F20101117_AABHKH miller_j_Page_058.jp2
767e90d43b33bab52111d34508c6e22b
604d57e17de3ed71f9ae6e2a76d3155b7dc6c9fc
F20101117_AABHJS miller_j_Page_040.jp2
7ffda2deb4aa37019d76c39fe30cb0ee
7bdfdf9229a79f5267520863b1c6b2d2a165dfb5
33427 F20101117_AABINK miller_j_Page_083.QC.jpg
11418d99037ede2237d571d4255151a5
f70fcd7b5a061042b7d43eadced36e25359e16a3
8654 F20101117_AABIMV miller_j_Page_057thm.jpg
bf6300b2459b6dbbae50221dd21fd79b
d4b984f0374007e8c5b004a526d5a2bacdc5c057
1051975 F20101117_AABHKI miller_j_Page_059.jp2
d7177b82f828dd98d4c61c8072d925fb
a57a6cd69eca3da5efd4c0704bfa242004fee371
964265 F20101117_AABHJT miller_j_Page_041.jp2
2f2a415abc696180bbc47f2c917253db
bff01c4e19c25b59ebbf4ee6f97e5620000e5cbe
8600 F20101117_AABINL miller_j_Page_083thm.jpg
1ab0816c93dc296483eab66be655ca5a
5b5a4b854ee1bd5d00a2d81ef26cccd5ac75056d
8705 F20101117_AABIMW miller_j_Page_059thm.jpg
d784a330ea75bddba50016ffa2f7a7f3
9dd5f1eee8e26508dd87fab40d7a42de228f4d32
F20101117_AABHKJ miller_j_Page_060.jp2
897c88c789eaf75009a4cf98c20bfeb0
cc5a05b273c87387f124143f22efe7e7edf8a0ed
938920 F20101117_AABHJU miller_j_Page_042.jp2
b34aa2100ff314fc5949a9fea0f90010
6cf4da36b31dd57460156a719c3b681396b5a525
36373 F20101117_AABIOA miller_j_Page_099.QC.jpg
1d6d1294f14edafb2a350f8a6d740430
f270329ed139f5157e2e1dd18b29e78df4b047e1
6857 F20101117_AABINM miller_j_Page_088thm.jpg
7b6bca713f1cc47536c53b2ba89d08cb
5632d373105f7fb7dce603d320c15be30d7db022
36518 F20101117_AABIMX miller_j_Page_060.QC.jpg
8898281b4d45fffdcbccd8ae0f9fc0d6
25bc423a28cacc2a4953edbc4014af290a8744a3
F20101117_AABHKK miller_j_Page_061.jp2
f0e9029e1e1b4a4cdc9e6e94cae39cdc
2df94c9763c38985d86ad0d6d0357667fb8d20b4
F20101117_AABHJV miller_j_Page_044.jp2
56068c7b1a887d479984f481b2cefba0
e85a8e642c2e7aff41c6860219d963bea02839e2
34208 F20101117_AABIOB miller_j_Page_101.QC.jpg
2ffd05cf05b3e3b854382b694fba53e0
d0c75f13c88e49fa395cbb09a5f43821321c43a5
6267 F20101117_AABINN miller_j_Page_089thm.jpg
1fdd366d257d5ea3913508778e2d0f01
df9cb51eb8d0f90f1ee00a24a9c5bc11020283fc
35623 F20101117_AABIMY miller_j_Page_061.QC.jpg
d8ea805140e91e792557499f59997e7c
f6caef60a7bb52f8c4b1d794cb40850fa218b1c7
1051935 F20101117_AABHKL miller_j_Page_062.jp2
9eb81848dedda8f74ab8891011a4a7a3
4a339b53c22aea88004543a3492a4248a6745422
F20101117_AABHJW miller_j_Page_045.jp2
363c6176926f7a13d78cfdfcb9608fd6
e320148f04de0746ec2ac5bd35d41de562655ab3
30173 F20101117_AABIOC miller_j_Page_102.QC.jpg
5415ee70510024ea55754031eca74b7f
64d8415741a2748428d49b044dfa94a138232a40
6442 F20101117_AABINO miller_j_Page_090thm.jpg
bb4ed8b60b654550139d4cca16c8eaee
ac2c452371211747578f5e312e1501a01151372b
37031 F20101117_AABIMZ miller_j_Page_063.QC.jpg
1aa3704925421f61801abd7c7e67cab6
6dc25e106da76403702046fe883fd126c53048ab
1005860 F20101117_AABHLA miller_j_Page_080.jp2
ccca29d391c016aa56910eda5a476adb
f90ceacdc391c90b359941b07a6827c0a0a87d71
F20101117_AABHKM miller_j_Page_063.jp2
93def7596c8dc939e3a33bed1e26548c
9de1aa8bd3ef6cdb6be020fe65d0a7d8b58f054d
996323 F20101117_AABHJX miller_j_Page_046.jp2
1e409f011caab0492d55bd92963a8324
eda4e4b8ca964dc3aa97f9839d1eb76dd1c59f0c
8567 F20101117_AABIOD miller_j_Page_102thm.jpg
965b61758ad1122ec0994ad52d337d8c
c5635cb60bb590b7f88ca2d8e92bedb183d4db8e
23909 F20101117_AABINP miller_j_Page_092.QC.jpg
6af280c149df0dd6c744dda802d78274
84b6456647d4ac0eb5cb8fd2a7b0723f9d5cfa5f
1050827 F20101117_AABHLB miller_j_Page_081.jp2
882bdcd2f2cb8ae53756cc7001f88cdf
276ef8550ea2bf582d92671537791831c27bb7b0
F20101117_AABHKN miller_j_Page_064.jp2
eadbffe43aa4ec741833ba40da274f0f
dfbec8c1c7a734902a83273cd8604a638a3829a9
792580 F20101117_AABHJY miller_j_Page_047.jp2
ed55321d7596f3a7b4ca8c4f0be9fb9f
1297eddda2c7352d6dda45f35cdc405c9cd5f634
33756 F20101117_AABIOE miller_j_Page_103.QC.jpg
5351fc8ea4af58aba2708dfaf559d357
8936c03d860dc54b0f685ab41db8e3b8724a57e5
1050678 F20101117_AABHLC miller_j_Page_082.jp2
0f65378bc7bd5a54a8767bbf57e1e1fd
a4b00850aa66fb8f51f2fa0b5f850d936b3027dc
1044273 F20101117_AABHJZ miller_j_Page_048.jp2
be4143155101d5c5215e2ce1bdae2d2b
3a0ea7b225dba7e4ae2f47522c247327d21e48db
36250 F20101117_AABIOF miller_j_Page_104.QC.jpg
b4e8d4cd9ac6fbaa16e84ce09fc3c2cd
63fa9af485b51d3b761153386e7332ae85f356b3
6669 F20101117_AABINQ miller_j_Page_092thm.jpg
55466478228b0214be9f7604c13901bc
e43fb6d79f1b1b29aeeac755bd873b9bfc642b45
1051958 F20101117_AABHLD miller_j_Page_083.jp2
c5919271d92f363cd93974906a467baf
bc128eb9f0f43201587e3a9c6201742c86c1adda
F20101117_AABHKO miller_j_Page_065.jp2
f98a926d2b81d9f19f3f2e46d42b0f6a
13f775efe83f15731e37b40e1fefad6b58c14d8e
34874 F20101117_AABIOG miller_j_Page_105.QC.jpg
55fcbf6e1d8f94bba899c8fab4f0e384
ccd5e28c3be2044a7802df06c6f594c771153c57
13933 F20101117_AABINR miller_j_Page_093.QC.jpg
c23e34fc44198d9000f6c4c8cecbe9a3
b06ddc6870c99656359d229a6c3b2c1d61311949
1051921 F20101117_AABHLE miller_j_Page_085.jp2
87403bf46a0a463ccb3891da5dc347ed
f853711277d581b8d74568e8bf38989fe2c0f153
F20101117_AABHKP miller_j_Page_067.jp2
f8b1f393e70feac3236f73f22d8a560b
84ff69f18010deb756b3b5384237b54bab6b517f
8699 F20101117_AABIOH miller_j_Page_105thm.jpg
e22762aced4344bd7e3fdeae2d9893c9
2ccb7f066e740ad3aa2a32e03470f9eb962e0464
3868 F20101117_AABINS miller_j_Page_093thm.jpg
7a4083b248b91c523dd41d082246b074
ce813d1083d321c6b47f0dba43b7ab5c12d46134
592248 F20101117_AABHLF miller_j_Page_086.jp2
7b6ab8768f82cf761ec8e68fe32c21d6
49017a6ae18c754b7bb833b8cad367797327e693
F20101117_AABHKQ miller_j_Page_068.jp2
6c96bd295b4069de46820f0439ae8829
93cc70cd072d2c297d19f051b6e854b6423bffa9
9009 F20101117_AABIOI miller_j_Page_106thm.jpg
80752e74a364e9f8f5210fed1638da4d
301561f777ba2a9a8c840b12a33ad8e3117a854b
8759 F20101117_AABINT miller_j_Page_094thm.jpg
5d8a386645bf4ee01e605c10f9cf1abf
51e2612c541973d0fddee1d055e245435c14e230
821899 F20101117_AABHLG miller_j_Page_087.jp2
182179a37492734fccca081844abc3eb
7bc408339f45b997407c20c4d18ffce753935aad
F20101117_AABHKR miller_j_Page_069.jp2
8a71e208548ac8301e369c8166f99d20
28c3a446166cc5873f4a0c7894c075d699b90e0b
36495 F20101117_AABIOJ miller_j_Page_109.QC.jpg
7405665300eab5cfadbe748b30f74ff5
ac2a7fea12e3036b4f55d101d5488f62ac9de137
33889 F20101117_AABINU miller_j_Page_095.QC.jpg
8d37f8967e786b25d01982b9292d8c15
8d8e1191b38c7dddf067facb98a473e4df8b17fc
816883 F20101117_AABHLH miller_j_Page_088.jp2
428f62182631c544cc26255a9bfcf441
bd98c39ab326b4b37716804a44e24e6175619527
434520 F20101117_AABHKS miller_j_Page_070.jp2
9d7455ff5978aff8126169447ed44d4c
cf47bda4c2b79060af6a10a9b71e617b3227eb71
8566 F20101117_AABIOK miller_j_Page_110thm.jpg
c34b594f7a600b6dba89ee47f568e43f
cf9cb3f2b401f36d25bebe433636ab40ced5a28e
8413 F20101117_AABINV miller_j_Page_095thm.jpg
05c0819c42380d0603cb98f8886f90b2
a24602e517d52b1d744c7f86f843a19bfd1a146d
781402 F20101117_AABHLI miller_j_Page_091.jp2
0c821e43acf8ca1304b72f822164d078
1a9797fcf0002d6cb0a45a0251a8a48404e5b4fa
179711 F20101117_AABHKT miller_j_Page_071.jp2
e15bb8ebc65302a5d87e671a099be819
ab79c023b661188d3a3eaaf5cc68950cb6d241ce
8396 F20101117_AABIOL miller_j_Page_111thm.jpg
64f51a1757ea8d706cd4f4b510c55837
461f605c21a1b3a77785ad5d650aab3e80e9e154
8677 F20101117_AABINW miller_j_Page_096thm.jpg
03213f8f4a07de9ae3f4c91e61c426db
f1f95d413c4140d85cab929f385354dd2f9dd64e
740161 F20101117_AABHLJ miller_j_Page_092.jp2
db50a272e310e3b9ec64958c97012b11
6ba5c8abd6bf4c22a84c8b75c028c2d41a137b69
854914 F20101117_AABHKU miller_j_Page_072.jp2
e4b8e4c61b849eeb59511259ae8b5e89
51ce0ae3bfbf2035903016453e6799cdca14e963
8638 F20101117_AABIPA miller_j_Page_124thm.jpg
f5857447ca91668fca7a1113cf816914
bc98692b87a308d5745b562f995a98f732916e7a
8316 F20101117_AABIOM miller_j_Page_113thm.jpg
f051ff0e91dcae5a486496c6b1410b9c
6ac19d3c7e271e02dc573266d3235d78aefefa78
8937 F20101117_AABINX miller_j_Page_097thm.jpg
317fe6781d0bf3f78f4d12ef3e1255b7
1a95b7161b2b606f07778bc2d7dc6eb281aecbfc
F20101117_AABHLK miller_j_Page_094.jp2
7852206331c651b17922585c89c7808b
6a4a39d47025aab84c5d3628bf5ac2475ed525d1
1051966 F20101117_AABHKV miller_j_Page_074.jp2
e67e9aa24053db14c1ba153d2cdd8e40
7c471aeeb5c738d77d9817c8a13e9c0eb9b3ebd4
35371 F20101117_AABIPB miller_j_Page_125.QC.jpg
0ceb8065a21c62d684e2c8fce9c12567
6e9c111f5aa783d3c57eda5108d0b39f29c706d0
8435 F20101117_AABION miller_j_Page_114thm.jpg
00e250d76a3e79ce164b9a86183fea78
dce5cdc863116e78e806afa5598a5285f6cb2757
33206 F20101117_AABINY miller_j_Page_098.QC.jpg
cd939d1b62740df7e7a2282e8035dbfc
d3b14101c3bb3a222ac379cfa4403b0d6988fc86
F20101117_AABHLL miller_j_Page_095.jp2
34fb3a7ae862d2daadc02407081a49f7
6e98588575dc6c87ee0cfc3e25c3053143cc42c2
F20101117_AABHKW miller_j_Page_075.jp2
da5b2c2314a2f163cd169c16c5f13009
f1f8c21c30b6e767521e3faa8da49bf1b4925be2
8745 F20101117_AABIPC miller_j_Page_125thm.jpg
fe057633b984c94d532e4f008594514f
7ea6fe728f200ac1593dd57952101535ee114d24
8787 F20101117_AABIOO miller_j_Page_115thm.jpg
5e60c2b872f92461abc1de3ad12a2a03
dc724361db06d22eb72a473e6253e28a3bdb0ca6
8516 F20101117_AABINZ miller_j_Page_098thm.jpg
c237c8c0a558bf96982dc002d91d4e8b
4b52c728351c582a9fb92f9de2a11b0908901ff6
F20101117_AABHLM miller_j_Page_096.jp2
59c59aff691564f6fffc3cad6e8ceaaf
a6f7fe415a5a35d497ea500a710a9f1e977674da
1051981 F20101117_AABHKX miller_j_Page_076.jp2
cca719f728afcb8bb3e8232a19d40830
52e0cbe388f1000cdd2de72bc47a053c095e1f5a
1051978 F20101117_AABHMA miller_j_Page_116.jp2
80ed3fbe9454ea923b47d2da1f5fdb13
a53aff2bccfd950fa3db2d040030944fd54d134e
32828 F20101117_AABIPD miller_j_Page_126.QC.jpg
2a8f13f06541f3da7648854766f31be9
f48007f698c84678670f56d2fc72d4fb729fa01d
34626 F20101117_AABIOP miller_j_Page_116.QC.jpg
9fcf052d47917865c1caeb4df6b6c23b
02fbe9ee649181883242ae909b41c5eda7ec8f47
F20101117_AABHLN miller_j_Page_097.jp2
bf60ab656121ec7275e3458ea6e0fd58
860acdc2da59ef6bf8014a4ae1c84cd8308a1aa9
F20101117_AABHKY miller_j_Page_077.jp2
4961e288fcde70cce8bbd46cbff81ab3
ee963bbde875119993b76f0c5eae074fd531cde7
1051953 F20101117_AABHMB miller_j_Page_117.jp2
76277816482e650b23ed686fb8e4c540
8f62d569f3ab968decce83197468e370a6984f3c
8747 F20101117_AABIPE miller_j_Page_127thm.jpg
dfd3afd1da667b7850fcfe0cec31ca1e
1fba5c0cf041c9b238ff690763b09d4ea6ab3423
8833 F20101117_AABIOQ miller_j_Page_117thm.jpg
eefba13d82f7389afb6257070d18c662
a920762320295bf1b49bd9db73c8df5b00ddd489
F20101117_AABHLO miller_j_Page_098.jp2
481a90d44ffbd3e5c0657128cf45af9f
1e8bc73d98e6613bf999420cb351b30a604ec71c
F20101117_AABHKZ miller_j_Page_079.jp2
b4cd58320803bdcdc80ea9c51825cca1
d9a7ac6eecb557690670597c2400859e7d9300e1
F20101117_AABHMC miller_j_Page_118.jp2
275a17555289096d68bde4110ad07057
917f2bc87e95f0b79d0dbb78916a1094f6a50346
33673 F20101117_AABIPF miller_j_Page_128.QC.jpg
9ab8b17c52ed359785dd12c3c4316b04
d61cbb60cdfa5c66e06ec87136f84cccbb7e9224
F20101117_AABHMD miller_j_Page_119.jp2
ee21804acc7420e2fed5054151879444
479759ab1930a3a89afb420ae80c09a7156d03ba
F20101117_AABIPG miller_j_Page_129thm.jpg
ece452636123f94d5305c560ee960493
fc7cd9bd3538ec5a7c994675ba19e6ebb00d07c6
33879 F20101117_AABIOR miller_j_Page_118.QC.jpg
46e7efc03ae708bd4a1c36862ff46094
4de96366a3ee7df64acd40cefef58310bbf4f2fd
1051917 F20101117_AABHLP miller_j_Page_101.jp2
f814a60b783c713e9ae06255bc364a5a
8ff617e76121b4cb1fdbe596dcd95270a3685896
F20101117_AABHME miller_j_Page_120.jp2
e4ce441d6bb1cacb93554f3f4916ca45
41d2d2e265181951b9289543a902b32e96f10c7e
8819 F20101117_AABIPH miller_j_Page_130thm.jpg
cae2ba224917a90a6c5acd4c46c3790a
6447b5d4ffee1c71ad25959a7b4d64836b89f0fd
8607 F20101117_AABIOS miller_j_Page_118thm.jpg
870f18903f1e1f6b28be4ee0c0a37e2c
c883e294aaaca1603e2a9275d4c6e8eade4de04d
F20101117_AABHLQ miller_j_Page_102.jp2
620058ea90c65f103ebb573cbec9d8f8
409c14211efdb7ba8922651051f0ddef7afd784e
F20101117_AABHMF miller_j_Page_121.jp2
15abc1b05e6bcd280b7bd3c65a82d024
818ef3b4b06fa6d57e1df4b2e8f42ce0992bea1c
8746 F20101117_AABIPI miller_j_Page_132thm.jpg
8ce9ae816fb1e7844289b69c9897d096
395021ca0d459f49a83de5ddd92b56daff6ccf6f
8841 F20101117_AABIOT miller_j_Page_119thm.jpg
fb1d2d34c00ed958c50c012ea9d6b7a6
c6cc96a40581d136785255b93981b1110d8247ed
F20101117_AABHLR miller_j_Page_103.jp2
c9a497340ac7d2c000a28e6ba0e46a83
c1b09862c0589e8c1a0ed05608c9fc4bd024f5c8
F20101117_AABHMG miller_j_Page_122.jp2
9d9f2072cd1aa7d03daba5418933aed8
db253198b8cf83c60f4e4dcd957000ef0e9f55b0
8621 F20101117_AABIPJ miller_j_Page_133thm.jpg
206083c55202b2e281e461d08c4d20af
96abd2f61e1ce26c4919d83a99cfcb45918a090a
34692 F20101117_AABIOU miller_j_Page_120.QC.jpg
4de9bc8e6a6ea4c1eb1251c595542f77
b4978ef6532ebd20e640d4bc4745b52dd8e36cbe
F20101117_AABHLS miller_j_Page_105.jp2
d35c43a52a8443b278ce439891206ccb
f8122c7d2a26b1b6c91de558b660c9f25280cad2
1051971 F20101117_AABHMH miller_j_Page_124.jp2
5c8ac054aa1b625e7907e199bbd534f9
20a7c461a9d028b2f544d02423e441b8f34ebffc
36305 F20101117_AABIPK miller_j_Page_134.QC.jpg
2463bc7372fecba412172e2e41edf637
1abb3fcba30128b19421f1799b29fd930b5e289a
8850 F20101117_AABIOV miller_j_Page_120thm.jpg
9138bdbc935eddc5e1ddfeabae4b2ffb
9214deaf2efa177656511062ac36cf6400a57470
F20101117_AABHLT miller_j_Page_107.jp2
b5dcfb3c54c8c3cdf47b2e310f23036a
258842a39b7d1ca1aebb49fd0ae49bd61306e653
F20101117_AABHMI miller_j_Page_125.jp2
c550934de4dbcb0c3ccfa5d9dfddd77a
8d609a54b1ce42b152b9c481efd6a1e8cf24936d
8910 F20101117_AABIPL miller_j_Page_134thm.jpg
a66b4ea39ab2880b64f4ff607dc7a8d9
4e8eec8752b73de0e398577e185d0f67b2a1cab3
8410 F20101117_AABIOW miller_j_Page_121thm.jpg
09e11c8e9749301d21aa335944dc42c3
4c8555a871a6f763a48d94f0a4dcca9a84cee74a
F20101117_AABHLU miller_j_Page_108.jp2
f6cddb4cc088fb8fea3d58c97f80ede3
60510771391a16084c5edebb554d6551e31d007f
F20101117_AABHMJ miller_j_Page_126.jp2
a3aa7644fb7f571da840bcf5b4928a88
cc73142747a25a24a008f2da34640d2e7aaaca94
30020 F20101117_AABIQA miller_j_Page_153.QC.jpg
9a4c1678b62fa870d07ff01782f0c2b5
8e0f120da09204f32d32582d9128ba6497e05f84
34973 F20101117_AABIPM miller_j_Page_135.QC.jpg
2657cf094d6ac631e796681d3dfc727c
b0a0b35376bf00d194ee0a808ef7aeab7588718d
8657 F20101117_AABIOX miller_j_Page_122thm.jpg
4b79d796b18c69f0aaf32e033abfdbd8
1e624915552d23b5e4e11802746a4f13d4b34ea8
1051877 F20101117_AABHLV miller_j_Page_110.jp2
cd7c04fa046428720329a9fee38dbef9
a307cfd85a965566120f947397a077645dc6d6b1
1051906 F20101117_AABHMK miller_j_Page_127.jp2
5cc41d4541cabe64ceb4f106b9df9207
ca903e2537701119de14a0b1782c624d81471a2f
7757 F20101117_AABIQB miller_j_Page_153thm.jpg
978827447502b607b61ff5a0c4a85284
f743b7b79d20b4024bf955c1b80e3f9c1693a520
8789 F20101117_AABIPN miller_j_Page_135thm.jpg
a893736faf446a3eb93c68187ea62f55
cd0215fd92c4c3fd965a84486eaeaf03c5440d33
32446 F20101117_AABIOY miller_j_Page_123.QC.jpg
2393d2c93c756e8e2381c5f5b371934e
28d65ff7ad017984b3a8da7fb3e45b25ebb69928
F20101117_AABHLW miller_j_Page_111.jp2
a5d9ee38c498f5da77ec43444c56fd15
b76ff3880e3e26cc4e13bb7fd350b2a6046dc054
F20101117_AABHML miller_j_Page_128.jp2
820943a2a70f3899da36cb83c34a57de
7b86774bda47e2bfaeb6ac243a2c0c8292157a25
8103 F20101117_AABIQC miller_j_Page_157.QC.jpg
510e0a059f7bf0230f26489a936ff582
69ede1d1f0fcfe0c9a7fbad9c63ef42fd330ab82
14458 F20101117_AABIPO miller_j_Page_137.QC.jpg
dd70ad97d854e2a2bf6ae55371c49ca6
176c9eb50f82b866d0d3157467c9aa6d3104f912
36727 F20101117_AABIOZ miller_j_Page_124.QC.jpg
10e4ceb107a823d667e4c324ed4fb13e
f71c2b19037d4f163537cebcbd2804bf5fa77956
F20101117_AABHLX miller_j_Page_112.jp2
f149bae089bfe7381f32c4653e701886
5d058434ad299786c0ab9c91a4d1855e1c7cb454
358866 F20101117_AABHNA miller_j_Page_146.jp2
4039ad9f5ccaf26bd67ce630d6de734d
d8bd144cdcf077a1d51a0483b7da715c3f53c52a
F20101117_AABHMM miller_j_Page_129.jp2
b7a98850757d9ff46507a125b510b60e
52bf03e2bab139676b10a8e5bc6bfb7a16b624b2
2895 F20101117_AABIQD miller_j_Page_158thm.jpg
96fa06686c82f52ae87dcca83fb49eb9
d006f80ee2e9295e8975aa4a7482bb904d5daab6
20878 F20101117_AABIPP miller_j_Page_138.QC.jpg
28b9199faf6046c18b82d0a1190767c4
8682c21a027a1f7f1e63e60b03c51918e244a8e9
1051919 F20101117_AABHLY miller_j_Page_113.jp2
7867c6fe1842a24e727c35342b2033eb
f757d5bf0453d93bb17a8089f86bb141d637eff3
810056 F20101117_AABHNB miller_j_Page_147.jp2
b2430eb252ef7b876fefd2436d2be16d
bdcac07ab421653ae3b3ffca320bc6d358ad9012
1051951 F20101117_AABHMN miller_j_Page_131.jp2
8969120a6009d6c06ceab5422ccd0cb3
6794052482ecdd8101636278178e61bc96a09d7d
9657 F20101117_AABIQE miller_j_Page_160.QC.jpg
52c35036fc3e9859bd273b3b17a3c4e1
55122567c20e5666ebb9c4d74c55d0e68a040518
5820 F20101117_AABIPQ miller_j_Page_138thm.jpg
5e7514f8034a32d18fb2bb7f3b9ad5a9
d74a88af34bd943119074edf35837d716cb19b2c
F20101117_AABHLZ miller_j_Page_114.jp2
8c38d1b34f9f81021579438aecb6b89e
f086eac86c24085a66dcc8f17c913e878df83420
812210 F20101117_AABHNC miller_j_Page_148.jp2
7af3ed12a73c3390e2f1d44af128a38f
b9e7348c42ff656292a81a319dedc331a48c52e3
1051949 F20101117_AABHMO miller_j_Page_134.jp2
fc90dd13c2e925aba6ce7b649691c812
c468bfb7dcb0bfee5d22aa45e86bf992e9da3cd5
3431 F20101117_AABIQF miller_j_Page_160thm.jpg
08977cd46a00a6090cb6e2cfe736cccc
570d60fb3027bf95514bb936297cc6e75b74c5d0
7641 F20101117_AABIPR miller_j_Page_141thm.jpg
e3b2b56788040da174e91994b93e4846
32d09aff66cd169b51748f86b1d9b8df4b95fd47
812435 F20101117_AABHND miller_j_Page_149.jp2
0896abc32f8371b202c11bc67849e04a
e8f7e0460219f9b258d6d27b7f26e0fb5cb7b922
F20101117_AABHMP miller_j_Page_135.jp2
bbb22813de53904f822ce9e6e436b88c
21e006c14a416090f66f23582bb9522f327b4977
9100 F20101117_AABIQG miller_j_Page_161.QC.jpg
8d9cb6c772107b4c855dd75f2a40538f
004b77ddccfcecce6afbaca0af9eb5c6a062d3f7
780739 F20101117_AABHNE miller_j_Page_151.jp2
5f2ab0c330718e142463ed195b444b59
69e14d24bda6f1f24d41e4d0dde0456429054a37
7675 F20101117_AABIQH miller_j_Page_163.QC.jpg
6abb7cff7cec947062ed654b786c1e48
e6ea6e0c2d5e1433854c7fc8ebf56b01ea2648ba
F20101117_AABIPS miller_j_Page_142thm.jpg
fdcb0c44ca20b4343cb7d77d2706e706
c6b437c39299ee3f0358bc373d442e8e759b7e3c
776243 F20101117_AABHNF miller_j_Page_152.jp2
945c263113d422ef9d5a7aa0bb8e8f9f
dfbe63307f02e0a052d2811a1cdee11df463fae0
F20101117_AABHMQ miller_j_Page_136.jp2
39a1ec974bd3faacdf6d46b9a16a7cf3
70348fd30d03de0d3c19dd0153948d7ee9214279
2997 F20101117_AABIQI miller_j_Page_164thm.jpg
9346d891232cf86685e06963b9f0f8a4
8122ec06cd69012bf2c7d193d7dbef0a4f3fb16e
28784 F20101117_AABIPT miller_j_Page_143.QC.jpg
6c4a4480d93f83e06b28c8d5fb1f4f00
7b3e300beebc8bdbd3b2c8a6a89811e693e80737
774846 F20101117_AABHNG miller_j_Page_153.jp2
41e5df48ad0abd0944e51baadeb4c5b9
af7ef37dc44ddd406c4cfebc4ea0151d0d76c994
459313 F20101117_AABHMR miller_j_Page_137.jp2
c114b947f6697ca93d0950690bbafe69
4d9e94b723c41dda744a17126a548cc9521dcf1d
2901 F20101117_AABIQJ miller_j_Page_167thm.jpg
0ea83f4a7d72e55f34f920e8b1e3b8a3
652ca5085a0c8f0c1e4f97e926ef3222d1382d12
28127 F20101117_AABIPU miller_j_Page_144.QC.jpg
0bb9deaf11d0ebabea5e4188b32494a6
d09283f5925a27bb57392c77ba776d6f0e9c111a
748870 F20101117_AABHNH miller_j_Page_154.jp2
e7b1ea3c9371a05d9a6430685df3cf17
b716daaa1414f31557c0729e3c19f14bcd3c475f
659191 F20101117_AABHMS miller_j_Page_138.jp2
96c36abc32aae375b8044d9b3158f5f1
7e83df7ff40d3a8ac2bc189c3e8bf5c3c1120c2c
F20101117_AABIQK miller_j_Page_169.QC.jpg
a1a71584da8f603f11a77fef6bb92a33
fee74f56dff3eaab736e47f5e245d7d9d81cfd5b
28724 F20101117_AABIPV miller_j_Page_145.QC.jpg
49f63f72b0d9361d018fa037da83fd8d
e5448e6bbae24c20c088719b6e099760d0193276
917437 F20101117_AABHNI miller_j_Page_155.jp2
58051f0032dd47de75d889e3d87e8857
7a088f9a67235a6497032b37ac2b228a573812ba
665729 F20101117_AABHMT miller_j_Page_139.jp2
1a9a5519adaf070031120f161328a6a0
a9dd0360bce207a142c975ff9d775aa33f9169c1
2661 F20101117_AABIQL miller_j_Page_169thm.jpg
b4fdf54b290da82ba4925978e00781d7
dd3688e7d11a9452a294bda70bda87d27b3dc046
12281 F20101117_AABIPW miller_j_Page_146.QC.jpg
1b8fe782d3ea223321e9fbc3e7780691
40ce5a41799bb1c8ae368691419b69f2146372ce
912550 F20101117_AABHNJ miller_j_Page_156.jp2
e3604c9503edd4c9572d17122c06d96f
3d3835373fb72aa7fae243d2df3ba55d1cd22c0d
819228 F20101117_AABHMU miller_j_Page_140.jp2
9e8112e20e14b93eb95832cc88fde507
6ed7b0b0f4e35c0c9b6349553bde69dcddc361bf
36417 F20101117_AABIRA miller_j_Page_183.QC.jpg
9a355121e8bf3fa119895e1d4167b293
824a5e8afb345009541350c35ee9d59d175810f2
7289 F20101117_AABIQM miller_j_Page_170.QC.jpg
173e9e3d8f449c00654f795f0c290b71
1732d83922f35da434ad833ffd15bafe7c3f2250
2862 F20101117_AABIPX miller_j_Page_146thm.jpg
07dbf227aba68091e956ecf59068c581
2c7103eaf8bd0ad55e9e6f330bdf15aa4b2d32a7
232131 F20101117_AABHNK miller_j_Page_157.jp2
481afff429b541d092efed2062c020d1
ed849d809302acb0ccd47c1005148ff949089555
852919 F20101117_AABHMV miller_j_Page_141.jp2
6c2fa9c08832fa520399ee643fb1fe53
500b50acf2ae8315b67a8c1db64b6bd7aea6ec86
8861 F20101117_AABIRB miller_j_Page_184thm.jpg
06ae3af42f208650b13be22069b56b0b
508e849c3da0467f46ec50fda5c78cdace6a4340
2555 F20101117_AABIQN miller_j_Page_170thm.jpg
c13274b3e34798731a32151ccb5f0140
e5b7b19e0adf2206846b7ab2e86ed708fee1fe01
29751 F20101117_AABIPY miller_j_Page_150.QC.jpg
8c1e6a7d819f0e8e5e10a801fc07e56b
e7e497d418e79a51116bcd2884ea7d8e434fa452
156888 F20101117_AABHNL miller_j_Page_158.jp2
ec6baf80aee8db512e75eab2c43aeadc
12fe44e7e936897548e765a1d60d990aa7784bc2
845949 F20101117_AABHMW miller_j_Page_142.jp2
a34b0825bc254d97b711649628cea703
cf1a1ac92218159c493d268745082b6436a83a43



PAGE 1

COMPARING POISSON, HURDLE, AND ZIP MODEL FIT UNDER VARYING DEGREES OF SKEW AND ZERO-INFLATION By JEFFREY MONROE MILLER A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLOR IDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2007 1

PAGE 2

Copyright 2007 by Jeffrey Monroe Miller 2

PAGE 3

To the memory of my grandfather, Rev. Harold E. Cato. 3

PAGE 4

ACKNOWLEDGMENTS Several people helped make this study possibl e. I would like to thank my stepfather and mother, Dr. Daniel and Gail Jacobs as well as my father and stepmother, Jerry and Darnelle Miller for their many years of encouragement. I would also like to thank my supervisory committee chair, M. David Miller, for his unyieldi ng guidance, patience, and support. I thank Dr. Jon Morris for the numerous training experiences Many professors are appreciated for providing the educational foundations for the dissertation topic including Dr. James Algina and Dr. Alan Agresti. The idea to research zero-inflation was in spired by experiences with data while consulting on projects. To this extent, I thank those clients Dr. Courtney Zmach and Dr. Lori Burkhead. Undergraduate faculty that I would like to acknowle dge for their inspiration and direction include Blaine Peden, Patricia Quinn, and Lee Anna Ra sar. Several friends have been a source of encouragement including Matt Grezik and Rachael Wilkerson. Finally, I thank those who made it financially possible to complete this dissertation including my consulting clients, the University of Florida College of Educa tion, the Lastinger Cent er, and Adsam, LLC. 4

PAGE 5

TABLE OF CONTENTS page ACKNOWLEDGMENTS...............................................................................................................4 LIST OF TABLES................................................................................................................. ..........8 LIST OF FIGURES.......................................................................................................................11 ABSTRACT...................................................................................................................................12 CHAPTER 1 INTRODUCTION..................................................................................................................14 Statement of the Problem....................................................................................................... .15 Rationale for the Study........................................................................................................ ...15 Purpose and Significance of the Study...................................................................................16 Research Questions............................................................................................................. ....16 2 REVIEW OF THE LITERATURE........................................................................................18 Zero-Inflated Count Data........................................................................................................18 Count Data.......................................................................................................................18 Zero-Inflation..................................................................................................................19 The Sources of Zero-Inflation.........................................................................................20 Impact of Zero-Inflation on Analyses.............................................................................21 Simple Solutions to Zero-Inflation..................................................................................22 Deleting zeros...........................................................................................................22 Assuming normality.................................................................................................22 Transforming Zeros..................................................................................................23 Generalized Linear Models.............................................................................................24 The Binomial Distribution and the Logit Link................................................................28 Evaluating Model Fit.......................................................................................................31 The Poisson Distribution and the Log Link....................................................................35 Iterative Estimation.........................................................................................................38 Interpretation of Coefficients..........................................................................................38 Hypothesis testing....................................................................................................39 Overdispersion.........................................................................................................39 Poisson and Negative Binomial Models with Zero-Inflation..........................................42 The Hurdle model.....................................................................................................45 The Negative Binomial Hurdle model.....................................................................48 The Zero-Inflated Poisson (ZIP) model...................................................................49 The Negative Binomial Zero-Inflated Poisson model..............................................51 Model Comparison Testing for Zero-Inflated Data.........................................................52 Review of Research Pertaining to and Using Zero-Inflated Count Data................................53 Hurdle Model...................................................................................................................53 5

PAGE 6

Statistical..................................................................................................................53 Applications.............................................................................................................54 Zero-Inflated Poisson Model...........................................................................................55 Statistical..................................................................................................................55 Applications.............................................................................................................59 ZIP and Hurdle Model-Comparisons..............................................................................63 Statistical..................................................................................................................64 Applications.............................................................................................................64 Discrepant Findings............................................................................................................ ....66 3 METHODOLOGY.................................................................................................................73 Research Questions............................................................................................................. ....73 Monte Carlo Study Design.....................................................................................................73 Monte Carlo Sampling............................................................................................................75 Pseudo-Population...........................................................................................................75 The Prespecified Zero Proportions..................................................................................76 Pre-Specified Skew.........................................................................................................76 Random Number Generation...........................................................................................77 Sample Size.....................................................................................................................78 Simulation Size................................................................................................................ 78 Iteration Size....................................................................................................................79 Distribution Generation...................................................................................................80 Monte Carlo Models...............................................................................................................80 Monte Carlo Analysis Procedures..........................................................................................82 Analysis Design......................................................................................................................84 4 RESULTS...............................................................................................................................94 Pseudo-Population Results.....................................................................................................94 Pseudo-Population Poisson Models................................................................................95 Pseudo-Population Hurdle Models..................................................................................96 Hurdle vs. Negative Binomial Hurdle......................................................................96 Poisson vs. Hurdle....................................................................................................97 Negative Binomial Poisson vs. Negative Binomial Hurdle.....................................98 Pseudo-Population ZIP Models.......................................................................................98 Comparing AICs For All Models...................................................................................99 Monte Carlo Simulation Results...........................................................................................101 Positively Skewed Distribution.....................................................................................101 Normal Distribution.......................................................................................................111 Negatively Skewed Distribution....................................................................................122 Review of Positively Skewed Distribution Findings............................................................133 Review of Normal Distribution Findings.............................................................................135 Review of Negatively Skew ed Distribution Findings..........................................................136 5 DISCUSSION.......................................................................................................................176 6

PAGE 7

The Impact of the Event Stage Distribution.........................................................................176 Positively Skewed Event-Stage Distributions...............................................................176 Normal Event-Stage Distributions................................................................................181 Negatively Skewed Event-Stage Distributions.............................................................183 Summary of Findings....................................................................................................185 Limitations.................................................................................................................... ........186 Discrete Conditions.......................................................................................................186 Convergence and Optimization.....................................................................................186 Underdispersion.............................................................................................................187 Other models.................................................................................................................18 8 Validity of Model-Fitting and Model-Comparisons.............................................................188 Suggestions for Future Research..........................................................................................190 Application in Educational Research....................................................................................190 Major Contribution of Findings............................................................................................191 LIST OF REFERENCES.............................................................................................................195 BIOGRAPHICAL SKETCH.......................................................................................................201 7

PAGE 8

LIST OF TABLES Table page 2-1 Five pairs of nested models valid for statistical comparison.............................................71 2-2 Summary of literature on zero-inflation............................................................................72 3-1 Proportions of counts as a function of zeros and skew......................................................87 3-2 Frequencies of counts as a function of zeros and skew.....................................................87 3-3 Descriptive statisti cs for each distribution.........................................................................88 3-4 Poisson model: pseudo-population parameters..................................................................88 3-5 Negative Binomial Poisson model: pseudo-population parameters..................................89 3-6 Hurdle model (zeros): pseudo-population parameters.......................................................89 3-7 Hurdle model (events) : pseudo-population parameters.....................................................90 3-8 Negative Binomial Hurdle model (zeros): pseudo-population parameters........................90 3-9 Negative Binomial Hurdle model (events): pseudo-population parameters......................91 3-10 ZIP model (zeros): pseudo-population parameters............................................................91 3-11 ZIP Model (events): pseudo-population parameters..........................................................92 3-12 Negative Binomial ZIP model ( zeros): pseudo-population parameters.............................92 3-13 Negative Binomial ZIP model (events): pseudo-population parameters...........................93 4-1 Deviance statistics comparing Poiss on and negative binomial Poisson models..............138 4-2 Deviance statistics comparing Hurdle and negative binomial Hurdle models................138 4-3 Deviance statistics comparing Poisson and Hurdle models.............................................138 4-4 Deviance statistics comparing NB Poisson and NB Hurdle models................................138 4-5 Deviance statistics comparing ZIP and negative binomial ZIP models..........................139 4-6 Log-likelihood comparisons for posit ively skewed distribution with .10 zeros..............139 4-7 AICs for positively skewed distribu tion models with a .10 proportion of zeros............140 4-8 Log-likelihood comparisons for posit ively skewed distribution with .25 zeros..............140 8

PAGE 9

4-9 AICs for positively skewed distribu tion models with a .25 proportion of zeros............141 4-10 Log-likelihood comparisons for posit ively skewed distribution with .50 zeros..............141 4-11 AICs for positively skewed distribu tion models with a .50 proportion of zeros............142 4-12 Log-likelihood comparisons for posit ively skewed distribution with .75 zeros..............142 4-13 AICs for positively skewed distribu tion models with a .75 proportion of zeros............143 4-14 Log-likelihood comparisons for posit ively skewed distribution with .90 zeros..............143 4-15 AICs for positively skewed distribu tion models with a .90 proportion of zeros............144 4-16 Log-likelihood comparisons fo r normal distribution with .10 zeros...............................144 4-17 AICs for normal distribution models with a .10 proportion of zeros.............................145 4-18 Log-likelihood comparisons fo r normal distribution with .25 zeros...............................145 4-19 AICs for normal distribution models with a .25 proportion of zeros.............................146 4-20 Log-likelihood comparisons fo r normal distribution with .50 zeros...............................147 4-21 AICs for normal distribution m odels with a .50 proportion of zeros.............................147 4-22 Log-likelihood comparisons fo r normal distribution with .75 zeros...............................148 4-23 AICs for normal distribution models with a .75 proportion of zeros.............................148 4-24 Log-likelihood comparisons fo r normal distribution with .90 zeros...............................149 4-25 AICs for normal distribution models with a .90 proportion of zeros.............................149 4-26 Log-likelihood comparisons for negatively skewed distribution with .10 zeros.............150 4-27 AICs for negatively skewed models with a .10 proportion of zeros..............................150 4-28 Log-likelihood comparisons for negatively skewed distribution with .25 zeros.............151 4-29 AICs for negatively skewed models with a .25 proportion of zeros..............................151 4-30 Log-likelihood comparisons for negatively skewed distribution with .50 zeros.............152 4-31 AICs for negatively skewed models with a .50 proportion of zeros..............................152 4-32 Log-likelihood comparisons for negatively skewed distribution with .75 zeros.............153 4-33 AICs for negatively skewed models with a .75 proportion of zeros..............................153 9

PAGE 10

4-34 Log-likelihood comparisons for negatively skewed distribution with .90 zeros.............154 4-35 AICs for negatively skewed models with a .90 proportion of zeros..............................154 4-36 Positively skewed distribution: percentage of simulations favoring complex model......155 4-37 AICs: positively skewed distribution (all conditions)....................................................155 4-38 Normal distribution: percentage of simulations favoring complex model......................155 4-39 AICs: normal dist ribution (all conditions)......................................................................155 4-40 Negatively skewed distribution: percen tage of simulations favoring complex model....156 4-41 AICs: negatively skewed distribution (all conditions)...................................................156 4-42 Convergence frequencies: positively skewed distribution...............................................156 4-43 Convergence frequencies: normal distribution................................................................156 4-44 Convergence frequencies: negatively skewed distribution..............................................157 10

PAGE 11

LIST OF FIGURES Figure page 4-1 Boxplot of AICs for all models for a .10 proportion of zeros........................................158 4-2 Boxplot of AICs for all models for a .25 proportion of zeros........................................159 4-3 Boxplot of AICs for all models for a .50 proportion of zeros........................................160 4-4 Boxplot of AICs for all models for a .75 proportion of zeros........................................161 4-5 Boxplot of AICs for all models for a .90 proportion of zeros........................................162 4-6 Boxplot of AICs for all models for a .10 proportion of zeros........................................163 4-7 Boxplot of AICs for all models for a .25 proportion of zeros........................................164 4-8 Boxplot of AICs for all models for a .50 proportion of zeros........................................165 4-9 Boxplot of AICs for all models for a .75 proportion of zeros........................................166 4-10 Boxplot of AICs for all models for a .90 proportion of zeros........................................167 4-11 Boxplot of AICs for all models for a .10 proportion of zeros........................................168 4-12 Boxplot of AICs for a ll models for a .25 proportion of zeros........................................169 4-13 Boxplot of AICs for a ll models for a .50 proportion of zeros........................................170 4-14 Boxplot of AICs for a ll models for a .75 proportion of zeros........................................171 4-15 Boxplot of AICs for a ll models for a .90 proportion of zeros........................................172 4-16 AIC rank order for positively skewed distribution models.............................................173 4-17 AIC rank order for normal distribution models...............................................................174 4-18 AIC rank order for negatively skewed distribution models............................................175 5-1 Poisson, NB Poisson, and Hurdle over all proportions of zeros......................................192 5-2 Hurdle, NB Hurdle, and NB Poiss on over all proportions of zeros.................................193 5-3 ZIP, NB ZIP, Hurdle, and NB Hu rdle over all proportions of zeros...............................194 11

PAGE 12

Abstract of Dissertation Pres ented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy COMPARING POISSON, HURDLE, AND ZIP MODEL FIT UNDER VARYING DEGREES OF SKEW AND ZERO-INFLATION By Jeffrey Monroe Miller May 2007 Chair: M. David Miller Major Department: Educational Psychology Many datasets are characterized as count data with a prepond erance of zeros. Such data are often analyzed by ignoring the zero-inflation and assuming a Poisson distribution. The Hurdle model is more sophisticated in that it cons iders the zeros to be completely separate from the nonzeros. The zero-inflated Poisson (ZIP) mode l is similar to the Hurdle model; however, it permits some of the zeros to be analyzed along with the nonzeros. Both models, as well as the Poisson, have negative binomial formulations for use when the Poisson assumption of an equal mean and variance is violated. The choice between the models should be guided by the researchers beliefs about the source of the zeros. Beyond this substantive co ncern, the choice should be based on the model providing the closest fit between the observed and predicted va lues. Unfortunately, the literature presents anomalous findings in terms of model superiority. Datasets with zero-inflation may vary in terms of the proportion of zeros. They may also vary in terms of the distribution for the nonzeros Our study used a Monte Carlo design to sample 1,000 cases from positively skewed, normal, a nd negatively skewed distributions with proportions of zeros of .10, .25, .50, .75, and .90. The da ta were analyzed with each model over 12

PAGE 13

2,000 simulations. The deviance statistic and Akai kes Information Criterion (AIC) value were used to compare the fit between models. The results suggest that the literature is not entirely anomalous; however, the accuracy of the findings depends on the proportion of zeros and the distribution for the nonzeros. Although the Hurdle model tends to be the superior model, there are situations when others, including the negative binomial Poisson model, are superior. The findings sugge st that the researcher should consider the proportion of zeros and the distribution for the nonzer os when selecting a model to accommodate zero-inflated data. 13

PAGE 14

CHAPTER 1 INTRODUCTION Analyzing data necessitates determination of the type of data being analyzed. The most basic assumption is that the data follows a nor mal distribution. However, there are many other types of distributions. The validity of the results can be affected by the dissimilarity between the distribution of the data and the di stribution assumed in the analysis. As such, it is imperative that the researcher choose a me thod for analyzing the data that maintains a distribution similar to that of the observed data. Counts are an example of data which does not readily lend itself to the assumption of a normal distribution. Counts are bounded by their lowest value, which is usually zero. A regression analysis assuming a normal distribution would permit results below zero. Further, counts are discrete integers while the normal distribution a ssumes continuous data. Finally, counts often display positive skew such that the frequency for low counts is considerably higher than the frequencies as the count levels increase. It is not uncommon to find count data an alyzed in a more appropriate manner than assuming a normal distribution. Typically, more a ppropriate analysis includes specification of a Poisson distribution with a log link, rather than a normal dist ribution with a Gaussian link. However, this does not guarantee accurate and valid results as other features of the data may warrant an even more sophisticated model. An example of data requiring a more rigorous treatment of the data is the case of zeroinflation. In this scenario, ther e are far more zeros than would be expected using the Poisson distribution. As such, a number of methods including the zero-i nflated Poisson (ZIP) model and the Hurdle model are available. Further, there ar e negative binomial variations of these for use when particular assumptions appear to be violated. The choice between the models depends on 14

PAGE 15

whether the researcher believes the zeros are all a complete lack of the quantity being measured or that at least some of the zeros are purely random error. Statement of the Problem The results from both simulated and actual data sets in the zero-inflation literature are in much disagreement. Lambert (1992) found the ZIP m odel to be superior to the negative binomial Poisson model, which was superior to the Poisson model. Greene (1994) found the negative binomial Poisson model to be superior to the ZIP model, which was superior to the Poisson model. Slymen, Ayala, Arredondo, and Elder (2006) found the ZIP and negative binomial ZIP models to be equal. Welsh, Cunningham, Donne lly, and Lindenmayer found the Hurdle and ZIP models to be equal while Pardoe and Durham (2 003) found the negative binomial ZIP model to be superior to both the Poisson and Hurdle models. One striking characteristic of these articles and others is their diffe rences in terms of the proportion of zeros and the distribution for the nonzeros. Some research (Bhning, Dietz, Schlattmann, Mendona, and Kirchner, 1999) analy zed data in which the proportion of zeros was as low as .216 while others (Zorn, 1996) used proportions as high as .958. Further, the nonzeros varied in terms of their distributions from highly positively skewed to normal to uniform. It is possible that different models yi eld different results depending on the proportion of zeros and the distribution for the nonzeros. Rationale for the Study The best model is the one that appropriat ely answers the resear chers question. Beyond this, a superior model is one that has close proximity between the observed data and that predicted by the model. In other words, a supe rior model is one with good fit to the data. This study compared the fit between the Poiss on, ZIP, and Hurdle models as well as their negative binomial formulations. Each analysis wa s performed for five different proportions of 15

PAGE 16

zeros and three different amounts of skew for the nonzero distribution. The intended results would clarify the discrepant fi ndings of previous research. Purpose and Significance of the Study The primary purpose of this study was to determine superiority of fit for various models under varying proportions of zero-in flation and varying levels of skew. As such, determination can be made as to which model has better fit give n data with a particular proportion of zeros and a particular distribution. The s econdary purpose was to elucidat e the reasons for discrepant findings in previous research. The superior model is the appropriate mode l given the research question. However, there are situations in which the appropriate model is unknown or unclear. Further, there may be situations in wh ich a simpler model such as the Poisson may be used in lieu of the more sophisticated Hurdle an d ZIP models. This resear ch provides results that aid researchers in determining the appropriate model to us e given zero-inflated data. Research Questions Model comparisons in this research were based on two measures. One is the deviance statistic, which is a measure of the difference in log-likelihood between two models, permitting a probabilistic decision as to whether one model is adequate or whether an alternative model is superior. This statistic is appropr iate when one model is nested within another model. The other measure is Akaikes Information Criterion (AIC). This statistic penalizes for model complexity and permits comparison of nonnested models; howe ver, it can only be used descriptively. These two measures of model fit were used to comp are results from data simulations where each dataset included 2,000 cases a nd each model was analyzed 1,000 times. Specifically, the measures of model fit were used to an swer the following research questions: 16

PAGE 17

Given one two-level categorical covariate with known values and one continuous covariate with known values, what is the difference in the estimated log-likelihood between a) the Negative binomial Poisson model vs. Poiss on model; b) the Hurdle model vs. Poisson model?; c) the Negative binomial Hurdle model vs. negative binomial Poisson model?; d) the Negative binomial Hurdle model vs. Hurdle model; and, e) the Negative binomial ZIP model vs. ZIP model? Given one two-level categorical covariate with known values and one continuous covariate with known values, what is the difference in the estimated AIC between all models? 17

PAGE 18

CHAPTER 2 REVIEW OF THE LITERATURE Zero-Inflated Count Data Count Data As the name implies, count data is data that arises from counting. They are the "realization of a nonnegative integer-valued ra ndom variable (Camer on & Travedi, 1998, p.1). As such, the response values take the form of discrete integers (Zorn, 1996). Although the lower boundary can feasibly be any integer, it is usua lly the case that its value is zero. Strictly speaking, there can be no nonnegative numbers. Hence, the data are constrained by this lower bound of zero and no upper bound. Acknowledgment of concerns over zero-inflatio n, ignoring covariat es, likely dates to Cohen (1954). Cameron and Triverdi (1989, p.1011) identified many areas in which special models have been used to analyze count data including models of counts of doctor visits and other types of health care utilization; occupational injuries and illnesses; absenteeism in the workplace; recreational or shopping trips; automobile insurance rate making; labor mobility; entry and exits from industry; takeover activity in business; mortgage prepayments and loan defaults; bank failures; patent registration in connection wi th industrial research and development; and frequency of airline accidents . as well as in many disciplines including demographic economics, in crime victimology, in marketing, political science and government, [and] sociology. Surprisingly, there was no menti on of research in education. Examples of variables in educational research that yield count data include a students numb er of days absent, number of test items scored correct or incorrect, and number of referrals for disciplinary action. 18

PAGE 19

The lower bound constraint of zero presents the biggest obstacle to ward analyzing count data when assuming a normal distribution. It is common for count data to have a skewed distribution that is tr uncated at the lower bound. 3 1 3() () (1)N i i YYY skewY Ns (2-1) Hence, the data are heteroscedastic with variance increasing as the count increases. Therefore, standard models, such as ordinary le ast squares regression, are not appropriate since they assume that the residuals are distribute d normally with a mean of zero and a standard deviation of one (Slymen, Ayala, Arredondo, & Elder, 2006). Cameron and Triverdi (1998) clarify that the use of standard OLS regression l eads to significant deficiencies unless the mean of the counts is high, in which case normal a pproximation and related regression methods may be satisfactory (p.2). An example of a count data variable is the number of household members under the age of 21 reported by respondents in the Adult Education for Work-Rel ated Reasons (AEWR) survey administered by National Council for Educationa l Statistics in 2003 (Hagedorn, Montaquila, Vaden-Kiernan, Kim & Chapman, 2004). The sample size was 12,725. This variable has a lower count boundary of zero and an upper count boundary of six. The count distribution is positively skewed at 1.971. The distribution mean of 0.54 is certainly not an accurate measure of central tendency; the median and mode are both zero (i.e., the lower-bound itself), and the standard deviation of 0.999 permits negative values in the lower 68% confidence interval. Zero-Inflation It is not uncommon for the outcome variable in a count data distribution to be characterized by a preponderance of zeros. As Tooze, Grunwald, & Jones (2002, p.341) explain, 19

PAGE 20

Typically, [for count data] the outcome variable measures an amount that must be nonnegative and may in some cases be zero. The positive values are generally skewed, often extremely so . Distributions of data of this type follow a common form: there is a spike of discrete probability mass at zero, followed by a bump or ramp describing positive values. The occurrence is primarily in the case of interva l/ratio count data and sometimes ordinal data (Bhning, Dietz, Schlattmann, Mendona, & Kirchner, 1999). Regarding continuous data, Hall and Zhang (2004) explain that these distributions have a null probability of yielding a zero . there is little motivation for a model such as [zero-inflated] normal, because all observed zeros are unambiguous . (p.162). If continuo us zeros are inflated and those zeros are of concern, they can be analyzed separately from the nonzeros. The null probability of continuous zeros is evident in measures such as height and age. The condition of excessive zeros is known as zero-inflation (Lachenbruch, 2002) or as a probability mass that clumps at zeros (Tooze, Gr unwald, & Jones, 2002). It has been recognized as an area of research in the mid-60s1 (Lachenbruch, 2002) when Weiler (1964) proposed a method for mixing discrete and continuous distribu tions. Min and Agresti (2005) formally define zero-inflation as data for which a generali zed linear model has lack of fit due to disproportionately many zeroes (p.1). There are simply a greater number of zero counts than would be expected under the Poisson or some of its variations (Zorn, 1996, p.1). The Sources of Zero-Inflation The zeros can be classified as being either true zeros or sampling zeros. True zeros represent responses of zero that are truly null. Suppose an educational inventory item states How many college preparatory wo rkshops have you attended? Some of the respondents in the sample may have no intentions to apply for college. Hence, the number of preparatory 1 Alternatively, if a scale is bound, it is reasonable to consider an inflated u pper bound. In this case, scale reversal and subsequent appropriate analysis if justified (Lachenbruch, 2002). 20

PAGE 21

workshops attended may never be greater than zer o. Sampling zeros, on the other hand, arise as a probability. There are a proportion of college-bo und students who have not attended a workshop due to the possibility that the workshop was not (o r is not yet) available. Alternatively, some college-bound students may feel prepared and have no reason to participate in a workshop. Hence, the mechanism underlying zero-infla tion can arise from one or both of 1) a possibility that no other respons e is probabilistic, or 2) that the response is within a random sample of potential count res ponses. Martin, Brendan, Wintle Rhodes, Kuhnert, Field, LowChoy, Tyre, and Possingham (2005) term the sampling zeros as false zeros and include error as a source of zeros. They state, Zero inflation is often the result of a large number of true zero observations caused by the real . effect of intere st . However, the term [zero-inflation] can also be applied to data sets with false zero observations because of sa mpling or observer errors in the course of data collection (p.1235). Often, the data contains both types of zeros. This is the result of a dual data generating process (Cameron & Trivedi, 1998). For example, some adults in the AEWR sample may have had true-zero household members under the age of 21 because they are unable to bear children or desire to bear children. Alternatively, they may have random-zero household members under the age of 21 because these adults do have such ch ildren but not as members of the household. Impact of Zero-Inflation on Analyses Much of the interest in count data modeli ng appears to stem from the recognition that the use of continuous distri butions to model integer out comes might have unwelcome consequences including inconsistent parameter estimates (Mullahy, 1986, p.341). In the typical count data scenario, the zero left-bound implies heteroscedasticity (Zorn, 1996). An even greater problem with zero-inflated distri butions, beyond this inadequacy of analyzing such a skewed and heteroscedastic distribution as if it were normal (Tooze, Grunwald, & Jones, 2002) is that they 21

PAGE 22

yield surprisingly large ineffi ciencies and nonsensical results (King, 1989, p126). Martin, Brendan, Wintle, Rhodes, Kuhnert, Field, Low-Choy, Tyre, and Possingham (2005) and McCullagh and Nelder (1989) explai n that zero-inflation is a speci al case of overdispersion in which the variance is greater than it should be given a particular distributional shape and measure of central tendency. The impact is bi ased/inconsistent parameter estimates, inflated standard errors and invalid inferences (Jang, 2005; Martin, Brendan, Wintle, Rhodes, Kuhnert, Field, Low-Choy, Tyre, and Possingham, 2005). Simple Solutions to Zero-Inflation Deleting zeros The simplest of solutions is to delete all ca ses having responses of zero on the variable of interest. A large proportion of total responses would then be remove d from the total dataset. This would then result in a loss of valuable inform ation impacting statistical conclusion validity (Tooze, Grunwald, & Jones, 2002). The sample size may also then be too small for analyses of the non-zero values. Assuming normality Another simple solution is to ignore the zer o-inflation, assume asymptotic normality, and analyze the data using standard techniques such as ordinary least squares regression. 012 iihhmSexAgei i (2-2) According to this model, the number of hous ehold members under the age of 21 for adult respondent i is predicted from the overall mean, a co efficient relating the respondents sex to hhm, a coefficient relating the respondents age to hhm, and error. The model assumes that the residuals for hhm are distributed normally with a m ean of zero and a common variance,2 For the first equation, y is a vector of responses, X is a design matrix for the explanatory variable 22

PAGE 23

responses, is a vector of regression coefficients relating y to X, and is a vector of residuals measuring the deviation between the observed values of the design matrix and those predicted from the fitted equation. Transforming Zeros Another simple solution is to transform the counts to coerce a mo re normal distribution (Slymen, Ayala, Arredondo, & Elder, 2006). Sinc e count distributions often appear to be positively skewed, one reasonable transformation involves taking the natu ral logarithm of the responses to the predic tor variables. However, assuming the zeros havent been deleted, the transformation will not work since the natural logarithm of zero is undefined (Zhou & Tu, 1999; King, 1989). Sometimes natural log transformations for zer o are handled by adding a small value, such as .001, to the zeros. However, this then leads to an inflation of that tran sformed adjusted value. If 70% of the scores are zero, the resulting transformed distri bution will have a 70% abundance of the transformed value (Delucchi & Bostrom, 2004).2 Further, since the transformation is linear, this technique has been shown to yield biased parameter estimates that differ as a function of the adjustment quantity (King, 1989). Although the undefined log zero problem has been handled, the original problems pervade. As Welsh, Cunningham, Donnelly, & Linenmayer (1996) state, It is clear for data with many zero va lues that such an approa ch will not be valid as the underlying distributional assumptions (linear ity, homoscedasticity and Gaussianity) will [still] be violated (p.298). Finally, for any technique, transformations sometimes create a new problem while solving the old one; a transform that produces constant variance may not produce normality . (Agresti, 1996, p.73). 2 This implies then that, beyond the dual generating proce ss for zeros, the problem can be generalized from inflated zeros to inflated lower boundaries for count data. 23

PAGE 24

Generalized Linear Models Bryk and Raudenbush (1996) state, There ar e important cases . for which the assumption of linearity and normality are not realistic, and no transformation can make them so (p.291). As previously demonstrated, count data is likely to be one such case. Instead of deleting cases or transforming the data, it is more reas onable to specify a diffe rent distribution. As explained by Hox (2002), although it is nice to be able to transform data, modeling inherently nonlinear functions directly is sometimes pref erable, because it may reflect some true developmental process (pp. 93-94). In order for a model to be inhe rently nonlinear (Hox, 2002), there must be no transfor mation that makes it linear.3 These nonlinear models belong to the class of generalized linear models (GLM). The following explanation of generalized lin ear models based on the seminal work of McCullagh and Nelder (1989) with additional clarif ication by Lawal (2003) and Agresti (1996). Lawal (2003) explains that generalized linear models are a subset of the traditional linear models that permit other possibilities th an modeling the mean as a linear function of the covariates. All GLM possess a random component, a systematic co mponent, and a link function. As explained by Agresti (1996), the random compone nt requires the specificati on of the distri bution for the outcome variable. One could specif y this distribution to be norma l; hence, classical models such as ordinary least squares regres sion and analysis of variance m odels are included within this broader class of generalized lin ear models. Other possible random components that could be specified include the binomial distribution, negative-binomial di stribution, gamma distribution, and Poisson distribution. Speci fying the random component depends on the expected population distribution of the outcome vari able. Given both zero-inflation a nd truncated count data yielding 3 Transforming covariates (e.g., including polynomial terms) may graphically appear to be nonlinear while still be linear in the parameters (Singer & Willett, 2003). 24

PAGE 25

an odd-shaped skewed distribution, the random co mponent plays an importa nt part in obtaining valid results. In order to better understand the formulation of the three components, it is necessary to clarify the theoretical foundations of distributions. The probabili ty density function for a normal distribution is 2 21( (;,)exp() 2 2 y fy ) (2-3) which, given random variable 2~(,) XN reduces to the standard normal probability density function 21 ()exp() 2 2 y fy (2-4) which when transformed to the cumulative density function yields 21 ()()exp() 2 2yu yFy du. (2-5) A convenient method for obtaining the paramete rs is to use the distributions moment generating function (Rice, 1995). For the normal distribution, this function is 22()[exp()]exp() 2yt MtEtYt (2-6) The logarithm of the moment generating function yields the cumulant generating function, which then yields the moments of the distribution. For the normal distribution, the first moment is the mean ( ), and the second moment is the variance (2 ). Strictly speaking, a requirement for GLM is that the outcome has a distribution within the exponential family of models (EFM) (McCullagh and Nelder, 1996). Th ese distributions are 25

PAGE 26

defined primarily by a vector of natural parameters ( ) and a scale parameter ( ). The formulation is given by (()) (;,)exp{()(,)} ()Yyb fy cy a (2-7) At first glance, it seems odd to include the nor mal distribution in the EFM; however, first recall the probability density function 2 21( (;,)exp() 2 2 y fy ) (2-8) Algebraic manipulation reveals th at the normal distribution is indeed an EFM formulation. 22 2 (0,1) 22()/2)1 (,)exp{( )(log(2)} 2NYyy EFMfy (2-9) Here, the natural (i.e., canonical) parameter is and the scale parameter is 2 These parameters need to be estimated McCullagh and Nelder (1996) explain the estimation as follows: In the case of generalized linear models, estimation proceeds by defining a measure of goodness of fit between the observed data and the fitted values that minimizes the goodness-of-fit criterion. We shall be concerned primarily with estimates obtained by maximizing the likelihood or log likelihood of the parameters for the data observed (p. 23-24). This turns out to be the log of the EFM function. (,;)log((;,))Yyfy (2-10) The natural and scale parameters are estimated by derivations revealing the mean function ()'() EYb (2-11) and the variance function var()''()() Yba (2-12) 26

PAGE 27

Note that the mean function depends on onl y one parameter. However, as McCullagh and Nelder (1989) explain, . the variance of Y is the product of two functions; one, ''() b depends on the [canonical] parameter (and hen ce on the mean) only and will be called the variance function [denoted V( )], while the other [() a ] is independent of and depends only on . The function () a is commonly of the form()/ aw (p.29) and is commonly c the dispersion parameter. For the normal distribution, the natural parameter is the mean ( alled ); th variance function, V ( e ); equals 1.0, and the dispersion parameter is 2 The systematic component is simply the mode l for the predictors established as a linear combination and is denoted The link function, g ( ), brings together the random component and the systematic component hence linking the function for the mean, and the function for the systematic component, as = g ( ). In other words, it specif ies how the population mean of the outcome variable with a particular distribution is related to the predictors in the model. If g ( ) redundantly equals then the population mean itself is related to the predictors. This is termed the identity link and is exactly the f unction used to link the mean of the normal distribution to its covariates. The key advantage of GLM is that they are no t restricted to one pa rticular link function. Many other links are available. For example, one could specify the log link as g ( ) = log( ) or the logit link as g( ) = log[ / (1)]. However, each random component has one common canonical link function that is best suited to the random component (McCullagh & Nelder. 1996). Alternatively, Each potentia l probability distri bution for the random component has one special function of the mean that is called its natural parameter (A gresti, 1996, p.73). For example, a normal random component usually co rresponds to an ident ity link, a Poisson 27

PAGE 28

distributed random component usua lly corresponds to a log link, and a binomial distributed random component usually corresponds to a logit link. In sum, th e canonical link and natural link are two equivalent terms for specifying the most suitable link connecting a particular distribution covariate function. The B l a different random compone nt and link can be specified to accommodate these e di stribution with parameters n for the number of trials and for the outcome variable with its lin ear systematicinomial Distribution and the Logit Link Suppose we were interested in the differ ences between households with zero children under age 21 and households with one or more children over the age of 21. We could feasibly collapse all nonzero responses in the AEWR data in to a value of one. Now, 71.58% of the values are zeros, and 28.42% of the values are ones. This distribution is obviously not normal. We have now introduced both a lower bound (zero), an up per bound (one), and an inherently nonnorma distribution; hence constraints. Variables that take on only one of tw o values are known as binary, or Bernoulli, variables, and the distribution of multiple i ndependent trials for th ese variables is termed binomial. Bernoulli responses are modeled in term s of the probability (Pr) that the outcome variable ( Y ) is equal to either zero or one. The random component over multiple indpendent trials is thus a binomial for the probability that Y = 1. ~(,) YBn (2-13) The binomial distribution assumes that the res ponses are dichotomous, mutually exc independent, and randomly selected (Agresti, 1996). Since the responses are discrete, lusive, the probability density function is termed the probability mass function and is defined as (;,)[(1)] !()!knn fknp pp knk k. (2-14) 28

PAGE 29

This function gives the p probability of k ones (i.e., heads, hits, successes) over n trials. Rice (1995) clarifies, Any particular sequence of k successes occurs with probability [( from the multiplication principle [i.e., inde pendent probabilities of realizations being multiplicative]. The total number of such sequences is 1)]knppk !()! n knk [permutations], since there are !()! n knk ways to assign k successes to n trials (p.36). The mome nt generating function is {1exp()}n and the cumulant generating function is log{1exp()} n The log-likelihood function is virt ually the same as the probability mass function. However, now we are determining the value of p as a function of n and k (rather than determining k as a function of p and n ) while taking the log of this maximum at [(;,)][(1)] !()!knn fpnk pp knk k. (2-15) The estimates are obtained through derivations of the likelihood func tion as was previously discussed for the normal distribution. Just as the normal distribution population mean, has the best maximum likelihood estimates of X the binomial distribu tion population probability, has the best maximum likelihood estimate of k divided by n which is the proportion of ones, hits, or successes. This greatly reduces calculations when a quick estimate is needed and the random component is not lin ked to any predictors. The binomial distribution even tually converges to a normal distribution. However, the speed of this convergence is primarily a function of skew, (12) () (1) Skewy n (2-16) with p = .50 yielding the fastest conver gence (McCullagh and Nelder, 1996). 29

PAGE 30

The link function should account for the binomial outcome vari able. If linear predictors are used to predict a probability, then we have predicted values in an infinite range rather than constrained to be between zero and one. What is needed is a link function that will map a bounded zero-one probability onto this range of infinite values. The canonical link for the binomial distribution is the logit link. ()log(/(1)) g (2-17) A logit is the natural log of an odds ratio or log[ p / (1-p )]. An odds ratio is equal to a probability divided by one minus that probability. He nce, if the probability is .5, then the odds are .5 / (1-.5) = 1 meaning that the odds of a one are the same as the odds of a zero [i.e., the odds of success and failure are identical] If the probability is .75, th en the odds are .75 / (1-.75) = 3 meaning that a response of one is three times more likely than a response of zero. The reciprocal odds of 1/3 means that a r esponse of one is three times less likely than a response of zero, which is e quivalent to stating that a response of zero is three times more likely than a response of one. Wh en using the logit link to connect the binomial random distribution and the systematic co mponent, the generalized linear model is logit( ) = 01log ... 1pXX (2-18) A probability of .50, which is an odds of one corresponds to a logit of zero. Odds favoring a response of one yield a positive logit, and odds favoring a response of zero yield a negative logit. Hence, the mapping is satisfied since th e logit can be any real number (Agresti, 1996). The regression parameters are slope -like in that they determine the relative rate of change of the curve. The exact rate of change depends on each probability with the best approximation 30

PAGE 31

at that probability being (1) with the steepest rate of change being at = 0.50, which is where X = / (Agresti & Finlay, 1997). Since natural logs can be re versed through exponen tiation and since odds can be converted to probabilities by dividing the odds by the sum of the odds and one, the fitted equation can be used to predict probabilities via 11 11exp( ) 1[exp( )]kk kkXX XX (2-19) It is more common to interpret logistic regression coefficients by only exponentiating them. Then, the coefficient has a slope, rather than slope-like, interpretation; however, the relationship is multiplicative rather than addi tive. Specifically, the expected outcome is multiplied by exp( ) for each one-unit increase in X. Evaluating Model Fit The next step in interpreting generalized linear model results is to determine how well the estimated model fits the observed data, where fit is the degree of di screpancy between the observed and predicted values. McCullagh and Nelder (1989) explain, In general the s will not equal the y s exactly, and the question then arises of how discrepant they are, because while a small discrepancy might be tolerable a large di screpancy is not (p. 33). The goodness of fit improves as the observed values and predicted va lues approach equality. For example, if a scatterplot reveals that all point s fall on a straight line, then the predictive power of the regression equation would be perfect, a nd the subsequent f it would be perfect. The comparison is usually performed through so me statistical comparison of the observed outcome values and the predicted (i.e., fitted) outcome values. Rather than compare and summarize the actual observed and predicted values, it is common to gain summary information 31

PAGE 32

by inspecting the log-likelihood va lue produced from the estimation procedure. Since the model parameters are estimated are from the data, pe rfect fit (i.e., observed-fitted = 0) is rare.4 Hence, the goodness of the fit is measure d to determine whether the difference is small enough to be tolerated. There are many measures of model fit. Typically, the model is compared either to a null model in which the only parameter is the mean or a full model in which the number of parameters is equal to the sample size. It is well-known that minus twice the LR statistic has a limiting chi-square distribution under the null hypothesis (Vuong, 1989, p.308). McCullagh and Nelder (1989) equivalently st ate, The discrepancy of a f it is proportiona l to twice the difference between the maximum log likelihood achievable and that achieved by the model under investigation (p.33). This deviance statistic (G2) is then considered to be asymptotically distributed chi-square with degrees of freedom equal to the number of parameters subtracted from the sample size.5 A significant p -value indicates that the devi ance is greater than what would be expected under a null hypothesis that th e model with less parameters is adequate; hence, the observed model with an additional parame ter or parameters is c onsidered a significant improvement over the null model. Another measure of model fit is Pearsons X2; however, unlike G2, it is not additive for nested models. Yet another measure of model fit is Akaikes Information Criterion (AIC), which penalizes the deviance for the number of parameters in the model.6 The notion is that increasing 4 Perfect fit is always obtained if th e number of parameters and the sample size are identical (McCullagh & Nelder, 1989). 5 The relationship is not always exact since sometimes the deviance is scaled and/or the likelihood is more difficult to estimate than in the simple logistic regression scenario presented here (McCullagh & Nelder, 1989). 6 Other measures such as the Bayesian Information Criteri on (BIC) penalize for both the number of parameters and the sample size. 32

PAGE 33

the number of parameters will increase the log-lik elihood regardless of the model and the data. Hence, the AIC penalizes the loglikelihood with regard to the number of parameters. There are two variations that provide furt her penalties. The Bayesian In formation Criterion (BIC) penalizes for sample size; the Consistent Akaike Inform ation Criterion (CAIC) penalizes even further by considering sample size and adding a small adjustment (Cameron & Trivedi, 1998). These indices can be compared to those of comp eting models; however, this must be done descriptively, not inferentially. The disadvantage is that the AIC can not be compared to a statistical distribution resulting in probabilities for significance testin g; however, the advantage is that, as a descriptive statistic, it can be used to compare nonnested models. The explanation thus far points to the fact that models can be compared to null, full, or other models. Statistical comparis on is valid to the extent that one model is nested within the other, which is to say that both models share the same parameters, and one model has at least one parameter that is not included in the other. Alternatively, Clar ke (2001) defines the models as follows: Two models are nested if one model can be reduced to the other model by imposing a set of linear restrictions on the parameter vector . Two models are nonnested, either partially or strictly, if one model cannot be reduced to the other model by imposing a set of linear restrictions on the parameter vector (p.727). The deviance for comparing two models is calculated as the difference in log likelihoods between the two models then multiplied by -2. This quantity is asymptotically distributed chisquare with degrees of freedom equal to the difference in parameters between the tw o models (Agresti, 1996). A significant p -value indicates that the deviance is greater th an what would be expected und er a null hypothesis of model equivalence; hence, the more complex model wi th an additional parameter or parameters is considered a significant improvement over the nested model. 33

PAGE 34

The difference in log-likelihood statistics (i.e., deviance) can not be used to statistically test nonnested models. This is due to the fact th at neither of the models can be considered the simple or more complex models with additional variables leading to a probabilistically higher log-likelihood. A t -test (or F -test) is a sensible al ternative that eliminat es concern for nesting. However, Monte Carlo simulations have demons trated that, for model comparison tests, the F test is lacking in sufficient power and can result in multicollinearity (Clarke, 2001). The motivation for the AIC statistic is that, al l else being equal, the greater the number of coefficients, the greater the log-likelihoods (C larke, 2001, p.731). Hence, model fit becomes impacted by the number of variables in the mode l along with the effects of those variables. Hence, the AIC penalizes for the number of parameters. The formula is 2()2 A ICLLK (2-20) where LL is the log-likel ihood estimate and K is number of parameters in the model including the intercept. Hence, now the l og-likelihood is adjusted to acco mmodate simplicity and parsimony (Mazerolle, 2004). In actuality, one could compare log-like lihoods between nonnested models. However, beyond the lack of parameter penalty, this tech nique might lead to th e statistical hypothesis testing associated with log-likelihood statistics (i.e., test for the deviance approximated by2 ).The AIC, on the other hand, should not be us ed in a formal statis tical hypothesis test regardless of whether the model is nested or nonnested (Clarke, 2001). Generally, the researcher looks at several AIC indices and decides which model fits best based on a lower-is-better criterion. Mazerolle (2004) states, The AIC is not a hypothesis test, does not have a p -value, and does not use notions of significan ce. Instead, the AIC focuses on the strength of evidence . and gives a measure of uncertainty for each model (p.181). 34

PAGE 35

Logistic modeling necessitated treating all nonzero numbers of children as a value of one. Depending on the research question, this may be a loss of valuable information (Slymen, Ayala, Arredondo, & Elder, 2006). Although sometimes it is necessary to model zero-inflated binomial data (Hall, 2000), specifying a binary distribut ion and logit link is not an ideal method for handling zero-inflated count data. The genera lized linear model that specifies a binomial distribution and a logit link beco mes more relevant when discu ssing the technique of splitting zero-inflated data into a model for the probability of zero separate from or combined with a model for the counts. The Poisson Distribution and the Log Link McCullagh and Nelder (1989), La wal (2003), and Rice (1995) are the key references for the technical underpinnings for this model and dist ribution. The generalized linear Poisson model is considered to be the benchmark mode l for count data (Cameron & Triverdi, 1998).7 This is primarily attributed to the fact that the Poisson distribution has a nonnegative mean (Agresti, 1996). If y is a nonnegative random variable, the Poi sson probability mass function is given by (;)Pr() y e fkYk k k = 0, 1, 2, . (2-21) where is standard Poisson notation for the mean ( ) and k is the range of counts. Derivations by Rice (1995) show that th e expected value of a random Poisson variable is ; hence, the parameter of the Poisson distribution can thus be interpreted as the av erage count (p.113). Alternatively, lambda ( ) represents the unobserved expected ra te or occurrence of events . (Zorn, 1996, p.1). The moment generating function is 7 It is also commonly used to model event count data, which is data composed of counts of the number of events occurring within a specific observation pe riod . [taking the] form of non-negative integers (Zorn, 1996, p.1). 35

PAGE 36

(exp()1)()tX tEee. (2-22) The resulting cumulant generating function is(exp()1) t which, with a variance function of and a dispersion parameter equal to one, lead s to mean and variance both being equal to and the skew equal to one divided by the square root of This equivalence of the mean and variance defined by a single parameter (Cameron & Triverdi, 1998; Agresti, 1996) is the result of a function that yields residuals that sum to zero (Jang, 2005); hence, the systematic portion of a Poisson GLM has no error term. The Poisson distribution is a generalization of a sequence of binomial distributions. Rodrguez (2006) explained that the Poisson distribution can be derived as a limiting form of the binomial distribution if you consider the di stribution of the number of successes in a very larger number of Bernoulli trials with a small proba bility of success in each trial. Specifically, if ~(,) YBn then the distribution of Y as n and 0 with n remaining fixed approaches a Poisson distribution with mean Thus, the Poisson distribution provides an approximation to the binomial for the analyses of rare events, where is small and n is large (p.3). Rice (1995) clarified, The Poisson dist ribution can be derived as the limit of a binomial distribution as the number of trials, n approaches infinity and th e probability of success on each trial, p approaches zero in such a way that np (p.43). Scheaffer (1995) and Rice (1995) have derived the generalization. Further, just as a binomial distribution conve rges to a normal distri bution given sufficient trials, the Poisson distributi on converges to a normal distribution given a large mean. The log-likelihood function for the Poisson distribution is (,)logii iyy (2-23) 36

PAGE 37

with the maximum likelihood estimate of simply being the sample mean (Rice, 1995) and with the related devi ance function being (,)2{log(/)()}iiiiiDevianceyyyy k. (2-24) McCullagh and Nelder (1989) state that the second term is often ignored. Provided that the fitted model includes a constant term, or intercept, the sum over the units of the second term is identically zero, justifying its omi ssion (McCullagh & Nelder, 1989, p.34). The systematic portion of the genera lized linear model takes the form 1122exp()exp()exp()...exp()iiiik ixxx xi (2-25) which is often equivalently expressed as log()'iX (2-26) with derived by solving the equation 1(exp())n iii iyxx0 (2-27) by using iterative computations such as the Newton-Raphson. The canonical link for a generalized linear model with a Poisson random component specifica tion is the log link (Stokes, Davis, & Koch, 1991). log() ~() YP (2-28) The Poisson distribution is not limited to count variates. Cameron and Triverdi (1998) explain that, although count s are usually in the purview of di rectly observable cardinal numbers, they may also arise through a latent process. In other words, ordinal rankings such as school course grades may be discretized as pseudocounts and assumed to have a Poisson distribution. Hence, the results of an analysis based on a Pois son distribution and the r esults using an ordinal analytic technique are often comparable. 37

PAGE 38

As is almost always the case, it is common to identify other variables associated with the count variable (i.e., misspecification). However, th e Poisson model has in interesting feature in that it assumes that there are no variables excluded from the model that are related to the count variable. In other words, there is no stochastic variation (i.e., no error term) (Cameron & Trivedi, 1998). Modifications must be made when one wi shes to use a Poisson model with stochastic variation. Iterative Estimation Agresti (1996) clarifies the iterative estimation procedure. The Newton-Raphson algorithm approximates the log-likelihood function in a neighborhood of the initial guess by a simpler polynomial function that has shape of a concave (mound-shaped) parabola. It has the same slope and curvature location of the maxi mum of this approximating polynomial. That location comprises the second guess for the ML estimates. One then approximates the loglikelihood function in a neighborhood of th e second guess by another concave parabolic function, and the third guess is the location of its maximum. The successive approximations converge rapidly to the ML estimates, often within a few cycles (p.94). The most common methods for estimating standard errors includ e Hessian maximum likelihood (MLH) (i.e., second partial derivative based) standard errors and maximum likelihood outer products (MLOP) (i.e., summed outer product of first derivative) estimation. Interpretation of Coefficients Standard ordinary linear squares re gression lends an in terpretation of as the predicted additive change in the response variable per one-unit change in the predictor variable. However, as was the case with the binomial distribution, the interpre tation differs when considering exponential distributions. For th e Poisson distribution, a on e-unit increase in X has a multiplicative of exp( ) on the The mean of Y at x +1 equals the mean of Y at x multiplied 38

PAGE 39

by exp( ) (Agresti, 1996, p.81).8 Due to the inherent difficulty in interpretation, it is common to express in one of three alternative wa ys. First, the direction of the sign of indicates a positive or negative effect of the predictor on the count variable. Second, the fitted value can be calculated at the mean.9 Third, some interpret the coefficien t in terms of percent change; hence, if =1.64, then as X increases to X+1, the predicted probability incr eases by 64% (Agresti, 1996). Hypothesis testing After conducting the analysis and estimati ng parameters, hypotheses can be tested in several ways as explained by Agresti (1996). One could test the hypothesis that =0 using the traditional Wald z -test via/b z bse Some programs provide Wald test results that are actually z2; this is the Wald 2 statistic with one degree of freedom and appropriate only for a two-tailed test. A third method, the Score test, is based on the behavior of the loglikelihood function at the null value for =0 (Agresti, 1996, p.94) yielding a chi-s quare distributed st atistic with one degree of freedom. Overdispersion In practice, the assumption of an equal mean a nd variance is the exception rather than the norm (McCullagh & Nelder, 1989). It is often the case that the sample variance is greater than or less than the observed sample mean with thes e two seldom being statistically equivalent (Cameron & Trivedi, 1998), especially for zero-inflated data (Welsh, Cunningham, Donnelly, 8 This is similar to the interpretation for the binomial distribution with logit link; however, now the multiplicative effect is directly on rather than on the odds of 9 This can be particularly troublesome since that fitted va lue will only hold at the mean. It provides no valid inference for values greater than or less than the mean since the function is a curve with steepness that can vary drastically between separate values for the predictors. 39

PAGE 40

and Lindenmayer, 1996). This condition is known as overdispersion10 (underdispersion) and is a violation of a major tenet of the Poisson dist ribution that the conditional mean and conditional variance of the dependent variab le are equal (i.e., equidispersion, nonstochastic ity) (Jang, 2005; Zorn, 1996).11 This assumption of equidispersion is th e analog of the ordinary least squares regression assumption of homoscedasticity. The overdispersion has been explained as heterogeneity that has not been accounted for [that is] unobserved (i.e., the population consis ts of several subpopulatio ns, in this case of Poisson type, but the sub population membership is not observed in the sample (Bhning, Dietz, Shlattman, Mendonca, & Kirchner, 1999, p.195). The impact of violation is one of incorrect conclusions due to inaccurate t -statistics and standard erro rs (Cameron & Triverdi, 1998; Agresti, 1996). The estimates of the coefficients can still be consistent using Poisson regression, but the standard errors can be biased and they will be too small (Jewell & Hubbard, 2006, p.14). Alternatively, Slymen, Ayala, Arredondo, and Elder (2006) state that Confidence intervals for regression estimates may be too narrow and tests of association may yield p -values that are too small (p.2). The underlying mechanism for overdispersion is explained as unobserved heterogeneity in responses. It is apparent that some modification to the variance to accommodate over-dispersion is ideal. Typically, maximum likelihood procedures are used to estimate parameters in the model. The term pseudo(or quasi) maximum likelihood estimation is used to describe the situation in which the assumption of correct sp ecification of the density is re laxed. Here the first moment [i.e., the mean] of the specified linear exponent ial family density is assumed to be correctly 10 Overdispersion is sometimes referred to as extra-Poisson variation (Bhning, Dietz, Shlattman, Mendonca, & Kirchner, 1999). 11 The underor overdispersion may disappear when predictors are added to the model; however, this is likely not the case if the variance is more than tw ice the mean. (Cameron & Trivedi, 1989). 40

PAGE 41

specified, while the second [i.e., the variance] and other moments are permitted to be incorrectly specified (Cameron & Triverdi, 1998, p.19). He nce, the Poisson distribution as a baseline (Ridout, Demtrio, & Hinde, 1998) can be m odified to accommodate overdispersion and underdispersion (Cameron & Triverdi, 1998). Rice (1995) states that ga mma densities provide a fairly flexible class for modeling nonnegative random variables (p. 52). One way to accommodate overdispersion is to consider the unobserved heterogeneity as a gamma distri buted disturbance added to the Poisson distributed count data (Jang, 2005). In other words, an individu al score may be distributed Poisson with a mean of but then this mean is regarded as a random variable which we may suppose in the population to have a gamma distribution with mean and index / (McCullagh & Nelder, 1989, p.199). This mixture l eads to the negative bi nomial distribution. Given gamma function, and count y the negative binomial probability mass function is () (;,) !()(1)yy prYy y (2-29) Cameron and Trivedi (1998) provide a formulati on where the negative binomial extends from the Poisson rather than being explained as a mi xture distribution. Gi ven a set of predictor variables, we can define p ii i (2-30) where is scalar parameter to be specified or estimated, and p is a pre-specified power term. If the scalar parameter is set to zero then the r esulting variance is equal to the mean, and the Poisson distribution holds. Hence, the Poisson model is nested within the negative binomial model. 41

PAGE 42

The standard formulation for the negativ e binomial formulation of the function, sometimes called NB2 (Cameron & Trivedi, 1998), leads to a variance that is quadratic by setting p to 2 2 ii i (2-31) This is the formulation seen in most textbooks (Rodrguez, 2006; Schea ffer, 1996). Its mean is the same as that of the Poisson distribution; how ever, its variance is derived from the gamma distribution (Cameron & Trivedi, 1998). Just as the Poisson distribution converges to a binomial distribution, the negative binom ial distribution converges to a Poisson distribution (Jewell & Hubbard, 2006). Poisson and Negative Binomial Models with Zero-Inflation As elaborated upon previously, th e zero-inflation problem is tw o-fold. First, the proportion of zeros is higher than expected given the spec ified population distribution shape resulting in an excess zeros problem. This can be descriptively determined by calculating the expected number of zeros as (())()*(exp()) EfqYfqYY = (2-32) e nFor example, Zorns example had a frequency of 4,052 with a = 0.11. The expected frequency of zeros would be (())4052*(exp(0.109))3,634 EfqY (2-33) which is less than the 3,882 zeros observed in the data. It turns out to be (3882/3634)*100 = 107% of the expectation (Zorn, 1996). Second, the zeros can be a mixture of stru ctural (i.e., true) zero s (Ridout, Demtrio, & Hinde, 1998) and sampled zeros reflecting the mu ltiple sources of zeros problem. Shankar, Milton, and Mannering (1997) state that if a two-state process is modeled as a single process . 42

PAGE 43

if applying traditional Poisson and NB distributions, the estimated models will be inherently biased because there will be an over-representation of zero observations in the data, many of which do not follow the assumed distribution of [the ] frequencies (p.830) Shankar, Milton, and Mannering (1997) note that the ne gative binomial model can spuriously indicate overdispersion when the underlying process actually consists of a zero-altered splitting mechanism (p.835836). In sum, the sources of zeros arise from a dual generating pro cess (i.e., structural and sampled) leading to two sources of unequal m ean/variance dispersion (i.e., that due to unobserved heterogeneity of responses and that due to zero-inflation). Most complex methods for analyzing zero-infl ated count data mode l a mixture of two different distributions. The justification for splitting the distribution into two pieces is wellreasoned by Delucci and Bostrom (2 004). If it is deemed more r easonable to consider the zeros as indicators of cases without a proble m, a more appropriate approach is to ask two questions: is there a difference in the proportion of subjects without the problem [i.e., structural true zeros], and, for those who have a problem [sampled fal se zeros], is there a difference in severity (p. 1164). Zorn (1996) refers to dual regime models wherein an observation experiences a first stage in which there is some probability that its count will move from a zero-only state to one in which it may be something othe r than zero (p.2). T ypically, the dual-regime is composed of a transition stage based on a binomial distribution and an events stage based on some type of Poisson distribution. There are many ways to model two-part di stributions. For example, Mullahy (1986) and King (1989) proposed a Hurdle model in which th e zeros are analyzed separately from the nonzeros. Lambert (1992) proposed a zero-inflated Poisson (ZIP) model in which different 43

PAGE 44

proportions of zeros are analy zed separately and along with the nonzeros. Another early formulation (Heilborn, 1989) w as the zero-altered Poisson (ZAP ) model. Arbitrary zeros are introduced by mixing point mass at 0 with a posit ive Poisson that assigns no mass to 0 rather than a standard Poisson (Lambert, 1992, p.1). Mullahy (1986) presented a variation of the Hurdle model based on a geometric distribution12 for use when specifying a Poisson distribution is not reasonable. Another possib ility is to specify a log-gamma distribution for the event stage (Moulton, Curriero, & Barruso, 2002). Lambert ( 1989) presented a variat ion to the ZIP model known as ZIP( ), which introduced a multiplicative constant to the event stage covariance matrix in order to account for the re lationship between the two models Gupta, Gupta, and Tripathi (1996) derived an adjusted generalized Poisson regression model for hand ling both zero-inflation and zero-deflation; however, accuracy was suggested to be contingent on the amount of inflation or deflation.13 It is also possible to form ulate the model with differen t link functions. Lambert (1989) mentions the possibility of usi ng the log-log link, complementary log-log link (Ridout, Demtrio, & Hinde, 1998), and additive log-log link whil e Lachenbruch (2002 ) mentions the lognormal and log-gamma distributions. Hall (2000) formulated a two-part model for zero-inflated binomial data. Gurmu (1997) describes a semi-parametric approach that avoids some distributional assumptions (Ridout, Demtrio, & Hinde, 1998). There is some research on the extension of two-part models to accommodate random effects (Min & Agresti, 2005; Ha ll & Zhang, 2004; Hall, 2004; Hall, 2002; Olsen, 1999). Halls (2000) model for zero-inflated binomial data pe rmits both fixed and random effects. Dobbie and 12 This distribution is an extension of the binomial distribution where the sequence in infin ite. It is typically used in cases where the researcher is concer ned with probability up to and including the first success (Rice, 1995). 13 Min (2004) stated that the Hurdle model also has this feature. 44

PAGE 45

Walsh (2001) permit correlated count data. Fi nally, Crepon and Duguet (1997) consider the cases where the variables are latent and correlated. The evolution of the research to date has led to an emphasis on the standard Hurdle model and ZIP models (along with their negative binomial extensions) with a binary distribution for the transition stage and a Poisson dist ribution for the events stage and with fixed covariates. For both models, estimates are obtained from maximum likelihood procedures, a lthough there has been some research on the use of generalized estim ating equations (GEE) (Hall & Zhang, 2004; Dobbie & Welsh, 2001). The models are primarily distinguished by whether zeros are permitted in the event stage. In other words, their diffe rences are a reflection of the researchers notions about the potentially multiple sources of zeros in the data and their relationship to excess zeros. They also differ in terms of the transition stage cumulative probability function. To be clarified in the sections that follow, Zorn ( 1996) summarizes the differences as follows: . the hurdle model has asymmetric hurdle proba bility while in the ZIP specification pi is symmetrical. Also, the hurdle model does not permit zero values to occur once the hurdle has been crossed, while in the ZIP model zeros may occur in either stage (p.4). Choosing be tween the models is a matter of validity; hence, the choice rests on substantive ground as well as statistical considerations. As Martin, Brendan, Wintle, Rhodes, Kuhnert, Field, Low-Choy, Tyre, and Possingham (2005) note, it is impe rative that the source of zero observations be considered and modeled accordingly, or we risk making incorrect inferences . (p.1243-1244). The Hurdle model The Hurdle model was developed separate ly by Mullahy (1986) in economics and King (1989) in political science, although the term itself was most likely coined by Cragg (1971). 14 14 Cragg (1971) proposed the basic two-step process in wh ich the probability of occurrence is modeled separately from frequencies of occurrence. 45

PAGE 46

Welsh, Cunningham, Donnelly, and Lindenmayer (1996) refer to it as a conditional Poisson model. The idea underlying the hurdle formula tions is that a binomial probability model governs the binary outcome whethe r a count variate has a zero or a positive realization [i.e., a transition stage]. If the reali zation is positive the hurdle is crossed, and the conditional distribution of the positives is governed by a truncated-at-zero15 count data model [i.e., events stage] (Mullahy, 1986, p.345) su ch as a truncated Poisson or truncated negative binomial distribution (Min & Agresti, 2005). In other wo rds, one distribution addresses the zeros while another distribution addresses the positive nonzero counts. For example, for grade-retention data, there would be a model for schools with no dropo uts and a model for school with at least one dropout. It is a finite mixture genera ted by combining the zeros generated by one density with the zeros and positives generated by a second zero-truncated density separately . (Cameron & Trivedi, 1998, p.124). Log-likelihood values are estimated separately for each density. A key feature of the transition model is asymmetry in that the probability of crossing the hurdle increases more quickly as the co variates increase than it d ecreases as the covariates decrease. The function is then as ymmetrical (King, 1989) leading to validity concerns supporting or refuting the substantive theo ry underlying the model. The two-part distributio n of the dependent variable is given first by the transition stage g1 probability mass function 1(0)(iPYg 0) 16 (2-34) 15 The lower bound is then one. 16 This is an alternative representation of the aforementionedPr(0)1 Y 46

PAGE 47

modeling whether the response crosses the hurdle of zero. Assuming a Poisson distribution and log link, Zorn (1996) expands the cumulative distribution function to include covariates as 001exp[exp()]ipX (2-35) The basic model for the event stage is then th e probability for a nonzero realization multiplied by the probability for the counts. 17 2 1 2() ()(1(0))...1,2,... 1(0)igj PYjg j g (2-36) Greene (1994) notates the models with a binomia l distribution and logit link for the transition stage and a Poisson distribution with a log link for the event stage as18 Transition: Pr(0)i y p (2-37) Event: 1 Pr()()(),1,2,... 1!i ik i ie p yk k ek (2-38) Here, p is the probability of a count of zero while is the truncated Poisson mean for the counts greater than zero. The generalized linear m odels as a function of covariates is then Transition Stage: 11log() 1i i ip p x (2-39) Event Stage: (2-40) 22log()ii xThe two vectors of parameters are estimated jointly. 1[{exp[exp()]}{1exp[exp()]} XX (2-41) 2[exp()/({exp[exp()]1}!)] yX Xy (2-42) An alternative notation provided by Min (2003) is 17 The second term is an alternativ e representation of the aforementioned (;)Pr() !ye fkYk k 18 Note that p is used to representative the probability of a zero rather than the conventional representation of a probability of one. 47

PAGE 48

111 11 00()[log(0;,)][log(1(0;,))]iiii i yyPyx Pyx1 1 ie2 (2-43) 1111 01,log(1)i in x i yix' 22 22' 22 2 00()[ log(1)]log(!)x i i i ix e ii i yyyxee y (2-44) Since the two models are func tionally independent, the likelih ood functions can be maximized separately (Min & Agresti, 2005; Min, 2003, Cameron & Trivedi, 1998; Mullahy, 1986). 1log()log()hurdle (2-45) 1212(,)()() (2-46) This is because the large sample covariance be tween the two sets of parameters is zero so the joint covariance matrices can be obtained from the separate fits (Welsh, Cunningham, Donnelly, & Lindermayer, 1996, p.300). Solving the likelihood equations uses either the Newton-Raphson algorithm or the Fisher scoring al gorithm, both giving equi valent results (Min, 2003). The Poisson model is nested within the Hu rdle model (Zorn, 1996). Hence, fit of these two models can be compared statistically.19 The Negative Binomial Hurdle model In the case of zero-inflated data, it is possibl e to have two sources of overdispersion. The variance can be greater th an the mean due to the preponderance of zeros. However, there is now the possibility of unobserved heterogeneity in th e event stage (Mazerolle, 2004; Min & Agresti, 2004). The former scenario has been referred to as zero-driven overdispersion (Zorn, 1996); the latter is Poisson overdispersion. Just as was the c ase with the Poisson model, it is possible to nest the Hurdle model within a more general nega tive binomial framework. Further, the negative binomial Poisson model is nested within the negative binomial Hurdle model. Hence, the fit of a) 19 Zhou and Tu (1999) developed likelihood ratio for count da ta with zeros; however, it was not generalized to any particular zero-inflation model. 48

PAGE 49

the Poisson model and the Hurdle model, b) the negative binomial Poisson model and the negative binomial Hurdle model, and c) the Hurdle model and the negative binomial Hurdle model can be compared using statistical tests.20 Estimation is typically performed by solving the maximum likelihood equations usin g the Newton-Raphson algorithm. The Zero-Inflated Poisson (ZIP) model The Zero-Inflated Poisson, or ZIP, model is another model that one can use when the zeros in a dataset are argued to be caused by both chance and syst ematic factors (Min & Agresti, 2005). The transition stage addresses zero-infla tion while the event stage addresses unobserved heterogeneity of responses including zeros ( Jang, 2005). Welsh, Cunningham, Donnelly, and Lindenmayer (1996) refer to it as a mixture model. This two-part model, developed by Lambert (1992) permits zeros to occur in both the transition stage and event stage (Cameron & Triv edi, 1998); crossing the hurdle in the ZIP model does not guarantee a positive realization of Y (Zorn, 1996, p.4). Further, the probability function in the transition stage is now symmetrical (Zorn, 1996) Lachenbruch (2002) explains that ZIP regression inflates th e number of zeros by mixing point mass at 0 with a Poisson distribution (p. 12). Zorn (1996, p.4) clarifies the distinction between the ZIP and Hurdle models as follows: As a special case of the general model, the ZIP regression is thus seen to make substantially different assumptions about the na ture of the data gene rating process than the hurdle model. Whether parameterized as a logit or a probit, the probabi lity exiting the zeroonly stage is assumed to follow a symmetric cumulative distribution. Likewise, even those cases which make the transition to the events stage may nevertheless have zero counts; crossing the hurdle in the ZIP model does not guarantee a positive realization of Y . The sole difference in assumptions here is that the hurdle models count distribution is assumed to be truncated at zero whereas the ZIP specification count data may take on zero 20 It is not the case that the negative binomial Poisson model is nested within the Hurdle model; hence, one can not statistically compare all fo ur models collectively. 49

PAGE 50

values in the event stage. Another difference is that, unl ess the ZIP model is overly pa rameterized, only the Hurdle model can handle zero defla tion (Min & Agresti, 2005). Compared to the Hurdle model, the equations for the event stage are very similar. The exception is that (1pi) is divided by (1ei ) in the Hurdle model before being multiplied by the remaining elements of the equation. However, the transition stage equations are strikingly different. For the Hurdle m odel, the equation is Pr(yi = 0) = p ; the ZIP model includes the addition of the probability of a nonzero multiplied by the exponentiated Poisson mean. This is the mathematical characteristic that distinguishe s the Hurdle models excl usion of zeros in the event stage and the ZIP models potential inclusion of zeros in the event stage. Rather than model the probability of a zero in the transition stage, the ZIP also models the probability that the counts have a Poisson distribution hence permitting zeros from both a perfect state and a Poisson state (Hur, 1999). Given this, parameterizes the mean of this Poisson distribution (Welsh, Cunningham, Donnelly, & Lindenmayer, 1996). When adding covariates, the Hurdle and ZIP generalized linear model appear the same. Transition Stage: 11log() 1i i ip p x (2-47) Event Stage: (2-48) 22log()ii xUnlike the Hurdle model, the ZIP model likelihood function can not be maximized separately for the transition and event stage. Hence, the Hurdle model has the attractive advantage of an orthogonal parameterization which makes it simpler to fit and interpret than the mixture model (Welsh, Cunningham, Donnelly, & Lindenmayer, 1996) with the disadvantage of asymmetrical transiti on to counts. The likelihood func tion derived by Lambert (1992) is 50

PAGE 51

0010log(exp())()log(1)log(!)ii ii ii in GBBG ii i yy iLeeyBeeyy (2-49) where B is the vector of coefficients and matrix of scores for the event stage and G is the vector of coefficients and matrix of scores for th e transition stage, and where iterations are based on the EM or Newton-Raphson algo rithms (Min, 2003; Lambert, 1992)21. Strictly speaking, the Poisson m odel is not nested within the ZIP model; therefore, it would not be wise to conduct a formal model fit test (Zorn, 1996; Greene, 1994). However, it is interesting to note that the log-likelihood of 10,607 is slightly lower than that produced by the negative binomial Hurdle model. This is in lin e with Greenes (1994) observation when using the Vuong statistic as an alternat ive for testing nonnested models. For present purposes, the important question is whether the ZIP models . provide any improvement over the basic negative binomial . The log-likelihood functio ns are uniformly higher, but as noted earlier, since the models are not nested these are not directly compar able. The Vuong statistic, however, is consistent with the observa tion (Greene, 1994, p.26). The Negative Binomial Zero-Inflated Poisson model The ZIP model can be extended to the negati ve binomial model just as the Poisson was extended to the negative binomial and as the Hu rdle was extended to the Hurdle negative binomial. This may be necessary as Min (2003) expl ains that Sometimes such simple models for overdispersion are themselves inadequate. For instance, the data might be bimodal, with a clump at zero and a separate hump around some c onsiderably higher value. This might happen for variables for which a certain fraction follows some distribution have positive probability of a zero outcome (p.13). He further explains, The equality of the mean and variance assumed by 21 Alternatively, Martin, Brendan, Wintle, Rhodes, Kuhnert, Field, Low-Choy, Tyre, and Possingham (2005), compared relative means and credible intervals estimate from a bootstrap procedure. 51

PAGE 52

the ZIP model . is often not realistic. Zero -inflated negative binomia l models would likely often be more appropriate than ZIP models (p .15). Unfortunately, current iterative techniques lead to a greater risk of nonc onvergence than when using the ot her two-part models (Fanoye & Singh, 2006; Lambert, 1992). The key assumption in accepting these results ov er those produced from the Hurdle models is that there are some zeros that belong w ith the counts representing no household members under the age of 21 for reasons other than thei r never having such household members at all. Although it is not valid to statistically compare the fit of the ZIP model with the Hurdle and Poisson models, it is reasonable to test the fit of the ZIP model within the negative binomial ZIP model. In sum, the choice between Hurdle models and ZIP models is ultimately guided by the assumptions one makes about the data generating process. Min (2004) states, The zero-inflated models are more natural when it is reasonable to think of the population as a mixture, with one set of subjects that necessarily has a 0 response. However, they are more complex to fit, as the model components must be fitted simultaneously. By contrast, one can separately fit the two components in the hurdle model. The hurdle model is also suitable for modeling data with fewer zeros than would be expected under standard distributi onal assumptions (p.20). Model Comparison Testing for Zero-Inflated Data The Poisson model is nested within th e negative binomial Poisson differing only by the dispersion parameter. In fact, the two models ar e equivalent when one uses the negative binomial model and restricts the dispersion parameter to 1.0 achieving Poisson equidispersion (Cameron & Trivedi, 1998). Likewise, the Hurdle model is nested within the negative binomial Hurdle model, and the ZIP model is nested within the negative binomial ZIP model. However, since each involves estimation of a transition stage and event stage model, the n esting rule applies to 52

PAGE 53

both equations. In other words, if the event stag e contains 3 parameters and the transition stage 4 parameters for the nested model, then the more complex model must contain at least these same 3 parameters in the event stage and at least th e same 4 parameters in the transition stage. According to Zorn (1996), th e Poisson model is nested within the Hurdle model, and the negative binomial Poisson model is nested within the negative bi nomial Hurdle model. This is reasonable given that the Hurdle models are estim ating the Poisson models in the event stage and that these likelihood statistics are independent of those produced in the transition stage. The ZIP models, on the other hand, are not estimated in th is manner; it is not reasonable to assume that the Poisson models are nested wi thin the ZIP Models (Greene, 1994) This leads to the hierarchy of permissible model-testing displayed in Table 2-1. Other models can be compared descriptively using the aforementioned Akaikes In formation Criterion (AIC). The remainder of the dissertation will use the deviance statistic for model comparison inference and the AIC for model comparison description. Review of Research Pertaining to and Using Zero-Inflated Count Data Hurdle Model Statistical As previously discussed, Mullahy (1986) pr esented the underlying statistical foundations for the Hurdle model. He specified it in term s of extending from both th e Poisson and geometric distributions. He also presented an extension to the Hurdle model that he termed the with-zeros (WZ) model. The WZ model adjusts for zero-in flation by augmenting or reducing the probability by an additive constant (rather than having them specified by the parent di stribution). Subsequent research has focused on the Hurdle model rather th an the WZ model; this is most likely due to the fact that the WZ turns out to be a special case that collapses into a Hurdle model in some specifications, and estimates are often similar between the models (Mullahy, 1986). Finally, 53

PAGE 54

Mullahy presented tests for specifi cations including a technically complex information matrix test as well as the Score test typically provided in soft ware output. Kings (1989) key contribution to the Hurdle model was a Monte Carlo study confirming that the counts can be viewed as arising from a Poisson process (Civetinni & Hines, 2005). Min and Agresti (2004) conducted a simulati on to compare a zero-inflation and a zerodeflation condition. They found th at the estimates were reasonable for the ZIP model under zeroinflation. However, the coefficient and standard error for the event stage were both very large under zero-deflation. They explain that logit mo dels simply can not accommodate too few zeros. The zero-inflated model is only suitable for zer o-inflation problems. However, the hurdle model is also suitable for modeling data with fewe r zeros than would be expected under standard distributional assumptions. In fact, when a data set is zero-deflated at some levels of the covariates, the zero-inflation model ma y fail (Min & Agresti, 2004, p.5). In contrast, Mullahy (1986) stat ed that, a particularly intere sting feature of the modified count data specifications consider ed here [i.e., Hurdle model] is that they provide a natural means for modeling overdispersion or underdispersion of the data. Specifically, overdispersion and underdispersion are viewed as arising from a misspecification of the maintained parent [data generating process] in which the re lative probabilities of zero and nonzero (positive) realizations implied by the parent distribution are not supported by the data (Mullahy, 1986, p.342). Applications Mullahy (1986) researched daily consumption of coffee, tea, and milk where each is a count variable. Covariates included age, years of completed schooling, family income, sex, race, and marital status. For the Hurdle model, the test statistics [ p -values] for the coffee-model were substantially smaller than those for tea-model a nd milk-model. Mullahy argued that this is not surprising given the ratio of estimates and standard errors. However, closer inspection of the data 54

PAGE 55

reveals that it is the coffee vari able that has the low est proportion of zeros at only 26.26%. It is possible that the Hurdle model was unable to adequately account for the additional overdispersion witnessed in the ot her two models (61.63% and 40.37% zeros). In other words, a two-part model may be a necessary but not suffi cient condition for handling overdispersion in zero-inflated count models, and negative binom ial formulations may be increasingly necessary as the proportion of zeros increases. King (1989) applied the Hurdle model to data for the relationship between the number of nations entering war in a period of time as a function of those in formal interna tional alliances. The hurdle model was formulated based on the premises of Mullahy (1986) and justified due to there being some countries who will not go to war and others who will not at first but will later be dragged in by alliances. Hence, this is a classic example of the justification for true zeros and event-driven zeros. This w as demonstrated statistically by comparing Hurdle and Poisson results. The Poisson alliance co efficient of .007 was significant, the Hurdle model event-stage coefficient of .007 was significant, and the Hurdle model transition stage coefficient of .001 was not significant. Hence, the Poisson interpretati on would be that increas ed alliances lead to increased war frequency. However, the Hurdle results clarify that this is on ly true after the onset of war (i.e., the hurdle has been crossed). Furthe r statistical evidence supported the Hurdle model based on the likelihood-ratio model comparison test. Zero-Inflated Poisson Model Statistical As previously discussed, Lambert (1989) presen ted the original formulation for the ZIP model and its ZIP( ) formulation. She also presented the models extension from the Poisson and negative binomial as well as the derivation of the maximum likelihood (EM) estimates. She ran several simulations to test the adequacy of the model. The first simulation varied sample size, 55

PAGE 56

with one covariate taking on a fixed coefficient and a fixed variance in both parts of the model. The result was an average of 50% zeros in the tr ansition stage and 23% zeros in the event stage. The results suggest that the ZIP model consistently converges at n =25 when using EM and at n =100 when using the Newton-Raphson algorithm. An examination of confidence intervals revealed that the normal-theory intervals are not reliable at n =100; however, almost all simulated likelihood-ratio confidence intervals contained the true mean even at n =25. To summarize, these simulations with one covariate for both and p are encouraging. The ZIP and ZIP( ) regressions were not difficult to compute, and as long as infere nce was applied only when the observed information matrix was nonsingular, estimated coefficien ts, standard errors based on observed information, and estimated properties of Y could be trusted (Lambert, 1992, p.7). Warton (2005) compared 20 datasets of va rying sample sizes, pr oportions of zeros, and factors/levels. The ordinary least squares version included the addi tion of one to all counts before taking the logarithm. The other models not acco mmodating zero-inflations were the Poisson and four formulations of the nega tive binomial Poisson (including the aforementioned quasi-Poisson where the variance is set to ). The zero-inflated models included the ZIP model and the negative binomial ZIP model. The Akaike Info rmation Criterion (AIC) values were calculated for a total of 1,672 variables averaged over data sets and rescaled to a minimum AIC of zero for each dataset. As expected, when overdispersion was pr esent, the negative binomial formulations outperformed the models without these formula tions. However, when overdispersion was not present, the reverse was true for 53% of the variables. This suggests that the level of skew in the model interacts with zero-inflation when measuring model adequacy. 56

PAGE 57

When the proportion of zeros is very small, the distribution looks more like a Poisson distribution truncated at zero. In other words, it shares f eatures modeled by the event stage of a Hurdle model. This led to estimation problems in which the negative bi nomial model ZIP model rarely converged. When it did conve rge, its fit was bett er than that of th e ZIP model for only 11% of the datasets. Possibly, as previously discussed, the zero-inflated Hurdle would converge more often since it can handle both zero-inflatio n and zero-deflation (Min & Agresti, 2005; Min & Agresti, 2004). A very interesting finding pertained to the transformed OLS fit indices. Although transformed least squares was not the best fitti ng model for data, it fitted the data reasonably well. Surprisingly, transformed least squares appeared to fit data about as well as the zeroinflated negative binomial model . The AIC for transformed least squares was not as small as for the negative binomial model overall, although it was smaller for 20 per cent of the variables considered here (Warton, 2005, p.283). Averaging over all datasets, the AIC was lowest for all negative binomial models followed by a close tie between the transfor med OLS model and the negative binomial ZIP model, which was followed by the ZIP model. A ll models were a drastic improvement over the standard Poisson model. The implications are th at, although the Poisson is rarely adequate when the data is not equidispersed and/or is inflated or deflated, an intui tive climb up the ladder of models may not be reasonable. There were feat ures of these datasets including varying degrees of zero-inflation and overall dist ributions that warrant further i nvestigation toward appropriate model selection. If one were to fit a zero-infl ated model, it would be advisable to present quantitative evidence that the zero-inflation term was required. Based on th e present results, it is likely that a term for extra zeros is not needed and a simpler model will usually suffice . 57

PAGE 58

special techniques are not gene rally necessary to account for th e high frequency of zeros. The negative binomial was found to be a good model for the number of zeros in counted abundance datasets, suggesting that a good approach to analyz ing such data will often be to use negative binomial log-linear models (Warton, 2005, p.287-288). In regard to their problems w ith zero-inflated negative bino mial convergence, Fanoye and Singh (2006) developed an extension that im proves convergence termed the zero-inflated generalized Poisson regression (ZIGP) model. Thei r recent research revealed convergence in less than 20 iterations for all trials. However paramete r estimates and standard errors were often very different than those produced by the ZIP m odel. They conclude, Even though the ZIGP regression model is a good competitor of ZINB regression model, we do not know under what conditions, if any, which one will be better. The on ly observation we have in this regard at this time is that in all of the datasets fitted to bot h models, we successfully fitted the ZIGP regression model to all datasets. However, in a few cases, th e iterative technique to estimate the parameters of ZINB regression model did not converge (p.128). Greene (1994) used the Vuong statistic when comparing the Poisson, negative binomial Poisson, and ZIP models. It w as noted that the rank order for the Vuong statistics and the loglikelihood estimates were in alignment. The c onclusion suggested futu re research using the Vuong statistic. The use of Vuongs statistic to test the specifica tion seems not to have appeared in the recent literature . We conjecture that the Vuong testi ng procedure offers some real potential for testing th e distributional assumption in the di screte data contex t. In the cases examined, it appears to perform well and in li ne with expectations (Greene, 1994, p.30). Shankar, Milton, and Mannering (1997) used the Vuong statistic to decide between the negative binomial, ZIP, and nega tive binomial ZIP model for traffic accident data. They clarify 58

PAGE 59

the interpretation of the statistic stating, A value >1.96 (the 95% confidence level of the t -test) for V favors the ZINB while a value < -1.96 favors the parent-NB (values in between 1.96 an 1.96 mean that the test is indecisive) . This test can also be applied for the ZIP( ) and ZIP cases (p.831). Civettini and Hines (2005) explored the effects of missp ecification on negative binomial ZIP models. This included misspecification by leavi ng a variable out of the event stage that was present in the event stage and misspecification by sh ifting a variable from the transition stage to the event stage. Applications Lambert (1992), in formulating the ZIP model, applied it to the anal ysis of defects in manufacturing. In terms of improperly soldered leads, 81% of circu it boards had zero defects relative to the 71% to be expect ed under a Poisson distribution linked to a model with a threeway interaction. This most complicated model had a log-likelihood of -638.20. This dropped to -511.2 for the ZIP model. Although comparing to a differe nt combination of covariates, the negative binomial Poisson model fit better than th e Poisson model but not as well as the ZIP model. Greene (1994) used credit-reporting data to investigate differences between the Poisson, negative binomial Poisson, ZIP, negative binomia l ZIP, as well as some of their aforementioned variants and the specification of a probit link rather than the logit link. The data consisted of 1,023 people who had been approved for credit car ds. The count variable of concern was the number of major derogatory reports (MDR), which is the numb er of payment delinquencies in the past 60 days. For this sample, 89.4% had zero MDR. Given a mean of 0.13, this frequency of 804 is nearly double the 418 we might expe ct in a Poisson distribution. Th e skew of 4.018 is reduced to 59

PAGE 60

2.77 when ignoring the zeros while the mean increases to 1.22. As expected, the negative binomial Poisson resulted in improved fit (based on the Vuong test statisti c), increased standard errors and different parameter estimates. The ZIP model resulted in sligh tly worse fit than the negative binomial Poisson while re maining much better compared to the Poisson model. If all of the overdispersion was due to unobserved response heterogeneity then the results should be similar for the negative binomial ZIP model. However, this model produced the best fit of all. It is interesting to note that, again, the standard errors increase while the parameter estimates are different relative to the ZIP mode l. In fact, of the 6 parameters, 4 estimates decreased, 2 increased, and 1 switc hed in sign. Hence, there are two implications. First, the negative binomial ZIP model was necessary to ac commodate two sources of overdispersion to adjust standard errors. Second, ignoring the negative binomial formulations would have led to nonsensical parameter estimates driv en by a sample mean of 0.13. Bhning, Dietz, Schlattmann, Mendona, and Kirchner (1999) compared preand postintervention scores on the decayed, missing, and filled teeth index (DMFT) for 797 children in one of six randomly assigned treatment conditions The results were not exhaustive; however, the log-likelihood did decrease fr om -1473.20 to -1410.27 when going from the Poisson model to the ZIP model. This study was somewhat unique in that all the covariates (sex, ethnicity, and condition) were categorical, and that the c onditions were dummy-coded represented as five parameters. Also, this data had features that mi ght suggest that zero-in flated models werent necessary. For pre-intervention, the proportion of zeros was 21.58%, which increased to only 28.99% at post-intervention. The means, with the zeros in the data, were 3.24 and 1.85, respectively. Ignoring the zeros changed these m eans to 4.13 and 2.61, respectively. The skew, with the zeros in the data, was 0.20 and 0.65, re spectively. Ignoring zeros changed the skew to 60

PAGE 61

0.08 and 0.63, respectively. In other words, many features of the data were consistent with what would be expected of a Poisson, and possibl y normal, distribution. Nonetheless, with these means and frequencies, the Poisson distributi on suggests overdispersion with 31 permissible zeros for the pre-intervention and 125 permissi ble for the post-intervention whereas the data revealed 173 zeros and 232 zeros, respectively. It then becomes a matter of whether the overdispersion was due to the proportion of zeros in each condition or unobserved heterogeneity in the event stage. The negative binomial ZI P model was not used to analyze this data. Xie, He, and Goh (2001) analyzed the number of computer hard disk read-write errors. Approximately 87% of the 208 cases were zeros. Given that the Poisson mean was 8.64, the authors noted that the ZIP model is to be preferred over the Poisson model. However, this mean is due to several values between 1 and 5, a few between 6 and 15, and 2 values of 75. These latter two values appear to be so far from the ot hers that they should have been treated as outliers and addressed in some other manner. Jang (2005) analyzed the numb er of non-home based trips per day from 4,416 households in Jeonju City, Korea. The provided bar graph suggested that approximately 45% of the cases were zeros. The Vuong statistic (Vuong, 1989) w as used for model selection given that the Poisson is not nested within the ZI P or negative binom ial ZIP models. The purpose of the article by Delucchi a nd Bostrom (2004) was to provide a brief introduction to many possible methods for ha ndling zero-inflation including standard t -tests, bootstrapping,22 and nonparameteric methods. In doing so they provided results from a study involving 179 patients with opioid dependence as signed to either a methadone-maintenance or methadone-assisted-detoxification treatment. Five out of seven ways to segment the sample 22 See Jung, Jhun, and Lee (2005) for bootstrap procedures and simulation results for Type I and Type II errors. 61

PAGE 62

resulted in zero-inflation ra nging from 17% to approximately 66% zeros. The only two-part model to be used was the ZIP model. The ta ble of results revealed that, in terms of p -values, the ZIP model performs either very similarl y or very differently from the Pearson2 test for the proportion of zero values, the Mann-Whitney-Wilc oxon test of nonzero values, and/or the MannWhitney-Wilcoxon test of difference in mean scores between treatment groups It is possible that these tests become more similar as the proportion of zeros declines but such conclusions are based purely on the table of p -values. Desouhant, Debouzie, and Menu (1 998) researched the frequenc y of immature weevils in chestnuts. One tree was measured over 16 years, another was measured over 11 years, and three trees were measured on 1 year. The means ranged fr om .06 to .63. None of the 30 distributions fits a Poisson, 2 values being always very significant . The ZIP distribution fits 25 out of 31 cases . The NB distribution fits 20 out of the 31 (Desouhant, Debouzie, & Menu, 1998, p.384). This led to the conclu sion that researchers should c onsider both true zeros and overdispersion (i.e., trees not ed as contagious and trees va rying in random oviposition behavior). Shankar, Milton, and Mannering (1997) anal yzed a 2-year summary of traffic accident frequencies. For principal arterials, they chose a negative binomial model with data ranging from 0 to 84 ( M = 0.294, SD = 1.09). For minor arterials, they chose the negative binomial ZIP model for data ranging from 0 to 7 ( M = 0.09, SD = 0.346). For collector arterials, they chose the ZIP model for data ranging from 0 to 6 ( M = 0.61, SD = 0.279). Model selection was based on the Vuong statistic. For example, they state, As suspected previous ly, inherent overdispersion in the data is due to the parent NB process and this was validated when the [negative binomial ZIP] 62

PAGE 63

specification failed to provide a statistically better fit (the Vuong statistic <1.96, which corresponds to the 95% c onfidence limit of the t -test (p.833). Slymen, Ayala, Arredondo, and Elder (2006) analyzed pe rcent calories from fat and number of days of vigorous physical activity from 357 females partic ipating in a baseline condition and one of three treatment conditi ons. Covariates included employment status, education, martial status, cigarette smoking, and self-reported health. The zero-inflation was 294 out of 357 (82.4%). They compared models usin g likelihood ratio tests between the Poisson and negative binomial Poisson and likewise between the ZIP and negative binomial ZIP. The AICs were inspected to compare the Poisson and ZIP models. Not surprisingly, the negative bi nomial model fit better than the Poisson model. However, the ZIP model did not fit better or worse than the negative binomial ZIP, and the parameter estimates and standard errors were nearly identical. This suggests almost no overdispersion in the data. Indeed, the nonzero percentages were as follows: 1 = 2. 8%, 2 = 3.4%, 3 = 4.8%, 5 = 2.0%, 6 = 0.0%, and 7 = 2.0%. This suggests strong eq uidispersion leaning to ward a uniform nonzero distribution. The AICs for both models were also nearly equal although both being considerably smaller than the AIC for the Poisson model and somewhat smaller than the AIC for the negative binomial Poisson model. Based on a smaller-is-b etter heuristic, the authors favored the ZIP model with an AIC of 562.5 over the zeroinflated ZIP model with an AIC of 565. ZIP and Hurdle Model-Comparisons The purpose of this section is to present a breadth of literature in which both the Hurdle and ZIP models were either compared statisticall y and/or used to analyze real data. This also includes extensions such as the negative binomial and tau form ulations (e.g., ZIP( )). Some authors presented alternatives that seem to depart from the ZIP and Hurdle models too drastically to be within the scope of this dissertation. For example, Lachenbruch (2001) used a two-part 63

PAGE 64

model; however, the splitting form ulation was not consistent with the literature. Further, the model was compared to atypica l formulations such as the Wilcoxon, Kolmogorov-Smirnov, and z tests. As such, these types of articles are not included in the subsequent review. One exception is Xie, He, and Goh (2001) who included a like lihood-ratio test for comparing the Poisson and ZIP models. Statistical Greene (1994) proposed several zero-altered count models fo r comparison. First, he took Mullahys with-zeros (WZ) adaptation of the Hu rdle model and included a scalar estimate for ease on computational burdens. Greene also pres ented an adaptation of Lamberts ZIP known as ZIP( ) and modified it for the negative binomia l formulations terming them ZINB and ZINB( ). The intention was to identify a procedure which will enable us to test the zero inflated model against the simple Poisson model or against the negative binomial model. The latter will allow us to make a statement as to wh ether the excess zeros are the consequence of the splitting mechanism or are a symptom of unobs erved heterogeneity (Greene, 1994, p.10). Greene developed a method for co mparing the models; however, he noted that there was no a priori reason to think that the Vuong statistic would be inferior. Applications Zorn (1996) examined the counts of actions taken by Congress addr essing Supreme Court decisions between 1953 and 1987. The zeros were seen to arise from two sources since many cases will not be addressed unless there are lobbyist s to pressure redress. Covariates included the year of the decision, the polit ical orientation of the decisi on, the presence of lower court disagreement, the presence of precedence altera tion, declaration of unconstitutionality, and unanimous vote. The number of actions ranged from 0 to 11 ( M = 0.11, SD = .64); however, 3,882 (95.8%) of the 4,052 counts were zeros. This contributed to an exceptionally high skew of 64

PAGE 65

7.97. When ignoring the zeros, th e skew was reduced to 1.86 ( M = 2.59, SD = 1.53). The observed zeros were 107% of that which would be Poisson-expected. Regardless, Poisson model results were in lin e with theory-driven expectations. However, the test of overdispersion w as significant when comparing the Poisson and negative binomial Poisson resulting in fewer significan t predictors than if ignoring overdispersion. The author also fitted a generalized negative binomial model in which the variance parameter is allowed to vary as an exponential function of the same independent variables included in the model of the count (Zorn, 1996, p.9), which led to ev en better model fit. However, due to zero-inflation, no model provided reasonable estimate sizes given the low mean count. Their analyses using the ZIP and Hurdle m odels yielded several findings. First, the probability of remaining a zero in the transition stage was considerably lower for the Hurdle model than for the ZIP model at lower levels of a predictor. This is a reflection of the asymmetry of the Hurdle model. Second, parameter estimates and standard errors we re similar between the two models. They concluded that at least in some circumstan ces the performance of ZIP and hurdle Poisson models will be quite similar. Th is suggests that, as a practical matter and barring any strong theoretical considera tions favoring one over the other, the choice betw een them may be made largely on the basis of conven ience of estimation (Zorn, 1996, p.11). Pardoe and Durham (2003) compared the Poisson, Hurdle, and ZIP models as well as their negative binomial formula tions using wine sal es data. Of the 1,425 counts, 1,000 (70.2%) were zeros. The authors noted that this is greate r than the 67.8% that a Poisson distribution is capable of predicting. Based on the AIC, the zeroinflated negative binomial ZIP performed best. Surprisingly, the Hurdle model fit more poorly than did the Poisson model. It is possible, given the unrealistically high AIC relative to the other models, that the value wasnt calculated 65

PAGE 66

correctly. Alternatively, the di stribution may not have been co rrectly specified since their analysis included Bayesian estimat es of the prior distributions. No discussion pertaining to the Hurdle model was included. However, they did provide a novel procedure for comparing the range of fit statistics across the zero-inflation models. This par allel coordinate plot for goodness of fit measures consists of an x-axis labeled Min on the left and Max on the right. The y-axis is a series of horizontal lines each pe rtaining to a fit sta tistic (e.g., AIC, BIC). The ceiling x-axis contains the labels for the models being comp ared. Then, points for each model are plotted on the lines for the fit statistics at their relative location between Min and Max. A similar procedure restricted to the AIC was used by Warton (2005). This technique was adapted to display the coverage for simulated fit statistics. Welsh, Cunningham, Donnelly, and Lindenmayer (1996) used zero-inflated rare species count data to compare the Poiss on, negative binomial Poisson, Hu rdle, negative binomial Hurdle, and ZIP models. Approximately 66% of the observa tions were zeros. They found little difference between the Hurdle, negative binomial Hurdle, and ZIP model results. Since there was no overdispersion, the authors recommended using the Po isson model. This is in line with Wartons (2005) assertion that the more complex zero-inflation models may not always be necessary; at least, this appears to be the case with 66% zero-inflation and equidispersion. Discrepant Findings What is exactly meant by zero-inflation? Min and Agresti (2005) define zero-inflated count data as data for which a generalized linear model has lack of fit due to di sproportionately many zero (p.1). This raises the question, At what point does the frequency of zeros become disproportionate to the frequenc y of non-zeros? One statistical definition states that the proportion of zeros is greater than that to be expected given th e posited distribution (Zorn, 1996). For example, for count data, the proportion of zeros should not be greater than that expected by a 66

PAGE 67

Poisson distribution. However, there are three problems with this. First, there may be many different proportions of zeros greater than that expected by a pa rticular Poisson distribution. Second, the definition assumes that the full model is distributed Poisson. Th ird, it ignores the two potential sources of overdispersion for Poisson zero-inflated data. The aforementioned AEWR example displayed a zero proportion of .7158 with a mean of .54, a standard deviation of 1, and a skew of 1.971. Ignoring th e zeros, although this event stage distribution remains negative skewed, the mean in creased to 1.91, and th e level of skew dropped to 0.96. The distribution for th e categorical sex variable was bi nomial with approximately 43% males and 47% females. The distribution for the age variable was roughly normal with a mean of 48.86, a standard deviation of 17. 41, and a skew 0.87. Hence, with 71% zeros, a heavily skewed distribution of 1.971, a moderate ly skewed nonzero distribution of 0.96, a normally distributed continuous predictor, and a tw o-level categorical predictor led to the following findings: 1) the Hurdle model fit better than Poisson model; 2) the negative binomial Hurdle fit better than negative binomial Poisson model; 3) the negative binomial Hurdle fit better than the Hurdle model; 4) the negative binomial ZIP fit better than ZIP model; 5) the negative binomial ZIP model descriptively fit better than all others; and, 6) the Hurdle and negative binomial Hurdle model yielded nearly identical estimates and p -values. Hence, findings between the zeroinflation models differed in terms of both fit and the significance of parameter estimates. Although not all research presented sufficient in formation (e.g., data necessary to calculate skew), there is clearly enough variation in results to warrant fu rther research. Mullahys (1986) Hurdle model analyses were impacted by zero-infl ation of .263 and not by zero-inflation of .616 or .404; however, this is in disagreement with findings that the Hurdle model adequately handles zero-deflation (Min, 2003). Lamberts ZIP analys is with 71.8% zeros favored the ZIP over the 67

PAGE 68

negative binomial ZIP. Greene s (1994) ZIP analyses resulted in nonsensical results under .894 zeros and heavy skew (4.02); the negative binomial ZIP corrected this. Slymen, Ayala, Arredondo, and Elders ( 2006) ZIP and negative binomial ZIP r esults were virtually identical; however, their event stage distribution was uniform. This w as confirmed by Wartons (2005) finding that the negative binomia l fits better than the ZIP only when zero-inflation and overdispersion both are indeed present. Extending from this is Bhning, Dietz, Schlattmann, Mendona, and Kirchners (1999) findi ngs that the ZIP actually fit better than the Poisson given .216 and .289 zero-deflation. This is again in contrast to the s uggestion that the Hurdle, and not the ZIP, is appropriate for zer o-deflated data. However, it c ould be argued that a normal distribution should have been assu med given that the event stag e distribution was relatively normal. When comparing the Hurdle model to the ZI P model, Zorn (1996) found similar results given .958 zero-inflation, skew of 7.97, and a reduction of skew to 1.86 for the event stage. These findings are in contrast to Zorns (1996) findings of greater zero-inflation and, subsequently, greater skew. Welsh, Cunningham, Donnelly, and Lindenmayer (1997) also found little difference between the Hurdle and ZIP mode ls. Table 2-2 summarizes the findings from the zero-inflation literature. There are three factors that may have caused the anomalous findings. The first possibility pertains to the effect of di fferent types, quantities, and values for predictors. The second possibility is the proportion of zer os for the outcome variable. The third possibility is the degree of skew in the event stage for the outcome variable. It has already been suggested that the Hurdle and ZIP mode ls should be chosen given a priori research about the source a nd nature of the zeros. Further, it has been established that the 68

PAGE 69

negative binomial formulations are meant to handl e additional overdispersion in the event stage. However, the previous findings suggest that there are additional considerations such as the proportion of zeros and the nature of the event st age distribution. The propor tion of zeros in this research ranged from as low as .20 (Delucch i & Bostrom, 1994) to .958 (Zorn, 1996). Distributions for the event stage included those that were heavily positively skewed (Greene, 1994), moderately positively skewed (AEWR exam ple), distributed norm ally (Bhning, Dietz, Schlattmann, Mendona, & Kirchner, 1999), and distributed uniformly (Slymen, Ayala, Arredondo, & Elder, 2006). The first possibility, pertaining to the cova riates, can only be tested by varying an incredibly large set of conditions ranging from small to large quantities of predictors as well as their types (e.g., nominal, ordinal), and distributions. However, given a particular set of covariates and corresponding values the other two possibilities pertaining to zero-inflation and skew can be explored. It is po ssible to vary the proportion of zer os for the outcome variable and to simultaneously vary the degree of skew of the nonzeros for this outcome variable. Hence, the following research questions are presented: Given one two-level categorical covariate with known values and one continuous covariate with known values, what is the difference in the estimated log-lik elihood between a) the Negative binomial Poisson model vs. Poiss on model; b) the Hurdle model vs. Poisson model?; c) the Negative binomial Hurdle model vs. negative binomial Poisson model?; d) the Negative binomial Hurdle model vs. Hurdle model; and, e) the Negative binomial ZIP model vs. ZIP model? Given one two-level categorical covariate with known values and one continuous covariate with known values, what is the difference in the estimated AIC between all models? These questions were answered by establishing several levels of skew and zero-inflation. The covariates and their values were fixed as one continuous variable fr om a standard normal distribution and one binary variab le. Data for the outcome variable, for all levels of skew and 69

PAGE 70

zero-inflation, were simulate d to estimate log-likelihood va lues, AIC indices, covariate coefficients, standard errors, and p -values. The details are delineat ed in the following chapter. The objective is consistent with Zorns (1996) advice: First and fore most, work should be undertaken to better ascertain the st atistical properties of the various estimators outlined here. It is important that we determine the robustness of these techniques to skewness in the dependent variable, model misspecification, and the host of ot her problems that all too frequently plague political science researchers . perhaps using Monte Carlo methods to assess under what circumstances the results of the tw o may diverge (Zorn, 1996, p.12). 70

PAGE 71

Table 2-1. Five pairs of nested m odels valid for statistical comparison Valid Comparisons (Nested Models) 1 2 3 4 5 Poisson Simple Simple NB Poisson Complex Simple Hurdle Simple Complex NB Hurdle Complex Complex ZIP Simple NB ZIP Complex 71

PAGE 72

Table 2-2. Summary of lit erature on zero-inflation Researcher(s) Models Compared Zeros Superior Model Comments Min & Agresti (2004) Hurdle vs. ZIP Simulation Hurdle Zero-deflation Min & Agresti (2004) Hurdle vs. ZIP Simulation Equal Zero-inflation Mullahy (1986) Hurdle .26, .62, .41 Hurdle .26 Lambert (1992) ZIP vs. NB Poisson vs. Poisson .718 ZIP over NB Poisson over Poisson Greene (1994) ZIP vs. NB ZIP vs. NB Poisson .894 NB ZIP over ZIP; NB Poisson over ZIP over Poisson Heavy skew; Probit link Slymen, Ayala, Arredondo, and Elder (2006) Poisson vs. NB Poisson; ZIP vs. NB ZIP .824 NB Poisson; Equal Uniform event stage; Overall, AICs favor ZIP Xie, He, and Goh (2001) ZIP vs. Poisson .87 ZIP Outliers Bhning, Dietz, Schlattmann, Mendona, and Kirchners (1999) Poisson vs. ZIP .216, .289 ZIP Zero-deflation; normal event stage Zorn (1996) Hurdle vs. ZIP .958 Equal Heavy skew Pardoe and Durham (2003) Poisson vs. Hurdle vs. ZIP .702 NB ZIP over Poisson over Hurdle Based on AICs Warton (2005) ZIP vs. NB ZIP Vari ous ZIP or NB ZIP Overdispersion favored NB ZIP Warton (2005) ZIP vs. NB ZIP Very low ZIP Rare convergence for NB ZIP Warton (2005) NB ZIP vs. OLS vs. NB ZIP vs. Poisson Various NB ZIP over OLS/NB ZIP over Poisson Based on AICs Welsh, Cunningham, Donnelly, and Lindenmayer (1997) Hurdle vs. ZIP .66 Equal No overdispersion 72

PAGE 73

CHAPTER 3 METHODOLOGY The previous chapter fulfilled three objectives. First, it described the statistical models and methods for analyzing count data including that which is zeroinflated. Second, it presented research, both technical and app lied, pertaining to three models (Poisson, Hurdle, and ZIP) as well as their negative binomial form ulations. Third, it was concluded that there is considerable divergence in findings between models and th at such differences s hould be explored by examining different levels of zero-inflation and sk ew for the count outcome variable. This led to the following research questions: Research Questions Given one two-level categorical covariate with known values and one continuous covariate with known values, what is the difference in the estimated log-lik elihood between a) the Negative binomial Poisson model vs. Poiss on model; b) the Hurdle model vs. Poisson model?; c) the Negative binomial Hurdle model vs. negative binomial Poisson model?; d) the Negative binomial Hurdle model vs. Hurdle model; and, e) the Negative binomial ZIP model vs. ZIP model? Given one two-level categorical covariate with known values and one continuous covariate with known values, what is the difference in the estimated AIC between all models? As recommended by Zorn (1996), these questio ns were answered using a Monte Carlo study in which the proportion of ze ros and the skew for the distribution of the event stage counts varied between simulations. Monte Carlo Study Design Monte Carlo studies begin in the same manner as other research methods. First, a problem is identified as a re search question. Second, the problem is made concrete in the form of a hypothesis or set of hypotheses. Third, the hypotheses are tested using rigorous methods. Fourth, conclusions are drawn from these results. Fifth, the implications and limitations are elucidated for future researchers and practitioners. 73

PAGE 74

For this study, the problem was zero-inflation for count out comes. The problem was then clarified in the form of the rese arch questions stated above. It is at this point that Monte Carlo studies differ from most other me thods. For the Monte Carlo study, no real data is gathered. Rather, a set of samples are generated based on given parameter specifications resulting in a sampling distribution that is consid ered to be equivalent to that which would have been obtained had this many real participants been available. Th is is a clear advantage ov er the true experiment where statistics are used to in fer from one sample to an en tire population (Mooney, 1997). The Monte Carlo study is not limited to a finite number of participants and is subsequently not prone to violations of asymptotic th eory (Paxton, Curran, Bollen, Kirby, & Chen, 2001). Paxton, Curran, Bollen, Kirby, and Chen (2001, p.287) provide a succinct expl anation as follows: The researcher begins by cr eating a model with known population parameters (i.e., the values are set by the research er). The analyst then draws repeated samples of size N from that population and, for each sample, estimates th e parameters of interest. Next, a sampling distribution is estimated for each populat ion parameter by collecting the parameter estimates from all the samples. The properties of that sampling distribution, such as its mean or variance, come from this estimated sampling distribution. Similarly, Mooney (1997, p.2) explains, Monte Carlo simulation offers an alternative to analytical mathematics for understanding a statistics sampling distributi on and evaluating its behavior in random samples. Monte Carlo simulation does this empirically using random samples from known populations of simulated data to track a statistics be havior. The basic concept is straightforward: If a statistics sampling distribution is the densit y function of the values it could take on in a given population, then its estimate is the rela tive frequency distribution of the values of that statistic that were actually observed in many samples drawn from that population. The Monte Carlo simulations were performed using the R programming language (R Development Core Team, 2006). R is an open-source language based on the commercially 74

PAGE 75

available S-Plus program (Insightful, 2005) and is just one of many programs that can be used to model zero-inflated data.23 Monte Carlo Sampling Pseudo-Population Mooney (1997) explains that defining th e population parameters requires defining the pseudo-population. In the pseudo-population, the values for a cat egorical factor, X1, with two levels were constant at either 0 or 1; this is representative of a typical two-level categorical factor such as sex. In the dataset, these values alternat ed. The Excel 2003 random number generator was used to draw a random selection of 1,000 normally di stributed values to represent the continuous variable, X2 ~ N (0,1). This resulted in a pseudo-population of N = 1,000 where the values for categorical X1 and continuous X2 were known. The values for X2 ranged from 2.945 to 3.28 with a mean of 0, a median of 0.05, a standard deviation of 0.986, and skew of 0.127. These two sets of covariates and their distributions were chosen as a parsimonious generalization to the basic ANCOVA general linear model that exte nds from analyses with either quantitative or qual itative predictors.24 The simulations varied in terms of a) the amount of zeroinflation present in the outcome variable scores; b) the amount of skew present in the event stage outcome variable scores, and c) the generalized linear model. The outcome variable, Y was established as a deterministic variable in that it varied systematically as a function of the specified distributions. As clar ified by Mooney (1997), Deterministic variables are vect ors of numbers that take on a ra nge of values in a prespecified, nonrandom manner (p.6). The regression coefficients for X1 and X2 are random variables that 23 Others include Stata, LIMDEP, COUNT, MATLAB, and SAS. Preliminary analyses compared results between SAS-code incorporated in research (Min & Agresti, 2004; Min, 2003) to publicly available R-code written by Simon Jackman of the Stanford Political Science Computing Laboratory to verify the comparability of results. 24 The model is similar to that of the AEWR examples. 75

PAGE 76

take on their realizations as a result of the relationship between deterministic Y and the two random X covariates. The Prespecified Zero Proportions Justification for generating values with prespecified proportions of event counts with frequencies also determined by prespecified proportions of zero s is justified by Mooney (1997). If we know (or are willing to make assumptions about) the components that make up a statistic, then we can simulate these components, calculate the statistic, and explore the behavior of the resulting estimates (Mooney, 1997, p.67). In his case, the concern was bias determined by calculating statistics and inspecti ng graphs as a result of Monte Ca rlo simulations. For this study, the concern was goodness-of-fit for six models by comparing log-likeli hoods and AICs and inspecting graphs as a result of M onte Carlo simulations over three le vels of skew and five levels of zero proportions. Previous research displayed zero-inflation ra nging from .20 (Mullahy, 1986) to .96 (Zorn, 1996). To reflect a range including both zero-deflation and zero-i nflation, six pseudo-populations were established differing in the proportion of zeros present in the count outcome variable. The pseudo-populations contained either 0.10, 0.25, 0.50, 0.75, or 0.90 proportions of zeros.25 Pre-Specified Skew To manipulate skew, the even t stage distributions for the count outcome variable were varied over three conditions. For each condition, proportions were specif ied and values were drawn randomly from a multinomial di stribution such that the freque ncies of the values added to the frequency for those already drawn to represe nt zero-inflation summed to 1,000. In other 25 Originally, a 0.00 proportion of zeros was included as a control condition. However, for the case of negative skew, this is simply a count model truncated at one. And, for a normal distribution event-stage, this is semicontinuous data often tested with different methods from those for zero-inflation (Min, 2002). 76

PAGE 77

words, if .50 (or 500) observations were assigned a value of zero then the remaining .50 (or 500) observations had values drawn from the prespecified multinomial distribution. Event stage values ranging from one to five were sampled in order to represent a range small enough to distinguish the distribution from one that might be analyzed as continuous given a particular shape. The prespecif ied probabilities for each of the five counts were determined primarily to achieve a particular level of skew and secondarily in order to achieve a convergent and admissible solution. Hence, th e proportions were not always exactly equal to .10, .25, .50, .75, and .90; the approximates were selected to achieve convergence leading to more trust in convergence for the Monte Carlo simulations. Table 3-1 displays the proportions of each count as a function of the three levels of skew and five levels of zer os. Table 3-2 displays this same information in terms of frequenc ies instead of proportions. Table 3-3 displays the descriptive statistics for each distribution. Random Number Generation By definition, a random number is one in which there is no way possible to a priori determine its value. Most statistical analys is software packages include random number generators. However, these generated random numbers are not truly random. Usually, one specifies a seed; when re plications are performed using the same seed, the generated numbers are identical to the first. Hence, the values are pseudo-random (Bonate, 2001). However, this limitation is actually an advantage in that th e researcher can check for errors in model programming and run the anal ysis again with the same generated sample (Mooney, 1997).26 26 Technically, this is only true for most random number generators. The R programming language, by default, bases its generation on the Wichman-Hill algorithm and the system clock resulting in a 626-integer seed (R Development Core Team, 2006). 77

PAGE 78

Another feature of Monte Carl o random sampling pertains to the desired distributions. Typically, the random numbers are drawn from a uniform distributi on, which is then followed by a transformation to the desired di stribution. As Mooney (1997) e xplains, In its standard form, U(0, 1), the uniform distribution is the building block of all Monte Carlo simulation work in that from it, in one way or another, variables with all other distributi on functions are derived. This is because the U(0, 1) distribution with its 0 x 1 range, can be used to simulate a set of random probabilities, which are used to generate ot her distribution functi ons through the inverse transformation and acceptance-re jection methods (p.10). The random number generation was performed using R 2.3.1 (R Development Core Team, 2006). The procedure requires the generic sample command in which the following were specified: 1) a range of counts, which in this study, was from one to five (not zero to five since proportions of zeros were alrea dy drawn from the pseudo-populati on), 2) the number of values to draw, which in this study was one minus the pres pecified proportion of ze ros, 3) the proportions for each value in the range, which in this case was one of three possibi lities determining skew, and 4) the specification to sample with replacem ent. The seed was arbitrarily set at 6770. Sample Size Determining the appropriate sample size for each simulate is an important concern. This could range from zero to infinity. However, if the sample size is too small then it is not safe to assume that estimates are asymptotically normal. On the other hand, computer time and burden increases as sample size increases. The sample size was based on the highest found in the literature pertaining to zero-i nflated count data, which was n = 1,000 (Civentti & Hines, 2005). Simulation Size Determining the number of simulations is also an important concern since too few replications may result in inaccurate estimat es and too many replications may unnecessarily 78

PAGE 79

overburden computer time and performance (Bonate 2001). Hurs (1999) r esearch pertaining to the ZIP model with random eff ects set the number of simulations at 200. Min and Agresti (2005) were able to sufficiently compare the goodness of fit for several competing models using 1,000 simulations. Likewise, Civettini and Hines ( 2005) selected 1,000 simulations when researching misspecification in negative binom ial ZIP models. Lambert (1989) set the number of simulations at 2,000 when researching the asympto tic properties of the ZIP model. Mooney (1997) states that The best practical advice on how many trials are needed for a given experiment is lots! Mo st simulations published recently report upward from 1,000 trials, and simulations of 10,000 and 25,000 trials are common (p.58). Given the previously noted problems with convergence for the negative binomi al ZIP model, it seem s prudent to minimize the number of simulations as much as possible. Howe ver, it is also important to simulate under conditions already found to produce asymptotic results. Hence, similar to Lamberts (1989) seminal study and equal to the maximum found in the literature, the number of simulations was set at 2,000 for each condition ( S = 2,000). Iteration Size Iteration size is not much of a concern for the basic Poisson and negative binomial Poisson model. However, obtaining valid estimates for the Hurdle model, ZIP model, and their negative binomial counterparts re quires selecting an appropriate maximum number of iterations. Too few iterations can lead to incorrect estimat es or, even worse, premature declaration of nonconvergence. Too many iterations results in u nnecessary computer time and burden. Various procedures for analyzing these models in R have maximum iterations of 500, 5,000, and 50,000. It was important to determine an iteration size that would be equal ac ross analyses and large enough given some models potential for nonconve rgence. These concerns were deemed more 79

PAGE 80

important than excessive computational burden. He nce, the procedure with the largest iteration size was selected for all procedures leading to 50,000 iterations per analysis. Distribution Generation The following describes the procedure fo r generating the distributions for each simulation. The total proportion of counts out of 1,000 to be sampled was reduced by the prespecified proportion of zeros. Hence, if the proportion of zeros was .50 then the proportion of event stage counts was 1.00-0.50 = .50. Translated into frequencies, this is 1,000-(0.50 1,000) = 500 The generic R sample procedure was used to sample with replacement from the event stage counts according the specified proportions de pending on the skew condition. The seed was set at 6770. The values were sampled over N = 1,000. Each sample was simulated S = 2,000 times. The data over all S = 2,000 at N = 1,000 were then stored in separate files as they were created. The filenames conformed to the labelin g format where the model was replaced by an underscore (e.g., _25Pos w as the filename for the S = 1,000 datasets at N =2,000 where the proportion of zeros was .25 and the sk ew for the event stage was positive). Monte Carlo Models As previously discussed, generalized lin ear models include a random component, a systematic component, and a link function. The X1 and X2 constants form the systematic component in the pseudo-populations genera lized linear model. The random component specification for the distribution of the outcome mean varied from pseudo-population to pseudopopulation. The base level genera lized linear model assuming a normally distributed outcome is given by 01122()()iiYXXi i (3-1) 80

PAGE 81

Subsequent models extended from this base model to form the six distributions for deterministic Y The first model, which was the Poisson generalized linear model with a log link, is given by 01122log()()()iiXXi (3-2) Table 3-4 displays the parameters for this model over all conditions of skew and zeroinflation. For example, the analysis with a .10 proportion of zeros and a positively skewed event stage distribution yielded 1log().450.004().037()iX2X (3-3) For both predictors, the coeffici ent near zero transforms to an exponentiation near one. The coefficient for X1 is lowest (.001) for the negatively distributed data with a .25 proportion of zeros and highest (.007) for the pos itively distributed data with a .50 proportion of zeros. The coefficient for X2 is lowest (-.007) for the negatively di stributed data with a .10 proportion of zeros and highest (-.165) for the normally distribute d data with a .90 proportion of zeros. Hence, for the two-level categorical X1 variable, changing from zero to one multiplies the mean of the outcome variable by approximately exp(0.00), whic h equals one. For the simulated data, the test for the coefficient estimates co rresponding to these pseudo-p opulation parameter values is approximately H0: = 0. The second model, which was the negative binomial formulation of the Poisson model, is the same as the Poisson model with the addition of a dispersion parameter. Table 3-5 displays the parameters for this model over all conditions of sk ew and zero-inflation. Like the Poisson model, the test for the simulated data is H0: = 0. The dispersion parameter is also included in Table 35. For the simulated data, the test that this pa rameter equals zero has equivalent results to the model comparison tests conducted. 81

PAGE 82

The third and fifth models were the general formulations fo r the Hurdle and ZIP models given by logit 01122()log()()() 1i i ipi i p ZZ p (3-4) 45162log()()()iiXXi (3-5) while the fourth and six models were their negative binomial fo rmulations. Tables 3-6 through 313 display the pseudo-population coefficients, standard errors log-likelihood values, and AIC values over all conditions of proportions of zero and skew. A lthough there are four models (i.e., Hurdle, negative binomial Hurdle, ZIP, and negative binomial ZIP), there are eight tables since, for each model, there are separat e results for the transition (zer os) stage and events (nonzero counts) stage. Monte Carlo Analysis Procedures A generic looping procedure was written in the R programming language to retrieve each of the 15 sets of simulated data. Each dataset was analyzed with each of the six models. The Poisson model used the generic glm procedure in R, which requires specification of the model, Poisson distribution, and l og link. The negative binomial Poisson model used the glm.nb procedure in R from the generic MASS library. This procedure requires only the specification of the model. The R Core Development Team (2006) describes the procedure as follows: An alternating iteration process is used For given 'theta' [dispersion] the GLM is fitted using the same process as used by 'glm()'. For fixed means th e 'theta' parameter is estimated using score and information iterations. The two are alternat ed until convergence of both is achieved (R Core Development Team, 2006). The Hurdle model used the hurdle procedure in the pscl library authored by Simon Jackman, PhD of the Political Science Computi ng Laboratory at Stanford University. The 82

PAGE 83

procedure requires specification of the count response variable (Y), the Poisson distribution for the event stage, the logit link function for the transition stage, and the models for both the transition stage and event stage, which for this study were X1 and X2. The negative binomial Hurdle model also used the hurdle procedure bu t specified a negative binomial rather than a Poisson distribution for the event stage. Th is hurdle model procedure maximizes the loglikelihood using either the Br oyden-Fletcher-Goldfarb-Shanno (BFGS) or Nelder-Mead methods. The Nelder-Mead (default) w as selected for solution optimization. The ZIP model used the zicounts procedure in the zicounts library authored by Samuel Mwalili, doctoral student in biostatistics at Katholieke Universeteit Leuven (Netherlands). Similar to the hurdle procedure, zicounts requires specification of th e count outcome variable and models for both the transition st age and event stage. The distri bution is specified as ZIP. Optimization procedures include the BFGS, Nelder-Mead, and conj oint gradient (CG) methods; for consistency, the Nelder-Mead was chosen. The negative binomial ZIP used the same procedure with the specification of the ZINB distribution. The specific procedure for analyzing the data with the models was as follows: Three separate loops were established for the negatively skewed, normal, and positively skewed distributions. Arrays were created to store the log-likelihood, AIC, coeffici ent estimates, standard errors, and p -values. The outermost loop for each dist ribution pertained to the five conditions of varying zero proportions; at each loop, the data correspondi ng to this zero-conditi on and distribution was loaded. Within these loops, another loop ing procedure pertained to the six models that analyzed the data. Within these loops another looping procedure was defined for the number of simulations in the data set. 83

PAGE 84

It was at this point that th e data corresponding to a particular distribution, proportion of zeros, model, and simulation were analyzed with the calculated AIC, log-likelihood, coefficient estimates, standard errors, and p -values transferred to the appropriate array. Models that failed to converge were automatically coded as NA.27 The results for each statistic over all simulati ons for a particular distribution, model, and proportion of zeros were then exported in comma -separated format for subsequent analyses. Hence, the three distributions by five proportion of zeros conditions by six models yielded 90 data files each containing column s pertaining to a particular stat istic or set of statistics and 2,000 rows of simulated results. Analysis Design A series of tests were conduc ted using the simulated data. The results and graphical output were created using R 2. 01 (R Development Core Team, 2006) and SPSS 14.0. The design for a particular analysis depended on the res earch question. These questions were as follows: Given one two-level categorical covariate with known values and one continuous covariate with known values, what is the difference in the estimated log-lik elihood between a) the Negative binomial Poisson model vs. Poiss on model; b) the Hurdle model vs. Poisson model?; c) the Negative binomial Hurdle model vs. negative binomial Poisson model?; d) the Negative binomial Hurdle model vs. Hurdle model; and, e) the Negative binomial ZIP model vs. ZIP model? Given one two-level categorical covariate with known values and one continuous covariate with known values, what is the difference in the estimated AIC between all models? The first question was answered by calculating log-likelihood values for the six models over the six zero proportion conditions and the thr ee skew conditions. The deviance statistic was then calculated as -2(LLS-LLC) where LLS is the model with less parameters (i.e., the simple model) than the LLC (i.e., the more complex model). Since there were 5 conditions for the proportions of zeros and 3 conditi ons for skew, this led to a) a total of 15 separate analyses comparing the fit of the negative binomial Pois son model and the Poisso n model; b) 15 separate analyses comparing the fit of th e Poisson model and the Hurdle model; c) 15 separate analyses comparing the fit of the negative binomial Po isson model and the negative binomial Hurdle 27 This number was chosen to reduce the probability of obtaining a valid result that would be identified as a nonconvergant solution 84

PAGE 85

model; and d) 15 separate analyses comparing the fit of ZIP model and the negative binomial ZIP model. Test statistics for the deviance were assume d asymptotically chi-square with degrees of freedom equal to the difference in the number of parameters between the two models. Some models differed only by the dispersion parameter; these were a) the comparison of the Poisson model and negative binomial Po isson model; b) the comparison of the Hurdle model and negative binomial Hurdle model; and c) the comp arison of the ZIP model and negative binomial ZIP model. The log-likelihood statistic for th e Hurdle model is based on the log-likelihood statistics from each of its two parts. Given thre e parameters in the Poisson model (i.e., the intercept and the two predictor coefficients), th ere are six parameters in the Hurdle model. Hence, the degrees of freedom for the model compar ison test are the difference, which are three. The same is true for comparing the negative bino mial Poisson model to the negative binomial Hurdle model where including the dispersion parame ter leads to subtracting 4 parameters from 7 parameters. Each of the 2,000 goodness-of-fit statistics for the simpler mode l was subtracted from each of the 2,000 goodness-of-fit statisti cs for the more complex model. These results were then multiplied by -2. Each of these values was then co mpared to the chi-square distribution with the appropriate degrees of fr eedom. This yielded a p -value representing the probability of obtaining a statistic this high or higher given that the simpler model adequately fits the data. Values exceeding this critical chisquare statistic based on a Type I error rate of = .05 were coded with these results suggesting that the more complex model fits th e data better than the Poisson model. Values failing to exceed the critical valu e were coded with these results suggesting adequate fit for the simpler model. 85

PAGE 86

The results were thus the pr oportion of simulated datasets favoring the more complex model over the simpler model for a particular proportion of zeros and le vel of skew. Output included 1.) descriptive statisti cs for the goodness-of-fit for each model and 2.) boxplots for the difference in goodness-of-fit between the two models. Answering the second question was done in a similar fashion to answering the first question. This is due to the fact that the AI C is a linear transforma tion of the log-likelihood statistic with a result that is positive in sign and is interpreted in a lower-is-better fashion. However, these analyses did not involve compar isons to the chi-square distribution. As previously explained, th e AIC should not be used in this manner. However, the advantage is that the AIC can be used to descriptively compare all models regardless of whether one is nested or not within another. Boxplots were created to display the range of AICs produced over simulations for each analysis. 86

PAGE 87

Table 3-1. Proportions of counts as a function of zeros and skew Study Proportion of Zeros Proportion of Remaining Nonzero Values Ones Twos Threes Fours Fives _Pos10 0.10 0.504 0.227 0.088 0.054 0.027 _Norm10 0.099 0.198 0.306 0.198 0.099 _Neg10 0.027 0.054 0.091 0.227 0.501 _Pos25 0.25 0.418 0.190 0.075 0.046 0.021 _Norm25 0.083 0.166 0.254 0.166 0.081 _Neg25 0.022 0.045 0.075 0.188 0.420 _Pos50 .50 0.280 0.125 0.050 0.030 0.015 _Norm50 0.053 0.107 0.175 0.110 0.055 _Neg50 0.015 0.030 0.050 0.125 0.280 _Pos75 0.75 0.140 0.062 0.025 0.015 0.008 _Norm75 0.028 0.052 0.089 0.053 0.028 _Neg75 0.008 0.015 0.025 0.062 0.140 _Pos90 0.90 0.056 0.025 0.010 0.006 0.003 _Norm90 0.011 0.021 0.035 0.024 0.009 _Neg90 0.004 0.005 0.010 0.025 0.056 Table 3-2. Frequencies of count s as a function of zeros and skew Study Frequency of Zeros Frequency of Individual Nonzero Values Ones Twos Threes Fours Fives _Pos10 100 504 227 88 54 27 _Norm10 99 198 306 198 99 _Neg10 27 54 91 227 501 _Pos25 250 418 190 75 46 21 _Norm25 83 166 254 166 81 _Neg25 22 45 75 188 420 _Pos50 500 280 125 50 30 15 _Norm50 53 107 175 110 55 _Neg50 15 30 50 125 280 _Pos75 750 140 62 25 15 8 _Norm75 28 52 89 53 28 _Neg75 8 15 25 62 140 _Pos90 900 56 25 10 6 3 _Norm90 11 21 35 24 9 _Neg90 4 5 10 25 56 87

PAGE 88

Table 3-3. Descriptive statistics for each distribution Range = 0 to 5 Range = 1 to 5 Mean Std.Dev. Skew Mean Std.Dev. Skew _Pos10 1.570 1.127 1.155 1.750 1.051 1.45 _Norm10 2.700 1.414 -0.310 3.000 1.150 0.08 _Neg10 3.820 1.619 -1.356 4.250 1.053 -1.43 _Pos25 1.310 1.181 1.092 1.750 1.046 1.42 _Norm25 2.250 1.635 -0.069 2.990 1.149 0.08 _Neg25 3.190 2.054 -0.673 4.840 0.367 -1.86 _Pos50 0.880 1.149 1.514 1.750 1.053 1.44 _Norm50 1.510 1.710 0.587 3.010 1.140 -0.01 _Neg50 2.130 2.253 0.236 4.250 1.053 -1.44 _Pos75 0.440 0.928 2.577 1.760 1.064 1.44 _Norm75 0.750 1.422 1.660 3.000 1.149 -0.00 _Neg75 1.060 1.914 1.349 4.240 1.064 -1.44 _Pos90 0.180 0.622 4.490 1.750 1.058 1.46 _Norm90 0.300 0.965 3.268 2.990 1.124 -0.06 _Neg90 0.420 1.318 2.908 4.240 1.084 -1.51 Table 3-4. Poisson model: pseudo-population parameters 0 0s 1 1s 2 2s LL AIC _Pos10 .450 .036 .004 .050 -.037 .026 -1475.3 2956.6 _Norm10 .992 .027 .003 .038 -.015 .020 -1806.7 3619.3 _Neg10 1.340 .023 .002 .032 -.007 .016 -2019.2 4044.5 _Pos25 0.270 .039 .003 .055 -.024 .028 -1468.1 2942.3 _Norm25 .807 .030 .004 .042 -.016 .021 -1913.9 3833.8 _Neg25 1.159 .025 .001 .035 -.009 .018 -2265.35 4534.7 _Pos50 -.138 .048 .007 .068 -.047 .034 -1334.3 2674.5 _Norm50 .407 .036 .004 .052 -.038 .026 -1888.5 3783.0 _Neg50 .752 .031 .003 .043 -.030 .022 -2370.5 4746.9 _Pos75 -.829 .068 .005 .095 -.077 .048 -972.9 1951.7 _Norm75 -.291 .052 .003 .073 -.079 .037 -1461.8 2929.6 _Neg75 .057 .043 .002 .061 -.044 .031 -1919.7 3845.3 _Pos90 _Norm90 -1.77 -1.23 .109 .083 .036 022 .151 116 -.158 -.165 .076 058 -546.5 -852.6 1099.0 1711.1 _Neg90 -.875 .069 .011 .097 -.149 .049 -1152.1 2310.2 88

PAGE 89

Table 3-5. Negative Binomial Poisson model: pseudo-population parameters 0 0s 1 1s 2 2s LL AIC _Pos10 .450 .032 .004 .045 -.037 .023 .45 -1594.5 _Norm10 .991 .023 .003 .033 -.015 .017 .32 -1998.9 _Neg10 1.340 .019 .002 .027 -.007 .014 .24 -2277.9 _Pos25 .269 .040 .004 .057 -.025 .029 .64 -1509.4 _Norm25 .807 .033 .004 .046 -.016 .023 .56 -1931.5 _Neg25 1.159 .029 .001 .041 -.009 .021 .51 -2235.5 _Pos50 -.139 .059 .008 .083 -.048 .042 1.05 -1290.1 _Norm50 .041 .051 .005 .072 -.039 .036 1.11 -1723.0 _Neg50 .751 .048 .004 .067 -.030 .034 1.16 -2041.3 _Pos75 -.829 .095 .005 .134 -.078 .068 1.61 -910.4 _Norm75 -.292 .085 .004 .120 -.078 .061 1.95 -1275.6 _Neg75 .057 .081 .003 .114 -.044 .058 2.26 -1567.4 _Pos90 orm90 -1.77 -1.23 .161 .146 .037 020 .225 204 -.159 -.165 .113 103 2.04 2.70 -511.6 -751.1 _N _Neg90 -.874 .140 .008 .197 -.149 .010 3.38 -957.3 6. Hurdlezer eudpulaaras Table 3model ( os): ps o-po tion p meter 0 0s 1 1s 2 2s LL AIC _Pos10 -2.198 .149 .003 .211 .044 .107 -1409.5 2830.9 _Norm10 -2.198 .149 .001 .210 .044 .107 -1797.1 3606.1 _Neg10 -2.198 .149 .001 .211 .044 .107 -1928.3 3868.7 _Pos25 -1.099 .103 .000 .146 -.019 .074 -1465.5 2943.0 _Norm25 -1.098 .103 -.001 .146 -.020 .074 -1787.6 3587.2 _Neg25 -1.099 .103 .000 .146 -.020 .074 -1897.9 3807.9 _Pos50 .000 .089 .000 .127 .037 .064 -1296.6 2605.1 _Norm50 -.005 .089 .005 .127 .035 .064 -1510.1 3023.2 _Neg50 .000 .089 .000 .127 .036 .064 -1583.9 3179.7 _Pos75 1.099 .103 .002 .146 .046 .074 -865.2 1742.4 _Norm75 1.100 .103 .000 .146 .047 .074 -970.3 1952.6 _Neg75 1.099 .103 .000 .146 .045 .074 -1008.1 2028.3 _Pos90 2.205 .150 -.001 .211 .137 .107 -444.9 901.8 _Norm90 2.206 .150 -.001 .211 .137 .107 -486.4 984.8 _Neg90 2.204 .150 -.004 .211 .138 .106 -503.0 1018.1 89

PAGE 90

Table 3-7. Hurdle model (event s): pseudo-population parameters 0 0s 1 1s 2 2s LL AIC _Pos10 .212 .051 .008 .072 -.063 .036 -1409.5 2830.9 _Norm10 1.035 .030 .003 .042 -.012 .021 -1797.1 3606.1 _Neg10 1.430 .024 .002 .033 -.003 .017 -1928.3 3868.7 _Pos25 .214 .056 .008 .078 -.062 .041 -1465.5 2943 _Norm25 1.033 .033 .005 .047 -.027 .024 -1787.6 3587.2 _Neg25 1.431 .026 .001 .037 -.016 .019 -1897.9 3807.9 _Pos50 .208 .068 .018 .096 -.060 .049 -1296.6 2605.1 _Norm50 1.035 .040 .011 .057 -.023 .029 -1510.1 3023.2 _Neg50 1.430 .032 .004 .045 -.012 .023 -1583.9 3179.7 _Pos75 .213 .096 .010 .135 -.087 .067 -865.2 1742.4 _Norm75 1.032 .057 .005 .081 -.054 .040 -970.3 1952.6 _Neg75 1.428 .045 .002 .064 -.011 .032 -1008.2 2028.3 _Pos90 .179 .156 .060 .215 -.070 .107 -444.9 901.8 _N _Neg90 1.420 .071 .009 .101 -.028 .051 -503.0 orm90 1.106 .091 .018 .128 -.052 .064 -486.4 984.8 1018.1 Table 3-8. Neinomrdle l (zerseudoulatiom gative B ial Hu m ode os): p -pop n para eters 0 0s 1 1s 2 2s LL AIC _Pos10 -2.197 .149 -.001 .211 .044 .107 -1395.7 2803.4 _Norm10 -2.197 .149 .000 .211 .044 .107 -1800.2 3612.3 _Neg10 -2.205 .150 -.001 .212 .047 .108 -1936.1 3884.1 _Pos25 -1.099 .103 .001 .146 -.019 .074 -1455.3 2922.5 _Norm25 -1.099 .103 .000 .146 -.019 .074 -1790.2 3592.3 _Neg25 -1.099 .103 .000 .146 -.019 .074 -1904.4 3820.8 _Pos50 .000 .089 .000 .127 .036 .064 -1288.9 2589.8 _Norm50 .000 .089 -.001 .127 .037 .064 -1511.9 3035.8 _Neg50 .001 .089 -.001 .127 .036 .064 -1588.1 3188.2 _Pos75 1.099 .103 .000 .146 .045 .074 -861.1 1734.3 _Norm75 _Neg75 1.099 1.099 .103 .103 .000 .000 .146 .146 .045 .045 .074 .074 -971.2 -1010.3 1954.4 2032.5 _Pos90 2.206 .150 -.001 .211 .137 .107 -443.3 898.7 _Norm90 2.244 .152 -.029 .213 .181 .107 -487.3 986.6 _Neg90 2.205 .150 -001 .211 .138 .107 -503.9 1019.8 90

PAGE 91

Table 3-9. Negative Binomial Hurdle m odel (events): pseudo-population parameters 0s1s2s0 1 2 Theta LL AIC _Pos10 -.067 .108 .009 .091 -.071 .046 2.039 -1395.7 2803.4 _Norm10 1.033 .030 .004 .043 -.012 .022 165.718 -1800.2 3612.3 _Neg10 1.429 .024 .002 .034 -.002 .017 164.970 -1936.1 3884.1 _Pos25 -.042 .114 .009 .098 -.067 .051 2.234 -1455.3 2922.5 _Norm25 1.030 .033 .005 .047 -.027 .025 166.348 -1790.2 3592.3 _Neg25 1.431 .026 .001 .037 -.015 .019 166.006 -1904.4 3820.8 _Pos50 -.072 .146 .020 .122 -.065 .062 2.030 -1288.9 2589.8 _Norm50 1.037 .041 .006 .057 -.025 .029 166.574 -1511.9 3035.8 _Neg50 1.429 .032 .004 .046 -.012 .023 166.526 -1588.1 3188.2 _Pos75 -.075 .208 .012 .173 -.097 .088 1.950 -861.1 1734.3 _Norm75 1.031 .058 .005 .081 -.054 .041 166.445 -971.2 1954.4 _Neg75 1.427 .046 .002 .064 -.011 .032 166.377 -1010.3 2032.5 _Pos90 orm90 -.109 1.051 .332 091 .078 -.051 .273 130 -.081 -.048 .140 .066 2.047 91.979 -443.3 -487.3 898.7 986.6 _N _Neg90 1.421 .072 .007 .102 -.028 .052 166.606 -503.9 1019.8 Table 3-10. Zl ulaet IP mode (zeros): pse do-popu tion param ers 0 0s1s2s 1 2 LL AIC _Pos10 -27.010 4434.000 11.490 44 1 35.000 -.363 50.700 -1478.3 2968.6 _Norm10 -3.114 .038 .002 .527 .050 .253 -1800.1 3612.1 _Neg10 -2.365 .175 .002 .246 .049 .125 -1931.3 3874.6 _Pos25 -15.460 38 _Pos90 1.797 .183 .032 .255 .104 .129 -447.9 907.7 _Norm90 2.133 .152 .002 .214 .128 .108 -489.4 990.8 _Neg90 2.188 .071 .007 .101 -.028 .051 -506.0 1024.1 34.000 -3.418 .002 1.213 .001 -1471.1 2954.3 _Norm25 -1.379 .132 .008 .186 -.056 .098 -1790.6 3593.1 _Neg25 -1.162 .109 .001 .154 -.025 .078 -1900.9 3813.9 _Pos50 -.890 .198 .051 .270 -.050 .141 -1299.5 2611.1 _Norm50 -.126 .097 .002 .137 .030 .070 -1513.1 3038.3 _Neg50 -.031 .091 .000 .129 .035 .065 -1586.9 3185.7 _Pos75 .610 .142 .010 .198 -.013 .010 -868.2 1748.5 _Norm75 1.032 .057 .005 .081 -.054 .040 -973.3 1958.5 _Neg75 1.079 .110 -.001 .147 .044 .075 -1011.2 2034.3 91

PAGE 92

Table 3-11. ZIP Model (event s): pseudo-population parameters 0s 1s 2s0 1 2 LL AIC _Pos10 .450 .004 .004 .050 -.037 .003 -1478.3 2968.6 _Norm10 1.035 .003 .004 .043 -.013 .002 -1800.1 3612.1 _Neg10 1.430 .024 .002 .033 -.003 .002 -1931.3 3874.6 _Pos25 .270 .040 .003 .055 -.024 .028 -1471.1 2954.3 _Norm25 1.032 .033 .005 .047 -.028 .025 -1790.6 3593.1 _Neg25 1.431 .026 .001 .037 -.016 .019 -1900.9 3813.9 _Pos50 6.0 1024.1 .205 .069 .024 .096 -.063 .050 -1299.5 2611.1 _Norm50 1.039 .040 .007 .057 -.025 .029 -1513.1 3038.3 _Neg50 1.429 .032 .004 .045 -.012 .023 -1586.9 3185.7 _Pos75 .213 .096 .001 .135 -.085 .067 -868.2 1748.5 _Norm75 1.015 .107 .001 .151 .033 .076 -973.3 1958.5 _Neg75 1.428 .045 .002 .064 -.011 .032 -1011.2 2034.3 _Pos90 orm90 .180 1.017 .155 091 .057 .017 .216 128 -.069 -.052 .108 065 -447.9 -489.4 907.7 990.8 _N _Neg90 1.422 .071 .007 .101 -.028 .051 -50 Table 3-12. Ni P ze do-atet egative B nomi al ZI model ( ros): p seu popul ion param ers LL AIC 0s 1s 2s 0 1 2 _Pos10 -15.280 142.300 -3.880 830.400 .719 102.100 -1479.3 2972.6 _Norm10 -3.110 .375 .0165 .527 .0510 .253 -1801.1 3616.1 _Neg10 -2.370 .175 -.001 .247 .047 .125 -1932.3 3878.7 _Pos25 -17.090 501.400 2.390 473.600 1.700 133.500 -1471.4 2956.7 _Norm25 -1.380 .132 -.002 .186 -.059 .983 -1791.6 3597.1 _Neg25 -1.162 .109 .000 .037 -.016 .019 -1901.9 3817.9 _Pos50 -4.895 3.373 1.352 2.211 -1.086 1.304 -1292.4 2598.8 _Norm50 -.131 .097 .008 .014 .031 .069 -1514.1 3042.3 _Neg50 -.031 .091 .000 .129 .035 .065 -1587.9 3189.7 _Pos75 .130 .364 .015 .271 -.040 .136 -865.1 1744.3 _Norm75 1.015 .107 .001 .151 .033 .076 -974.3 1962.6 _Neg75 1.080 .104 -.002 .147 .044 .075 -1012.2 2038.3 _Pos90 1.462 .386 .047 .298 .098 .151 -447.3 908.7 _Norm90 2.133 .152 .002 .214 .128 .108 -490.4 994.8 _Neg90 2.187 .150 -.001 .212 .136 .107 -507.0 1028.1 92

PAGE 93

Table 3-13. Negative Bino mial ZIP model (events): pseudo-population parameters Theta LL AIC 0s1s2s0 1 2 _Pos10 .450 .036 .004 .050 -.037 .026 14.69 -1479.3 2972.6 _Norm10 1.040 .030 .004 .043 -.013 .021 12.64 -1801.1 3616.1 _Neg10 1.430 .024 .001 .033 -.003 .017 13.02 -1932.3 3878.7 _Pos25 .270 .040 .003 .057 -.024 .029 3.16 -1471.4 2956.7 _Norm25 1.030 .033 .005 .047 -.027 .025 10.88 -1791.6 3597.1 _Neg25 1.431 .026 .001 .037 -.016 .019 15.94 -1901.9 3817.9 _Pos50 -.127 .067 .045 .108 -.080 .052 .57 -1292.4 2598.8 _Norm50 1.039 .040 .007 .057 -.026 .029 11.70 -1514.1 3042.3 _Neg50 1.429 .032 .004 .045 -.012 .023 14.19 -1587.9 3189.7 _Pos75 -.066 .204 .010 .172 -.099 .087 .69 -865.1 1744.3 _Norm75 1.033 .057 .005 .081 -.054 .040 12.02 -974.3 1962.6 _Neg75 1.430 .045 .000 .064 -.010 .032 12.99 -1012.2 2038.3 _Pos90 _Norm90 -.102 1.016 .327 091 .072 017 .273 128 -.080 -.052 .140 065 .73. 12.76 -447.3 -490.4 908.7 994.8 _Neg90 1.422 .071 .006 .101 -.028 .051 14.73 -507.0 1028.1 93

PAGE 94

CHAPTER 4 RESULTS This chapter presents the results based on the previously discussed methods and procedures for analyzing data with varying proportions of zeros a nd varying event stage distributions. First, the results are presented using the data for the pseudo-population in which the proportions for each count level are exactly that which was rando mly sampled from in the simulations. Tables and figures ar e included to support interpretation. Second, the results are presented outlined by the skew level (i.e., positiv e, normal, and negative) and by the proportion of zeros within that skew level (i.e., .10, 25, .50, .75, and .90). For each combination of conditions, the results are presented for the five model comparisons. Third, the results are summarized separately for the negative, normal, and positive event coun t distributions. Tables and figures are included to suppor t interpretation. The primary purpose of the r esults was to assist in answering the research que stions, which were as follows: Given one two-level categorical covariate with known values and one continuous covariate with known values, what is the difference in the estimated log-lik elihood between a) the Negative binomial Poisson model vs. Poiss on model; b) the Hurdle model vs. Poisson model?; c) the Negative binomial Hurdle model vs. negative binomial Poisson model?; d) the Negative binomial Hurdle model vs. Hurdle model; and, e) the Negative binomial ZIP model vs. ZIP model? Given one two-level categorical covariate with known values and one continuous covariate with known values, what is the difference in the estimated AIC between all models? Pseudo-Population Results Before addressing the results of the Monte Ca rlo simulations, it is necessary to discuss the results when each model was analyzed with the pseudo-population data. The prespecified proportions for each count were displayed in Table 3-1. Table 3-2 displayed this same information in terms of frequenc ies instead of proportions. Table 3-3 displayed the descriptive statistics for each distribution. Tables 3-4 through 3-13 displaye d the coefficients, standard 94

PAGE 95

errors, log-likelihood values, and AIC values with each table pertaining to either a specific model or one of the two stages for a specific model. The following sections provide the results when comparing models using the pseudopopulation data. First, the Poisson model results were compared to the negative binomial Poisson model results, both descriptively via AICs and inferentially vi a the deviance model comparison test. Second, the results are presented in a likewise manne r for the three Hurdle model comparisons. These comparisons were a) the Hurdle model vs. the negative binomial Hurdle model, b) the Poisson model vs. the Hurdle model, and c) the ne gative binomial Poisson model vs. the negative binomial Hurdle model. Third, the results are presented for the comparison of the ZIP model and the negative binomial ZIP mo del. Fourth, descriptive results are presented comparing all models for each of the five proportions of zeros. Pseudo-Population Poisson Models Based on the AIC, for all proportions of zeros the data fit the Poisson model better when the distribution was positively skewed than when nor mally distributed. In addition, this data fit the Poisson model better when the distribution was normally di stributed than when it was negatively skewed. For example, for the negativel y skewed distribution wi th a .10 proportion of zeros, the AIC was 4,044.5. As the curve shifted le ft to a normal distribution, the AIC dropped to 3,619.3, and as the curve shifted further left to a positively skewed distribution, the AIC dropped to 2,956.3. The same pattern emerged for the negative binom ial Poisson models. For example, for the negatively skewed distribution w ith a .10 proportion of zeros, the AIC was 4,563.8. As the curve shifted left to a normal distribut ion, the AIC dropped to 4,005.9, and as the curve shifted further left to a positively skewed distri bution, the AIC dropped to 3,196.9. 95

PAGE 96

For the .10 proportion of zeros condition, the Poisson models had a lower AIC than that calculated for the negative binomial Poisson model. For the .25 proportion of zeros condition, the AICs were approximately equal between the two models. For the .50, .75, and .90 proportions of zeros conditions, th e AIC was lower for the negative binomial Poisson model than for the Poisson model. The deviance statistic was cal culated comparing the Poisson log-likelihood and negative binomial Poisson log-likelihood for each skew condition and zero proportion condition. These statistics are displayed in Table 4-1. For all analyses, the Poiss on model is nested within its negative binomial Poisson formulation differing by one degree of freedom (i.e., the dispersion parameter in the negative binomial Poisson model) Hence, at the .05 T ype I error rate and assuming deviance statistics asymp totically distributed chi-square, a deviance exceeding 3.84 suggests better fit for the more complex negative binomial Poisson model. This was the result for all analyses in which the proporti on of zeros was .50 or greater. However, for all analyses with .10 and .25 proportions of zeros, the statistic was significant in favor of the Poisson model. Pseudo-Population Hurdle Models Hurdle vs. Negative Binomial Hurdle For the Hurdle models, regardless of the proportion of zeros, the positively skewed distributions had a lower AIC than did the normal distributions. These, in turn, had a lower AIC than the negatively skewed distribution. The sam e was true for the negative binomial Hurdle models. However, for both models, the differenc e between AICs for the three skew conditions decreased as the proportion of zer os decreased (i.e., the models became more similar in fit). When comparing the Hurdle models and negative binomial Hurdle models, the AICs appear to be similar regardless of the proportion of zeros and regardless of skew. 96

PAGE 97

The deviance statistic was cal culated comparing the Hurdle log-likelihood and negative binomial Hurdle log-likelihood for each skew condition and zero proportion condition. These statistics are displayed in Table 4-2. For all analyses, the Hurdle model is nested within its negative binomial Hurdle formulation differing by one degree of freedom (i.e., the dispersion parameter in the negative binomial Hurdle model) Hence, at the .05 T ype I error rate and assuming deviance statistics asymp totically distributed chi-square, a deviance exceeding 3.84 suggests better fit for the more comp lex negative binomial Hurdle model. For the positively skewed distributions, the negative binomial Hurdle model fit significantly better than th e Hurdle model, except when the proportion of zeros was .90. In this case the deviance statistic did not exceed 3.84. For the normal distributions, the Hurdle model fit significantly better than the negative binomial Hurdle mode l when the proportion of zeros was .10 or .25. When the proportion of zeros was .50, .75, or .90, the deviance did not exceed 3.84. For the negatively skewed distributions, the Hu rdle model fit significantly better than the negative binomial Hurdle model when the proportion of zeros was .10, .25, .50, or .75. When the proportion of zeros of .90, the devi ance statistic did not exceed 3.84. Poisson vs. Hurdle Deviance statistics were also calculated to compare the Po isson log-likelihood and Hurdle log-likelihood for each skew condition and zer o proportion condition. These statistics are displayed in Table 4-3. For all analyses, the Po isson model is nested within the Hurdle model differing by three degrees of freedom. Hence, at the .05 Type I error rate and assuming deviance statistics asymptotically distribut ed chi-square, a deviance exceedi ng 7.82 suggests better fit for the more complex Hurdle model. The deviances were large supporting the Hurd le model fit over the Po isson model fit. In fact, several of the deviances were over 1,000. There was only one analysis that did not favor the 97

PAGE 98

Hurdle model fit over the Poisson model fit. Th is was the deviance calculated for the positively skewed distribution with a .25 proportion of zeros. Negative Binomial Poisson vs. Negative Binomial Hurdle Deviance statistics were also calculated to compare the negative binomial Poisson loglikelihood and the negative binomial Hurdle log-likelihood for each skew condition and zero proportion condition. These statistics are displayed in Table 4-4. For all analyses, the negative binomial Poisson model is nested within the ne gative binomial Hurdle model differing by three degrees of freedom (i.e., the duplication of paramete rs in the negative binomial Hurdle model to represent both a transition stage and an event stage).28 Hence, at the .05 Type I error rate and assuming deviance statistics asymp totically distributed chi-square, a deviance exceeding 7.82 suggests better fit for the more comp lex negative binomial Hurdle model. As was the case when comparing the Poisson model and the Hurdle model, the deviances were large. Likewise, there was one deviance that did not exceed the critical value. However, it was for the positively skewed distribution with .50 zeros rather than the positively skewed distribution with .25 zeros. Pseudo-Population ZIP Models The pseudo-population results for the ZIP models were rather similar to those obtained for the Hurdle models. Graphically, regardless of skew condition and proportions of zeros, the AICs for the ZIP models and negative bi nomial ZIP models were very similar. Additionally, the AICs appeared to be equal between the .10 a nd .25 proportions of zeros conditions. The deviance statistic was calculated comparing the ZI P log-likelihood and negative binomial ZIP log-likelihood for each skew c ondition and zero propor tion condition. These 28 The presence of a dispersion parameter is now redundant between models. 98

PAGE 99

statistics are displayed in Table 45. For all analyses, the ZIP model is nested within its negative binomial ZIP formulation differing by one degree of freedom (i.e., the di spersion parameter in the negative binomial ZIP model). Hence, at th e .05 Type I error rate and assuming deviance statistics asymptotically distribut ed chi-square, a deviance exceedi ng 3.84 suggests better fit for the more complex negative binomial ZIP model. The deviance was significant for only two comparisons. When the distribution was positively skewed, the fit was significantly better for the negative binomial ZIP model than for the ZIP model when the proportion of zeros was either .50 or .75. All other results suggested that the ZIP model fit was adequate. Comparing AICs For All Models For the .10 proportion of zeros conditi on with negative skew, the AICs were approximately equal for all models except for the Poisson models. The AIC for the Poisson model was higher than for the other models, and the AIC for the negative binomial Poisson model was highest of all. When the distribution was normal, the only model to have a noticeably different AIC was the negative binomial Hurdle which was again highest of all. For the positively skewed distribution, there was some separation between the ZIP and Hurdle models, with the Hurdle models having the lower AIC. The differences between these models and their negative binomial formulations appeared to be tr ivial. The Poisson model appeared to have the same AIC as those displayed for the ZIP models The AIC for the negative binomial was again highest of all. Between the th ree distributions, the AIC declined from the negatively skewed distribution to the normal distribution to the positivel y skewed distribution. For the .25 proportion of zeros, there was lit tle distinction between the Poisson models and negative binomial Poisson models for all di stributions. Further, th ere was no distribution displaying a nontrivial distinc tion between the Hurdle models and the ZIP models. For the 99

PAGE 100

positively skewed distribution, all six distributions appeared approximately equal with a slightly higher AIC apparent for the negative binomial Po isson model. Between the three distributions, the AICs appeared equal for the negatively skew ed distribution and the normal distribution; However, the AICs for the normal distribution were considerably smaller than the AICs for the other two distributions. For the .50 proportion of zeros condition with negative skew, the r esults for the Poisson model and the negative binomial Poisson model re versed. For the negativel y skewed distribution, the AIC for the negative binomial Poisson was highe r than those for the ZIP and Hurdle models, while the AIC for the Poisson model was higher ye t. As in the .50 propor tion of zeros condition, the AIC appeared equal for the ZIP and Hurdle models. Also, as in the .50 proportion of zeros condition, there appeared to be no difference between any of the models for the positively skewed distribution. Between th e three distributions, the AICs declined from the negatively skewed distribution to the normal distribution to the positively sk ewed distribution with declines being most rapid for the Poisson and negative binomial Poisson models. Beyond the fact that the overa ll AIC was lower for the .90 proportion of zeros condition than for the .75 condition, these last two c onditions displayed similar results. For each distribution, the Hurdle and ZIP model AICs were approximately equal. They declined slightly from the negatively skewed distribution to the no rmal distribution and declined a bit more from this distribution to the positively skewed distribution. For the negatively skewed distribution, the negative binomial Poisson AIC w as considerably higher than the AICs for the ZIP and Hurdle models; the AIC for the Poisson model was highest of all. For the normal distribution, both the Poisson and negative binomial Poisson model AICs declined by approximately 50% of their value for the negatively skewed distribution. For the positively skewed distribution, they 100

PAGE 101

declined further yet revealing a Poisson model AIC slightly hi gher than that of the negative binomial AIC model, which was slightly higher than that of the nearly equal Hurdle and ZIP model AICs. Monte Carlo Simulation Results Positively Skewed Distribution These results pertain to the data in which 100 of the 1,000 observations in each of the 2,000 simulated datasets were constant at zero. The remaining 900 values were sampled from the positively skewed pseudo-population distribution with the count proportions and frequencies displayed in Table 31 and Table 3-2. For the .10 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from 271.5 to -201.98 with a mean of -238.24 and a standard deviation of 10.43. Given a Type I error rate of .05, th e deviance statistic was significant for all of the 2,000 model comparisons. The average AIC for the Poisson model was 2957.06, while the average AIC for the negative binomial Poisson model was 3193.3, which descriptively supports the inferential findings of better fit for the Poisson model. For the .10 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from 5.14 to 69.8 with a m ean of 27.93 and a standard deviation of 9.0. Given a Type I error rate of .05, the deviance st atistic was significant for all of the 2,000 model comparisons. The average AIC for the Hurdle model was 2838.81 while the average AIC for the negative binomial Hurdle model was 2804.88, which descriptively suppo rts the inferential findings of better fit for the negative binomial Hurdle model. For the .10 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from 86. 36 to 187.64 with a mean of 132.25 and a standard deviation of 14.58. Given a Type I error rate of .05, the devi ance statistic was significa nt for all of the 2,000 101

PAGE 102

valid model comparisons. The average AIC for the Poisson model was 2957.06 while the average AIC for the Hurdle model was 2838.81, which descriptively supports the inferential findings of better fit for the Hurdle model. For the .10 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from 332.81 to 470.81 w ith a mean of 398.42 and a standard deviation of 19.86. Given a Type I error rate of .05, th e deviance statistic was significant for all of the 2,000 model comparisons. The average AIC for th e negative binomial Poisson model was 3193.3 while the average AIC for the negative binomial Hurdle model was 2804.88, which descriptively supports the inferential findings of better fit for the negative binomial Hurdle model. For the .10 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from -2. 02 to -2.00 with a mean of -2 .00 and a standard deviation of .001 Given a Type I error rate of .05, the de viance statistic was signifi cant for none of the 112 valid model comparisons. The average AIC for the ZIP model was 2967.85 the average AIC for the negative binomial ZIP model was 2984.79, which descriptively suggests be tter fit for the ZIP model. Table 4-6 summarizes the log-likelihood statistics for the positively skewed distribution model comparisons with a .25 proportion of zeros. In this table, LL represents the mean loglikelihood for the pseudo-population, L LminLLmaxLLLLsrepresents the mean log-likelihood for the simulated dataset samples, and represent the log-likelihood minimum and maximum for the simulated dataset sample, and LLs represent the standard deviat ion and standard error for the simulated dataset samples, and n represents the sample size for the simulated dataset samples. 102

PAGE 103

A comparison of AICs over all models revealed a minimum of 2622.53 and a maximum of 3096.30 across all 9,678 valid model comparis ons. Between the six m odels, the minimum was 2804.88 (i.e., the negative binomial Hurdle mo del), and the maximum was 3193.30 (i.e., the negative binomial Poisson model). The rank or der from lowest AIC to highest AIC was as follows: negative binomial Hurdle, Hurdle, Poisson, ZIP, negative binomial ZIP, negative binomial Poisson. These rankings are in line with the rankings for th e pseudo-population AICs. Table 4-7 displays the descriptive statistics for these two models. Figure 4-1 is a boxplot illustrating these results. The following set of results pertains to the data in which 250 of the 1,000 observations in each of the 2,000 simulated datase ts were fixed at zero. The re maining 750 values were sampled from a positively skewed pse udo-population distribution with the count proportions and frequencies displayed in Table 3-1 and Table 3-2. For the .25 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from -126. 64 to -31.15 with a mean of -83. 2 and a standard deviation of 12.3. Given a Type I error rate of .05, the devi ance statistic was significa nt for all of the 2,000 model comparisons. The average AIC for the Po isson model was 2939.65 while the average AIC for the negative binomial Poisson model w as 3020.85, which descriptively supports the inferential findings of better fit for the Poisson model. For the .25 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from .87 to 22.71 with a mean of 5.76 and a standard deviation of 3.47. Given a Type I error rate of .05, the deviance st atistic was significant for 1,994 of the 2,000 valid model comparisons. The average AIC for the Hu rdle model was 2947.9 while the average AIC 103

PAGE 104

for the negative binomial Hurdle model was 2921, which descriptively supports the 99.7% inferential findings of better fit for th e negative binomial Hurdle model. For the .25 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from .07 to 22.7 with a mean of 5.76 a nd a standard deviation of 3.47. Given a Type I error rate of .05, the deviance st atistic was significant for 503 of the 2,000 valid model comparisons. The average AIC for the Po isson model was 2939.65 while the average AIC for the Hurdle model was 2947.9, which descriptivel y supports better fit for the Poisson model. For the .25 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from 75.94 to 156.11 with a mean of 109.85 and a standard deviation of 12.33. Given a Type I error rate of .05, the devi ance statistic was significa nt for all of the 2,000 valid model comparisons. The average AIC for th e negative binomial Poisson model was 3020.9, while the average AIC for the negative binom ial Hurdle model was 2921, which descriptively supports the inferential findings of better fit for the negative binomial Hurdle model. For the .25 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from 10.55 to 17.14 with a mean of 54 and a standard deviation of 3.86. Given a Type I error rate of .05, the de viance statistic was signifi cant for 344 of the 1,850 valid model comparisons. The average AIC fo r the ZIP model was 2951.12 while the average AIC for the negative binomial ZIP model was 2956.15 which descriptively supports better fit for the ZIP model. Table 4-8 summarizes the log-likelihood statistics for the positively skewed distribution model comparisons with a .25 proportion of zeros. In this table, LL represents the mean loglikelihood for the pseudo-population, L LminLLmaxLLrepresents the mean log-likelihood for the simulated dataset samples, and represent the log-likelihood minimum and maximum for the 104

PAGE 105

simulated dataset sample, and LLs LLs represent the standard deviat ion and standard error for the simulated dataset samples, and n represents the sa mple size for the simulated dataset samples. A comparison of AICs over all models revealed a minimum of 2754.05 and a maximum of 3127.33 across all valid model comparisons Between the six models, the minimum was 2921.00 (i.e., the negative binomial Hurdle mo del), and the maximum was 3020.85 (i.e., the negative binomial Poisson model). The rank or der from lowest AIC to highest AIC was as follows: negative binomial Hurdle, Poisson, Hurdle, ZIP, negative binomial ZIP, negative binomial Poisson. These rankings are in line with the rankings for th e pseudo-population AICs. Table 4-9 displays the descriptive statistics for these two models. Figure 4-2 is a boxplot illustrating these results. The following set of results pertains to the data in which 500 of the 1,000 observations in each of the 2,000 simulated datase ts were fixed at zero. The re maining 500 values were sampled from a positively skewed pse udo-population distribution with the count proportions and frequencies displayed in Ta ble 3-1 and Table 3-2. For the .50 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from 31.23 to 142.04 with a mean of 87.9 and a standard deviation of 16.37. Given a Type I error rate of .05, the devi ance statistic was significa nt for all of the 2,000 valid model comparisons. The average AIC for the Poisson model was 2672.88 while the average AIC for the negative binomial Poisson model was 2582.98, which descriptively supports the inferential findings of better fit for the negative binomial Poisson model. For the .50 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from .84 to 43.88 with a mean of 15.69 and a standard deviation of 6.8. Given a Type I error rate of .05, the deviance st atistic was significant for 1,961 of the 2,000 valid 105

PAGE 106

model comparisons. The average AIC for the Hu rdle model was 2601.93 while the average AIC for the negative binomial Hurdle model was 2588. 24, which descriptively supports the inferential findings of better fit for the negative binomial Hurdle model. For the .50 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from 35.5 to 142.74 with a mean of 76.95 and a standard deviation of 17.08. Given a Type I error rate of .05, the devi ance statistic was significa nt for all of the 2,000 valid model comparisons. The average AIC for the Poisson model was 2674.5 while the average AIC for the Hurdle model was 2601.93, which descr iptively supports the inferential findings of better fit for the Hurdle model. For the .50 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from .12 to 18.73 with a mean of 4.74 and a standard deviation of 2.87. Given a Type I error rate of .05, the deviance st atistic was significant for 268 of the 2,000 valid model comparisons. The average AIC for the negative binomial Poisson model was 2582.98 while the average AIC for the negative binomial Hurdle model was 2588.24, which descriptively supports better fit for the negative binomial Poisson model. For the .50 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from -1.16 to 38.97 with a mean of 13.61 and a standard deviation of 6.51. Given a Type I error rate of .05, the devi ance statistic was significa nt for 1,894 of the 2,000 valid model comparisons. The average AIC for the ZIP model was 2607.84 while the average AIC for the negative binomial ZIP model was 2596.23 which descriptively supports the 94.7% inferential findings of better fit for the negative binomial ZIP model. Table 4-10 summarizes the log-likelihood statis tics for the positively skewed distribution model comparisons with a .50 proportion of zeros. In this table, LL represents the mean log106

PAGE 107

L likelihood for the pseudo-population, LminLLmaxLLLLsrepresents the mean log-likelihood for the simulated dataset samples, and represent the log-likelihood minimum and maximum for the simulated dataset sample, and LLs represent the standard deviat ion and standard error for the simulated dataset samples, and n represents the sample size for the simulated dataset samples. A comparison of AICs over all models revealed a minimum of 2451.80 and a maximum of 2862.94 across all 10,000 valid model comparis ons. Between the six models, the minimum was 2582.09 (i.e., the negative binomial Poiss on model), and the maximum was 2672.88 (i.e., the Poisson model). The rank order from lowest AIC to highest AIC was as follows: negative binomial Poisson, negative binomial Hurdle, negative binomial ZIP, Hurdle, ZIP, Poisson. These rankings are in line with the ra nkings for the pseudopopulation AICs. Table 4-11 displays the descriptive statistics for these two models. Figure 4-3 is a boxplot illustrating these results. The following set of results pertains to the data in which 750 of the 1,000 observations in each of the 2,000 simulated datase ts were fixed at zero. The re maining 250 values were sampled from a positively skewed pse udo-population distribution with the count proportions and frequencies displayed in Ta ble 3-1 and Table 3-2. For the .75 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from 68.65 to 180.23 with a mean of 125.85 and a standard deviation of 16.52. Given a Type I error rate of .05, th e deviance statistic was significant for all of the 2,000 valid model comparisons. The average AIC for the Poisson model was 1955.22, while the average AIC for the negative binomial Poisson model was 1827.36, which descriptively supports the inferential findings of better fit for the negative binomial Poisson model. For the .75 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from -1.45 to 32.0 with a mean of 8.3 and a standard deviation of 4.77. 107

PAGE 108

Given a Type I error rate of .05, the deviance st atistic was significant for 1,963 of the 2,000 valid model comparisons. The average AIC for the Hu rdle model was 1749.71 while the average AIC for the negative binomial Hurdle model was 1735.4, which descriptiv ely supports the 98.1% inferential findings of better fit for th e negative binomial Hurdle model. For the .75 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from 95.23 to 363.75 with a mean of 219.51 and a standard deviation of 35.63. Given a Type I error rate of .05, the devi ance statistic was significa nt for all of the 2,000 valid model comparisons. The average AIC for the Poisson model was 1955.22 while the average AIC for the Hurdle model was 1749.7, which descriptively s upports the in ferential findings of better fit for the Hurdle model. For the .75 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from 35.6 to 186.3 with a mean of 101.96 and a standard deviation of 19.64. Given a Type I error rate of .05, the devi ance statistic was significa nt for all of the 2,000 valid model comparisons. The average AIC for the negative binomial Poisson model was 1827.36 while the average AIC for the negative binomial Hurdle model was 1735.4, which descriptively supports the inferential findings of better fit for the negative binomial Hurdle model. For the .75 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from -2.0 to 29.66 with a mean of 6.27 and a standard deviation of 4.73. Given a Type I error rate of .05, the deviance st atistic was significant for 1,304 of the 2,000 valid model comparisons. The average AIC for the ZI P model was 1747.72, while the average AIC for the negative binomial ZIP model was 1743.44, which descriptively supports the 65.2% inferential findings of better fit for the negative binomial ZIP model. 108

PAGE 109

Table 4-12 summarizes the log-likelihood statis tics for the positively skewed distribution model comparisons with a .75 proportion of zeros. In this table, LL represents the mean loglikelihood for the pseudo-population, L LminLLmaxLLLLsrepresents the mean log-likelihood for the simulated dataset samples, and represent the log-likelihood minimum and maximum for the simulated dataset sample, and LLs represent the standard deviat ion and standard error for the simulated dataset samples, and n represents the sample size for the simulated dataset samples. A comparison of AICs over all models revealed a minimum of 1617.27 and a maximum of 2188.38 across all valid model comparisons Between the six models, the minimum was 1735.40 (i.e., the negative binomial Hurdle mo del), and the maximum was 1827.36(i.e., the negative binomial Poisson model). The rank or der from lowest AIC to highest AIC was as follows: negative binomial Hurdle, negative binomial ZIP, ZIP, Hurdle, negative binomial Poisson, Poisson. These rankings are not in line with the rank ings for the pseudo-population AICs; in the pseudo-population, the Hurdle mo del has a lower AIC than the negative binomial ZIP model. Table 4-13 displays the descriptive statistics for these two mo dels. Figure 4-4 is a boxplot illustrating these results. The following set of results pertains to the data in which 900 of the 1,000 observations in each of the 2,000 simulated datase ts were fixed at zero. The re maining 100 values were sampled from a positively skewed pse udo-population distribution with the count proportions and frequencies displayed in Ta ble 3-1 and Table 3-2. For the .90 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from 29.24 to 188.86 with a mean of 69.82 and a standard deviation of 12.5. Given a Type I error rate of .05, the devi ance statistic was significa nt for all of the 2,000 valid model comparisons. The average AIC for the Poisson model was 1098.79 while the 109

PAGE 110

average AIC for the negative binomial Poisson model was 1026.98, which descriptively supports the inferential findings of better fit for the negative binomial Poisson model. For the .90 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from -4.1 to 17.15 with a mean of 2.96 and a standard deviation of 2.69. Given a Type I error rate of .05, the deviance st atistic was significant for 611 of the 1,992 valid model comparisons. The average AIC for the Hu rdle model was 906.36 while the average AIC for the negative binomial Hurdle model was 897. 5, which descriptively supports the 30.7% inferential findings of better fit for th e negative binomial Hurdle model. For the .90 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from 86.47 to 367.79 with a mean of 206.43 and a standard deviation of 41.18. Given a Type I error rate of .05, the devi ance statistic was significa nt for all of the 2,000 valid model comparisons. The average AIC for the Poisson model was 1098.79 while the average AIC for the Hurdle model was 906.36, which descriptively s upports the in ferential findings of better fit for the Hurdle model. For the .90 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from 55.21 to 251.73 with a mean of 139.66 and a standard deviation of 28.92. Given a Type I error rate of .05, the devi ance statistic was significa nt for all of the 1,992 valid model comparisons. The average AIC for the negative binomial Poisson model was 1026.98 while the average AIC for the negative binomial Hurdle model was 897.5, which descriptively supports the inferential findings of better fit for the negative binomial Hurdle model. For the .90 proportion of zeros condition with a positively skewed distribution, the deviance statistic ranged from -2.11 to 14.85 with a mean of .96 and a standard deviation of 2.66. 110

PAGE 111

Given a Type I error rate of .05, the deviance st atistic was significant for 265 of the 1,992 valid model comparisons. The average AIC for the ZI P model was 904.34 while the average AIC for the negative binomial ZIP model was 905.38, which descriptively supports better fit for the ZIP model. Table 4-14 summarizes the log-likelihood statis tics for the positively skewed distribution model comparisons with a .90 proportion of zeros. In this table, LL represents the mean loglikelihood for the pseudo-population, L LminLLmaxLLLLsrepresents the mean log-likelihood for the simulated dataset samples, and represent the log-likelihood minimum and maximum for the simulated dataset sample, and LLs represent the standard deviat ion and standard error for the simulated dataset samples, and n represents the sample size for the simulated dataset samples. A comparison of AICs over all models reveal ed a minimum of 834.28 and a maximum of 1313.41 across all valid model comparisons. Be tween the six models, the minimum was 897.50 (i.e., the negative binomial Hurdle model), and the maximum was 1098.79 (i.e., the Poisson model). The rank order from lowest AIC to highest AIC was as follows: negative binomial Hurdle, ZIP, negative binomial ZIP, Hurdle, ne gative binomial Poisson, Poisson. These rankings are in line with the rankings for the pseudo-popul ation AICs. Table 4-15 displays the descriptive statistics for these two models. Figure 4-5 is a boxplot illustrating these results. Normal Distribution The following set of results pertains to the data in which 100 of the 1,000 observations in each of the 2,000 simulated datase ts were fixed at zero. The re maining 900 values were sampled from a normal pseudo-population distribution with the count proportions and frequencies displayed in Table 31 and Table 3-2. 111

PAGE 112

For the .10 proportion of zeros condition with a normal distribution, the deviance statistic ranged from -425.53 to -347.98 with a mean of -3 85.1 and a standard deviation of 11.5. Given a Type I error rate of .05, the deviance statistic was significant for none of the 2,000 valid model comparisons. The average AIC for the Poison model was 3618.76 while the average AIC for the negative binomial Poisson model was 4005.86, which descriptively supports the inferential findings of better fit fo r the Poisson model. For the .10 proportion of zeros condition with a normal distribution, the deviance statistic ranged from -40.51 to -4.94 with a mean of -6.61 and a standard deviation of 2.75. Given a Type I error rate of .05, the devi ance statistic was significant for none of the 1,503 valid model comparisons. The average AIC for the Hurdle model was 3605.03 while the average AIC for the negative binomial Hurdle model was 3613.55, which descriptively suppo rts the inferential findings of better fit for the Hurdle model. For the .10 proportion of zeros condition with a normal distribution, the deviance statistic ranged from 9.87 to 31.88 with a mean of 19.73 and a standard deviation of 2.94. Given a Type I error rate of .05, the devi ance statistic was significant fo r all of the 2,000 valid model comparisons. The average AIC for the Poisson m odel was 3618.76 while the average AIC for the Hurdle model was 3605.03, which descriptively suppor ts the inferential findi ngs of better fit for the Hurdle model. For the .10 proportion of zeros condition with a normal distribution, the deviance statistic ranged from 359.25 to 439.66 with a mean of 397. 78 and a standard deviation of 12.5. Given a Type I error rate of .05, the deviance statistic was significant for all of the 1,503 valid model comparisons. The average AIC for the negative binomial Poisson model was 4005.86 while the 112

PAGE 113

average AIC for the negative binomial Hurdle model was 3613.55, which descriptively supports the inferential findings of better fit for the negative binomial Hurdle model. For the .10 proportion of zeros condition with a normal distribution, the deviance statistic ranged from -16.62 to -1.83 with a mean of -2.04 and a standard de viation of .7 Given a Type I error rate of .05, the devian ce statistic was significant (ZIP) for 7 of the 1,982 valid model comparisons. The average AIC for the ZIP m odel was 3611.05 while the average AIC for the negative binomial ZIP model was 3615 which descriptively supports the 99.6% inferential findings of better fit for the ZIP model. Table 4-16 summarizes the log-likelihood statistics for the normal distribution model comparisons with a .10 proportion of zeros. In this table, LL represents the mean log-likelihood for the pseudo-population, L LminLLmax LLsrepresents the mean log-likeli hood for the simulated dataset samples, and LL represent the loglikelihood minimum and maximum for the simulated dataset sample, and LLs represent the standard deviation and standard error for the simulated dataset samples, and n represents the sample size for the simulated dataset samples A comparison of AICs over all models revealed a minimum of 3530.45 and a maximum of 4071.95 across all valid mode l comparisons. Between the six models, the minimum was 3611.05 (i.e., the Hurdle model), and the ma ximum was 4005.86 (i.e., th e negative binomial Poisson model). The rank order from lowest AI C to highest AIC was as follows: Hurdle, ZIP, negative binomial Hurdle, negative binomial ZIP, Poisson, negative binomial Poisson. These rankings are in line with the ra nkings for the pseudopopulation AICs. Table 4-17 displays the descriptive statistics for these two models. Figur e 4-19 is a boxplot illustrating these results. 113

PAGE 114

The following set of results pertains to the data in which 250 of the 1,000 observations in each of the 2,000 simulated datase ts were fixed at zero. The re maining 750 values were sampled from a normal pseudo-population distribution with the count proportions and frequencies displayed in Table 31 and Table 3-2. For the .25 proportion of zeros condition with a normal distribution, the deviance statistic ranged from -69.34 to -5.87 with a mean of -35.75 and a standa rd deviation of 9.23. Given a Type I error rate of .05, the deviance statistic was significant for none of the 2,000 valid model comparisons. The average AIC for the Poisson m odel was 3832.24 while the average AIC for the negative binomial Poisson model was 3869.99, wh ich descriptively supports the inferential findings of better fit fo r the Poisson model. For the .25 proportion of zeros condition with a normal distribution, the deviance statistic ranged from -19.83 to -3.93 with a mean of -5.2 and a standard de viation of .82. Given a Type I error rate of .05, the devi ance statistic was significant fo r all of the 1,634 valid model comparisons. The average AIC for the Hurdle model was 3586.47 while the average AIC for the negative binomial Hurdle model was 3593.63 which descriptively suppo rts the inferential findings of better fit for the Hurdle model. For the .25 proportion of zeros condition with a normal distribution, the deviance statistic ranged from 202.52 to 312.73 with a mean of 251. 77 and a standard deviation of 14.35. Given a Type I error rate of .05, the deviance statistic was significant for all of the 2,000 valid model comparisons. The average AIC for the Poisson m odel was 3832.24 while the average AIC for the Hurdle model was 3586.47, which descriptively suppor ts the inferential findi ngs of better fit for the Hurdle model. 114

PAGE 115

For the .25 proportion of zeros condition with a normal distribution, the deviance statistic ranged from 228.78 to 330.2 with a mean of 282. 08 and a standard deviation of 14.95. Given a Type I error rate of .05, the deviance statistic was significant for all of the 1,634 valid model comparisons. The average AIC for the negative binomial Poisson model was 3869.99 while the average AIC for the negative binomial Hurdle model was 3593.63, which descriptively supports the inferential findings of better fit for the negative binomial Hurdle model. For the .25 proportion of zeros condition with a normal distribution, the deviance statistic ranged from -3.09 to -1.94 with a mean of -2.01 and a standard de viation of .05. Given a Type I error rate of .05, the devi ance statistic was significant fo r none of the 2,000 valid model comparisons. The average AIC for the ZIP m odel was 3592.44 while the average AIC for the negative binomial ZIP model was 3596.45, whic h descriptively suggests better fit for the ZIP model. Table 4-18 summarizes the log-likelihood statistics for the normal distribution model comparisons with a .25 proportion of zeros. In this table, LL represents the mean log-likelihood for the pseudo-population, L LminLLmax LLsrepresents the mean log-likeli hood for the simulated dataset samples, and LL represent the log-likelihood minimu m and maximum for the simulated dataset sample, and LLs represent the standard deviation a nd standard error for the simulated dataset samples, and n represents the sample size for the simulated dataset samples. A comparison of AICs over all models revealed a minimum of 3529.93 and a maximum of 2966.65 across all 11,634 valid model comparis ons. Between the six models, the minimum was 3586.47 (i.e., the Hurdle model), and the ma ximum was 3869.99 (i.e., the negative binomial Poisson model). The rank order from lowest AI C to highest AIC was as follows: Hurdle, ZIP, negative binomial Hurdle, negative binomial ZIP, Poisson, negative binomial Poisson. These 115

PAGE 116

rankings are not in line with the rankings for the pseudo-pop ulation AICs; in the pseudopopulation, the negative binomial Hurdle model h as a lower AIC (i.e., better fit) than the ZIP model Table 4-19 displays the descriptive st atistics for these two models. Figure 4-20 is a boxplot illustrating these results. The following set of results pertains to the data in which 500 of the 1,000 observations in each of the 2,000 simulated datase ts were fixed at zero. The re maining 500 values were sampled from a normal pseudo-population distribution with the count proportions and frequencies displayed in Table 31 and Table 3-2. For the .50 proportion of zeros condition with a normal distribution, the deviance statistic ranged from 286.84 to 381.11 with a mean of 331. 24 and a standard deviation of 13.74. Given a Type I error rate of .05, the deviance statistic was significant for all of the 2,000 valid model comparisons. The average AIC for the Poisson m odel was 3642.82 while the average AIC for the negative binomial Poisson model was 3356.12, wh ich descriptively supports the inferential findings of better fit for the nega tive binomial Poisson model. For the .50 proportion of zeros condition with a normal distribution, the deviance statistic ranged from -41.05 to 3.15 with a mean of -3.86 and a standard de viation of 2.19 Given a Type I error rate of .05, the devi ance statistic was significant for none of the 1,602 valid model comparisons. The average AIC for the Hurdle model was 2978.57 while the average AIC for the negative binomial Hurdle model was 2994.57, which descriptively suppo rts the inferential findings of better fit for the Hurdle model. For the .50 proportion of zeros condition with a normal distribution, the deviance statistic ranged from 655.17 to 857.57 with a mean of 758. 23 and a standard deviation of 31.92. Given a Type I error rate of .05, the deviance statistic was significan t all of the 2,000 valid model 116

PAGE 117

comparisons. The average AIC for the Poisson m odel was 3642.82 while the average AIC for the Hurdle model was 3023.52, which descriptively suppor ts the inferential findi ngs of better fit for the Hurdle model. For the .50 proportion of zeros condition with a normal distribution, the deviance statistic ranged from 356.28 to 489.60 with a mean of 423. 12 and a standard deviation of 21.81. Given a Type I error rate of .05, the deviance statistic was significant for all of the 1,601 valid model comparisons. The average AIC for the mode l was 3356.12 while the average AIC for the negative binomial Poisson model was 2994.57, wh ich descriptively supports the inferential findings of better fit for the negative binomial Hurdle model. For the .50 proportion of zeros condition with a normal distribution, the deviance statistic ranged from -2.57 to -1.93 with a mean of -2.01 and a standard deviation of .029 Given a Type I error rate of .05, the devi ance statistic was significant for none of the 2,000 valid model comparisons. The average AIC for the ZIP m odel was 2992.55 while the average AIC for the negative binomial ZIP model was 2996.55, which descriptively suggests better fit for the ZIP model. Table 4-20 summarizes the log-likelihood statistics for the normal distribution model comparisons with a .50 proportion of zeros. In this table, LL represents the mean log-likelihood for the pseudo-population, L LminLLmax LLsrepresents the mean log-likeli hood for the simulated dataset samples, and LL represent the log-likelihood minimu m and maximum for the simulated dataset sample, and LLs represent the standard deviation a nd standard error for the simulated dataset samples, and n represents the sample size for the simulated dataset samples. A comparison of AICs over all models revealed a minimum of 2992.55 and a maximum of 3926.85 across all 12,000 valid model comparis ons. Between the six models, the minimum 117

PAGE 118

was 3023.52 (i.e., the Hurdle model), and the ma ximum was 3783.75 (i.e., the Poisson model). The rank order from lowest AIC to highest AI C was as follows: Hurdle, ZIP, negative binomial Hurdle, negative binomial ZIP, negative binomial Poisson, Poisson. Thes e rankings are not in line with the rankings for the pseudo-population AI Cs; in the population, the negative binomial Hurdle model has a lower AIC (i.e., better fit) than the ZIP model. Ta ble 4-21 displays the descriptive statistics for these two models. Figur e 4-21 is a boxplot illustrating these results. The following set of results pertains to the data in which 750 of the 1,000 observations in each of the 2,000 simulated datase ts were fixed at zero. The re maining 250 values were sampled from a normal pseudo-population distribution with the count proportions and frequencies displayed in Table 31 and Table 3-2. For the .75 proportion of zeros condition with a normal distribution, the deviance statistic ranged from 313.34 to 435.26 with a mean of 372. 68 and a standard deviation of 17.63. Given a Type I error rate of .05, the deviance statistic was significant for all of the 2,000 valid model comparisons. The average AIC for the Poisson m odel was 2930.85 while the average AIC for the negative binomial Poisson model was 2561.17, wh ich descriptively supports the inferential findings of better fit for the nega tive binomial Poisson model. For the .75 proportion of zeros condition with a normal distribution, the deviance statistic ranged from -36.75 to .10 with a mean of -2.29 a nd a standard deviation of .24. Given a Type I error rate of .05, the devi ance statistic was significant fo r none of the 1,491 valid model comparisons. The average AIC for the Hurdle model was 1952.86 while the average AIC for the negative binomial Hurdle model was 1957.07, which descriptively suppo rts the inferential findings of better fit for the Hurdle model. 118

PAGE 119

For the .75 proportion of zeros condition with a normal distribution, the deviance statistic ranged from 827.03 to 1132.06 with a mean of 983. 00 and a standard deviation of 47.54. Given a Type I error rate of .05, the deviance statistic was significant for all of the 2,000 valid model comparisons. The average AIC for the Poisson m odel was 2930.85 while the average AIC for the Hurdle model was 1952.86, which descriptively suppor ts the inferential findi ngs of better fit for the Hurdle model. For the .75 proportion of zeros condition with a normal distribution, the deviance statistic ranged from 499.96 to 701.73 with a mean of 608. 46 and a standard deviation of 30.96. Given a Type I error rate of .05, the deviance statistic was significant for all of the 1,490 valid model comparisons. The average AIC for the negative binomial Poisson model was 2560.17 while the average AIC for the negative binomial Hurdle model was 1957.07, which descriptively supports the inferential findings of better fit for the negative binomial Hurdle model. For the .75 proportion of zeros condition with a normal distribution, the deviance statistic ranged from -5.86 to -.64 with a mean of -2.01 and a standard deviation of .10. Given a Type I error rate of .05, the devi ance statistic was significant fo r none of the 2,000 valid model comparisons. The average AIC for the ZIP m odel was 1958.85 while the average AIC for the negative binomial ZIP model was 1962.87, which suggests adequate fit for the ZIP model. Table 4-22 summarizes the log-likelihood statistics for the normal distribution model comparisons with a .75 proportion of zeros. In this table, LL represents the mean log-likelihood for the pseudo-population, L LminLLmax LLsrepresents the mean log-likeli hood for the simulated dataset samples, and LL represent the log-likelihood minimu m and maximum for the simulated dataset sample, and LLs represent the standard deviation a nd standard error for the simulated dataset samples, and n represents the sample size for the simulated dataset samples. 119

PAGE 120

A comparison of AICs over all models revealed a minimum of 1918.91 and a maximum of 3103.14 across all valid model comparisons Between the six models, the minimum was 1952.86 (i.e., the Hurdle model), and the maximum was 2930.85 (i.e., the Poisson model). The rank order from lowest AIC to highest AIC w as as follows: Hurdle, negative binomial Hurdle, ZIP, negative binomial ZIP, negative binomial Po isson, Poisson. These rankings are in line with the rankings for the pseudo-popula tion AICs. Table 4-23 displays the descriptive statistics for these two models. Figure 4-22 is a b oxplot illustrating these results. The following set of results pertains to the data in which 900 of the 1,000 observations in each of the 2,000 simulated datase ts were fixed at zero. The re maining 100 values were sampled from a normal pseudo-population distribution with the count proportions and frequencies displayed in Table 31 and Table 3-2. For the .90 proportion of zeros condition with a normal distribution, the deviance statistic ranged from 157.17 to 249.62 with a mean of 204. 06 and a standard deviation of 14.38. Given a Type I error rate of .05, the deviance statistic was significant for all of the 2,000 valid model comparisons. The average AIC for the Poisson m odel was 1716.21 while the average AIC for the negative binomial Poisson model was 1514.15, wh ich descriptively supports the inferential findings of better fit for the nega tive binomial Poisson model. For the .90 proportion of zeros condition with a normal distribution, the deviance statistic ranged from -20.63 to 13.74 with a mean of -1.11 and a standa rd deviation of 1.62 Given a Type I error rate of .05, the deviance statistic was significant for none of the 1,537 valid model comparisons. The average AIC for the Hurdle model was 984.24 while the average AIC for the negative binomial Hurdle model was 987.32, which descriptively supports the inferential findings of better fit for the Hurdle model. 120

PAGE 121

For the .90 proportion of zeros condition with a normal distribution, the deviance statistic ranged from 583.77 to 893.39 with a mean of 737. 97 and a standard deviation of 49.17. Given a Type I error rate of .05, the deviance statistic was significant for all of the 2,000 valid model comparisons. The average AIC for the Poisson model was 1716.2 while the average AIC for the Hurdle model was 1514.15, which descriptively suppor ts the inferential findi ngs of better fit for the negative binomial Hurdle model. For the .90 proportion of zeros condition with a normal distribution, the deviance statistic ranged from 425.97 to 642.72 with a mean of 532. 39 and a standard deviation of 34.87. Given a Type I error rate of .05, the deviance statistic was significant for all of the 1,537 valid model comparisons. The average AIC for the model negative binomial Poisson was 1514.15 while the average AIC for the negative binomial Hurdle model was 984.24, which descriptively supports the inferential findings of better fit for the negative binomial Hurdle model. For the .90 proportion of zeros condition with a normal distribution, the deviance statistic ranged from -2.53 to -1.97 with a mean of -2.01 and a standard de viation of .02. Given a Type I error rate of .05, the devi ance statistic was significant fo r none of the 2,000 valid model comparisons. The average AIC for the ZIP m odel was 990.22 while the average AIC for the negative binomial ZIP model was 994.23 which descriptively suggests better fit for the ZIP model. Table 4-24 summarizes the log-likelihood statistics for the normal distribution model comparisons with a .90 proportion of zeros. In this table, LL represents the mean log-likelihood for the pseudo-population, L LminLLmaxrepresents the mean log-likeli hood for the simulated dataset samples, and LL represent the log-likelihood minimu m and maximum for the simulated 121

PAGE 122

dataset sample, and LLs LLs represent the standard deviation a nd standard error for the simulated dataset samples, and n represents the sample size for the simulated dataset samples. A comparison of AICs over all models revealed a minimum of 962.46 and a maximum of 1881.02 across all 11,536 valid model comparisons Between the six models, the minimum was 984.24 (i.e., the Hurdle model), and the maximu m was 1716.20 (i.e., the Poisson model). The rank order from lowest AIC to highest AIC w as as follows: Hurdle, negative binomial Hurdle, ZIP, negative binomial ZIP, negative binomial Po isson, Poisson. These rankings are in line with the rankings for the pseudo-popula tion AICs. Table 4-25 displays the descriptive statistics for these two models. Figure 4-23 is a b oxplot illustrating these results. Negatively Skewed Distribution The following set of results pertains to the data in which 100 of the 1,000 observations in each of the 2,000 simulated datase ts were fixed at zero. The re maining 900 values were sampled from a negatively skewed p seudo-population distribution wi th the count proportions and frequencies displayed in Ta ble 3-1 and Table 3-2. For the .10 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from -572.46 to -470.93 with a mean of -517.56 and a standard deviation of 13.66. Given a Type I error rate of .05, the deviance statistic was significant for all of the 2,000 model comparisons. The average AIC for the Poisson model was 4044.50 while the average AIC for the negative binomial Poisson model was 4563.58, which descriptively supports the inferential findings of better fit for the Poisson model. For the .10 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from 13.4 to 61.02 with a mean of 15.75 and a standard deviation of .28. Given a Type I error rate of .05, the devi ance statistic was signifi cant for all of the 1,626 122

PAGE 123

valid model comparisons. The average AIC for the Hurdle model was 3868.06 while the average AIC for the negative binomial Hurdle model was 3886.31, which descri ptively support the inferential findings of better fit for the Hurdle model. For the .10 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from 162.77 to 202.59 w ith a mean of 181.96 and a standard deviation of 5.72. Given a Type I error rate of .05, the deviance statistic was significant for all of the 2,000 model comparisons. The average AIC for the Po isson model was 4044.02 while the average AIC for the Hurdle model was 3868.06, which descriptiv ely support the inferen tial findings of better fit for the Hurdle model. For the .10 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from 609.07 to 754.25 w ith a mean of 682.91 and a standard deviation of 17.72. Given a Type I error rate of .05, th e deviance statistic was significant for all of the 1,626 valid model comparisons. The average AIC for the negative binomial Poisson model was 4563.58 while the average AIC for the negative binomial Hurdle model was 3886.31, which descriptively supports the inferential findings of better fit for the negative binomial Hurdle model. For the .10 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from -2.49 to -1.92 with a mean of -2.01 and a standard deviation of 0.03. Given a Type I error rate of .05, the devi ance statistic was significa nt for none of the 2,000 valid model comparisons. The average AIC fo r the ZIP model was 3874.07 while the average AIC for the negative binomial Hurdle model was 3878.08, which descriptivel y better fit for the ZIP model. 123

PAGE 124

Table 4-26 summarizes the log-likelihood statistics for the negatively skewed distribution model comparisons with a .10 proportion of zeros. In this table, LL represents the mean loglikelihood for the pseudo-population, L LminLLmaxLLLLsrepresents the mean log-likelihood for the simulated dataset samples, and represent the log-likelihood minimum and maximum for the simulated dataset sample, and LLs represent the standard deviat ion and standard error for the simulated dataset samples, and n represents the sample size for the simulated dataset samples. A comparison of AICs over all models revealed a minimum of 3819.85 and a maximum of 4598.21 across all 9,626 valid model comparis ons. Between the six m odels, the minimum was 3868.06 (i.e., the Hurdle model), and the ma ximum was 4563.58 (i.e., th e negative binomial Poisson model). The rank order fr om lowest AIC to highest AIC was as follows: Hurdle model, ZIP model, negative binomial ZI P model, negative binomial Hurd le model, Poisson model, and negative binomial Poisson model. These rankings are in line with the rankings for the pseudopopulation AICs. Table 427 displays the descriptive statisti cs for these two models. Figure 4-24 is a boxplot illustrating these results. The following set of results pertains to the data in which 250 of the 1,000 observations in each of the 2,000 simulated datase ts were fixed at zero. The re maining 750 values were sampled from a negatively skewed p seudo-population distribution wi th the count proportions and frequencies displayed in Ta ble 3-1 and Table 3-2. For the .25 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from 30.72 to 86.61 with a mean of 57.02 and a standard deviation of .99. Given a Type I error rate of .05, the devi ance statistic was signifi cant for of the all 2,000 model comparisons. The average AIC for the Po isson model was 4534.49 while the average AIC 124

PAGE 125

for the negative binomial Poisson model w as 4479.47, which descriptively supports the inferential findings of better fit for th e negative binomial Poisson model. For the .25 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from -48. 44 to -10.95 with a mean of -13. 01 and a standard deviation of 1.47. Given a Type I error rate of .05, the devi ance statistic was significa nt for of the all 1,517 valid model comparisons. The average AIC for the Hurdle model was 3807.17 while the average AIC for the negative binomial Hurdle model was 3822.26 which descriptively the inferential findings of better fit for the Hurdle model. For the .25 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from 674.99 to 784.98 w ith a mean of 733.31 and a standard deviation of 15.50. Given a Type I error rate of .05, th e deviance statistic was significant for of the all 2,000 valid model comparisons. The average AIC for the Poisson model was 4534.49, while the average AIC for the Hurdle model was 3807.17 which descriptively supports the inferential findings of better fit for the Hurdle model. For the .25 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from 602.99 to 723.43 w ith a mean of 663.28 and a standard deviation of 19.81. Given a Type I error rate of .05, th e deviance statistic was significant for all of the 1,517 valid model comparisons. The average AIC for the negative binomial Poisson model was 4479.47 while the average AIC for the negative binomial Hurdle model was 3822.26, which descriptively the inferential findings of better f it for the negative binomial Hurdle model. For the .25 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from -4.51 to -2.01 with a mean of -1.95 and a standard deviation of .06. Given a Type I error rate of .05, the devi ance statistic was signifi cant for none of the 2,000 125

PAGE 126

valid model comparisons. The average AIC fo r the ZIP model was 3813.16 while the average AIC for the negative binomial ZIP model was 38 17.17, which descriptively supports better fit for the ZIP model. Table 4-28 summarizes the log-likelihood statistics for the negatively skewed distribution model comparisons with a .25 proportion of zeros. In this table, LL represents the mean loglikelihood for the pseudo-population, L LminLLmaxLLLLsrepresents the mean log-likelihood for the simulated dataset samples, and represent the log-likelihood minimum and maximum for the simulated dataset sample, and LLs represent the standard deviat ion and standard error for the simulated dataset samples, and n represents the sample size for the simulated dataset samples. A comparison of AICs over all models revealed a minimum of 3765.15 and a maximum of 4576.05 across all 11,517 valid model comparis ons. Between the six models, the minimum was 3,807.17 (i.e., the Hurdle model), and the maximum was 4534.49 (i.e., the Poisson model). The rank order from lowest AIC to highest AI C was as follows: Hurdle model, ZIP model, negative binomial ZIP model, negative binom ial Hurdle model, negative binomial Poisson model, and Poisson model These rankings ar e in line with the rankings for the pseudopopulation AICs. Table 429 displays the descriptive statisti cs for these two models. Figure 4-25 is a boxplot illustrating these results. The following set of results pertains to the data in which 500 of the 1,000 observations in each of the 2,000 simulated datase ts were fixed at zero. The re maining 500 values were sampled from a negatively skewed p seudo-population distribution wi th the count proportions and frequencies displayed in Ta ble 3-1 and Table 3-2. For the .50 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from 618.37 to 694.3 with a mean of 658.45 and a standard deviation of 126

PAGE 127

9.97. Given a Type I error rate of .05, the de viance statistic was signifi cant for all 2,000 valid model comparisons. The average AIC for th e model Poisson model was 4747.63 while the average AIC for the negative binomial Poisson model was 4091.18, which descriptively supports the inferential findings of better fit for the negative binomial Poisson model. For the .50 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from -4.36 to 57.05 with a mean of 8.65 and a standard deviation of 2.17. Given a Type I error rate of .05, the devi ance statistic was significa nt for 1,665 of the 1,672 valid model comparisons. The average AIC for the Hurdle model was 3179.05 while the average AIC for the negative binomial Hurdle mode l was 3189.83 which descriptively the 99.58% inferential findings of better fit for the Hurdle model. For the .50 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from 1472.82 to 1677. 02 with a mean of 1574.58 and a standard deviation of 31.48. Given a Type I error rate of .05, the deviance statistic was significant for of all 2,000 valid model comparisons. The average AIC for the Poisson model was 4747.63 while the average AIC for the Hurdle model was 3179. 05 which descriptively supports the inferential findings of better fit for the Hurdle model. For the .50 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from 823. 6 to 984.79 with a mean of 907.16 and a standard deviation of 24.35. Given a Type I error rate of .05, the de viance statistic was signifi cant for all 1,672 valid model comparisons. The average AIC for the negative binomial Poisson model was 4091.18 while the average AIC for the negative binomia l Hurdle model was 3189.83, which descriptively supports the inferential findings of better fit for the negative binomial Hurdle model. 127

PAGE 128

For the .50 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from -2.36 to -1.87 with a mean of -2.00 and a standard deviation of .02. Given a Type I error rate of .05, the devi ance statistic was signifi cant for none of the 1,998 valid model comparisons. The average AIC for the ZIP model was 3184.98 while the average AIC for the negative binomial ZIP model was 31 80.98, which descriptively supports better fit for the negative binomial ZIP model. Table 4-30 summarizes the log-likelihood statistics for the negatively skewed distribution model comparisons with a .50 proportion of zeros. In this table, LL represents the mean loglikelihood for the pseudo-population, L LminLLmaxLLLLsrepresents the mean log-likelihood for the simulated dataset samples, and represent the log-likelihood minimum and maximum for the simulated dataset sample, and LLs represent the standard deviat ion and standard error for the simulated dataset samples, and n represents the sample size for the simulated dataset samples. A comparison of AICs over all models revealed a minimum of 3145.34 and a maximum of 4850.32 across all valid model comparisons Between the six models, the minimum was 3179.05 (i.e., the Hurdle model), and the maximum was 4747.63 (i.e., the Poisson model). The rank order from lowest AIC to highest AIC was as follows: Hurdle model, negative binomial ZIP model, ZIP model, negative binomial Hurdle model, negative binomial Poisson model, and Poisson model. These rankings are in line with the rankings for the pseudo-population AICs. Table 4-31 displays the descriptive statistics for these two models. Figure 4-26 is a boxplot illustrating these results. The following set of results pertains to the data in which 750 of the 1,000 observations in each of the 2,000 simulated datase ts were fixed at zero. The re maining 250 values were sampled 128

PAGE 129

from a negatively skewed p seudo-population with the count proportions a nd frequencies displayed in Table 31 and Table 3-2. For the .75 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from 647.7 to 760.47 with a mean of 704.52 and a standard deviation of 16.27. Given a Type I error rate of .05, the deviance statistic was significant for all 2,000 valid model comparisons. The average AIC for the Po isson model was 3845.4 while the average AIC for the negative binomial Poisson model w as 3142.87, which descriptively supports the inferential findings of better fit for th e negative binomial Poisson model. For the .75 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from -171. 64 to 9.79 with a mean of -4.89 and a standard deviation of 5.61. Given a Type I error rate of .05, the de viance statistic was signi ficant for 1418 of the 1563 valid model comparisons. The average AIC for the Hurdle model was 2027.47 while the average AIC for the negative binomial Hurdle model was 2034.39, which descri ptively supports the 90.72% inferential findings of better fit for the Hurdle model. For the .75 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from 1472.82 to 1677. 82 with a mean of 1574.58 and a standard deviation of 31.48. Given a Type I error rate of .05, the deviance statistic was significant for all of the 2,000 valid model comparisons. The averag e AIC for the Poisson model was 3845.4 while the average AIC for the Hurdle model was 2027. 47 which descriptively supports the inferential findings of better fit for the Hurdle model. For the .75 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from 976.32 to 1218.4 w ith a mean of 1114.55 and a standard deviation of 31.27. Given a Type I error rate of .05, th e deviance statistic was significant for all of the 129

PAGE 130

1,562 valid model comparisons. The average AIC for the negative binomial Poisson model was 3142.87 while the average AIC for the negative binomial Hurdle model was 2034.39 which descriptively supports the findings of better fit for the negative binomial Hurdle model. For the .75 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from -8.54 to -1.59 with a mean of -2.01 and a standard deviation of .15. Given a Type I error rate of .05, the devi ance statistic was signifi cant for 1of the 1,999 valid model comparisons. The average AIC for the ZI P model was 2033.45 while the average AIC for the negative binomial ZIP model was 2037.46, whic h descriptively supports the one inferential finding of better fit for the ZIP model. Table 4-32 summarizes the log-likelihood statistics for the negatively skewed distribution model comparisons with a .75 proportion of zeros. In this table, LL represents the mean loglikelihood for the pseudo-population, L LminLLmaxLLLLsrepresents the mean log-likelihood for the simulated dataset samples, and represent the log-likelihood minimum and maximum for the simulated dataset sample, and LLs represent the standard deviat ion and standard error for the simulated dataset samples, and n represents the sample size for the simulated dataset samples. A comparison of AICs over all models revealed a minimum of 2000.2 and a maximum of 3994.94 across all 11,561 valid model comparisons Between the six models, the minimum was 2027.47 (i.e., the Hurdle model), and the maxi mum was 3845.4 (i.e., the Poisson model ). The rank order from lowest AIC to highest AIC was as follows: Hurdle model, ZIP model, negative binomial Hurdle model, negative binomial ZIP model, negative binomial Poisson model, and Poisson model. These rankings are, for the most part, in line with the rankings for the pseudopopulation AICs. However, in the pseudo-population, the AICs for the Hurdle model and the 130

PAGE 131

negative binomial Hurdle model were tied at 2020.3. Table 4-33 displays the descriptive statistics for these two models Figure 4-27 is a boxplot i llustrating these results. The following set of results pertains to the data in which 900 of the 1,000 observations in each of the 2,000 simulated datase ts were fixed at zero. The re maining 100 values were sampled from a negatively skewed p seudo-population distribution wi th the count proportions and frequencies displayed in Table 3-1 and Table 3-2. For the .90 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from 334.64 to 438.18 w ith a mean of 389.67 and a standard deviation of 14.71 Given a Type I error rate of .05, th e deviance statistic was significant for all of the 2,000 valid model comparisons. The average AIC for the Poisson model was 2311.08, while the average AIC for the negative binomial Poiss on model was 1923.4, which descriptively the inferential findings of better fit for th e negative binomial Poisson model. For the .90 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from -40. 33 to -1.06 with a mean of -2.03 and a standard deviation of 2.11 Given a Type I error rate of .05, the de viance statistic was signifi cant for 57 of the 1,478 valid model comparisons. The average AIC for the Hurdle model was 1017.47, while the average AIC for the negative binomial Hurdle model was 1021.39 which descript ively supports findings of better fit for the Hurdle model; inferentially, better fit was found for the negative binomial Hurdle model for 3.9% of the simulations. For the .90 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from 1119.3 to 1443.11 with a mean of 1299.61 and a standard deviation of 48.19. Given a Type I error rate of .05, the deviance statistic was significant for all of the 2,000 valid model comparisons. The av erage AIC for the Poisson model was 2311.08 131

PAGE 132

while the average AIC for the Hurdle model was 1017.47, which descri ptively supports the inferential findings of better fit for the Hurdle model. For the .90 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from 783.55 to 997.99 w ith a mean of 907.75 and a standard deviation of 33.26. Given a Type I error rate of .05, th e deviance statistic was significant for all of the 1,478 valid model comparisons. The average AIC for the negative binomial Poisson model was 1923.40 while the average AIC for the negative binomial Hurdle model was 1021.39 which descriptively supports the inferential findings of better fit for the negative binomial Hurdle model. For the .90 proportion of zeros condition w ith a negatively skewed distribution, the deviance statistic ranged from -3.33 to -1.97 with a mean of -2.01 and a standard deviation of .04. Given a Type I error rate of .05, the devi ance statistic was signifi cant for none of the 2,000 valid model comparisons. The average AIC fo r the ZIP model was 1023.46 while the average AIC for the negative binomial ZIP model was 10 27.47, which descriptively supports better fit for the ZIP model. Table 4-34 summarizes the log-likelihood statistics for the negatively skewed distribution model comparisons with a .90 proportion of zeros. In this table, LL represents the mean loglikelihood for the pseudo-population, L LminLLmaxLLLLsrepresents the mean log-likelihood for the simulated dataset samples, and represent the log-likelihood minimum and maximum for the simulated dataset sample, and LLs represent the standard deviat ion and standard error for the simulated dataset samples, and n represents the sample size for the simulated dataset samples. A comparison of AICs over all models revealed a minimum of 997.99 and a maximum of 2454.31 across all 11,478 valid model comparisons Between the six models, the minimum was 132

PAGE 133

1017.47 (i.e., the Hurdle model), and the maxi mum was 2311.08 (i.e., the Poisson model ). The rank order from lowest AIC to highest AIC w as as follows: Hurdle model, negative binomial Hurdle model, negative binomial Poisson, ZIP m odel, negative binomial ZIP model, and Poisson model. These rankings are in line with the rank ings for the pseudo-popula tion AICs. Table 4-35 displays the descriptive statistics for these two models. Figure 4-28 is a boxplot illustrating these results. Review of Positively Skewed Distribution Findings For the .50, .75, and .90 proportions of zer os conditions, the negative binomial Poisson model displayed significantly better fit than the Poisson model; however, for the .10 and .25 proportions of zeros conditions, the Poisson model fit significantly better than the negative binomial Poisson model. For all conditions, the Hurdle model displayed significantly better fit than the Poisson model except at the .25 propor tion of zeros condition where the Hurdle model displayed better fit for 25.2% of the simulations. The negative binomial Hurdle model displayed significantly better fit than the negative binomial Poi sson model for all proportions zeros conditions and all simulations except for the .50 proportion of zeros condition in which the negative binomial Hurdle disp layed significantly be tter fit for .133 of the simulations. There was considerable variability in result s comparing the Hurdle model and negative binomial Hurdle model as well as results compar ing the ZIP model and negative binomial ZIP model. The negative binomial Hurdle displayed significantly better fit for all simulations when the proportion of zeros was .10 and .50. The proportion of simulations favoring the negative binomial Hurdle model dropped to .831 when the proportion of zeros of .75. This proportion significant dropped further to .307 for the .90 proportion of zeros conditions and dropped further yet to .252 for the .25 proportion of zeros conditions. 133

PAGE 134

There were no conditions in which the negative binomial ZIP model displayed significantly better fit than th e ZIP model. The proportion of such cases was .947 at the .50 proportion of zeros condition. The proportion dro pped to .652 at the .75 proportion of zeros. There were only .133 and .186 simu lations favoring the negative binomial ZIP model at .25 and .90 zeros, respectively. Finally, there were no simulations favoring the negative binomial ZIP model when the proportion of zeros was .10. However, as displayed in Table 4-42, there were convergence problems when the proportion of zero s was .10 and .25. Table 4-36 displays the percentage of simulations favoring the more complex model. In terms of AIC, the results were similarl y inconsistent. The negative binomial Hurdle model typically displayed the best fit; however the negative binomial Poisson model, which displayed the worst fit for the .10 and .25 proporti on of zeros conditions then displayed best fit for the .50 proportion of zeros condition. The Hu rdle model displayed the second best fit when the proportion of zeros was .10. However, its ra nk decreased over subsequent simulations from third to fourth best-fitting model. The Poisson model also displayed inconsistent findings. For the .10 proportion of zeros condition, this mode l was the third best-fi tting model. It then improved in fit at the .25 proportion of zeros condition but then became the worst fitting model for the remaining proportion of zeros conditions. The ZIP model displayed fourth best fit when the proportion of zeros was .10 and .25; this rank was superseded by the negativ e binomial ZIP model. However, the results reversed for the .50 and .75 proportion of zeros conditions and re versed again at the .90 proportion of zeros conditions. It is important to note that these resu lts are based on AICs; inspection of previous AIC figures suggests that the ra nk order is purely for compar ison since the actual AIC values may be nearly equal between models and/or pr oportion of zeros conditio ns. For example, the 134

PAGE 135

superiority of the Hurdle model over the ZIP mo del .25 zeros is based on an AIC difference of 3.22 while the superiority of the Hurdle model over the ZIP model at .10 zeros is based on an AID difference of 129.04. Table 4-37 presents th e AICs for the positively skewed distribution over all models and proportions of zeros. Figur e 4-29 illustrates the rank-ordering of these AICs. Review of Normal Distribution Findings The results for the normally distributed event stage were consistent for all proportion of zeros. For all proportions, the Hurdle model displa yed the best fit. Ther e were no significant findings of support for the negative binomial Hu rdle model. The two models did become more similar in terms of AICs at the .75 and .90 pr oportions of zeros. This same pattern was found for the ZIP models. However, it is important to not e that at the .25 and higher proportion of zeros conditions, Hurdle model comparisons converged for approximately 75% of the simulations while the ZIP model comparisons converged for 100% of the model comparisons. The convergence frequencies are displa yed in Table 4-43. Both infere ntially and descriptively, the Poisson model was superior to the negative binomial Poisson model when the proportion of zeros was .10 or .25; the reverse was true when the proportion of zeros was .50, .75, or .90. Table 4-38 displays the percentage of simulations significantly favoring the more complex model. It is important to note that these results ar e based on AICs; inspect ion of previous AIC figures suggests that the rank orde r is purely for comparison since the actual AIC values may be nearly equal between models and/or proporti on of zeros conditions. For example, for all proportions of zeros, the difference in AICs between the ne gative binomial Hurdle, ZIP, and ZIP models are quite small. Table 4-39 presents the AICs for the normal distribution over all models and proportions of zeros Figure 4-30 illustrates the rank-ordering over conditions. 135

PAGE 136

Review of Negatively Skew ed Distribution Findings For all conditions except for the .10 proporti on of zeros, the negative binomial Poisson model displayed significantly be tter fit than the Po isson model. Between all zero proportion conditions, the Hurdle model consistently displaye d significantly better fit when compared to the Poisson model while the negative binomial Hurd le model consistently displayed significantly better fit when compared to the negative binomial Poisson model. In general, the Hurdle model displayed significantly better fit than the negative binomial Hurdle model; however, the results should be inte rpreted cautiously since for all conditions, the negative binomial Hurdle had convergence problems of roughly 25%. This is displayed in Table 4-44. The negative binomial ZIP model never di splayed significantly better fit than the ZIP model. Table 4-40 displays the percentage of simulations significantly favoring the more complex model. Descriptively, the Hurdle model displayed the best fit regardless of the proportion of zeros. The Poisson model displayed the worst fit excep t for the .10 proportion of zeros condition in which the negative binomial Poisson model displaye d the worst fit. Consequently, the next worst fitting model was this negative binomial Poisson m odel except for the .10 proportion of zeros. In between the best fitting and tw o worst fitting models, for the .10 and .25 proportions of zeros conditions, the ZIP model fit better than the negative binomial ZIP model, which fit better than the negative binomial Hurdle model. For the .50 pr oportion of zeros conditi on, the rank order for the ZIP and negative binomial ZIP was reversed with the negative binomial ZIP model displaying better fit. However, for the .75 propo rtion of zeros condition the ZIP model returned to being the second best-fitting model. Finally, for the .90 proportion of zeros condition the negative binomial Hurdle model gained in rank again to become the second best-fitting model. 136

PAGE 137

The negative binomial Poisson also rose in rank to be the third best-fitting model. This was followed by the ZIP and negativ e binomial ZIP models. It is important to note that these results ar e based on AICs; inspect ion of previous AIC figures suggests that the rank orde r is purely for comparison since the actual AIC values may be nearly equal between models and/or proportion of zeros conditions. For example, although the AIC is lower for the Hurdle models than for the negative binomial Hurdle models, the difference between the AICs declines sharply as the propor tion of zeros increases. Table 4-41 presents the AICs for the negatively skewed distribution over all models and proportions of zeros. Figure 431 illustrates the rank-ordering over conditions. 137

PAGE 138

Table 4-1. Deviance statistics comparing Po isson and negative binomial Poisson models Distribution Zeros Positive Normal Negative .10 -238.40 -384.40 -517.40 .25 -82.60 -35.20 -1534.80 .50 88.40 331.00 658.40 .75 125.00 372.40 704.60 .90 69.80 203.00 389.60 Table 4-2. Deviance statistics comparing Hu rdle and negative binomial Hurdle models Distribution Zeros Positive Normal Negative .10 27.60 -6.20 -15.60 .25 20.40 -5.20 -13.00 .50 15.40 -3.60 -8.40 .75 8.20 -1.80 -4.40 .90 3.20 -1.80 -1.80 Table 4-3. Deviance statistics comp aring Poisson and Hurdle models Distribution Zeros Positive Normal Negative .10 131.60 19.20 181.80 .25 5.20 252.60 859.60 .50 75.40 756.80 1573.20 .75 215.40 983.00 1823.20 .90 203.20 732.40 1298.20 Table 4-4. Deviance statistics compari ng NB Poisson and NB Hurdle models Distribution Zeros Positive Normal Negative .10 397.60 397.40 683.60 .25 108.20 282.60 662.20 .50 2.40 422.20 906.40 .75 98.60 608.80 1114.20 .90 136.60 527.60 906.80 138

PAGE 139

Table 4-5. Deviance statistics compari ng ZIP and negative binomial ZIP models Distribution Zeros Positive Normal Negative .10 -2.00 -2.00 -2.00 .25 -0.60 -2.00 -2.00 .50 14.20 -2.00 -2.00 .75 6.20 -2.00 -2.00 .90 1.20 -2.00 -2.00 Table 4-6. Log-likelihood comparisons for positively skewed distri bution with .10 zeros Model LL L LminLLmaxLLLLs LLs n Poisson -1475.3 -1475.53 -1539.15 -1409.23 19.27 0.430995 2000 1 NB Poisson -1594.5 -1594.65 -1650.38 -1539.67 15.79 0.353131 2000 Hurdle -1409.5 -1409.40 -1495.97 -1315.41 26.27 0.587599 2000 2 NB Hurdle -1395.7 -1395.44 -1483.98 -1304.26 25.68 0.574254 2000 Poisson -1475.3 -1475.53 -1539.15 -1409.23 19.27 0.430995 2000 3 Hurdle -1409.5 NB -1409.40 -1495.97 -1315.41 26.27 0.587599 2000 Poiss on -1594.5-1594.65 -1650.38-1539.6715.790.3531312000 4 le 1 NB Hurd -1395.7 -1395.44 -1483.98 -1304.26 25.68 0.574254 2000 ZIP -1478.3 -1477.92 -1542.15 -1412.24 19.8380 0.505194 542 5 NB ZIP -1479.3-1485.34 -1531.90-1439.61 18.7300 0 1.570000136 139

PAGE 140

Table 4-7. AICs for positively skewed distri bution models with a .10 proportion of zeros Model AIC A ICmin A ICmax A ICAICs A ICs n Poisson 2956.6 2957.06 2824.47 3084.30 38.55 0.86 2000 NB Poisson le 2 IP 2962.6 2967.852836.473096.3039.68 1.01 1542 136 3194.9 3193.30 3083.34 3304.77 31.58 0.71 2000 Hurdle 2822.9 2838.81 2650.83 3011.94 52.56 1.18 2000 NB Hurd 2975.4 2804.88 2622.53 2981.96 51.36 1.15 000 Z NB ZIP 2964.6 2984.793077.792984.6936.55 3.13 Table 4ics vd d ti zeros Model LL-8. Log-l kelihood omparison for positi ely skewe istri bu on with .25 L L min max LLs LL LL LLn s Poisson -1468.1 -1466.83-1542.31-1388.7722.760.509050 2000 1 NB Poisson -1509.4 -1508.43-1561.66-1448.0817.540.392323 2000 Hurdle -1465.5 -1463.95-1542.28-1377.4124.290.543358 2000 2 NB Hurdle -1455.3 -1453.50-1523.70-1370.0323.670.529329 2000 Poisson -1468.1 -1466.83-1542.31-1388.7722.760.509050 2000 3 Hurdle -1465.5 -1463.95-1542.28-1377.4124.290.543358 2000 NB Poisson -1509.4 -1508.43-1561.66-1448.0817.540.392323 2000 4 NB Hurdle -1455.3 -1453.50-1523.70-1370.0323.670.529329 2000 ZIP -1471.1 -1469.56-1544.96-1391.7722.710.511613 1972 5 NB ZIP -1471.4 -1471.07-1537.28-1407.3620.360.470331 1874 140

PAGE 141

Table 4-9. AICs for positively skewed distri bution models with a .25 proportion of zeros Model AIC A ICmin A ICmax A ICAICs A ICs n Poisson 2942.3 2939.654 2783.538 3090.621 45.53082 1.018100 2000 NB Poisson 3026.8 3020.851 2900.160 3127.329 35.09046 0.784647 2000 Hurdle 2943.0 2947.899 2774.825 3104.551 48.59943 1.086716 2000 NB Hurdle 2922.5 2921.001 2754.053 3061.392 47.34459 1.058657 2000 ZIP 2954.3 2951.124 2795.538 3101.917 45.43863 1.023227 1972 NB ZIP 2956.7 2956.147 2828.716 3088.553 40.72097 0.940661 1874 Table 4-10. Log-likelihood comparisons for pos itively skeed distrLLw i bution with .50 zeros Model L LminLLmaxLLLLs LLs n Poisson -1334.3 -1333.44 -1428.47 -1239.52 29.11 0.650 2000 1 -1290.1 -1289.49-1361.70-1223.9021.340.477 2000 Hurdle -1296.6 -1294.96-1357.10-1221.9020.800.465 2000 NB Poisson 2 NB Hurdle -1288.9 -1287.12-1352.33-1220.4220.260.906 2000 Poisson -1334.3 -1333.44-1428.47-1239.5229.110.650 2000 3 Hurdle -1296.6 -1294.96-1357.10-1221.9020.800.465 2000 NBPoisson -1290.1 -1289.49-1361.70-1223.9021.340.477 2000 4 NB Hurdle -1288.9 -1287.12-1352.33-1220.4220.260.906 2000 ZIP -1299.5 -1297.92-1360.09-1224.6320.800.465 2000 5 NB ZIP -1292.4 -1291.12-1356.33-1224.0020.170.451 2000 141

PAGE 142

Table 4-11. AICs for positively skewed distribution models with a .50 proportion of zeros Model AIC A ICmin A ICmax A ICAICs A ICs n Poisson 2674.5 2672.876 2485.036 2862.935 58.22 1.301853 2000 NB Poisson e IP 2611.1 2607.8402461.2622732.18941.610.930524 2000 2000 2588.3 2582.975 2451.801 2727.394 42.68 0.954555 2000 Hurdle 2605.1 2601.929 2455.800 2726.197 41.60 0.930422 2000 NBHurdl 2589.8 2588.240 2454.830 2718.660 40.52 0.905000 2000 Z NB ZIP 2598.8 2596.2322462.0042726.65440.350.902392 Table 4-12. Log-likelihood comor p ewed distri bution with .75 zeros Model LLparisons f os itively sk L L minLL max LLs LL LLsn Poisson -972.9 -974.608-1088.690-853.76131.925 0.7132000 1 on -910.4 -911.681-999.070-819.43823.872 0.5332000 Hurdle -865.2 -864.852-907.808-806.14614.387 0.3212000 NB Poiss 2 NB on Hurdle -861.1 -860.701 -905.917 -801.635 14.148 0.316 2000 Poiss -972.9 -974.608 1088.690 -853.761 31.925 0.713 2000 3 Hurdle on -865.2 -864.852 -907.808 -806.146 14.387 0.321 2000 NB Poiss -910.4 -911.681 -999.070 -819.438 23.872 0.533 2000 4 NB Hurd le ZIP -868.2 -867.858-910.807-809.10514.382 0.3212000 -861.1 -860.701 -905.917 -801.635 14.148 0.316 2000 5 NB ZIP -865.1 -864.721-909.922-805.61814.135 0.3162000 142

PAGE 143

Table 4-13. AICs for positively skewed distribution models with a .75 proportion of zeros Model min A IC A IC s A ICsAIC A IC AIC max n Poisson 1951.7 1955.215 1713.523 2183.381 63.85 1.42776 2000 NB Poisson le IP 1750.5 1747.7161630.21001833.61528.76 0.643202000 2000 1828.7 1827.363 1642.875 2002.140 47.74 1.06762 2000 Hurdle 1742.5 1749.705 1632.292 1835.616 28.77 0.64343 2000 NB Hurd 1734.3 1735.402 1617.270 1825.834 28.29 0.63273 2000 Z NB ZIP 1744.3 1743.4421625.2371833.84428.27 0.63215 Table 4-14. Log-likelihood compa p skribu h .9 Model risons for os itively ewed dist t ion wit 0 zerosLL L L mi ma snLLxLLLL LLsn Poisson -546.5 -546.397-653.705-453.19029.33 0.6552000 1 on -511.6 -511.489-594.277-437.84823.20 0.5182000 Hurdle -444.9 -443.180-469.812-408.2419.020 0.2012000 NB Poiss 2 NB Hurdle -443.3 -441.750 -468.410 -410.220 8.848 0.198 1992 Poisson -546.5 -546.397 -653.705 -453.190 29.33 0.655 2000 3 Hurdle -444.9 -443.180 -469.812 -408.241 9.020 0.201 2000 NB Poisson -511.6 -511.489 -594.277 -437.848 23.20 0.518 2000 4 NB Hurd le ZIP -447.9 -446.170-472.716-411.1429.021 0.2012000 -443.3 -441.750 -468.410 -410.220 8.848 0.198 1992 5 NB ZIP -447.3 -445.691-472.314-412.1428.876 0.1982000 143

PAGE 144

Table 4-15. AICs for positively skewed distribution models with a .90 proportion of zeros Model A IC A IC s A ICsAIC A IC AIC min max n Poisson 1099.0 1098.794 912.398 1313.41 58.66007 1.311679 2000 NB Poisson 11 le IP 907.7 904.339834.283957.43218.04285 0.4034502000 399702000 031.2 026.978 879.696 1192.55 46.40218 1.037584 2000 Hurdle 901.8 906.361 836.481 959.623 18.04078 0.403404 2000 NB Hurd 898.7 897.499 834.439 950.819 17.69794 0.396532 1992 Z NB ZIP 908.7 905.381838.283958.62717.75302 0.6 Table 4-16. Log-likelihood com oistritos Model LLparisons f r normal d ibution w h .10 zer L L min maxLL LL LL s LLs n Poisson -1806.7 -1806.38-1834.64-1776.0810.080 0.2252000 1 on -1998.9 -1998.93-2031.97-1961.239.706 0.2172000 -1822.85-1768.77-1822.859.086 0.2032000 NB Poiss Hurdle -1797.1 2 NB Hurdle -1800.2 -1825.6 -1772.63 -1825.60 9.330 0.240 1503 Poisson -1806.7 -1806.38 -1834.64 -1776.08 10.080 0.225 2000 3 Hurdle -1797.1 -1822.85 -1768.77 -1822.85 9.086 0.203 2000 NB Poisson -1998.9 -1998.93 -2031.97 -1961.23 9.706 0.217 2000 4 NB Hurd le ZIP -1800.1 -1799.53-1825.87-1771.789.085 0.2032000 -1800.2 -1799.77 -1825.60 -1772.63 9.330 0.240 1503 5 NB ZIP -1801.1 -1800.63-1826.87-1772.789.042 0.2031982 144

PAGE 145

Table 4-17. AICs for normal distribution models with a .10 proportion of zeros Model A IC A IC s A ICsAIC A ICAIC min max n Poisson 3619.3 3618.760 3558.16 0 3675.27 4 20.17 0.45 2000 NB Poisson 4005.9 4005.863 3930.44071.9 19.410.43 2000 le Hurdle 0.40 1982 5 2 3 4 6 9 Hurd 3606.1 3605.028 3549.5 2 3657.6 5 18.17 0.40 2000 NB 3612.3 3613.546 3559.26 1 3665.19 5 18.66 0.48 1503 ZIP 3612.1 3611.051 3555.56 2 3663.73 3 18.17 0.40 2000 NB ZIP 3616.1 3615.268 3559.56 2 3667.73 9 18.08 TabLooi ormution w zeros el le 4-18. g-likelih od compar sons for n al distrib ith .25 ModLL L L minLL max LLLLs LLs n Poisson -1913.9 -1913.12 -1968.17 -1869.47 13.99 0.312 2000 1 NB Poisson -1931.5 -1930.99 -1979.33 -1889.82 11.77 0.263 2000 Hurdle -1787.6 -1787.24 -1813.49 -1758.97 8.104 0.181 2000 2 NB Hurdle -1790.2 -1789.81 -1816.07 -1764.27 8.152 0.201 1634 Poisson -1913.9 -1913.12 -1968.17 -1869.47 13.999 0.312 2000 3 Hurdle -1787.6 -1787.24 -1813.49 -1758.97 8.104 0.181 2000 NB Poiss on 11.775 -1931.5 -1930.99 -1979.33 -1889.82 0.263 2000 4 NB Hurdle -1790.2-1789.81 -1816.07-1764.278.1520.201 1634 ZIP -1790.6-1790.22 -1816.47-1761.968.1050.181 2000 5 NB ZIP -1791.6-1791.22 -1817.47-1762.978.1050.181 2000 145

PAGE 146

Table 4-19. AICs for normal distribution models with a .25 proportion of zeros Model A IC A ICAICs AIC A IC min max A ICs n Poisson 3833.8 3832.239 3744.945 3942.340 27.99438 0.625973 2000 NB Poisson le IP 3593.1 3592.4363535.9173644.93816.211450.362499 2000 NB ZIP 3597.1 3596.4483539.9443648.93916.211050.362490 2000 3871.0 3869.986 3787.645 3966.654 23.55270 0.526654 2000 Hurdle 3587.2 3586.473 3529.931 3638.976 16.20934 0.362452 2000 NB Hurd 3592.3 3593.629 3542.533 3646.150 16.30441 0.403347 1634 Z 146

PAGE 147

Table 4-20. Log-likelihood comparisons for normal distribution with .50 zeros LLsminLLmaxLLLLs Model LL L L n Poisson -1888.5 -1888.88 -1960.42 -1818.41 20.88 0.466 2000 1 NB Poisson -1723.0 -1723.25 -1769.87 -1674.06 14.60 0.326 2000 Hurdle -1510.1 -1509.76 -1532.97 -1487.29 6.62 0.148 2000 2 NB Hurdle -1511.9 -1509.76 -1532.97 -1487.29 6.62 0.148 1602 Poisson -1888.5 -1888.88 -1960.42 -1818.41 20.88 0.466 2000 3 Hurdle -1510.1 NB -1509.76 -1532.97 -1487.29 6.627 0.148 2000 Poiss on -1723.0 -1723.25-1769.87-1674.0614.60 0.326 2000 4 le NB Hurd -1511.9 -1509.76 -1532.97 -1487.29 6.62 0.148 1602 ZIP -1513.1 -1512.75 -1535.96 -1490.28 6.61 0.147 2000 5 NB 2000 ZIP -1514.1 -1513.75-1536.96-1491.286.61 0.147 Table 4-21. AICs f ode 50 propof ze el or normal distribution m ls with a ortion ros Mod A IC A IC min A IC A maxIC A ICs A IC sn Poisson 3783.0 3783.75 3642.81 3926.846 41.76 0.93387 2000 NB Poisson 3454.0 3454.50 3356.11 3547.741 29.21 0.653052 2000 Hurdle 3023.2 3023.52 2978.57 3069.939 13.25 0.296204 2000 NB Hurdle 3035.8 3039.52 2994.573085.93913.250.296204 1602 ZIP 3038.3 3037.49 2992.553083.92813.230.295936 2000 NB ZIP 3042.3 3041.50 2996.553087.92813.230.295941 2000 147

PAGE 148

Table 4-22. Log-likelihood comparisons for normal distribution with .75 zeros Model LL L Ln x LL miLL maLL s LLs n Poisson -1461.8 -1462.42 -1548.57 -1371.53 27.106 0.606 2000 1 NB Poisson -1275.6 -1276.08 -1332.19 -1214.35 18.507 0.413 2000 Hurdle -970.3 -970.42 -987.29 -953.45 4.779 0.106 2000 2 NB Hurdle -971.2 -971.53 -992.37 -954.37 4.993 0.129 1490 Poisson -1461.8 -1462.42 -1548.57 -1371.53 27.106 0.606 2000 3 Hurdle -970.3 -970.42 -987.29 -953.45 4.779 0.106 2000 NB Poisson -1275.6 -1276.08 -1332.19 -1214.35 18.507 0.413 2000 4 NB Hurdle -971.2 -971.53 -992.37 -954.37 4.993 0.129 1490 ZIP -990.3 -973.30 -956.46 -990.29 4.770 0.106 2000 5 NB ZIP -991.3 -974.30 -957.45 -991.29 4.770 0.106 2000 Table 4 Model -23. As for noral distribution m a .7 proporon of zeros IC m odels with 5 ti n A IC A ICmi n A IC ma x A IC A ICs A ICs Poiss NB on 2929.6 2930.846 2749.052 3103.141 54.21258 1.21223 2000 Poisson 2559.3 2560.165 2436.693 2 le 1 1958.5 1958.852 1924.9221992.589.5575590.21371 2000 NB ZIP 1962.6 1962.866 1928.9181996.5819.5594520.21375 2000 672.373 37.0152 0.82768 2000 Hurdle 1952.6 1952.855 1918.906 1986.58 9.559016 0.21374 2000 NB Hurd ZIP 1954.4 1957.073 1922.741 998.756 9.987633 0.25874 1490 148

PAGE 149

Table 4-24. Log-likelihood comparisons for normal distribution with .90 zeros Model LL L Ln max LLs miLL LL LLs n Poisson -852.6 -855.101 -937.508 -767.641 26.52 0.593 1999 1 NB Poisson -751.1 -753.075 -812.698 -689.054 19.45 0.434 2000 Hurdle -486.4 -486.118 -494.801 -475.228 2.89 0.064 2000 2 NB Hurdle -487.3 -486.659 -498.835 -475.636 2.91 0.074 1537 Poisson -852.6 -855.101 -937.508 -767.641 26.52 0.593 1999 3 Hurdle -486.4 NB -486.118 -494.801 -475.228 2.89 0.064 2000 Poiss on -751.1 -753.075-812.698-689.05419.45 0.4342000 4 le NB Hurd -487.3 -486.659 -498.835 -475.636 2.91 0.074 1537 ZIP -489.4 -489.111 -497.799 -478.242 2.89 0.064 2000 5 -490.4 -490.114-498.799-479.2432.89 0.0642000 NB ZIP Table 4-25. AICs a onith orro for norm l distributi models w a .90 prop tion of ze s Model A IC A IC min A IC max A IC A IC s A IC sn Poisson 1711.1 1716.202 1541.281 1881.015 53.0534 1.18660 1999 NB Poisson 1 NB Hurdle 986.6 987.3187965.27281011.675.82673 0.148621537 ZIP 990.8 990.2214968.48481007.5975.79280 0.129532000 NB ZIP 994.8 994.2271972.48541011.5975.79265 0.129582000 510.2 1514.149 1386.108 1633.397 38.9068 0.86998 2000 Hurdle 984.8 984.2365 962.4563 1001.602 5.79776 0.12964 2000 149

PAGE 150

Table 4-26. Log-likelihood comparisons for ne gatively skewed distribution with .10 zeros Model LL L L miLLn maxLLLLs LLs n Poisson -2019.20 -2019.01 -2035.41 -2001.08 5.26 .12 2000 1 NB Poisson -2277.90 -2277.79 -2296.11 -2258.54 5.97 .13 2000 Hurdle -1928.30 -1928.03 -1948.88 -1907.92 5.82 .13 2000 2 NB Hurdle -1936.10 -1936.16 -1965.98 -1918.98 5.64 .14 1626 Poisson -2019.20 -2019.01-2035.41-2001.085.26 .122000 3 le -1928.30 -1928.03-1948.88-1907.925.82 .132000 on -2277.90 -2277.79-2296.11-2258.545.97 .132000 Hurd NB Poiss 4 le NB Hurd -1936.10 -1936.16 1965.98 -1918.98 5.64 .14 1626 ZIP -1931.30 -1931.03 1951.87 -1910.92 5.82 .13 2000 5 ZIP -1932.30 -1932.041952.87-1911.925.82 NB .132000 Table 4-27ry s el AICs fo negativel kewed models with a .10 proportion of zeros ModAIC A IC min A IC max A ICAICs A ICn s Poisson 4044.50 4044.02 4008.15 4076.81 10.53 .24 2000 NB Poisson 4563.90 4563.58 4525.08 4600.21 11.94 .27 2000 Hurdle 3868.70 3868.06 3827.85 3909.75 11.25 .28 2000 NB urdle 3884.10 3886.31 3851.963945.9711.25.28 1626 ZIP 3874.60 3874.07 3833.843915.7411.65.26 2000 NB ZIP 3878.70 3878.08 3837.853919.7511.65.26 2000 H 150

PAGE 151

Table 4-28. Log-likelihood comparisons for ne gatively skewed distribution with .25 zeros Model LLsminLLmaxLLLLsLL L L n Poisson -2264.3 -2264.24 -2285.03 -2239.14 7.13 0.16 2000 1 NB Poisson -2235.5 -2235.74 -2258.49 -2208.87 7.18 0.16 2000 Hurdle -1897.9 -1897.59 -1917.18 -1880.57 5.24 0.12 2000 2 NB Hurdle -1904.4 -1904.13 -1922.89 -1887.67 5.12 0.13 1517 Poisson -2264.3 -2264.24 -2285.03 -2239.14 7.13 0.16 2000 3 Hurdle -1897.9 -1897.59 -1917.18 -1880.57 5.24 0.12 2000 NB Poiss on -2235.5 -2235.74-2258.49-2208.877.18 0.162000 4 le -1904.4 -1904.13-1922.89-1887.675.12 0.131517 NB Hurd ZIP -1900.9 -1900.58 -1920.18 -1883.57 5.24 0.12 2000 5 -1901.9 -1921.18-1884.57 0.122000 NB ZIP -1901.58 5.24 Table 4-29. AICs for negatively skewed Model mAICodels with a .25 proportion of zeros A IC min A IC max A IC AIC s A ICs n Poisson 4534.7 4534.49 4484.28 4576.05 14.25 0.32 2000 NB Poisson B Hurdle 3820.8 3822.263789.343859.7810.250.26 1517 ZIP 3813.9 3813.163783.143852.3610.480.23 2000 NB ZIP 3817.9 3817.173783.153856.3610.480.23 2000 4479.0 4479.47 4425.75 4524.99 14.37 0.32 2000 Hurdle 3807.9 3807.17 3773.15 3844.37 10.47 0.23 2000 N 151

PAGE 152

Table 4-30. Log-likelihood comparisons for ne gatively skewed distribution with .50 zeros LLs Model LL L Lx minLL maLLLLs n Poisson -2370.5 -2370.81 -2422.16 -2317.09 14.38 0.32 2000 1 NB Poisson -2041.3 -2041.59 -2075.01 -2007.90 9.95 0.22 2000 Hurdle -1583.9 -1583.52 -1597.10 -1570.67 4.44 0.10 2000 2 NB Hurdle -1588.1 -1587.91 -1615.10 -1576.02 4.42 0.11 1672 Poisson -2370.5 -2370.81 -2422.16 -2317.09 14.38 0.32 2000 3 Hurdle -1583.9 -1583.52 -1597.10 -1570.67 4.44 0.10 2000 NB Poisson -2041.3 -2041.59 -2075.01 -2007.90 9.95 0.22 2000 4 NB Hurdle -1588.1 -1587.91 -1615.10 -1576.02 4.42 0.11 1672 ZIP -1586.9 -1586.49 -1600.10 -1573.67 4.43 0.10 2000 5 NB ZIP -1587.9 -1587.49 -1601.10 -1574.67 4.44 0.10 1998 Table 4-31. AICs for negatively skewed models with a .50 proportion of zeros Model A IC A ICmin A ICmax A IC A ICs A ICs n Poisson 4746.9 4747.63 4640.17 4850.32 28.76 0.64 2000 NB Poisson 4090.7 4091.18 4023.80 4158.02 1 ZIP 3185.7 3184.98 3156.343212.208.870.20 2000 NB ZIP 3181.7 3180.98 3155.343208.208.870.20 1998 9.89 0.44 2000 Hurdle 3179.7 3179.05 3153.34 3206.21 8.88 0.20 1672 NB Hurdle 3188.2 3189.83 3166.04 3244.19 8.83 0.22 2000 152

PAGE 153

Table 4-32. Log-likelihood comparisons for ne gatively skewed distribution with .75 zeros Model LL L L minLL maLLLLsx LLs n Poisson -1919.7 -1919.70 -1994.47 -1845.46 21.89 0.49 2000 1 NB Poisson -1567.4 -1567.44 -1614.24 -1519.31 13.88 0.31 2000 Hurdle -1008.1 -1007.74 -1019.17 -998.10 3.14 0.07 2000 2 NB Hurdle -1008.2 -1010.19 -1085.59 -1000.67 4.05 0.10 1562 Poisson -1919.7 -1919.70 -1994.47 -1845.46 21.89 0.49 2000 3 Hurdle -1008.1 -1007.74 -1019.17 -998.10 3.14 0.07 2000 NB Poiss on -1567.4 -1567.44-1614.24-1519.3113.88 0.31 2000 4 le NB Hurd -1008.2 -1010.19 -1085.59 -1000.67 4.05 0.10 1562 ZIP -1011.2 -1010.73 -1022.17 -1001.10 3.13 0.07 2000 5 -1012.2 -1011.73-1023.17-1002.103.13 0.07 1999 NB ZIP Table 4-33. AICs fo of z el r negatively skewed models with a .75 prop rtion o eros Mod A IC A IC min A IC max A IC A IsC A ICs n Poisson 3845.3 3845.40 3696.92 3994.94 43.78 0.98 2000 NB Poisson 3142.8 3142.87 3046.63 3236.48 2 NB Hurdle 2028.3 2034.39 2015.342185.198.110.21 1562 ZIP 2034.3 2033.45 2014.202056.346.260.14 2000 NB ZIP 2038.3 2037.46 2018.202060.346.270.14 1999 7.77 0.62 2000 Hurdle 2028.3 2027.47 2008.20 2050.34 6.27 0.14 2000 153

PAGE 154

Table 4-34. Log-likelihood comparisons for ne gatively skewed distribution with .90 zeros LLsminLLmaxLLLLsModel LL L L n Poisson -1152.1 -1152.54 -1224.16 -1064.51 23.17 0.52 2000 1 NB Poisson -957.3 -957.70 -1005.07 -897.19 15.88 0.36 2000 Hurdle -503.0 -502.73 -510.67 -496.99 2.09 0.05 2000 2 NB Hurdle -503.9 -503.70 -524.86 -498.07 2.31 0.06 1478 Poisson -1152.1 -1152.54 -1224.16 -1064.51 23.17 0.52 2000 3 Hurdle -503.0 -502.73 -510.67 -496.99 2.09 0.05 2000 NB Poisson -957.3 -957.70 -1005.07 -897.19 15.88 0.36 2000 4 NB Hurdle -503.9 -503.70 -524.86 -498.07 2.31 0.06 1478 ZIP -506.0 -505.73 -513.67 -499.99 2.09 0.05 2000 5 NB ZIP -507.0 -506.74 -514.67 -501.00 2.09 0.05 2000 Table 4-35. As for negtively skewed odels with a .90 proportion of zeros IC a m Model AIC A IC min A IC max A IC s AIC A ICsn Poisson 2310.2 2311.08 2135.01 2454.31 46.33 1.04 2000 NB Poiss le on 31.77 le ZIP 1024.1 1023.46 1011.991039.344.180.09 2000 NB ZIP 1028.1 1027.47 1015.991043.344.180.09 2000 1922.5 1923.40 1802.37 2018.13 0.71 2000 Hurd 1018.1 1017.47 1015.99 1033.35 4.18 0.09 2000 NB Hurd 1019.8 1021.39 1010.15 1063.71 4.61 0.12 1478 154

PAGE 155

Table 4-36. Positively skewed distribution: perc entage of simulations favoring complex model. Simple vs. Complex .10 .25 .50 .75 .90 P vs. NBP 0.0 0.0 100.0 100.0 100.0 H vs. NBH 100.0 99.7 98.0 98.1 30.7 P vs. H 100.0 25.2 100.0 100.0 100.0 NBP vs. NBH 100.0 100.0 100.0 100.0 100.0 Z vs. NBZ 0.0 18.6 94.7 65.2 13.3 Table 4-37. AICs: pos itively skewed distri bution (all conditions) Proportion of Zeros Model 0.10 0.25 0.50 0.75 0.90 Poisson 2957.06 2939.65 2672.88 1955.22 1098.79 NB Poisson 3193.30 3020.85 2582.98 1827.36 1026.98 Hurdle 2838.81 2947.90 2601.93 1749.71 906.36 NB Hurdle 2804.88 2921.00 2588.24 1735.40 897.50 ZIP 2967.85 2951.12 2607.84 1747.72 904.34 NB ZIP 2984.79 2956.15 2596.23 1743.44 905.38 Table 4-38. Normal distribu tion: percentage of simulations favoring complex model. Simple vs. Complex .10 .25 .50 .75 .90 P vs. NB P 0.0 0.0 100.0 100.0 100.0 H vs. NBH 0.0 0.0 0.0 0.0 0.0 P vs. H 100.0 100.0 100.0 100.0 100.0 NBP vs. NBH 100.0 100.0 100.0 100.0 100.0 Z vs. NBZ .4 0.0 0.0 0.0 0.0 Table 4-39. AICs: normal distribution (all conditions) Proportion of Zeros Model 0.10 0.25 0.50 0.75 0.90 Poisson 3618.76 3832.24 3783.75 2930.85 1716.20 NB Poisson 4005.86 3869.99 3454.51 2560.17 1514.15 Hurdle 3605.03 3586.47 3023.52 1952.86 984.24 NB Hurdle 3613.55 3593.63 3039.52 1957.07 987.32 ZIP 3611.05 3592.44 3037.49 1958.85 990.22 NB ZIP 3615.27 3596.45 3041.50 1962.87 994.23 155

PAGE 156

Table 4-40. Negatively skewed di stribution: percentage of simu lations favoring complex model. Simple vs. Complex .10 .25 .50 .75 .90 P vs. NB P 0.0 100.0 100.0 100.0 100.0 H vs. NBH 100.0 100.0 99.9 90.7 100.0 P vs. H 100.0 100.0 100.0 100.0 100.0 NBP vs. NBH 100.0 100.0 100.0 100.0 100.0 Z vs. NBZ 0.0 0.0 0.0 0.0 0.0 Table 4-41. AICs: ne gatively skewed distri bution (all conditions) Proportion of Zeros Model 0.10 0.25 0.50 0.75 0.90 Poisson 4044.02 4534.49 4747.63 3845.40 2311.08 NB Poisson 4563.58 4479.47 4091.18 3142.87 1923.40 Hurdle 3868.06 3807.17 3179.05 2027.47 1017.47 NB Hurdle 3886.31 3822.26 3189.83 2034.39 1021.39 ZIP 3874.07 3813.16 3184.98 2033.45 1023.46 NB ZIP 3878.08 3817.17 3180.98 2037.46 1027.47 Table 4-42. Convergence frequencie s: positively skewed distribution Proportion of Zeros Model 0.10 0.25 0.50 0.75 0.90 Poisson 2000 2000 2000 2000 2000 NB Poisson 2000 2000 2000 2000 2000 Hurdle 2000 2000 2000 2000 2000 NB Hurdle 2000 2000 2000 2000 2000 ZIP 1542 1972 2000 2000 2000 NB ZIP 136 1874 2000 2000 1992 Table 4-43. Convergence fre quencies: normal distribution Proportion of Zeros Model 0.10 0.25 0.50 0.75 0.90 Poisson 2000 2000 2000 2000 2000 NB Poisson 2000 2000 2000 2000 2000 Hurdle 1503 2000 2000 2000 2000 NB Hurdle 1982 1634 1602 1490 1537 ZIP 2000 2000 2000 2000 2000 NB ZIP 2000 2000 2000 2000 2000 156

PAGE 157

Table 4-44. Convergence frequencie s: negatively skewed distribution Proportion of Zeros Model 0.10 0.25 0.50 0.75 0.90 Poisson 2000 2000 2000 2000 2000 NB Poisson 2000 2000 2000 2000 2000 Hurdle 2000 2000 2000 2000 2000 NB Hurdle 1627 1517 1672 1562 1478 ZIP 2000 2000 2000 2000 2000 NB ZIP 2000 2000 2000 1999 2000 157

PAGE 158

3250.0000 3000.0000 2750.0000 PosAIC NB ZIP ZIP NB Hurdle Hurdle NB Poisson PoissonModel Figure 4-1. Boxplot of AICs for all models for a .10 proportion of zeros 158

PAGE 159

3100.0000 3000.0000 2900.0000 2800.0000 PosAIC NB ZIP ZIP NB Hurdle Hurdle NB Poisson PoissonModel Figure 4-2. Boxplot of AICs for all models for a .25 proportion of zeros 159

PAGE 160

2800.0000 2600.0000 2400.0000 PosAIC NB ZIP ZIP NB Hurdle Hurdle NB Poisson PoissonModel Figure 4-3. Boxplot of AICs for all models for a .50 proportion of zeros 160

PAGE 161

NB ZIP ZIP NB Hurdle Hurdle NB Poisson Poisson Model 2200.0000 2100.0000 2000.0000 1900.0000 1800.0000 1700.0000 1600.0000 PosAIC Figure 4-4. Boxplot of AICs for al l models for a .75 proportion of zeros 161

PAGE 162

NB ZIP ZIP NB Hurdle Hurdle NB Poisson Poisson Model1400.0000 1300.0000 1200.0000 1100.0000 1000.0000 900.0000 800.0000 PosAIC Figure 4-5. Boxplot of AICs for all models for a .90 proportion of zeros 162

PAGE 163

NB ZIP ZIP NB Hurdle Hurdle NB Poisson Poisson Model 4100.0000 4000.0000 3900.0000 3800.0000 3700.0000 3600.0000 3500.0000 PosAIC Figure 4-6. Boxplot of AICs for all models for a .10 proportion of zeros 163

PAGE 164

NB ZIP ZIP NB Hurdle Hurdle NB Poisson Poisson Model 4000.0000 3800.0000 3600.0000 PosAIC Figure 4-7. Boxplot of AICs for all models for a .25 proportion of zeros 164

PAGE 165

NB ZIP ZIP NB Hurdle Hurdle NB Poisson Poisson Model 4000.0000 3800.0000 3600.0000 3400.0000 3200.0000 3000.0000 2800.0000 PosAIC Figure 4-8. Boxplot of AICs for all models for a .50 proportion of zeros 165

PAGE 166

NB ZIP ZIP NB Hurdle Hurdle NB Poisson Poisson Model 3000.0000 2700.0000 2400.0000 2100.0000 1800.0000 PosAIC Figure 4-9. Boxplot of AICs for all models for a .75 proportion of zeros 166

PAGE 167

NB ZIP ZIP NB Hurdle Hurdle NB Poisson Poisson Model2000.0000 1800.0000 1600.0000 1400.0000 1200.0000 1000.0000 800.0000 PosAIC Figure 4-10. Boxplot of AICs for al l models for a .90 proportion of zeros 167

PAGE 168

NB ZIP ZIP NB Hurdle Hurdle NB Poisson Poisson Model 4600.0000 4400.0000 4200.0000 4000.0000 3800.0000 AIC Figure 4-11. Boxplot of AICs for al l models for a .10 proportion of zeros 168

PAGE 169

NB ZIP ZIP NB Hurdle Hurdle NB Poisson Poisson Model 4500.0000 4250.0000 4000.0000 3750.0000 AIC Figure 4-12. Boxplot of AICs for al l models for a .25 proportion of zeros 169

PAGE 170

NB ZIP ZIP NB Hurdle Hurdle NB Poisson Poisson Model 5000.0000 4500.0000 4000.0000 3500.0000 3000.0000 AIC Figure 4-13. Boxplot of AICs for al l models for a .50 proportion of zeros 170

PAGE 171

NB ZIP ZIP NB Hurdle Hurdle NB Poisson Poisson Model 3900.0000 3600.0000 3300.0000 3000.0000 2700.0000 2400.0000 2100.0000 AIC Figure 4-14. Boxplot of AICs for al l models for a .75 proportion of zeros 171

PAGE 172

NB ZIP ZIP NB Hurdle Hurdle NB Poisson Poisson Model 2400.0000 2200.0000 2000.0000 1800.0000 1600.0000 1400.0000 1200.0000 1000.0000 AIC 54,984 50,641 50,322 48,641 48,322 56,731 58,731 52,731 Figure 4-15. Boxplot of AICs for al l models for a .90 proportion of zeros 172

PAGE 173

0 1 2 3 4 5 6 7 0.10.250.50.750.9Rank Poisson NB Poisson Hurdle NB Hurdle ZIP NB ZIP Figure 4-16. AIC rank order for posi tively skewed distribution models 173

PAGE 174

0 1 2 3 4 5 6 7 0.10.250.50.750.9Rank Poisson NB Pois son Hurdle NB Hurd le ZIP NB ZIP Figure 4-17. AIC rank order fo r normal distribution models 174

PAGE 175

0 1 2 3 4 5 6 7 0.10.250.50.750.9 Proportion of ZerosRan k Poisson NB Pois son Hurdle NB Hurd le ZIP NB ZIP Figure 4-18. AIC rank order for nega tively skewed distribution models 175

PAGE 176

CHAPTER 5 DISCUSSION This research was conducted to determine differences in fit betw een six models under five conditions of zero-infla tion and three conditions of skew for the nonzero distribution.29 A Monte Carlo study was conducted in which samples of size 1,000 were sampled from distributions with prespecified proportions of zeros and prespeci fied proportions for the nonzero count levels. Six models were used to analy ze the samples with each simulated 2,000 times. Each model was analyzed with each combination of the five proportions of zeros combined with the three levels of skew. The research questions to be answered were as follows: Given one two-level categorical covariate with known values and one continuous covariate with known values, what is the difference in the estimated log-lik elihood between a) the Negative binomial Poisson model vs. Poiss on model; b) the Hurdle model vs. Poisson model?; c) the Negative binomial Hurdle model vs. negative binomial Poisson model?; d) the Negative binomial Hurdle model vs. Hurdle model; and, e) the Negative binomial ZIP model vs. ZIP model? Given one two-level categorical covariate with known values and one continuous covariate with known values, what is the difference in the estimated AIC between all models? The Impact of the Event Stage Distribution Positively Skewed Event-Stage Distributions The positively skewed event stage distribu tion most resembles that which is normally treated as Poisson or negative binomial Poisson. However, departure from these distributions is possible under different conditions of zero-inflation. Furt her departure may arise depending on the model used to analyze the data. For example, the Hurdle and ZIP formul ations treat the event stage as distributed Poisson while treating th e transition stage as distributed binomial. 29 All conclusions from this research are based on these conditions. Hence, conclusions are provisional as there may be other important determinations (e.g., parameter estimates, standard errors, power) not addressed in this research. 176

PAGE 177

As seen in Figure 5-1, the negative binomial Poisson model displayed better fit than the Poisson model only when the proportion of zeros was .50, .75, or .90. This suggests that overdispersion in a positively skew ed distribution arises in medium to high proportions of zeros. However, an extremely low proportion of zeros offsets this effect leadin g to equidispersion and adequate fit for the Poisson model. In other wo rds, the need for a negative binomial model arises as the proportion of zeros increases hence increasing zero-driven overdispersion. The comparison of the Poisson and Hurdle mo dels presented anomalous findings. At .10 zero-inflation, the Hurdle model displayed s uperior fit supporting its ability to handle zerodeflation. The same results were found at .50, .75, and .90 zero-inflation supporting the Hurdle models ability to accommodate zero-inflation. However, as seen in Figure 5-1, only 25.2% of the simulations favored the Hurdle model over the Poisson model when the proportion of zeros was .25. It is possible that the pr oportions for all counts yields a distribution that is distributed Poisson more than a distribution that tr eats the zeros as distributed binomial. The comparison of the negative binomial Poisson and negative binomial Hurdle yielded similar findings. However, this time it was at th e .50 proportion of zeros that fewer simulations (i.e., 13.3%) favored the more complex model. The reason for this is the addition of the overdispersion parameter for the negative binomial formulations of the model. This parameter accounts for zero-driven overdispersion; however, at .50 zeros, the fit of the overall distribution is reduced by treating the zer os as distributed binomial. The decrease of zero-driven overdispersion at .10 and .25 zeros and the increase of zero-driven overdispersion at .75 and .90 zeros then lend support to the nega tive binomial Hurdle model. This is illustrated in Figure 5-2. The negative binomial Hurdle consistently fit better than the Hurdle model until the proportion of zeros was .90. This can be seen in Figure 5-2; at this proportion, the percent of 177

PAGE 178

simulations favoring the negativ e binomial Hurdle model was 30.7%. For low proportions of zeros, this finding supports Agresti and Mins (2004) assertion that the Hurdle models can accommodate zero-inflation as well as zero-deflation. However, as the proportion of zeros increases, the negative binomial Hurdle would be expected to display better fit due to the increase in zero-driven overdisper sion. One explanation is that th e zero-driven dispersion at high proportions of zeros is counteracted by the reduction in samp le size for the event-stage. For example, if 90 of 100 responses are zero then ther e are certainly far fewe r counts contributing to event-stage model fit. The fit of the ZIP model formulations, which permit zeros in the event stage distribution, were affected by both the proportions of zeros and the positively skew ed distribution. At extremely high or low proportions of zeros (e.g., .10, .90), the majority of the simulations suggest adequate fit for the ZIP model. This is reasonable since, for a .10 proportion of zeros, the probability of a zero in the event stage is low; th e results become similar to that of the Poisson model at a .10 proportion of zeros. As was th e case when comparing the Hurdle and negative binomial Hurdle models, there are not enough zeros in either of the two parts of the model to necessitate a negative binomial formulation to ha ndle zeroand event-driven overdispersion. For a .90 proportion of zeros, a portion of zeros transfe rred to the event-stage l eads to more skew in the event-stage. In between these conditions, over dispersion arises due to the proportion of zeros justifying the negative binomial formulation. In fact, as seen in Figure 5-3, the negative binomial ZIP displayed significantly better fit than the ZIP model for al most 95% of the simulations when the proportion of zeros was .50. An alternative explanation to these findi ngs comparing ZIP and negative binomial ZIP model fit pertains to convergence. As displaye d in Table 4-42, there we re convergence problems 178

PAGE 179

for the ZIP models, especially when the proportion of zeros was 10. Hence, the comparisons are less stable since they are based on a lower number of simulations. This supports Min and Agrestis (2004) findings of supe rior convergence rates for the Hurdle models over the ZIP models. The AICs revealed inconsistent findings compared to those of the deviance tests. This is most likely due to the similarity in AICs offset by their penalty for the number of parameters in the models. For the nested models, the deviance test provides the more valid interpretation of results. The AICs can be useful if a) the res earcher is unsure of the nature of a dual data generating process or b) whether a model refl ecting a dual generating process is necessary. Given this data, the negative binomial Hurdle is superior to the ZIP models. However, the superiority of the other models appears to depend on the proporti on of zeros. At a low proportion of zeros, the Hurdle model was superior to the ZIP models; although, at a high proportion of zeros, the ZIP model was superior to the Hurdle models. There were several situations in which the Po isson models were supe rior to the two-part models suggesting that the two-pa rt models may not be necessary in some conditions. At a very low proportion of zeros (i.e., .10), the Poisson model was superior to both ZIP models. At the .25 proportion of zeros, the Poisson model was superi or to both the ZIP models and the Hurdle model. The negative binomial Pois son model is typically a poor c hoice over the two-part models and the Poisson model with one major exception. At a proportion of zeros of .50, the negative binomial Poisson model can be considered the best fitting model based on the penalty for model complexity provided by the AICs. The results are in line with the findings of Lambert (1992). At a .718 proportion of zeros, her results revealed that the ZIP model was supe rior to the negative binomial Poisson model, 179

PAGE 180

which was superior to the Poisson model. The present study found these results at a .75 proportion of zeros. Overall, the results imply th at zero-inflation and zer o-deflation are best accommodated for by using a negative binomial Hurdle model; however, when half of the data is zeros, the negative binomial Poisson model will suffice. Greene (1994), using a heavily positively sk ewed distribution with a .894 proportion of zeros found similar results for preferring the ne gative binomial ZIP model over the ZIP model. However, for the present study, at a proportion of zeros of .90, this was only true for .133 of the simulations; this may be due to the convergence problems. Mullahy (1986) found adequate f it for the Hurdle model only at a .26 proportion of zeros. At proportions of .41 and .62, the fit was no l onger adequate. This is in agreement with the findings of the present research. At low proportions of zeros with a positively skewed distribution and in terms of AICs, the Hurdle model displayed fit only improved by the negative binomial Hurdle model. However, the fit for the Hurdle model became worse as the proportion of zeros increased. Overall, these results provide suggestions for analyzing data with various proportions of zeros and a positively skewed event-stage distribution. First, if there is no reason to justify a twopart model then it is not necessa rily prudent to assume a nega tive binomial Poisson distribution to accommodate zero-driven overdispersion. The r esults of this study suggest that the negative binomial Poisson model is superior to the Poi sson only when the proportion of zeros is greater than or equal to .50. Second, if justifying separate processes generati ng the zeros and nonzeros then both the Hurdle and nega tive binomial Hurdle should be analyzed and compared at proportions of zeros around .90 since here there appears to be a drop in the need for the negative binomial Hurdle model. Third, it appears that the need for the negative binomial ZIP model 180

PAGE 181

declines as the proportion of zeros increases from .50 to .90. Howe ver, all interpretations should be treated with caution due to the severe convergence pr oblems at the .10 and .25 proportion of zeros conditions. Hence, if the researcher ac knowledges a dual data generating process but is uncertain about the source, the ZIP models are mo re valid since they dont exclude zeros from the event-stage that may be present. However, ther e is a greater probability that the analysis will not converge. This again supports Min and Agrestis (2004) assertion of Hurdle model superiority in the condition of zero-deflation. Finally, under cert ain conditions, two-part models arent necessary. In particular a Poisson model may fit as well as a Hurdle model at .25 zeros, and a negative binomial Poisson model may fit as well as a negative binomial Hurdle model at .50 zeros. Normal Event-Stage Distributions Regardless of the proportion of zeros, the Hu rdle model fits best when the event stage distribution is normal. The negative binomial Hurdle is not necessary. However, the results should be interpreted cautiously approximately since, as displayed in Table 4-43, 25% of simulations failed to converge in the .25 and hi gher proportion of zeros conditions. This is an interesting finding since previous research (Min & Agresti, 2 004) and these findings displayed convergence problems for the posit ively skewed distribution. If opting for a two-part model in which zeros are permitted in the event stage, the superior model is the ZIP model. Hence, the statistical analyses suggest that, regardless of the proportion of zeros, negative binomial formulations are no t necessary for both the ZIP and Hurdle models. If a two-part model is not n ecessary then the model depends on the proportion of zeros. As was the case with a positively skewed distribut ion, negative binomial Poisson models are not necessary when the proportion of zeros is .10 or .25. It is at proportions of .50, .75, and .90 zeros that the negative binomial Poisson model fits the data better than the Poisson model. 181

PAGE 182

It is important to note that model superiority do es not inform researchers as to the extent of that superiority. The descriptive AIC values sugg est better fit for the Hurdle over the other twopart models; however, the AICs for the other tw o-part models do not differ by much. This is most prevalent at the .10 proportion of zeros condition; the relative ad vantage of the Hurdle model over the other two-part m odels declines as the proportion of zeros increases. At the .90 proportion of zeros condition there is barely any distinction between the four two-part models at all. The discrepancy between Poisson model a nd negative binomial Poisson models is consistently large across proporti ons of zeros. The only condition wh ere they approach the AICs of the two-part models is at the .10 proportion of zeros condition. Over all interpretations can then be made. First, the Hurdle model is the be st-fitting model of all two-part models; however, the extent of this fit is great est at low proportions of zeros. Second, negative binomial Poisson models are advantageous at medium to high propo rtions of zeros. Fina lly, when unsure about whether to use a two-part model, a Poisson model will suffice at low proportions of zeros. Slymen, Ayala, Arredondo, and Elder (2006) compared the Po isson model to the negative binomial Poisson model and the ZIP model to th e negative binomial ZI P model. The proportion of zeros was .824, and the event stage distri bution was uniform. The present study, at a .90 proportion of zeros, has an event stage distribution that appears uniform given the reduction in event stage frequencies from a possible 1,000 to 100. The results be tween the studies were similar in that the negative binomial model is su perior to the Poisson model. Also, similar to their findings, there was no drastic distinction between the ZIP and negative binomial ZIP models at this proportion of zeros. 182

PAGE 183

Bhning, Dietz, Schlattmann, Mendona, a nd Kirchner (1999) compared the Poisson and ZIP models under two proportions of zeros both close to .25 and w ith normally distributed event stages. There results are similar to those of the present study. The ZIP model is superior to negative binomial ZIP. However, this isnt surpri sing since the ZIP mode is consistently superior to the negative binomial ZIP model regardless of the proportion of zeros. Only when adjusting for model complexity via AICs do the model become more similar in fit. Table 4-38 summarizes these findings. Negatively Skewed Even t-Stage Distributions No previous research was found in which th e event stage distribution was negatively skewed. It is possible that resear chers begin with the notion of zer o-inflation, which, due to the count nature of the data, lead s to considering the Poisson di stribution. Introductory texts introduce a Poisson distribution with a low me an yielding a positively skewed distribution. Hence, negatively skewed zero-bounded dist ributions receive lit tle consideration. Distributions displaying both zero-inflation and a negatively skewed event stage are certainly feasible. For example, the response vari able may be students number of days absent from an honors course. A proportion of these students should have no absences. Beyond zero, there may be a relationship between enrolling in the honors course and participating in extracurricular activities requiring absences. Thes e students would then have more absences than the others leading to a negatively skewed distribution. The negative binomial Poisson model displaye d significantly better model fit than the Poisson mode for all conditions except when th e proportion of zeros was .10. The Hurdle model fit better than the Poisson model, and the nega tive binomial Hurdle model fit better than the negative binomial Poisson model for all simulations. The negative binomial Hurdle model fit better than the Hurdle model for all simulations except when the proportion of zeros was .75. In 183

PAGE 184

this condition, the negative binomial Hurdle model fit better than the Hurdle model for 90.7% of the simulations. As was found for the normal dist ribution conditions, conclusions regarding the negative binomial Hurdle model should be tr eated cautiously as convergence failures were approximately 25% over all proportion of zeros conditions. Finally, the negative binomial ZIP model never displayed significantly better fit than the ZIP model. These findings are interesting in that, although the proportion of ones is approximately four times the proportion of zer os, the overall distri bution (including the zeros) is negatively skewed. In other words, the serious zero-defla tion may be disguised as part of one negatively skewed distribution. Hence, the implications of these results apply to zero-bounded data that appear to be the inver se of the typical Poisson distribution. The negative binomial Hurdle model could be used to analyze such a distribution pr ovided that the zeros ca n be viewed as arising from a completely different mechanism than the nonzeros. However, the simpler ZIP model could be used if there is reason to consider tw o zero generating processes. Both of these models provide better fit over the Poisson. No previous research was found to address a ne gatively skewed event stage distribution. However, there did appear to be one incongrue nt finding. Several arti cles reported that the Hurdle was advantageous over the ZIP model for zero-deflation since, wh ile both give nearly equivalent results, the ZIP model often has problems with convergence (Min & Agresti, 2004; Min, 2004). For the negatively sk ewed distribution as well as for the normal distribution, the negative binomial ZIP models failed to converge only once out of a total of 20,000 simulations while the Hurdle models failed to converge 3,899 tim es. This is seen in Ta ble 4-44. It is possible that the convergence advantage for the Hurdle m odels only applies to a positively skewed eventstage distribution. However, the present re search only found convergence problems for the 184

PAGE 185

Hurdle model at low proportions of zeros with 112 successful convergences out of 2,000 for the .10 proportion of zeros and 1,850 successful conver gences out of 2,000 for the .25 proportion of zeros. Summary of Findings One of the most interesting findings pertai ned to the Poisson and negative binomial Poisson models. When the event-stage distribution is positively skewed or normal, the negative binomial Poisson model is not necessary until the proportion of zeros is .50. For the negatively skewed distribution, the negative binomial Pois son is not necessary when the proportion of zeros is .10. The results are not surprising or the positively skewed and normal distributions; the proportion of zeros is so low that the event-stage distribut ion not overdispersed. For the negatively distribution, the inte rpretation is suspect since, ig noring the zeros, the event-stage distribution has no resemblance to the Poisson di stribution. However, at least for the positively skewed and normal distributions, these results present some guidelines for selecting either the Poisson or more complicated negative binomial Poisson model. Another interesting finding pertains to the Hu rdle and negative binomial Hurdle models. For the positively skewed distribution, the negative binomial Hu rdle model should be chosen regardless of the proportion of zeros. This was also true for the negatively skewed distribution. However, for the normal distribution, the more complicated negative binomial Hurdle model is not necessary. This provides a guideline for c hoice between the Hurdle and negative binomial Hurdle models for the distributions if willing to trust the statistical conclusion validity of 75% successful convergences for the normal and pos itively skewed distribution comparisons. Finally, this research showed that the nega tive binomial ZIP model is never necessary when the event-stage distribution is normal or negatively skewed. In fact, even for the positively skewed distribution, it may not be necessary except at .10, .25, a nd .90 proportions of zeros. The 185

PAGE 186

negative binomial ZIP model appear s to be superior when the even t-stage distribution is positive and when there is moderate to moderately-hi gh zero-inflation but not extreme zero-inflation. Limitations Discrete Conditions The results generated by this Monte Carlo si mulation derive from prespecified discrete conditions. There were five presp ecified proportions of zeros and three prespecified levels of skew. The former was fixed for each dataset, and the latter was de termined by randomly selecting from prespecified pseudo-population proba bilities for each of the five count levels. The covariates were fixed as one continuous va riable distributed standard normal and one categorical variable with two levels each at n = 500 and alternating cons ecutively in the dataset. The interpretation of results is limited to these discrete conditions. For example, for the negatively skewed event count distribution, the Poisson model fit better than the negative binomial Poisson model at a .10 proportion of ze ros. However, the reverse was true at a .25 proportion of zeros. This is not to say that the Poisson model has better fit until the proportion of zeros is .25. It could be the case that the negative binomial Poisson has better fit at a proportion of .11 zeros. Further, the results are generalizable to the extent that ap plied research maintains the same proportions for event stag e counts (i.e., skew) and covariat e characteristics as used in this study. Convergence and Optimization Optimization is a mathematical process of finding the best solution to a function. Statistically, optimization is achieved through an iterative process until convergence is reached resulting in the optimum values. This topic was previously discu ssed in the context of maximum likelihood estimation. 186

PAGE 187

Different optimization procedures can produce different results and different rates of convergence. The Hurdle model requires complex op timization since there are two models to be optimized. The ZIP model is even more complex since the two models are not independent of one another. The negative binomial ZIP model is most complex since the event stage distribution requires the addition of a dispersion parameter. There is no perfect optimizati on procedure that finds the best solution within the most reasonable amount of time for all sets of data (SAS Institute, 2000). For the SAS programming language, these zero-inflated models would typica lly be analyzed using PROC NLMIXED. The default optimization procedure here is th e Quasi-Newton method which does not require computation of second-order derivatives. Hence, it has the advantage of finding solutions rather quickly for moderately large pr oblems. However, it does not consider the boundary constraints present in the zero-inflated data In fact, the Quasi-Newton met hod is not even included as an option in the R procedures for the Hurdle and ZI P models. Here, the default is the Nelder-Mead Simplex Optimization. This method does not require any derivatives and permits boundary constraints. Its process increases speed of convergence while large sample sizes (e.g., N > 40) increase convergence time and decrease the probab ility of generating a precise result (SAS Institute, 2000). This may then explain some of the discrepancy in fi ndings between this study and that of Min and Agrest i (2004) and Min (2004) who se studies used the NLMIXED procedure with its default Qu asi-Newton optimization method. Underdispersion Strictly speaking, conditions such as a .10 proportion of zeros with a negatively skewed event stage distribution can lead to a variance that is less than its mean. As such, it is expected that the models outperform their negative bi nomial counterparts. However, a dispersion 187

PAGE 188

parameter that accounts for underdispersion may reveal a model that has superior fit to these models. Further, the results are likely to vary be tween the Poisson, Hurdle, and ZIP formulations. Interpreting Model Fit This research used the deviance statistic and Akaikes Information Criterion to examine fit between different models. Th ere are certainly othe r alternatives. Vuong (1989) proposed an inferential likelihood ratio test for comparing nonnested models. Alternatives to the AIC include the Bayesian Information Criterion, which is ty pically used for unbalanced sample sizes and the Schwarz Criterion, which is typica lly used for hierarchical models. Other models ZIP and Hurdle models have become th e most common models for analyzing zeroinflated data. However, as previ ously discussed, there are many othe r models to choose. Some of these include Lamberts (1989) ZIP( ) to accommodate correlation between the zeros and nonzeros, other links functions for the eventstage (e.g., log-gamma, log-log), and Ridout, Demtrio, & Hindes (1998) method for avoidi ng some of the distributional assumptions. Another potential model to analyze zero-inflate d count data is the latent class model. The model is similar to a general linear model in that one may posit group membership. However, in latent class modeling, the groups are unobserved. Hence, it is able to provide estimates for potential classes that may explain heterogeneity in responses (Ding, 2006). It may be very useful for researching zero-inflated count data due to the two potential sources of zeros. In fact, dUva (2006) proposed a Hurdle variation for latent cl ass modeling in health care utilization research. Validity of Model-Fitting and Model-Comparisons The purpose of comparing models is to determ ine which has the best fit. This is done by selecting the model that minimizes th e loss of information. As previ ously mentioned, this is often done by comparing statistics such as the log-li kelihood, deviance, and Akaikes Information 188

PAGE 189

Criterion (AIC). However, any model will fit perfectly or nearly perfectly given enough parameters (Lord, Washington, & Ivan, 2006). Furt her, there is always the possibility that a different set or combination of parameters will yield the same fit statistics. When comparing the Poisson and zero-inflation models, it is important not to lose sight of validity concerns. Ultimately, the best model will not be the one with the best fit. Rather, it will be the one that leads to correct inferences, inte rpretations, and decisions. Although one may not know the mo del specification that will l ead to enhanced statistical conclusion validity, it is possible to maintain several inferential principles as clarified by Burnham and Anderson (2001) and re viewed by Mazerolle (2004). Firs t, the ideal model will be simple and parsimonious in line with Ockhams Razo r, which states that, a ll else being equal, the simplest model is usually the correct model. The use of Akaikes Information Criterion as a method for adjusting for model complexity is one example of striving for parsimony (Lord, Washington, & Ivan, 2006). Another example is favoring a simpler model over a more complex model when the fit is nearly identical: Our ch oice of the ZIP model over the [negative binomial ZIP] model was based on parsimony since they provide similar fit. The ZIP model does not include the random error term that allows the conditional variance of y to exceed the conditional mean. But the interpretations of the model regr ession parameters are the same for both models (Slymen, Ayala, Arredondo, & Elder, 2006, p.8). Second, the researcher should ma intain several hypotheses noting that there is no perfectly true model. Third, since validity is a process of accumulating evidence, conclusions should be substantiated by continuing efforts to confirm the strength of evidence using various techniques such as replication and cross-validation. 189

PAGE 190

Suggestions for Future Research The results of this study provide ample sugge stions for future research. As previously mentioned as limitations, it is suggested that future research consider ot her proportions of zeros and event stage distributions, underdispersion ad justments, different optimization procedures, and other models. This research should also be extended to ordinal data by developing a model that combines the ZIP and Hurdle models w ith a cumulative model (e.g., proportional odds model). The research should also be extended to latent outcomes possibly by specifying a probit distribution. Although recen t research has been extended to incorporate random effects (e.g., Min & Agresti, 2005; Hall & Zhang, 2004; Hall, 2004; Hall, 2002; Olsen, 1999), these findings should be examined in the context of varying pr oportions of zeros and ev ent stage distributions Application in Educational Research Models for zero-inflation have been applied in a wide variety of fi elds from the health sciences to economics to transportation. Howeve r, no research was found to apply models for zero-inflation to the field of education. There ar e plenty of opportunities to analyze data with such methods (e.g., retention, subs tance abuse, disciplinary action). In an era of high-stakes accountability, it is imperative that educational researchers use the most sophisticated methods available. This is especially important si nce the results of these studies lead to decisions direc tly affecting human lives. It wo uld be fallacious to treat zeroinflated count data as purely continuous analy zed via ordinary least squares regression. A higher level of sophistication would l ead to specification of a Poiss on distribution with a log link. However, this would result in responses of zero included with the other responses leading to incorrect inferences due to st ructural and/or ra ndom reasons for the responses of zero. It would be more judicious to first consider the nature of th e zeros and then consider the use of the two-part models. Unfortunately, the application of sophisticated methods tends to 190

PAGE 191

become more pervasive as the methods become si mple to apply. As such, educational research programs should promote the existence of these methods, and software programs should provide prepackaged functions for ease of application. Major Contribution of Findings The most important finding of this dissertation was that the results of the simulation were in agreement with previous fi ndings when the proportion of zeros and event stage distribution were similar. Hence, anomalous findings between previous research ca n be attributed to differences in proportions and differences in the event stage distribution. The intention is that researchers become more aware of how different proporti ons and distributions affect the findings when comparing models such as the Poisson, ZIP, and Hurdle models. For example, as discussed, the negative binomial Poisson model can be superior to the two-part models (e.g., positive event stage distribution with a .50 proporti on of zeros), and the relative fit of the ZIP model and negative binomial ZIP model can depe nd strongly on the proporti on of zeros (e.g., for a negative event stage distribution). Ultimately, the choice should be guided by the data generating process. Data with purely structural zeros should not permit zeros in the even t stage, data with stru ctural and random zeros should not have those zeros completely separate from the counts. The findings of this dissertation help guide researchers when uncertain or when questio ning the need for a particular model 191

PAGE 192

-10 10 30 50 70 90 110 0.100.250.500.750.90 Proportion of Zeros% Simulations Favoring Complex Model P vs NBP P vs H Figure 5-1. Poisson, NB Poisson, and Hurdle over all proportions of zeros 192

PAGE 193

-10 10 30 50 70 90 110 0.100.250.500.750.90 Proportion of Zeros% Simulations Favoring Complex Model H vs NBH NBP vs NBH Figure 5-2. Hurdle, NB Hurdle, and NB Poisson over all proportions of zeros 193

PAGE 194

-10 10 30 50 70 90 110 0.100.250.500.750.90 Proportion of Zeros% Simulations Favoring Complex Model Z vs NBZ H vs NBH Figure 5-3. ZIP, NB ZIP, Hurdle, and NB Hurdle over all proportions of zeros 194

PAGE 195

LIST OF REFERENCES Agresti, A., & Finlay, B. (1997). Statistical methods for the social sciences (3rd ed.). Upper Saddle River, NJ: Prentice Hall. Agresti, A. (1996). An introduction to categorical data analysis New York: John Wiley and Sons. Akaike, H. (1972). Information theory as an extension of the maximum likelihood principle. In Second International Sympos ium on Information Theory B. N. Petrov, & F. Csaki (eds.). Akademiai Kiado: Budapest. Atkinson, K. A. (1988). An Introduction to Numerical Analysis (2nd ed.). New York: John Wiley and Sons. Arulampalam, W. & Booth, A. (1997). Who gets over the training hurdle? A study of the training experiences of young men and women in Britain. Journal of Population Econometrics, 10, 197-217. Berk, K. N., & Lachenbruch, P. A. (2002). Repeated measures with zeroes. Statistical Methods in Medical Research, 11, 303-316. Bhning, D., Dietz, E., Schlattmann, P., Me ndona, L, & Kirchner, P. (1999). The zeroinflated Poisson model and the decayed, mi ssing and filled teeth index in dental epidemiology. Journal of the Royal Statistical Association, Series A, 162, 195-209. Bonate, P. L. (2001). A brief introdu ction to Monte Carlo simulation. Clinical Pharmacokinetics, 40, 15-22. Bryk, A. S., Raudenbush, S. W., & Congdon, R. T. (1996). HLM: Hierarchical linear and nonlinear modeling with the HLM/2L and HLM/3L programs Chicago, IL: Scientific Software International, Inc. Burnham, K. P., and Anderson, D. R. (2001). Kullback-Leibler information as a basis for strong inference in ecological studies. Wildlife Research, 28, 111-119. Cameron, A. C., & Trivedi, P. K. (1998). Regression analysis of c ount data. New York: Cambridge University Press. Civettini, A. J., & Hines, E. (2005, January). Misspecification effects in zero-inflated negative binomial regression models: Common cases Paper presented at annual meeting of the Southern Political Science A ssociation. New Orleans, LA. Clarke, K. A. (2001). Testing nonnested models of international rela tions: Reevaluating realism. American Journal of Political Science, 45, 724-744. 195

PAGE 196

Cohen, A. (1954). Estimation of the Poisson parameter for truncated samples and from censored samples. Journal of the American Statistical Association, 49, 158-168. Cragg, T. C. (1971). Some statistical mode ls for limited dependent variables with application to the demand for durable goods. Econometrica, 39, 829-844. Crepon, B., & Duguet, E. (1997) Research a nd development, competition and innovation: Pseudo-maximum likelihood and simulated maxi mum likelihood methods applied to count data with heterogeneity, Journal of Econometrics, 79, 355-378. Delucchi, K. L., & Bostrom, A. (2004). Methods for an alysis of skewed data distributions in psychiatric clinical studies: Working with many zero values. American Journal of Psychiatry, 161, 1159-1168. DUva, T. B. (2006). Latent class mode ls for utilisation of health care. Health Economics, 15, 329-342. Desouhant, E., Debouzie, D., & Menu, F. ( 1998). Oviposition patte rn of phytophagous insects: On the importance of host population heterogeneity. Oecologia, 114, 382-388. Ding, C. S. (2006). Using re gression mixture analysis in educational research. Practical Assessment, Research & Evaluation, 11, p.1-11. Dobbie, M. J., and Welsh, A. H. (2001). M odeling correlated zero-inflated count data. Australian and New Zealand J ournal of Statistics, 43, 431-444. Duan, N., Manning, W. G., Morris, C. N., & Newhouse, J. P. A comparison of alternative models for the demand of medical care. RAND corporation: Santa Monica, CA. El-Shaarawi, A. H. (1985) Some goodness-of -fit methods for the Poisson plus added zeros distribution, Applied and Environmental Microbiology, 5, 1304-1306. Famoye, F., & Singh, K. (2006). Zero-inflated generalized Poisson regression model with an application to domestic violence data, Journal of Data Science, 4, 117-130. Greene, W. H. (1994). Accounting for excess zeros and sample selection in Poisson and negative binomial regression models. (New York University Department of Economics Working Paper 94-10). Retrieved May 16, 2006, from http://ideas.repec.org/s/ste/nystbu.html Gupta, P. L., Gupta, R. C., & Tripathi, R. C. (1996). Analysis of zeroadjusted count data. Computational Statistics & Data Analysis, 23, 207-218. Gurmu, S. (1997). Semi-parametric estimati on of hurdle regression models with an application to Medicaid utilization. Journal of Applied Econometrics, 12, 225-242. 196

PAGE 197

Hagedorn, M., Montaquila, J ., Vaden-Kiernan, N., Kim, K., & Chapman, C. (2004). National Household Surveys of 2003: Data File Users Manual, Volume I. (NCES 2004101). Washington, DC: U.S. Department of Education. National Ce nter for Education Statistics. Hall, D. B. (2000). Zero-inflated Poisson and bi nomial regression with random effects: a case study. Biometrics, 56, 1030-1039. Hall, D. B., & Berenhaut, K. S. (2002). Score te st for heterogeneity a nd overdispersion in zero-inflated Poisson and binomial regression models. The Canadian Journal of Statistics, 30, 1-16. Hall, D. B., & Zhang, Z. (2004). Marginal models for zero inflated clustered data. Statistical Modelling, 4, 161-180. Heilborn, D. C. (1989). Genera lized linear models for alte red zero probabilities and overdispersion in count data. Unpublished techni cal report, University of California, San Francisco, Dept. of Epidemiology and Biostatistics. Insightful. (2005). S-Plus (Versi on 6.0). Seattle, WA: Insightful. Jackman, S. (2006, November). The pscl package. Retrieved May, 2006 from http://cran.r-project.org/doc/packages/pscl.pdf Jang, T. Y. (2005). Count data models for trip generation. Journal of Transportation Engineering, 131, 444-450. Jung, B. C., Jhun, M., & Lee, J. W. (2005). Bootstrap tests for overdispersion in zeroinflated Poisson regression model. Biometrics, 61, 626-629. King, G. (1989). Event count models for inte rnational relations: Generalizations and applications. International Studies Quarterly, 33, 123-147. Lachenbruch, P. A. (2002). Analys is of data with excess zeroes. Statistical Methods in Medical Research, 11, 297-302. Lachenbruch, P. A. (2001a). Comparisons of two-part models with competitors. Statistics in Medicine, 20, 1215-1234. Lachenbruch, P. A. (2001b). Power and sample size requirements for two-part models Statistics in Medicine, 20, 1235-1238. Lambert, D. (1992). Zero-inflated Poisson regress ion, with an applica tion to defects in manufacturing. Technometrics, 34, 1-14. 197

PAGE 198

Lawal, B. (2003). Categorical data analysis wi th SAS and SPSS applications Mahwah, NJ: Lawrence Erlbaum. Martin, T. G., Wintle, B. A., R hodes, J. R., Kunnert, P., Field, S. A., Low-Choy, S. J., Tyre, A. J., & Possingham, P. (2005). Zero tolerance ecology: Improving ecological inference by modeling the s ource of zero observations. Ecology Letters, 8, 1235-1246. Mazerolle, M. J. (2004). Mouvements et reproduction des amphibiens et tourbis perturbes Unpublished doctoral di ssertation,. Universit Laval, Qubec. McCullagh, P., & Nelder, J. A. (1989). Generlised Linear Models (2nd ed.). London: Chapman & Hall. Micceri, T. (1989). The unicorn, the normal cu rve, and other improbable creatures. Psychological Bulletin, 105, 156-166. Min, Y. (2003). Modeling data with clumps. Dissertation Abstracts Inte rnational B, 64(12). (UMI3117356). Min, Y., Agresti, A. (2004). Random effects models for repeated measures of zeroinflated count data. (Technical Report 2004-026), Department of Statistics, University of Florida, Retrieved May 8, 2006 from second author. Mooney, C. Z. (1997). Monte Carlo Simulation. Thousand Oaks, CA: Sage. Moulton, L. H., Curriero, F. C., & Barroso, P. F. (2002). Mixture models for quantitative HIV RNA data. Statistical Methods in Medical Research, 11, 317-325. Mullahy, J. (1986). Specificati on and testing of some modi fied count data models. Journal of Econometrics, 33, 341-365. Mwaliki, S. M. (January, 2005). The zicount s package. Retrieved May, 2006 from http://cran.r-project.org /doc/packages/zicounts.pdf Pardoe, I., & Durham, C. A. (2003). Model choice applied to consumer preferences. In Proceedings of the 2003 Join t Statistical Meetings, Alexandria, VA, American Statistical Association. Paxton, P., Curran, P. J., Bollen, K. A, Kirby, J, & Chen, F. (2001). Monte Carlo experiments: Design and implementation. Structural Equation Modeling, 8, 287-312. R Development Core Team (2006). R: A language and environment for statistical computing. Vienna, Austria: R Foundation for Statistical Computing. Rice, J. A. (1995). Mathematical Statisti cs and Data Analysis. (2nd ed.). Belmont, CA: Wadsworth. 198

PAGE 199

Ridout, M.S., Demtrio, C.G.B. and Hinde, J.P. (1998) Models for counts data with many zeroes. Proceedings of the XIXth International Biometric Conference Cape Town, Invited Papers, pp. 179-192. Paper retrieved March 13, 2006 from http://www.kent.ac.uk/ims/personal/msr/zip1.html SAS Institute Inc. (2000). SAS/STAT User's Guide V.8, Cary, NC: SAS Institute Inc. Scheaffer, R. L. (1995). Introduction to probabilit y and its applications (2nd ed.). Belmont, CA: Wadsworth. Shankar, V., Milton, J., & Mannering, F. ( 1997). Modeling accident frequencies as zeroaltered probability processes: an empirical inquiry. Accident Analysis and Prevention, 29, 829-837. Singer, J. D., & Willett, J.B. (2003). Applied longitudinal da ta analysis: Modeling change and event occurrence New York: Oxford University Press. Slymen, D. J., Ayala, G. X., Arredondo, E. M., & Elder, J. P. (2006). A demonstration of modeling count data with an a pplication to phys ical activity. Epidemiologic Perspectives & Innovations, 3, 1-9. SPSS (2006). SPSS for Windows, Rel. 14.0. Chicago: SPSS, Inc. Stokes, M. E., Davis, C. S., & Koch, G. G. (1991). Categorical data analysis using the SAS system (2nd ed.). Cary, NC: SAS Institute, Inc. Slymen, D. J., Ayala, G. X., Arredondo, E. M., & Elder, J. P. (2006). A demonstration of modeling count data with an a pplication to phys ical activity. Epidemiologic Perspectives & Innovations, 3, 1-9. Tooze, J. A., Grunwald, G. K., & Jones, R. H. (2002). Analysis of repeated measures data with clumping at zero. Statistical Methods in Medical Research, 11, 341-355. Vuong, Q. H. (1989). Likeli hood ratio tests for model selection and non-nested hypotheses. Econometrica, 57, 307-333. Warton, D. I. (2005). Many zeros does not m ean zero inflation: Comparing the goodnessof-fit of parametric models to multivariate abundance data. Environmetrics, 16, 275-289. Weiller, H. (1964). A significance test for simultaneous quantal and quantitative responses. Technometrics, 6, 273-285. Welsh, A. H., Cunningham, R. B., Donnelly C. F., & Lindenmayer, D. B. (1996). Modelling abundance of rare sp ecies: Statistical m odels for counts with extra zeros. Ecological Modelling, 88, 297-308. 199

PAGE 200

Zhou, X., & Tu, W. (1999). Comparison of several indepe ndent population means when their samples contain log-normal a nd possibly zero observations. Biometrics, 55, 645 651. Zorn, C. J. W. (April 18-20, 1996). Evalua ting zero-inflated a nd hurdle Poisson specifications. Midwest Political Science Association, 1-16. 200

PAGE 201

BIOGRAPHICAL SKETCH Jeffrey Monroe Miller was born in Greensboro, NC and raised in Jacksonville, FL. He earned a Bachelor of Arts degree from the Univ ersity of Wisconsin Eau Claire in 2001 majoring in Psychology with two minors in Music. As an undergraduate, he became fascinated by the methods used to analyze data. In 2002, he en rolled in the Research and Evaluation Methodology program at the University of Florida in the De partment of Educational Psychology (College of Education). In 2004, after author ing and defending a thesis pertai ning to item-order effects in surveys, he received the Master of Arts in Education degree in this program. In 2007, he completed the requirements for the degree of Doctor of Philosophy. 201


Permanent Link: http://ufdc.ufl.edu/UFE0018500/00001

Material Information

Title: Comparing Poisson, Hurdle, and Zip Model Fit under Varying Degrees of Skew and Zero-Inflation
Physical Description: Mixed Material
Copyright Date: 2008

Record Information

Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.
System ID: UFE0018500:00001

Permanent Link: http://ufdc.ufl.edu/UFE0018500/00001

Material Information

Title: Comparing Poisson, Hurdle, and Zip Model Fit under Varying Degrees of Skew and Zero-Inflation
Physical Description: Mixed Material
Copyright Date: 2008

Record Information

Source Institution: University of Florida
Holding Location: University of Florida
Rights Management: All rights reserved by the source institution and holding location.
System ID: UFE0018500:00001


This item has the following downloads:


Full Text





COMPARING POISSON, HURDLE, AND ZIP MODEL FIT
UNDER VARYING DEGREES OF SKEW AND ZERO-INFLATION























By

JEFFREY MONROE MILLER


A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY

UNIVERSITY OF FLORIDA

2007
































Copyright 2007

by

Jeffrey Monroe Miller



































To the memory of my grandfather, Rev. Harold E. Cato.









ACKNOWLEDGMENTS

Several people helped make this study possible. I would like to thank my stepfather and

mother, Dr. Daniel and Gail Jacobs as well as my father and stepmother, Jerry and Darnelle

Miller for their many years of encouragement. I would also like to thank my supervisory

committee chair, M. David Miller, for his unyielding guidance, patience, and support. I thank Dr.

Jon Morris for the numerous training experiences. Many professors are appreciated for providing

the educational foundations for the dissertation topic including Dr. James Algina and Dr. Alan

Agresti.

The idea to research zero-inflation was inspired by experiences with data while

consulting on proj ects. To this extent, I thank those clients Dr. Courtney Zmach and Dr. Lori

Burkhead. Undergraduate faculty that I would like to acknowledge for their inspiration and

direction include Blaine Peden, Patricia Quinn, and Lee Anna Rasar. Several friends have been a

source of encouragement including Matt Grezik and Rachael Wilkerson. Finally, I thank those

who made it financially possible to complete this dissertation including my consulting clients,

the University of Florida College of Education, the Lastinger Center, and Adsam, LLC.












TABLE OF CONTENTS


page

ACKNOWLEDGMENTS .............. ...............4.....


LIST OF TABLES ............ ...... .__. ...............8....


LIST OF FIGURES ........._.___..... .__. ...............11....


AB S TRAC T ............._. .......... ..............._ 12...


CHAPTER


1 INTRODUCTION ................. ...............14.......... ......


Statement of the Problem ................. ...............15................
Rationale for the Study ................... .......... ...............15......
Purpose and Significance of the Study .............. ...............16....
Research Questions............... ...............1


2 REVIEW OF THE LITERATURE ................. ...............18.......... ....


Zero-Inflated Count Data ................. ...............18........... ....
Count Data ................. ...............18.................
Zero-Inflation .................... ...............19.
The Sources ofZero-Inflation .............. ...............20....

Impact of Zero-Inflation on Analyses .............. ...............21....
Simple Solutions to Zero-Inflation ................. ...............22................
Deleting zeros............... ...............22.
Assuming normality .............. ...............22....
Transforming Zeros ................. ...............23.................
Generalized Linear Models .............. .. ...... .. ..............2
The Binomial Distribution and the Logit Link ................. ...............28..............
Evaluating M odel Fit .................. ... .......... .. ................3 1....
The Poisson Distribution and the Log Link .............. ...............35....
Iterative Estimation .............. ...............38....
Interpretation of Coefficients .............. ...............38....
Hypothesis testing .............. ...............39....
O verdispersion ............................ .... ....................3
Poisson and Negative Binomial Models with Zero-Inflation............... ..............4
The Hurdle model .................. ......__ ...............45...
The Negative Binomial Hurdle model .............. ...............48....
The Zero-Inflated Poisson (ZIP) model .............. ..... ...............49..
The Negative Binomial Zero-Inflated Poisson model............_ .. ......._ ........51
Model Comparison Testing for Zero-Inflated Data....................... ..............5
Review of Research Pertaining to and Using Zero-Inflated Count Data. .............. ..... ........._.53
Hurdle M odel ............ ..... .._ ...............53...












Statistical .............. ...............53....

Applications .................. ...............54..
Zero-Inflated Poisson Model ................. ...............55..._._._......
Statistical .............. ...............55....

Applications .............. ...............59..
ZIP and Hurdle Model-Comparisons .............. ...............63....
Statistical .............. ...............64....

Applications .............. ...............64....
Discrepant Findings ....___ ................ .........__..........6


3 METHODOLOGY .............. ...............73....


Research Questions............... ...............7
Monte Carlo Study Design .............. ...............73....
Monte Carlo Sampling............... ...............75
P seudo-P opul ati on ................... ...............75.......... ....
The Prespecified Zero Proportions ........._.. ...._._..... ...............76...
Pre-Specified Skew .............. ...............76....
Random Number Generation............... ...............7

Sam ple Size .............. ...............78....
Simulation Size............... ...............78..
Iteration Size............... ...............79..
Distribution Generation ........._..... ...._... ...............80.....
Monte Carlo Model s.........._...._ ......_.. ...............80.....
Monte Carlo Analysis Procedures .............. ...............82....
Analysis Design .........._..._.. ...............84........ ......

4 RE SULT S .............. ...............94....


Pseudo-Population Results .................. ...............94..
Pseudo-Population Poisson Models .............. ...............95....
Pseudo-Population Hurdle Models ..........._.._.. ...............96...._._. .....
Hurdle vs. Negative Binomial Hurdle ..........._.... ....._._ ....._.._.........9
Poisson vs. Hurdle............... ...... .. .. .. ............9

Negative Binomial Poisson vs. Negative Binomial Hurdle .............. ...................98
Pseudo-Population ZIP Models............... ...............98.
Comparing AIC's For All Models .........._...._. ...............99......... ..
Monte Carlo Simulation Results .........._...._. ...............101.._.... .....
Positively Skewed Distribution .........._...._ ......_. ...._... ............10
Normal Distribution. ..........._.._. ...............111....... ......

Negatively Skewed Distribution.................. ............12
Review of Positively Skewed Distribution Findings ..........._.._.. ..........._.. ...._.._ ...133
Review of Normal Distribution Findings .............. ...............135....
Review of Negatively Skewed Distribution Findings .............. ...............136....


5 DI SCUS SSION ........._..... ...._... ..............._ 176..












The Impact of the Event Stage Distribution ................. ......... ......... ...........7
Positively Skewed Event-Stage Distributions ................. ...............................176
Normal Event-Stage Distributions .............. ...............181....
Negatively Skewed Event-Stage Distributions .............. ...............183....
Summary of Findings .............. ...............185....
Limitations ................. ...............186................
Discrete Conditions .............. ...............186....
Convergence and Optimization ................. ......... ...............186 ....
Under di spersi on ................. ...............187................
Other m odels .............. .. ............... ...........18
Validity of Model-Fitting and Model-Comparisons ................. ..............................188
Suggestions for Future Research .............. ...............190....
Application in Educational Research............... ...............19
Maj or Contribution of Findings ........._.___..... .___ ...............191...

LIST OF REFERENCES ........._.___..... .___ ...............195....


BIOGRAPHICAL SKETCH .............. ...............201....










LIST OF TABLES


Table page

2-1 Five pairs of nested models valid for statistical comparison ................ ......................71

2-2 Summary of literature on zero-inflation ..........._ ..... ..__ ....___ ............._..72

3-1 Proportions of counts as a function of zeros and skew ....._____ ... .....___ ..............87

3-2 Frequencies of counts as a function of zeros and skew .......... ................ ...............87

3-3 Descriptive statistics for each distribution............... ..............8

3-4 Poisson model: pseudo-population parameters ................. ...............88...............

3-5 Negative Binomial Poisson model: pseudo-population parameters ................. ...............89

3-6 Hurdle model (zeros): pseudo-population parameters ................. ......... ................89

3-7 Hurdle model (events): pseudo-population parameters ......... ................ ...............90

3-8 Negative Binomial Hurdle model (zeros): pseudo-population parameters ................... .....90

3-9 Negative Binomial Hurdle model (events): pseudo-population parameters ................... ...91

3-10 ZIP model (zeros): pseudo-population parameters ................. ............... ......... ...91

3-11 ZIP Model (events): pseudo-population parameters ................. ................ ......... .92

3-12 Negative Binomial ZIP model (zeros): pseudo-population parameters.................... .........92

3-13 Negative Binomial ZIP model (events): pseudo-population parameters ................... ........93

4-1 Deviance statistics comparing Poisson and negative binomial Poisson models. ............138

4-2 Deviance statistics comparing Hurdle and negative binomial Hurdle models ................138

4-3 Deviance statistics comparing Poisson and Hurdle models............... ................13

4-4 Deviance statistics comparing NB Poisson and NB Hurdle models................ ...............138

4-5 Deviance statistics comparing ZIP and negative binomial ZIP models ..........................139

4-6 Log-likelihood comparisons for positively skewed distribution with .10 zeros ..............139

4-7 AIC's for positively skewed distribution models with a .10 proportion of zeros............ 140

4-8 Log-likelihood comparisons for positively skewed distribution with .25 zeros .............140










4-9 AIC's for positively skewed distribution models with a .25 proportion of zeros............ 141

4-10 Log-likelihood comparisons for positively skewed distribution with .50 zeros .............141

4-11 AIC's for positively skewed distribution models with a .50 proportion of zeros............1 42

4-12 Log-likelihood comparisons for positively skewed distribution with .75 zeros..........._...142

4-13 AIC's for positively skewed distribution models with a .75 proportion of zeros............ 143

4-14 Log-likelihood comparisons for positively skewed distribution with .90 zeros ..............143

4-15 AIC's for positively skewed distribution models with a .90 proportion of zeros............1 44

4-16 Log-likelihood comparisons for normal distribution with .10 zeros .............. ..... ........._.144

4-17 AIC's for normal distribution models with a .10 proportion of zeros ................... ..........145

4-18 Log-likelihood comparisons for normal distribution with .25 zeros .............. ................145

4-19 AIC's for normal distribution models with a .25 proportion of zeros ................... ..........146

4-20 Log-likelihood comparisons for normal distribution with .50 zeros .............. ................147

4-21 AIC's for normal distribution models with a .50 proportion of zeros ................... ..........147

4-22 Log-likelihood comparisons for normal distribution with .75 zeros .............. ................148

4-23 AIC's for normal distribution models with a .75 proportion of zeros ................... ..........148

4-24 Log-likelihood comparisons for normal distribution with .90 zeros .............. ................149

4-25 AIC's for normal distribution models with a .90 proportion of zeros ................... ..........149

4-26 Log-likelihood comparisons for negatively skewed distribution with .10 zeros.............1 50

4-27 AIC's for negatively skewed models with a .10 proportion of zeros .............. .... ...........150

4-28 Log-likelihood comparisons for negatively skewed distribution with .25 zeros.............1 51

4-29 AIC's for negatively skewed models with a .25 proportion of zeros .............. .... ...........151

4-30 Log-likelihood comparisons for negatively skewed distribution with .50 zeros.............1 52

4-3 1 AIC's for negatively skewed models with a .50 proportion of zeros .............. .... ...........152

4-32 Log-likelihood comparisons for negatively skewed distribution with .75 zeros.............1 53

4-33 AIC's for negatively skewed models with a .75 proportion of zeros .............. .... ...........153










4-34 Log-likelihood comparisons for negatively skewed distribution with .90 zeros.............1 54

4-3 5 AIC's for negatively skewed models with a .90 proportion of zeros .............. .... ...........154

4-36 Positively skewed distribution: percentage of simulations favoring complex model......155

4-37 AIC's: positively skewed distribution (all conditions) ................ .........................155

4-38 Normal distribution: percentage of simulations favoring complex model. .....................155

4-39 AIC's: normal distribution (all conditions)............... ..............15

4-40 Negatively skewed distribution: percentage of simulations favoring complex model....156

4-41 AIC's: negatively skewed distribution (all conditions) .............. ....................15

4-42 Convergence frequencies: positively skewed distribution ................. ............ .........156

4-43 Convergence frequencies: normal distribution ................ ...............156........... ...

4-44 Convergence frequencies: negatively skewed distribution ................. ......................157










LIST OF FIGURES


Figure page

4-1 Boxplot of AIC's for all models for a .10 proportion of zeros ........._._.... ......._. .....158

4-2 Boxplot of AIC's for all models for a .25 proportion of zeros ........._._.... ......._. .....159

4-3 Boxplot of AIC's for all models for a .50 proportion of zeros ........._._.... ......._. .....160

4-4 Boxplot of AIC's for all models for a .75 proportion of zeros ........._._.... ......._. .....161

4-5 Boxplot of AIC's for all models for a .90 proportion of zeros ........._._.... ......._. .....162

4-6 Boxplot of AIC's for all models for a .10 proportion of zeros ........._._.... ......._. .....163

4-7 Boxplot of AIC's for all models for a .25 proportion of zeros ........._._.... ......._. .....164

4-8 Boxplot of AIC's for all models for a .50 proportion of zeros ........._._.... ......._. .....165

4-9 Boxplot of AIC's for all models for a .75 proportion of zeros ........._._.... ......._. .....166

4-10 Boxplot of AIC's for all models for a .90 proportion of zeros ........._._.... ......._. .....167

4-11 Boxplot of AIC's for all models for a .10 proportion of zeros ........._._.... ......._. .....168

4-12 Boxplot of AIC's for all models for a .25 proportion of zeros ........._._..........._........169

4-13 Boxplot of AIC's for all models for a .50 proportion of zeros ........._._..........._........170

4-14 Boxplot of AIC's for all models for a .75 proportion of zeros ........._._..........._........171

4-15 Boxplot of AIC's for all models for a .90 proportion of zeros ........._._..........._........172

4-16 AIC rank order for positively skewed distribution models .............. ....................17

4-17 AIC rank order for normal distribution models............... ...............174

4-18 AIC rank order for negatively skewed distribution models .............. .....................7

5-1 Poisson, NB Poisson, and Hurdle over all proportions of zeros ................. ................. 192

5-2 Hurdle, NB Hurdle, and NB Poisson over all proportions of zeros ................. ...............193

5-3 ZIP, NB ZIP, Hurdle, and NB Hurdle over all proportions of zeros .............. ..... .......... 194









Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy

COMPARING POISSON, HURDLE, AND ZIP MODEL FIT
UNDER VARYING DEGREES OF SKEW AND ZERO-INFLATION

By

Jeffrey Monroe Miller

May 2007

Chair: M. David Miller
Major Department: Educational Psychology

Many datasets are characterized as count data with a preponderance of zeros. Such data

are often analyzed by ignoring the zero-inflation and assuming a Poisson distribution. The

Hurdle model is more sophisticated in that it considers the zeros to be completely separate from

the nonzeros. The zero-inflated Poisson (ZIP) model is similar to the Hurdle model; however, it

permits some of the zeros to be analyzed along with the nonzeros. Both models, as well as the

Poisson, have negative binomial formulations for use when the Poisson assumption of an equal

mean and variance is violated.

The choice between the models should be guided by the researcher' s beliefs about the

source of the zeros. Beyond this substantive concern, the choice should be based on the model

providing the closest fit between the observed and predicted values. Unfortunately, the literature

presents anomalous findings in terms of model superiority.

Datasets with zero-inflation may vary in terms of the proportion of zeros. They may also

vary in terms of the distribution for the nonzeros. Our study used a Monte Carlo design to sample

1,000 cases from positively skewed, normal, and negatively skewed distributions with

proportions of zeros of .10, .25, .50, .75, and .90. The data were analyzed with each model over










2,000 simulations. The deviance statistic and Akaike's Information Criterion (AIC) value were

used to compare the fit between models.

The results suggest that the literature is not entirely anomalous; however, the accuracy of

the findings depends on the proportion of zeros and the distribution for the nonzeros. Although

the Hurdle model tends to be the superior model, there are situations when others, including the

negative binomial Poisson model, are superior. The findings suggest that the researcher should

consider the proportion of zeros and the distribution for the nonzeros when selecting a model to

accommodate zero-inflated data.









CHAPTER 1
INTTRODUCTION

Analyzing data necessitates determination of the type of data being analyzed. The most

basic assumption is that the data follows a normal distribution. However, there are many other

types of distributions. The validity of the results can be affected by the dissimilarity between the

distribution of the data and the distribution assumed in the analysis. As such, it is imperative that

the researcher choose a method for analyzing the data that maintains a distribution similar to that

of the observed data.

Counts are an example of data which does not readily lend itself to the assumption of a

normal distribution. Counts are bounded by their lowest value, which is usually zero. A

regression analysis assuming a normal distribution would permit results below zero. Further,

counts are discrete integers while the normal distribution assumes continuous data. Finally,

counts often display positive skew such that the frequency for low counts is considerably higher

than the frequencies as the count levels increase.

It is not uncommon to find count data analyzed in a more appropriate manner than

assuming a normal distribution. Typically, more appropriate analysis includes specification of a

Poisson distribution with a log link, rather than a normal distribution with a Gaussian link.

However, this does not guarantee accurate and valid results as other features of the data may

warrant an even more sophisticated model.

An example of data requiring a more rigorous treatment of the data is the case of zero-

inflation. In this scenario, there are far more zeros than would be expected using the Poisson

distribution. As such, a number of methods including the zero-inflated Poisson (ZIP) model and

the Hurdle model are available. Further, there are negative binomial variations of these for use

when particular assumptions appear to be violated. The choice between the models depends on









whether the researcher believes the zeros are all a complete lack of the quantity being measured

or that at least some of the zeros are purely random error.

Statement of the Problem

The results from both simulated and actual data sets in the zero-inflation literature are in

much disagreement. Lambert (1992) found the ZIP model to be superior to the negative binomial

Poisson model, which was superior to the Poisson model. Greene (1994) found the negative

binomial Poisson model to be superior to the ZIP model, which was superior to the Poisson

model. Slymen, Ayala, Arredondo, and Elder (2006) found the ZIP and negative binomial ZIP

models to be equal. Welsh, Cunningham, Donnelly, and Lindenmayer found the Hurdle and ZIP

models to be equal while Pardoe and Durham (2003) found the negative binomial ZIP model to

be superior to both the Poisson and Hurdle models.

One striking characteristic of these articles and others is their differences in terms of the

proportion of zeros and the distribution for the nonzeros. Some research (Boihning, Dietz,

Schlattmann, Mendonga, and Kirchner, 1999) analyzed data in which the proportion of zeros was

as low as .216 while others (Zorn, 1996) used proportions as high as .958. Further, the nonzeros

varied in terms of their distributions from highly positively skewed to normal to uniform. It is

possible that different models yield different results depending on the proportion of zeros and the

distribution for the nonzeros.

Rationale for the Study

The best model is the one that appropriately answers the researcher' s question. Beyond

this, a superior model is one that has close proximity between the observed data and that

predicted by the model. In other words, a superior model is one with good fit to the data.

This study compared the fit between the Poisson, ZIP, and Hurdle models as well as their

negative binomial formulations. Each analysis was performed for five different proportions of









zeros and three different amounts of skew for the nonzero distribution. The intended results

would clarify the discrepant Eindings of previous research.

Purpose and Significance of the Study

The primary purpose of this study was to determine superiority of fit for various models

under varying proportions of zero-inflation and varying levels of skew. As such, determination

can be made as to which model has better fit given data with a particular proportion of zeros and

a particular distribution. The secondary purpose was to elucidate the reasons for discrepant

Endings in previous research.

The superior model is the appropriate model given the research question. However, there

are situations in which the appropriate model is unknown or

unclear. Further, there may be situations in which a simpler model such as the Poisson may be

used in lieu of the more sophisticated Hurdle and ZIP models. This research provides results that

aid researchers in determining the appropriate model to use given zero-inflated data.

Research Questions

Model comparisons in this research were based on two measures. One is the deviance

statistic, which is a measure of the difference in log-likelihood between two models, permitting a

probabilistic decision as to whether one model is adequate or whether an alternative model is

superior. This statistic is appropriate when one model is nested within another model. The other

measure is Akaike's Information Criterion (AIC). This statistic penalizes for model complexity

and permits comparison of nonnested models; however, it can only be used descriptively. These

two measures of model fit were used to compare results from data simulations where each

dataset included 2,000 cases and each model was analyzed 1,000 times. Specifically, the

measures of model fit were used to answer the following research questions:










* Given one two-level categorical covariate with known values and one continuous covariate
with known values, what is the difference in the estimated log-likelihood between a) the
Negative binomial Poisson model vs. Poisson model; b) the Hurdle model vs. Poisson
model?; c) the Negative binomial Hurdle model vs. negative binomial Poisson model?; d) the
Negative binomial Hurdle model vs. Hurdle model; and, e) the Negative binomial ZIP model
vs. ZIP model?

* Given one two-level categorical covariate with known values and one continuous covariate
with known values, what is the difference in the estimated AIC between all models?









CHAPTER 2
REVIEW OF THE LITERATURE

Zero-Inflated Count Data

Count Data

As the name implies, count data is data that arises from counting. They are the

"realization of a nonnegative integer-valued random variable" (Cameron & Travedi, 1998, p. 1).

As such, the response values take the form of discrete integers (Zorn, 1996). Although the lower

boundary can feasibly be any integer, it is usually the case that its value is zero. Strictly

speaking, there can be no nonnegative numbers. Hence, the data are constrained by this lower

bound of zero and no upper bound.

Acknowledgment of concerns over zero-inflation, ignoring covariates, likely dates to

Cohen (1954). Cameron and Triverdi (1989, p. 10-11) identified many areas in which special

models have been used to analyze count data including "models of counts of doctor visits and

other types of health care utilization; occupational injuries and illnesses; absenteeism in the

workplace; recreational or shopping trips; automobile insurance rate making; labor mobility;

entry and exits from industry; takeover activity in business; mortgage prepayments and loan

defaults; bank failures; patent registration in connection with industrial research and

development; and frequency of airline accidents .. as well as in many disciplines including

demographic economics, in crime victimology, in marketing, political science and government,

[and] sociology". Surprisingly, there was no mention of research in education. Examples of

variables in educational research that yield count data include a student' s number of days absent,

number of test items scored correct or incorrect, and number of referrals for disciplinary action.









The lower bound constraint of zero presents the biggest obstacle toward analyzing count

data when assuming a normal distribution. It is common for count data to have a skewed

distribution that is truncated at the lower bound.



skew(Y)= "= (2-1)
(N -1l)s,3

Hence, the data are heteroscedastic with variance increasing as the count increases.

Therefore, standard models, such as ordinary least squares regression, are not appropriate since

they assume that the residuals are distributed normally with a mean of zero and a standard

deviation of one (Slymen, Ayala, Arredondo, & Elder, 2006). Cameron and Triverdi (1998)

clarify that the use of standard OLS regression "leads to significant deficiencies unless the mean

of the counts is high, in which case normal approximation and related regression methods may

be satisfactory" (p.2).

An example of a count data variable is the number of household members under the age

of 21 reported by respondents in the Adult Education for Work-Related Reasons (AEWR) survey

administered by National Council for Educational Statistics in 2003 (Hagedorn, Montaquila,

Vaden-Kiernan, Kim & Chapman, 2004). The sample size was 12,725. This variable has a lower

count boundary of zero and an upper count boundary of six. The count distribution is positively

skewed at 1.971. The distribution mean of 0.54 is certainly not an accurate measure of central

tendency; the median and mode are both zero (i.e., the lower-bound itself), and the standard

deviation of 0.999 permits negative values in the lower 68% confidence interval.

Zero-Inflation

It is not uncommon for the outcome variable in a count data distribution to be

characterized by a preponderance of zeros. As Tooze, Grunwald, & Jones (2002, p.341) explain,










Typically, [for count data] the outcome variable measures an amount that must be non-
negative and may in some cases be zero. The positive values are generally skewed, often
extremely so .. Distributions of data of this type follow a common form: there is a spike
of discrete probability mass at zero, followed by a bump or ramp describing positive
values.

The occurrence is primarily in the case of interval/ratio count data and sometimes ordinal

data (Boihning, Dietz, Schlattmann, Mendonga, & Kirchner, 1999). Regarding continuous data,

Hall and Zhang (2004) explain that these distributions "have a null probability of yielding a zero

.. there is little motivation for a model such as [zero-inflated] normal, because all observed

zeros are unambiguous .. (p. 162). If continuous zeros are inflated and those zeros are of

concern, they can be analyzed separately from the nonzeros. The null probability of continuous

zeros is evident in measures such as height and age.

The condition of excessive zeros is known as zero-inflation (Lachenbruch, 2002) or as a

probability mass that clumps at zeros (Tooze, Grunwald, & Jones, 2002). It has been recognized

as an area of research in the mid-60'sl (Lachenbruch, 2002) when Weiler (1964) proposed a

method for mixing discrete and continuous distributions. Min and Agresti (2005) formally define

zero-inflation as "data for which a generalized linear model has lack of fit due to

disproportionately many zeroes" (p. 1). There are simply "a greater number of zero counts than

would be expected under the Poisson or some of its variations" (Zorn, 1996, p. 1).

The Sources of Zero-Inflation

The zeros can be classified as being either true zeros or sampling zeros. True zeros

represent responses of zero that are truly null. Suppose an educational inventory item states

"How many college preparatory workshops have you attended?" Some of the respondents in the

sample may have no intentions to apply for college. Hence, the number of preparatory


SAltemnatively, if a scale is bound, it is reasonable to consider an inflated upper bound. In this case, scale reversal
and subsequent appropriate analysis if justified (Lachenbruch, 2002).









workshops attended may never be greater than zero. Sampling zeros, on the other hand, arise as a

probability. There are a proportion of college-bound students who have not attended a workshop

due to the possibility that the workshop was not (or is not yet) available. Alternatively, some

college-bound students may feel prepared and have no reason to participate in a workshop.

Hence, the mechanism underlying zero-inflation can arise from one or both of 1) a

possibility that no other response is probabilistic, or 2) that the response is within a random

sample of potential count responses. Martin, Brendan, Wintle, Rhodes, Kuhnert, Field, Low-

Choy, Tyre, and Possingham (2005) term the sampling zeros as 'false zeros' and include error as

a source of zeros. They state, "Zero inflation is often the result of a large number of 'true zero'

observations caused by the real .. effect of interest .. However, the term [zero-inflation] can

also be applied to data sets with 'false zero' observations because of sampling or observer errors

in the course of data collection" (p. 1235).

Often, the data contains both types of zeros. This is the result of a dual data generating

process (Cameron & Trivedi, 1998). For example, some adults in the AEWR sample may have

had true-zero household members under the age of 21 because they are unable to bear children or

desire to bear children. Alternatively, they may have random-zero household members under the

age of 21 because these adults do have such children but not as members of the household.

Impact of Zero-Inflation on Analyses

"Much of the interest in count data modeling appears to stem from the recognition that

the use of continuous distributions to model integer outcomes might have unwelcome

consequences including inconsistent parameter estimates" (Mullahy, 1986, p.341). In the typical

count data scenario, the zero left-bound implies heteroscedasticity (Zorn, 1996). An even greater

problem with zero-inflated distributions, beyond this inadequacy of analyzing such a skewed and

heteroscedastic distribution as if it were normal (Tooze, Grunwald, & Jones, 2002) is that they










yield "surprisingly large inefficiencies and nonsensical results" (King, 1989, pl26). Martin,

Brendan, Wintle, Rhodes, Kuhnert, Field, Low-Choy, Tyre, and Possingham (2005) and

McCullagh and Nelder (1989) explain that zero-inflation is a special case of overdispersion in

which the variance is greater than it should be given a particular distributional shape and

measure of central tendency. The impact is biased/inconsistent parameter estimates, inflated

standard errors and invalid inferences (Jang, 2005; Martin, Brendan, Wintle, Rhodes, Kuhnert,

Field, Low-Choy, Tyre, and Possingham, 2005).

Simple Solutions to Zero-Inflation

Deleting zeros

The simplest of solutions is to delete all cases having responses of zero on the variable of

interest. A large proportion of total responses would then be removed from the total dataset. This

would then result in a loss of valuable information impacting statistical conclusion validity

(Tooze, Grunwald, & Jones, 2002). The sample size may also then be too small for analyses of

the non-zero values.

Assuming normality

Another simple solution is to ignore the zero-inflation, assume asymptotic normality, and

analyze the data using standard techniques such as ordinary least squares regression.

hhml = P, + fl,Sexl + PAge, + (2-2)

According to this model, the number of household members under the age of 21 for adult

respondent i is predicted from the overall mean, a coefficient relating the respondent's sex to

hhm, a coefficient relating the respondent' s age to hhm, and error. The model assumes that the

residuals for hhm are distributed normally with a mean of zero and a common variance, cr For

the first equation, y is a vector of responses, X is a design matrix for the explanatory variable










responses, p is a vector of regression coefficients relating y to X, and E is a vector of residuals

measuring the deviation between the observed values of the design matrix and those predicted

from the fitted equation.

Transforming Zeros

Another simple solution is to transform the counts to coerce a more normal distribution

(Slymen, Ayala, Arredondo, & Elder, 2006). Since count distributions often appear to be

positively skewed, one reasonable transformation involves taking the natural logarithm of the

responses to the predictor variables. However, assuming the zeros haven't been deleted, the

transformation will not work since the natural logarithm of zero is undefined (Zhou & Tu, 1999;

King, 1989).

Sometimes natural log transformations for zero are handled by adding a small value, such

as .001, to the zeros. However, this then leads to an inflation of that transformed adjusted value.

If 70% of the scores are zero, the resulting transformed distribution will have a 70% abundance

of the transformed value (Delucchi & Bostrom, 2004). 2 Further, since the transformation is

linear, this technique has been shown to yield biased parameter estimates that differ as a function

of the adjustment quantity (King, 1989). Although the undefined log zero problem has been

handled, the original problems pervade. As Welsh, Cunningham, Donnelly, & Linenmayer

(1996) state, "It is clear for data with many zero values that such an approach will not be valid as

the underlying distributional assumptions (linearity, homoscedasticity and Gaussianity) will

[still] be violated" (p.298). Finally, for any technique, transformations sometimes create a new

problem while solving the old one; "a transform that produces constant variance may not

produce normality .. (Agresti, 1996, p.73)".


2 This implies then that, beyond the dual generating process for zeros, the problem can be generalized from inflated
zeros to inflated lower boundaries for count data.









Generalized Linear Models

Bryk and Raudenbush (1996) state, "There are important cases .. for which the

assumption of linearity and normality are not realistic, and no transformation can make them so"

(p.291). As previously demonstrated, count data is likely to be one such case. Instead of deleting

cases or transforming the data, it is more reasonable to specify a different distribution. As

explained by Hox (2002), although it is nice to be able to transform data, "modeling inherently

nonlinear functions directly is sometimes preferable, because it may reflect some 'true'

developmental process" (pp. 93-94). In order for a model to be 'inherently nonlinear' (Hox,

2002), there must be no transformation that makes it linear.3 These nonlinear models belong to

the class of generalized linear models (GLM).

The following explanation of generalized linear models based on the seminal work of

McCullagh and Nelder (1989) with additional clarification by Lawal (2003) and Agresti (1996).

Lawal (2003) explains that generalized linear models are a subset of the traditional linear models

that permit other possibilities than modeling the mean as a linear function of the covariates. All

GLM possess a random component, a systematic component, and a link function. As explained

by Agresti (1996), the random component requires the specification of the distribution for the

outcome variable. One could specify this distribution to be normal; hence, classical models such

as ordinary least squares regression and analysis of variance models are included within this

broader class of generalized linear models. Other possible random components that could be

specified include the binomial distribution, negative-binomial distribution, gamma distribution,

and Poisson distribution. Specifying the random component depends on the expected population

distribution of the outcome variable. Given both zero-inflation and truncated count data yielding


3 Transforming covariates (e.g., including polynomial terms) may graphically appear to be nonlinear while still be
linear in the parameters (Singer & Willett, 2003).









an odd-shaped skewed distribution, the random component plays an important part in obtaining

valid results.

In order to better understand the formulation of the three components, it is necessary to

clarify the theoretical foundations of distributions. The probability density function for a normal

distribution is

1 (y U)2
f (y; u, a) = exp(- ), (2-3)


which, given random variable X ~ N(pu,o-'), reduces to the standard normal probability density

function

Sy2
f(y)= exp(- ),(2-4)
J2~i2

which when transformed to the cumulative density function yields

1 uZ
~(DUy = FOy) = exp(- )u.(2-5)


A convenient method for obtaining the parameters is to use the distribution's moment

generating function (Rice, 1995). For the normal distribution, this function is

0 t2
M, y(t) = E[exp(tY)] = exp(put + ) (2-6)


The logarithm of the moment generating function yields the cumulant generating function,

which then yields the moments of the distribution. For the normal distribution, the first moment

is the mean (pU), and the second moment is the variance (0 ).

Strictly speaking, a requirement for GLM is that the outcome has a distribution within the

exponential family of models (EFM) (McCullagh and Nelder, 1996). These distributions are









defined primarily by a vector of natural parameters (0) and a scale parameter (#). The

formulation is given by

(yB b(0))
f, (y; 8, #) = exp (( ) + c(y, #)) (2-7)
a(#)

At first glance, it seems odd to include the normal distribution in the EFM; however, first

recall the probability density function


f (y; u, a) = exp(- ) (2-8)


Algebraic manipulation reveals that the normal distribution is indeed an EFM formulation.

( y p U)2 / 2) 1 y
EF~Zw,,u) = f y(8, ~) = exp (( ) o(zc) (2-9)
o 20

Here, the natural (i.e., canonical) parameter is pu, and the scale parameter is a .

These parameters need to be estimated. McCullagh and Nelder (1996) explain the

estimation as follows: "In the case of generalized linear models, estimation proceeds by defining

a measure of goodness of fit between the observed data and the fitted values that minimizes the

goodness-of-fit criterion. We shall be concerned primarily with estimates obtained by

maximizing the likelihood or log likelihood of the parameters for the data observed" (p. 23-24).

This turns out to be the log of the EFM function.

e(0, #; y) =log( f, (y; 0, #)) (2-10)

The natural and scale parameters are estimated by derivations revealing the mean function

E(Y)= pu = b '(0), (2-11)

and the variance function

var(Y) = b "(0)a(#) (2-12)









Note that the mean function depends on only one parameter. However, as McCullagh and

Nelder (1989) explain, .. the variance of Y is the product of two functions; one, b "(0) ,

depends on the [canonical] parameter (and hence on the mean) only and will be called the

variance ju~nction [denoted V(pu)], while the other [a(#) ] is independent of I and depends only

on # .. The function a(#) is commonly of the form a(#) = / w (p.29) and is commonly called

the dispersion parameter. For the normal distribution, the natural parameter is the mean (pu); the

variance function, V( pu); equals 1.0, and the dispersion parameter is cr

The systematic component is simply the model for the predictors established as a linear

combination and is denoted r. The link function, g(-), brings together the random component

and the systematic component hence linking the function for the mean, pu, and the function for the

systematic component, r, as r = g( pu). In other words, it specifies how the population mean of

the outcome variable with a particular distribution is related to the predictors in the model. If

g( pu) redundantly equals pu, then the population mean itself is related to the predictors. This is

termed the identity link and is exactly the function used to link the mean of the normal

distribution to its covariates.

The key advantage of GLM is that they are not restricted to one particular link function.

Many other links are available. For example, one could specify the log link as g( pu) = log( pu) or

the logit link as g( pu) = log[ pu / (1- p)]. However, each random component has one common

'canonical' link function that is best suited to the random component (McCullagh & Nelder.

1996). Alternatively, "Each potential probability distribution for the random component has one

special function of the mean that is called its natural parameter" (Agresti, 1996, p.73). For

example, a normal random component usually corresponds to an identity link, a Poisson









distributed random component usually corresponds to a log link, and a binomial distributed

random component usually corresponds to a logit link. In sum, the canonical link and natural link

are two equivalent terms for specifying the most suitable link connecting a particular distribution

for the outcome variable with its linear systematic covariate function.

The Binomial Distribution and the Logit Link

Suppose we were interested in the differences between households with zero children

under age 21 and households with one or more children over the age of 21. We could feasibly

collapse all nonzero responses in the AEWR data into a value of one. Now, 71.58% of the values

are zeros, and 28.42% of the values are ones. This distribution is obviously not normal. We have

now introduced both a lower bound (zero), an upper bound (one), and an inherently nonnormal

distribution; hence, a different random component and link can be specified to accommodate

these constraints.

Variables that take on only one of two values are known as binary, or Bernoulli,

variables, and the distribution of multiple independent trials for these variables is termed

binomial. Bernoulli responses are modeled in terms of the probability (Pr) that the outcome

variable (Y) is equal to either zero or one. The random component over multiple independent

trials is thus a binomial distribution with parameters n for the number of trials and ai for the

probability that Y= 1.

Y ~ B(n, zi) (2-13)

The binomial distribution assumes that the responses are dichotomous, mutually exclusive,

independent, and randomly selected (Agresti, 1996). Since the responses are discrete, the

probability density function is termed the probability mass function and is defined as


f (k; n, p) = n!" k n-"k]i. (2-14)
k !(n k)!i









This function gives the p probability of k ones (i.e., heads, hits, successes) over n trials. Rice

(1995) clarifies, "Any particular sequence of k successes occurs with probability [pkl n-"k

from the multiplication principle [i.e., independent probabilities of realizations being


multiplicative]. The total number of such sequences is [permutations], since there are
k !(n k)!i


ways to assign k successes to n trials" (p.36). The moment generating function is
k !(n k)!i

{1- xi + xi exp(5)) ", and the cumulant generating function is n logt 1- ai + xi exp(5) } .

The log-likelihood function is virtually the same as the probability mass function. However, now

we are determining the value of p as a function of n and k (rather than determining k as a

function of p and n) while taking the log of this maximum at


e[f(p; n,k)] = n! k1 -k] (2-15)
k !(n k)!i

The estimates are obtained through derivations of the likelihood function as was previously

discussed for the normal distribution. Just as the normal distribution population mean, pu, has the

best maximum likelihood estimates of X the binomial distribution population probability, zi,

has the best maximum likelihood estimate of k divided by n, which is the proportion of ones,

hits, or successes. This greatly reduces calculations when a quick estimate is needed and the

random component is not linked to any predictors.

The binomial distribution eventually converges to a normal distribution. However, the

speed of this convergence is primarily a function of skew,


Men y) =(2-16)


with p = .50 yielding the fastest convergence (McCullagh and Nelder, 1996).









The link function should account for the binomial outcome variable. If linear predictors

are used to predict a probability, then we have predicted values in an infinite range rather than

constrained to be between zero and one. What is needed is a link function that will map a

bounded zero-one probability onto this range of infinite values. The canonical link for the

binomial distribution is the logit link.

r = g(s) = log(ir /(1- zi)) (2-17)

A logit is the natural log of an odds ratio or log[p / (1-p)]. An odds ratio is equal to a

probability divided by one minus that probability. Hence, if the probability is .5, then the odds

are .5 / (1-.5) = 1 meaning that the odds of a one are the same as the odds of a zero [i.e., the odds

of success and failure are identical]. If the probability is .75, then the odds are .75 / (1-.75) = 3

meaning that a response of one is three times more likely than a response of zero. The reciprocal

odds of 1/3 means that a response of one is three times

less likely than a response of zero, which is equivalent to stating that a response of zero is

three times more likely than a response of one. When using the logit link to connect the binomial

random distribution and the systematic component, the generalized linear model is


logity)= log( )= f, + fX, +...X (2-18)

A probability of .50, which is an odds of one, corresponds to a logit of zero. Odds favoring

a response of one yield a positive logit, and odds favoring a response of zero yield a negative

logit. Hence, the mapping is satisfied since the logit can be any real number (Agresti, 1996).

The regression parameters are slope-like in that they determine the relative rate of change

of the curve. The exact rate of change depends on each probability with the best approximation









at that probability being Pfr(1 zi) with the steepest rate of change being at ai = 0.50, which is

where X = -a /p (Agresti & Finlay, 1997).

Since natural logs can be reversed through exponentiation and since odds can be converted

to probabilities by dividing the odds by the sum of the odds and one, the fitted equation can be

used to predict probabilities via

exp(a + ,X, + ---+ PkXk)
a = (2-19)
1+ [exp(a + ,X, + ---+ PkXkl

It is more common to interpret logistic regression coefficients by only exponentiating

them. Then, the coefficient has a slope, rather than slope-like, interpretation; however, the

relationship is multiplicative rather than additive. Specifically, the expected outcome is

multiplied by exp( P) for each one-unit increase in X.

Evaluating Model Fit

The next step in interpreting generalized linear model results is to determine how well the

estimated model fits the observed data, where fit is the degree of discrepancy between the

observed and predicted values. McCullagh and Nelder (1989) explain, "In general the pus will

not equal the y' s exactly, and the question then arises of how discrepant they are, because while a

small discrepancy might be tolerable a large discrepancy is not" (p.33). The goodness of fit

improves as the observed values and predicted values approach equality. For example, if a

scatterplot reveals that all points fall on a straight line, then the predictive power of the

regression equation would be perfect, and the subsequent fit would be perfect.

The comparison is usually performed through some statistical comparison of the observed

outcome values and the predicted (i.e., fitted) outcome values. Rather than compare and

summarize the actual observed and predicted values, it is common to gain summary information










by inspecting the log-likelihood value produced from the estimation procedure. Since the model

parameters are estimated are from the data, perfect fit (i.e., observed-fitted = 0) is rare.4 Hence,

the goodness of the fit is measured to determine whether the difference is small enough to be

tolerated.

There are many measures of model fit. Typically, the model is compared either to a null

model in which the only parameter is the mean or a full model in which the number of

parameters is equal to the sample size. "It is well-known that minus twice the LR statistic has a

limiting chi-square distribution under the null hypothesis" (Vuong, 1989, p.308). McCullagh

and Nelder (1989) equivalently state, "The discrepancy of a fit is proportional to twice the

difference between the maximum log likelihood achievable and that achieved by the model

under investigation" (p.33). This deviance statistic (G2) iS then considered to be asymptotically

distributed chi-square with degrees of freedom equal to the number of parameters subtracted

from the sample size.' A significant p-value indicates that the deviance is greater than what

would be expected under a null hypothesis that the model with less parameters is adequate;

hence, the observed model with an additional parameter or parameters is considered a significant

improvement over the null model.

Another measure of model fit is Pearson's X2; however, unlike G2, it is not additive for

nested models. Yet another measure of model fit is Akaike's Information Criterion (AIC), which

penalizes the deviance for the number of parameters in the model.6 The notion is that increasing



4 Perfect fit is always obtained if the number of parameters and the sample size are identical (McCullagh & Nelder,
1989).
5 The relationship is not always exact since sometimes the deviance is scaled and/or the likelihood is more difficult
to estimate than in the simple logistic regression scenario presented here (McCullagh & Nelder, 1989).

6 Other measures such as the Bayesian Information Criterion (BIC) penalize for both the number of parameters and
the sample size.









the number of parameters will increase the log-likelihood regardless of the model and the data.

Hence, the AIC penalizes the log-likelihood with regard to the number of parameters. There are

two variations that provide further penalties. The Bayesian Information Criterion (BIC) penalizes

for sample size; the Consistent Akaike Information Criterion (CAIC) penalizes even further by

considering sample size and adding a small adjustment (Cameron & Trivedi, 1998). These

indices can be compared to those of competing models; however, this must be done

descriptively, not inferentially. The disadvantage is that the AIC can not be compared to a

statistical distribution resulting in probabilities for significance testing; however, the advantage is

that, as a descriptive statistic, it can be used to compare nonnested models.

The explanation thus far points to the fact that models can be compared to null, full, or

other models. Statistical comparison is valid to the extent that one model is nested within the

other, which is to say that both models share the same parameters, and one model has at least one

parameter that is not included in the other. Alternatively, Clarke (2001) defines the models as

follows: "Two models are nested if one model can be reduced to the other model by imposing a

set of linear restrictions on the parameter vector .. Two models are nonnested, either partially

or strictly, if one model cannot be reduced to the other model by imposing a set of linear

restrictions on the parameter vector" (p.727). The deviance for comparing two models is

calculated as the difference in log likelihood between the two models then multiplied by -2.

This quantity is asymptotically distributed chi-square with degrees of freedom equal to the

difference in parameters between the two models (Agresti, 1996). A significant p-value indicates

that the deviance is greater than what would be expected under a null hypothesis of model

equivalence; hence, the more complex model with an additional parameter or parameters is

considered a significant improvement over the nested model.









The difference in log-likelihood statistics (i.e., deviance) can not be used to statistically

test nonnested models. This is due to the fact that neither of the models can be considered the

simple or more complex models with additional variables leading to a probabilistically higher

log-likelihood. A t-test (or F-test) is a sensible alternative that eliminates concern for nesting.

However, Monte Carlo simulations have demonstrated that, for model comparison tests, the F-

test is lacking in sufficient power and can result in multicollinearity (Clarke, 2001).

The motivation for the AIC statistic is that, all else being equal, "the greater the number of

coefficients, the greater the log-likelihoods" (Clarke, 2001, p.731). Hence, model fit becomes

impacted by the number of variables in the model along with the effects of those variables.

Hence, the AIC penalizes for the number of parameters. The formula is

AIC = -2(LL) + 2K (2-20)

where LL is the log-likelihood estimate and K is number of parameters in the model including the

intercept. Hence, now the log-likelihood is adjusted to accommodate simplicity and parsimony

(Mazerolle, 2004).

In actuality, one could compare log-likelihoods between nonnested models. However,

beyond the lack of parameter penalty, this technique might lead to the statistical hypothesis

testing associated with log-likelihood statistics (i.e., test for the deviance approximated

by X2).The AIC, on the other hand, should not be used in a formal statistical hypothesis test

regardless of whether the model is nested or nonnested (Clarke, 2001). Generally, the researcher

looks at several AIC indices and decides which model fits best based on a lower-is-better

criterion. Mazerolle (2004) states, "The AIC is not a hypothesis test, does not have a p-value, and

does not use notions of significance. Instead, the AIC focuses on the strength of evidence .. and

gives a measure of uncertainty for each model" (p. 181).










Logistic modeling necessitated treating all nonzero numbers of children as a value of one.

Depending on the research question, this may be a loss of valuable information (Slymen, Ayala,

Arredondo, & Elder, 2006). Although sometimes it is necessary to model zero-inflated binomial

data (Hall, 2000), specifying a binary distribution and logit link is not an ideal method for

handling zero-inflated count data. The generalized linear model that specifies a binomial

distribution and a logit link becomes more relevant when discussing the technique of splitting

zero-inflated data into a model for the probability of zero separate from or combined with a

model for the counts.

The Poisson Distribution and the Log Link

McCullagh and Nelder (1989), Lawal (2003), and Rice (1995) are the key references for the

technical underpinnings for this model and distribution. The generalized linear 'Poisson' model

is considered to be the benchmark model for count data (Cameron & Triverdi, 1998).7 This is

primarily attributed to the fact that the Poisson distribution has a nonnegative mean (Agresti,

1996). If y is a nonnegative random variable, the Poisson probability mass function is given by

ef~Z
f (k; A2) = Pr(Y = k) = ,k = 0, 1, 2, .. (2-21)


where Ai is standard Poisson notation for the mean ( p) and k is the range of counts. Derivations

by Rice (1995) show that the expected value of a random Poisson variable is Ai; hence, "the

parameter Ai of the Poisson distribution can thus be interpreted as the average count" (p. 113).

Alternatively, lambda ( Ai) represents "the unobserved expected rate or occurrence of events ...'

(Zorn, 1996, p.1). The moment generating function is



SIt is also commonly used to model event count data, which is "data composed of counts of the number of events
occurring within a specific observation period .. [taking the] form of non-negative integers (Zorn, 1996, p.1)".










E(e") = e;"'exp't)-1). (2-22)

The resulting cumulant generating function is Al(exp(t)-1i), which, with a variance function of

Ai and a dispersion parameter equal to one, leads to mean and variance both being equal to Ai, and

the skew equal to one divided by the square root of Ai This equivalence of the mean and

variance defined by a single parameter (Cameron & Triverdi, 1998; Agresti, 1996) is the result

of a function that yields residuals that sum to zero (Jang, 2005); hence, the systematic portion of

a Poisson GLM has no error term.

The Poisson distribution is a generalization of a sequence of binomial distributions.

Rodriguez (2006) explained that "the Poisson distribution can be derived as a limiting form of

the binomial distribution if you consider the distribution of the number of successes in a very

larger number of Bernoulli trials with a small probability of success in each trial. Specifically, if

Y ~ B(n, zi), then the distribution of Y as n + oo and 71 + 0 with pu = nzi remaining Eixed

approaches a Poisson distribution with mean pu Thus, the Poisson distribution provides an

approximation to the binomial for the analyses of rare events, where ri is small and n is large

(p.3). Rice (1995) clarified, "The Poisson distribution can be derived as the limit of a binomial

distribution as the number of trials, n, approaches infinity and the probability of success on each

trial, p, approaches zero in such a way that np = Ai (p.43). Scheaffer (1995) and Rice (1995)

have derived the generalization.

Further, just as a binomial distribution converges to a normal distribution given sufficient

trials, the Poisson distribution converges to a normal distribution given a large mean.

The log-likelihood function for the Poisson distribution is

(A, y) = C y, log A, -1,, (2-23)









with the maximum likelihood estimate of Ai simply being the sample mean (Rice, 1995) and

with the related deviance function being

Devian~ce(A, y) = 2C { y, log(y, / A) -(y, A,~) }. (2-24)

McCullagh and Nelder (1989) state that the second term is often ignored. "Provided that the

fitted model includes a constant term, or intercept, the sum over the units of the second term is

identically zero, justifying its omission" (McCullagh & Nelder, 1989, p.34).

The systematic portion of the generalized linear model takes the form

Ai, = exp(x, p) = exp(x,l p,) exp(x,l p,)... exp(xk; pk) (2-25)

which is often equivalently expressed as

log(A~) =P' XI (2-26)

with p derived by solving the equation


(ex(y, )- epxu), = 0 (2-27)


by using iterative computations such as the Newton-Raphson. The canonical link for a

generalized linear model with a Poisson random component specification is the log link (Stokes,

Davis, & Koch, 1991).

9= lo(A),Y ~ (A).(2-28)

The Poisson distribution is not limited to count variates. Cameron and Triverdi (1998)

explain that, although counts are usually in the purview of directly observable cardinal numbers,

they may also arise through a latent process. In other words, ordinal rankings such as school

course grades may be discretized as pseudocounts and assumed to have a Poisson distribution.

Hence, the results of an analysis based on a Poisson distribution and the results using an ordinal

analytic technique are often comparable.










As is almost always the case, it is common to identify other variables associated with the

count variable (i.e., misspecification). However, the Poisson model has in interesting feature in

that it assumes that there are no variables excluded from the model that are related to the count

variable. In other words, there is no stochastic variation (i.e., no error term) (Cameron & Trivedi,

1998). Modifications must be made when one wishes to use a Poisson model with stochastic

variation.

Iterative Estimation

Agresti (1996) clarifies the iterative estimation procedure. "The Newton-Raphson

algorithm approximates the log-likelihood function in a neighborhood of the initial guess by a

simpler polynomial function that has shape of a concave (mound-shaped) parabola. It has the

same slope and curvature location of the maximum of this approximating polynomial. That

location comprises the second guess for the ML estimates. One then approximates the log-

likelihood function in a neighborhood of the second guess by another concave parabolic

function, and the third guess is the location of its maximum. The successive approximations

converge rapidly to the ML estimates, often within a few cycles" (p.94). The most common

methods for estimating standard errors include Hessian maximum likelihood (MLH) (i.e., second

partial derivative based) standard errors and maximum likelihood outer products (MLOP) (i.e.,

summed outer product of first derivative) estimation.

Interpretation of Coefficients

Standard ordinary linear squares regression lends an interpretation of P as the predicted

additive change in the response variable per one-unit change in the predictor variable. However,

as was the case with the binomial distribution, the interpretation differs when considering

exponential distributions. For the Poisson distribution, "a one-unit increase in X has a

multiplicative of exp(/7) on the pu The mean of Y at x+1 equals the mean of Y at x multiplied










by exp( P)" (Agresti, 1996, p.81).s Due to the inherent difficulty in interpretation, it is common

to express in one of three alternative ways. First, the direction of the sign of P indicates a

positive or negative 'effect' of the predictor on the count variable. Second, the fitted value can be

calculated at the mean.9 Third, some interpret the coefficient in terms of percent change; hence,

if f =1.64, then as X increases to X+1, the predicted probability increases by 64% (Agresti,

1996).

Hypothesis testing

After conducting the analysis and estimating parameters, hypotheses can be tested in

several ways as explained by Agresti (1996). One could test the hypothesis that P=0 using the

traditional Wald :-test via z = b / seb Some programs provide Wald test results that are actually

Z2; this is the Wald X2 Statistic with one degree of freedom and appropriate only for a two-tailed

test. A third method, the Score test, is "based on the behavior of the log-likelihood function at the

null value for p =0" (Agresti, 1996, p.94) yielding a chi-square distributed statistic with one

degree of freedom.

Overdispersion

In practice, the assumption of an equal mean and variance is the exception rather than the

norm (McCullagh & Nelder, 1989). It is often the case that the sample variance is greater than or

less than the observed sample mean with these two seldom being statistically equivalent

(Cameron & Trivedi, 1998), especially for zero-inflated data (Welsh, Cunningham, Donnelly,




SThis is similar to the interpretation for the binomial distribution with logit link: however. now the multiplicative
effect is directly on p rather than on the odds of p.

9 This can be particularly troublesome since that fitted value will only hold at the mean. It provides no valid
inference for values greater than or less than the mean since the function is a curve with steepness that can vary
drastically between separate values for the predictors.










and Lindenmayer, 1996). This condition is known as overdispersionl0 (underdispersion) and is a

violation of a maj or tenet of the Poisson distribution that the conditional mean and conditional

variance of the dependent variable are equal (i.e., equidispersion, nonstochasticity) (Jang, 2005;

Zorn, 1996). 1 This assumption of equidispersion is the analog of the ordinary least squares

regression assumption of homoscedasticity.

The overdispersion has been explained as heterogeneity that "has not been accounted for

[that is] unobserved (i.e., the population consists of several subpopulations, in this case of

Poisson type, but the subpopulation membership is not observed in the sample" (Boihning, Dietz,

Shlattman, Mendonca, & Kirchner, 1999, p. 195). The impact of violation is one of incorrect

conclusions due to inaccurate t-statistics and standard errors (Cameron & Triverdi, 1998;

Agresti, 1996). "The estimates of the coefficients can still be consistent using Poisson regression,

but the standard errors can be biased and they will be too small" (Jewell & Hubbard, 2006, p.14).

Alternatively, Slymen, Ayala, Arredondo, and Elder (2006) state that "Confidence intervals for

regression estimates may be too narrow and tests of association may yield p-values that are too

small" (p.2). The underlying mechanism for overdispersion is explained as unobserved

heterogeneity in responses.

It is apparent that some modification to the variance to accommodate over-dispersion is

ideal. Typically, maximum likelihood procedures are used to estimate parameters in the model.

"The term pseudo- (or quasi-) nzaxinsun likelihood estimation is used to describe the situation in

which the assumption of correct specification of the density is relaxed. Here the first moment

[i.e., the mean] of the specified linear exponential family density is assumed to be correctly

'o Overdispersion is sometimes referred to as extra-Poisson variation (Bdhning, Dietz, Shlattman, Mendonca, &
Kirchner, 1999).

11 The under- or overdispersion may disappear when predictors are added to the model: however, this is likely not
the case if the variance is more than twice the mean. (Cameron & Trivedi, 1989).










specified, while the second [i.e., the variance] and other moments are permitted to be incorrectly

specified" (Cameron & Triverdi, 1998, p.19). Hence, the Poisson distribution as a baseline

(Ridout, Demetrio, & Hinde, 1998) can be modified to accommodate overdispersion and

underdispersion (Cameron & Triverdi, 1998).

Rice (1995) states that "gamma densities provide a fairly flexible class for modeling

nonnegative random variables" (p. 52). One way to accommodate overdispersion is to consider

the unobserved heterogeneity as a gamma distributed disturbance added to the Poisson

distributed count data (Jang, 2005). In other words, an individual score may be distributed

Poisson with a mean of 2 but then this mean is "regarded as a random variable which we may

suppose in the population to have a gamma distribution with mean pu and index p / # "

(McCullagh & Nelder, 1989, p. 199). This mixture leads to the negative binomial distribution.

Given gamma function, r, and count y, the negative binomial probability mass function is


pr (Y = y; pu, #) = (2-29)


Cameron and Trivedi (1998) provide a formulation where the negative binomial extends from

the Poisson rather than being explained as a mixture distribution. Given a set of predictor

variables, we can define

co, = pu, +apu,P (2-3 0)

where a is scalar parameter to be specified or estimated, and p is a pre-specified power term. If

the scalar parameter is set to zero then the resulting variance is equal to the mean, and the

Poisson distribution holds. Hence, the Poisson model is nested within the negative binomial

model .









The standard formulation for the negative binomial formulation of the function,

sometimes called NB2 (Cameron & Trivedi, 1998), leads to a variance that is quadratic by

setting p to 2

co, = pu, + a p, (2 31)

This is the formulation seen in most textbooks (Rodriguez, 2006; Scheaffer, 1996). Its mean is

the same as that of the Poisson distribution; however, its variance is derived from the gamma

distribution (Cameron & Trivedi, 1998). Just as the Poisson distribution converges to a binomial

distribution, the negative binomial distribution converges to a Poisson distribution (Jewell &

Hubbard, 2006).

Poisson and Negative Binomial Models with Zero-Inflation

As elaborated upon previously, the zero-inflation problem is two-fold. First, the proportion

of zeros is higher than expected given the specified population distribution shape resulting in an

excess zeros problem. This can be descriptively determined by calculating the expected number

of zeros as

E(fq(Y))= fq(Y)*(exp(-Y)) = ne" (2-32)

For example, Zorn' s example had a frequency of 4,052 with a A = 0.11. The expected frequency

of zeros would be

E( fq(Y)) = 4052* (exp(-0. 109)) = 3, 634, (2-33)

which is less than the 3,882 zeros observed in the data. It turns out to be (3 882/3634)* 100 =

107% of the expectation (Zorn, 1996).

Second, the zeros can be a mixture of structural (i.e., true) zeros (Ridout, Demetrio, &

Hinde, 1998) and sampled zeros reflecting the multiple sources of zeros problem. Shankar,

Milton, and Mannering (1997) state that if "a two-state process is modeled as a single process ..










if applying traditional Poisson and NB distributions, the estimated models will be inherently

biased because there will be an over-representation of zero observations in the data, many of

which do not follow the assumed distribution of [the] frequencies" (p.830O)" Shankar, Milton, and

Mannering (1997) note that the negative binomial model "can spuriously indicate overdispersion

when the underlying process actually consists of a zero-altered splitting mechanism" (p.835-

836). In sum, the sources of zeros arise from a dual generating process (i.e., structural and

sampled) leading to two sources of unequal mean/variance dispersion (i.e., that due to

unobserved heterogeneity of responses and that due to zero-inflation).

Most complex methods for analyzing zero-inflated count data model a mixture of two

different distributions. The justification for splitting the distribution into two pieces is well-

reasoned by Delucci and Bostrom (2004). "If it is deemed more reasonable to consider the zeros

as indicators of cases without a problem, a more appropriate approach is to ask two questions: is

there a difference in the proportion of subjects without the problem [i.e., structural true zeros],

and, for those who have a problem [sampled false zeros], is there a difference in severity" (p.

1164).

Zorn (1996) refers to 'dual regime' models "wherein an observation experiences a first

stage in which there is some probability that its count will move from a 'zero-only' state to one

in which it may be something other than zero" (p.2). Typically, the dual-regime is composed of a

ttrttrttrtranstion~rtrt~t stage based on a binomial distribution and an events stage based on some type of

Poisson distribution.

There are many ways to model two-part distributions. For example, Mullahy (1986) and

King (1989) proposed a Hurdle model in which the zeros are analyzed separately from the

nonzeros. Lambert (1992) proposed a zero-inflated Poisson (ZIP) model in which different










proportions of zeros are analyzed separately and along with the nonzeros. Another early

formulation (Heilborn, 1989) was the zero-altered Poisson (ZAP) model. "Arbitrary zeros are

introduced by mixing point mass at 0 with a positive Poisson that assigns no mass to 0 rather

than a standard Poisson (Lambert, 1992, p.1)". Mullahy (1986) presented a variation of the

Hurdle model based on a geometric distributionl2 for use when specifying a Poisson distribution

is not reasonable. Another possibility is to specify a log-gamma distribution for the event stage

(Moulton, Curriero, & Barruso, 2002). Lambert (1989) presented a variation to the ZIP model

known as ZIP(z), which introduced a multiplicative constant to the event stage covariance matrix

in order to account for the relationship between the two models. Gupta, Gupta, and Tripathi

(1996) derived an adjusted generalized Poisson regression model for handling both zero-inflation

and zero-deflation; however, accuracy was suggested to be contingent on the amount of inflation

or deflation. 13

It is also possible to formulate the model with different link functions. Lambert (1989)

mentions the possibility of using the log-log link, complementary log-log link (Ridout, Demetrio,

& Hinde, 1998), and additive log-log link while Lachenbruch (2002 ) mentions the lognormal

and log-gamma distributions. Hall (2000) formulated a two-part model for zero-inflated binomial

data. Gurmu (1997) describes a semi-parametric approach that avoids some distributional

assumptions (Ridout, Demetrio, & Hinde, 1998).

There is some research on the extension of two-part models to accommodate random

effects (Min & Agresti, 2005; Hall & Zhang, 2004; Hall, 2004; Hall, 2002; Olsen, 1999). Hall's

(2000) model for zero-inflated binomial data permits both fixed and random effects. Dobbie and


12 This distribution is an extension of the binomial distribution where the sequence in infinite. It is typically used in
cases where the researcher is concerned with probability up to and including the first success (Rice, 1995).

13 Min (I 'li 1)Stated that the Hurdle model also has this feature.










Walsh (2001) permit correlated count data. Finally, Crepon and Duguet (1997) consider the

cases where the variables are latent and correlated.

The evolution of the research to date has led to an emphasis on the standard Hurdle model

and ZIP models (along with their negative binomial extensions) with a binary distribution for the

transition stage and a Poisson distribution for the events stage and with fixed covariates. For both

models, estimates are obtained from maximum likelihood procedures, although there has been

some research on the use of generalized estimating equations (GEE) (Hall & Zhang, 2004;

Dobbie & Welsh, 2001). The models are primarily distinguished by whether zeros are permitted

in the event stage. In other words, their differences are a reflection of the researcher' s notions

about the potentially multiple sources of zeros in the data and their relationship to excess zeros.

They also differ in terms of the transition stage cumulative probability function. To be

clarified in the sections that follow, Zorn (1996) summarizes the differences as follows: .. the

hurdle model has asymmetric hurdle probability while in the ZIP specification p, is symmetrical.

Also, the hurdle model does not permit zero values to occur once the 'hurdle' has been crossed,

while in the ZIP model zeros may occur in either stage" (p.4). Choosing between the models is a

matter of validity; hence, the choice rests on substantive ground as well as statistical

considerations. As Martin, Brendan, Wintle, Rhodes, Kuhnert, Field, Low-Choy, Tyre, and

Possingham (2005) note, "it is imperative that the source of zero observations be considered and

modeled accordingly, or we risk making incorrect inferences .. (p.1243-1244).

The Hurdle model

The Hurdle model was developed separately by Mullahy (1986) in economics and King

(1989) in political science, although the term itself was most likely coined by Cragg (1971). 14


14 Cragg (1971) proposed the basic two-step process in which the probability of occurrence is modeled separately
from frequencies of occurrence.









Welsh, Cunningham, Donnelly, and Lindenmayer (1996) refer to it as a 'conditional Poisson

model'. "The idea underlying the hurdle formulations is that a binomial probability model

governs the binary outcome whether a count variate has a zero or a positive realization [i.e., a

transition stage]. If the realization is positive the 'hurdle' is crossed, and the conditional

distribution of the positives is governed by a truncated-at-zerol5 count data model [i.e., events

stage]" (Mullahy, 1986, p.345) such as a truncated Poisson or truncated negative binomial

distribution (Min & Agresti, 2005). In other words, one distribution addresses the zeros while

another distribution addresses the positive nonzero counts. For example, for grade-retention data,

there would be a model for schools with no dropouts and a model for school with at least one

dropout.

It is a "finite mixture generated by combining the zeros generated by one density with the

zeros and positives generated by a second zero-truncated density separately .. (Cameron &

Trivedi, 1998, p.124). Log-likelihood values are estimated separately for each density. A key

feature of the transition model is asymmetry in that the probability of crossing the hurdle

increases more quickly as the covariates increase than it decreases as the covariates decrease.

The function is then "asymmetrical" (King, 1989) leading to validity concerns supporting or

refuting the substantive theory underlying the model.

The two-part distribution of the dependent variable is given first by the transition stage g;

probability mass function

P(Y I= 0) =g,(0) 16 (2-34)





'5 The lower bound is then one.

16 This is an alternative representation of the aforementioned Pr(Y = 0) = 1 Fr .









modeling whether the response crosses the hurdle of zero. Assuming a Poisson distribution and

log link, Zorn (1996) expands the cumulative distribution function to include covariates as

p,= 1- exp[- exp(Xof 0)] (2-3 5)

The basic model for the event stage is then the probability for a nonzero realization multiplied by

the probability for the counts. 1


P(Y I = j) = (1 -g,(0)) 2 ..j = 1, 2,... (2-3 6)
1- g2 (0)

Greene (1994) notates the models with a binomial distribution and logit link for the transition

stage and a Poisson distribution with a log link for the event stage asls

Transition: Pr(y, = 0) = p, (2-3 7)
1- p e A/2
Event: Pr(y, = k) = ( ( ') ,2..(2-38)
1-e k!

Here, p is the probability of a count of zero while ii is the truncated Poisson mean for the counts

greater than zero. The generalized linear models as a function of covariates is then


Transition Stage: log( ) = xllB, (2-3 9)
1 p,
Event Stage: log(il,) = x22B2 (2-40)

The two vectors of parameters are estimated j ointly.

A, = [n {exp[- exp(XP)]} C {-exp[-ex p(Xp)]}\1 (2-41)
A2 = On Xp(yXp) /({ex ep (X p)]-Y R1}y!)],, (2-42)

An alternative notation provided by Min (2003) is




17 The second term is an alternative representation of the aforementioned f (k; ii) = Pr(Y = k) =

1s Note that p is used to representative the probability of a zero rather than the conventional representation of a
probability of one.










8 (P>)= [logg(y, = 0; P,, x,l)]+C [log(1 I(yI = 0; P,,x,l))]








Since the two models are functionally independent, the likelihood functions can be maximized

separately (Min & Agresti, 2005; Min, 2003, Cameron & Trivedi, 1998; Mullahy, 1986).

A;,,,,vie = log(Az) + log(A,) (2-45)
e(p,, p,) = (p, )+ (p,) (2-46)

This is because "the large sample covariance between the two sets of parameters is zero so

the joint covariance matrices can be obtained from the separate fits" (Welsh, Cunningham,

Donnelly, & Lindermayer, 1996, p.300). Solving the likelihood equations uses either the

Newton-Raphson algorithm or the Fisher scoring algorithm, both giving equivalent results (Min,

2003). The Poisson model is nested within the Hurdle model (Zorn, 1996). Hence, fit of these

two models can be compared statistically. 19

The Negative Binomial Hurdle model

In the case of zero-inflated data, it is possible to have two sources of overdispersion. The

variance can be greater than the mean due to the preponderance of zeros. However, there is now

the possibility of unobserved heterogeneity in the event stage (Mazerolle, 2004; Min & Agresti,

2004). The former scenario has been referred to as zero-driven overdispersion (Zorn, 1996); the

latter is Poisson overdispersion. Just as was the case with the Poisson model, it is possible to nest

the Hurdle model within a more general negative binomial framework. Further, the negative

binomial Poisson model is nested within the negative binomial Hurdle model. Hence, the fit of a)


19 Zhou and Tu (1999) developed likelihood ratio for count data with zeros; however, it was not generalized to any
particular zero-inflation model.









the Poisson model and the Hurdle model, b) the negative binomial Poisson model and the

negative binomial Hurdle model, and c) the Hurdle model and the negative binomial Hurdle

model can be compared using statistical tests.20 Estimation is typically performed by solving the

maximum likelihood equations using the Newton-Raphson algorithm.

The Zero-Inflated Poisson (ZIP) model

The Zero-Inflated Poisson, or ZIP, model is another model that one can use when the

zeros in a dataset are argued to be caused by both chance and systematic factors (Min & Agresti,

2005). The transition stage addresses zero-inflation while the event stage addresses unobserved

heterogeneity of responses including zeros (Jang, 2005). Welsh, Cunningham, Donnelly, and

Lindenmayer (1996) refer to it as a mixture model.

This two-part model, developed by Lambert (1992) permits zeros to occur in both the

transition stage and event stage (Cameron & Trivedi, 1998); "crossing the 'hurdle' in the ZIP

model does not guarantee a positive realization of Y (Zorn, 1996, p.4). Further, the probability

function in the transition stage is now symmetrical (Zorn, 1996). Lachenbruch (2002) explains

that "ZIP regression inflates the number of zeros by mixing point mass at 0 with a Poisson

distribution" (p. 12). Zorn (1996, p.4) clarifies the distinction between the ZIP and Hurdle

models as follows:

As a special case of the general model, the ZIP regression is thus seen to make
substantially different assumptions about the nature of the data generating process than the
hurdle model. Whether parameterized as a logit or a probit, the probability exiting the zero-
only stage is assumed to follow a symmetric cumulative distribution. Likewise, even those
cases which make the transition to the events stage may nevertheless have zero counts;
crossing the "hurdle" in the ZIP model does not guarantee a positive realization of Y ..
The sole difference in assumptions here is that the hurdle model's count distribution is
assumed to be truncated at zero whereas the ZIP specification count data may take on zero



20 It is not the case that the negative binomial Poisson model is nested within the Hurdle model: hence, one can not
statistically compare all four models collectively.









values in the event stage.

Another difference is that, unless the ZIP model is overly parameterized, only the Hurdle

model can handle zero deflation (Min & Agresti, 2005).

Compared to the Hurdle model, the equations for the event stage are very similar. The

exception is that (1-p,) is divided by (1- e ) in the Hurdle model before being multiplied by the

remaining elements of the equation. However, the transition stage equations are strikingly

different. For the Hurdle model, the equation is PrOy, = 0) = p; the ZIP model includes the

addition of the probability of a nonzero multiplied by the exponentiated Poisson mean. This is

the mathematical characteristic that distinguishes the Hurdle model's exclusion of zeros in the

event stage and the ZIP model's potential inclusion of zeros in the event stage. Rather than

model the probability of a zero in the transition stage, the ZIP also models the probability that the

counts have a Poisson distribution hence permitting zeros from both a perfect state and a Poisson

state (Hur, 1999). Given this, 32 parameterizes the mean of this Poisson distribution (Welsh,

Cunningham, Donnelly, & Lindenmayer, 1996). When adding covariates, the Hurdle and ZIP

generalized linear model appear the same.


Transition Stage: log( ) = xllB, (2-47)
1 p,
Event Stage: log(il,) = x22B2 (2-48)

Unlike the Hurdle model, the ZIP model likelihood function can not be maximized

separately for the transition and event stage. Hence, the Hurdle model "has the attractive

advantage of an orthogonal parameterization which makes it simpler to fit and interpret than the

mixture model" (Welsh, Cunningham, Donnelly, & Lindenmayer, 1996) with the disadvantage

of asymmetrical transition to counts. The likelihood function derived by Lambert (1992) is










L=Clog(ewr +exp(-e ))+C(y:BS-e a)-[log(1+ear)-[loggy!) (2-49)
Y;=o y,>0 I=1 y,>0
where PB is the vector of coefficients and matrix of scores for the event stage and 7G is the

vector of coefficients and matrix of scores for the transition stage, and where iterations are based

on the EM or Newton-Raphson algorithms (Min, 2003; Lambert, 1992)21

Strictly speaking, the Poisson model is not nested within the ZIP model; therefore, it would

not be wise to conduct a formal model fit test (Zorn, 1996; Greene, 1994). However, it is

interesting to note that the log-likelihood of 10,607 is slightly lower than that produced by the

negative binomial Hurdle model. This is in line with Greene' s (1994) observation when using the

Vuong statistic as an alternative for testing nonnested models. "For present purposes, the

important question is whether the ZIP models .. provide any improvement over the basic

negative binomial .. The log-likelihood functions are uniformly higher, but as noted earlier,

since the models are not nested, these are not directly comparable. The Vuong statistic, however,

is consistent with the observation" (Greene, 1994, p.26).

The Negative Binomial Zero-Inflated Poisson model

The ZIP model can be extended to the negative binomial model just as the Poisson was

extended to the negative binomial and as the Hurdle was extended to the Hurdle negative

binomial. This may be necessary as Min (2003) explains that "Sometimes such simple models

for overdispersion are themselves inadequate. For instance, the data might be bimodal, with a

clump at zero and a separate hump around some considerably higher value. This might happen

for variables for which a certain fraction follows some distribution have positive probability of a

zero outcome" (p. 13). He further explains, "The equality of the mean and variance assumed by



21 Alternatively, Martin, Brendan, Wintle, Rhodes, Kuhnert, Field, Low-Choy, Tyre, and Possingham (2005),
compared relative means and credible intervals estimate from a bootstrap procedure.









the ZIP model .. is often not realistic. Zero-inflated negative binomial models would likely

often be more appropriate than ZIP models" (p.15). Unfortunately, current iterative techniques

lead to a greater risk of nonconvergence than when using the other two-part models (Fanoye &

Singh, 2006; Lambert, 1992).

The key assumption in accepting these results over those produced from the Hurdle models

is that there are some zeros that belong with the counts representing no household members

under the age of 21 for reasons other than their never having such household members at all.

Although it is not valid to statistically compare the fit of the ZIP model with the Hurdle and

Poisson models, it is reasonable to test the fit of the ZIP model within the negative binomial ZIP

model .

In sum, the choice between Hurdle models and ZIP models is ultimately guided by the

assumptions one makes about the data generating process. Min (2004) states, "The zero-inflated

models are more natural when it is reasonable to think of the population as a mixture, with one

set of subj ects that necessarily has a 0 response. However, they are more complex to fit, as the

model components must be fitted simultaneously. By contrast, one can separately fit the two

components in the hurdle model. The hurdle model is also suitable for modeling data with fewer

zeros than would be expected under standard distributional assumptions" (p.20).

Model Comparison Testing for Zero-Inflated Data

The Poisson model is nested within the negative binomial Poisson differing only by the

dispersion parameter. In fact, the two models are equivalent when one uses the negative binomial

model and restricts the dispersion parameter to 1.0 achieving Poisson equidispersion (Cameron

& Trivedi, 1998). Likewise, the Hurdle model is nested within the negative binomial Hurdle

model, and the ZIP model is nested within the negative binomial ZIP model. However, since

each involves estimation of a transition stage and event stage model, the nesting rule applies to









both equations. In other words, if the event stage contains 3 parameters and the transition stage 4

parameters for the nested model, then the more complex model must contain at least these same

3 parameters in the event stage and at least the same 4 parameters in the transition stage.

According to Zorn (1996), the Poisson model is nested within the Hurdle model, and the

negative binomial Poisson model is nested within the negative binomial Hurdle model. This is

reasonable given that the Hurdle models are estimating the Poisson models in the event stage and

that these likelihood statistics are independent of those produced in the transition stage. The ZIP

models, on the other hand, are not estimated in this manner; it is not reasonable to assume that

the Poisson models are nested within the ZIP Models (Greene, 1994). This leads to the hierarchy

of permissible model-testing displayed in Table 2-1. Other models can be compared

descriptively using the aforementioned Akaike's Information Criterion (AIC). The remainder of

the dissertation will use the deviance statistic for model comparison inference and the AIC for

model comparison description.

Review of Research Pertaining to and Using Zero-Inflated Count Data

Hurdle Model

Statistical

As previously discussed, Mullahy (1986) presented the underlying statistical foundations

for the Hurdle model. He specified it in terms of extending from both the Poisson and geometric

distributions. He also presented an extension to the Hurdle model that he termed the "with-zeros"

(WZ) model. The WZ model adjusts for zero-inflation by augmenting or reducing the probability

by an additive constant (rather than having them specified by the parent distribution). Subsequent

research has focused on the Hurdle model rather than the WZ model; this is most likely due to

the fact that the WZ turns out to be a special case that collapses into a Hurdle model in some

specifications, and estimates are often similar between the models (Mullahy, 1986). Finally,









Mullahy presented tests for specifications including a technically complex information matrix

test as well as the Score test typically provided in software output. King's (1989) key

contribution to the Hurdle model was a Monte Carlo study confirming that the counts can be

viewed as arising from a Poisson process (Civetinni & Hines, 2005).

Min and Agresti (2004) conducted a simulation to compare a zero-inflation and a zero-

deflation condition. They found that the estimates were reasonable for the ZIP model under zero-

inflation. However, the coefficient and standard error for the event stage were both very large

under zero-deflation. They explain that logit models simply can not accommodate too few zeros.

"The zero-inflated model is only suitable for zero-inflation problems. However, the hurdle model

is also suitable for modeling data with fewer zeros than would be expected under standard

distributional assumptions. In fact, when a data set is zero-deflated at some levels of the

covariates, the zero-inflation model may fail" (Min & Agresti, 2004, p.5).

In contrast, Mullahy (1986) stated that, "a particularly interesting feature of the modified

count data specifications considered here [i.e., Hurdle model] is that they provide a natural

means for modeling overdispersion or underdispersion of the data. Specifically, overdispersion

and underdispersion are viewed as arising from a misspecification of the maintained parent [data

generating process] in which the relative probabilities of zero and non-zero (positive) realizations

implied by the parent distribution are not supported by the data" (Mullahy, 1986, p.342).

Applications

Mullahy (1986) researched daily consumption of coffee, tea, and milk where each is a

count variable. Covariates included age, years of completed schooling, family income, sex, race,

and marital status. For the Hurdle model, the test statistics [p-values] for the coffee-model were

substantially smaller than those for tea-model and milk-model. Mullahy argued that this is not

surprising given the ratio of estimates and standard errors. However, closer inspection of the data









reveals that it is the coffee variable that has the lowest proportion of zeros at only 26.26%. It is

possible that the Hurdle model was unable to adequately account for the additional

overdispersion witnessed in the other two models (61.63% and 40.37% zeros). In other words, a

two-part model may be a necessary but not sufficient condition for handling overdispersion in

zero-inflated count models, and negative binomial formulations may be increasingly necessary as

the proportion of zeros increases.

King (1989) applied the Hurdle model to data for the relationship between the number of

nations entering war in a period of time as a function of those in formal international alliances.

The hurdle model was formulated based on the premises of Mullahy (1986) and justified due to

there being some countries who will not go to war and others who will not at first but will later

be 'dragged in' by alliances. Hence, this is a classic example of the justification for true zeros

and event-driven zeros. This was demonstrated statistically by comparing Hurdle and Poisson

results. The Poisson alliance coefficient of .007 was significant, the Hurdle model event-stage

coefficient of .007 was significant, and the Hurdle model transition stage coefficient of .001 was

not significant. Hence, the Poisson interpretation would be that increased alliances lead to

increased war frequency. However, the Hurdle results clarify that this is only true after the onset

of war (i.e., the hurdle has been crossed). Further statistical evidence supported the Hurdle model

based on the likelihood-ratio model comparison test.

Zero-Inflated Poisson Model

Statistical

As previously discussed, Lambert (1989) presented the original formulation for the ZIP

model and its ZIP(z) formulation. She also presented the models' extension from the Poisson and

negative binomial as well as the derivation of the maximum likelihood (EM) estimates. She ran

several simulations to test the adequacy of the model. The first simulation varied sample size,









with one covariate taking on a fixed coefficient and a fixed variance in both parts of the model.

The result was an average of 50% zeros in the transition stage and 23% zeros in the event stage.

The results suggest that the ZIP model consistently converges at n=25 when using EM and at

n=100 when using the Newton-Raphson algorithm. An examination of confidence intervals

revealed that the normal-theory intervals are not reliable at n=100; however, almost all simulated

likelihood-ratio confidence intervals contained the true mean even at n=25. "To summarize,

these simulations with one covariate for both 32 and p are encouraging. The ZIP and ZIP(z)

regressions were not difficult to compute, and as long as inference was applied only when the

observed information matrix was nonsingular, estimated coefficients, standard errors based on

observed information, and estimated properties of Y could be trusted" (Lambert, 1992, p.7).

Warton (2005) compared 20 datasets of varying sample sizes, proportions of zeros, and

factors/levels. The ordinary least squares version included the addition of one to all counts before

taking the logarithm. The other models not accommodating zero-inflations were the Poisson and

four formulations of the negative binomial Poisson (including the aforementioned quasi-Poisson

where the variance is set to
negative binomial ZIP model. The Akaike Information Criterion (AIC) values were calculated

for a total of 1,672 variables averaged over datasets and rescaled to a minimum AIC of zero for

each dataset.

As expected, when overdispersion was present, the negative binomial formulations

outperformed the models without these formulations. However, when overdispersion was not

present, the reverse was true for 53% of the variables. This suggests that the level of skew in the

model interacts with zero-inflation when measuring model adequacy.









When the proportion of zeros is very small, the distribution looks more like a Poisson

distribution truncated at zero. In other words, it shares features modeled by the event stage of a

Hurdle model. This led to estimation problems in which the negative binomial model ZIP model

rarely converged. When it did converge, its fit was better than that of the ZIP model for only

1 1% of the datasets. Possibly, as previously discussed, the zero-inflated Hurdle would converge

more often since it can handle both zero-inflation and zero-deflation (Min & Agresti, 2005; Min

& Agresti, 2004).

A very interesting finding pertained to the transformed OLS fit indices. "Although

transformed least squares was not the best fitting model for data, it fitted the data reasonably

well. Surprisingly, transformed least squares appeared to fit data about as well as the zero-

inflated negative binomial model .. The AIC for transformed least squares was not as small as

for the negative binomial model overall, although it was smaller for 20 per cent of the variables

considered here" (Warton, 2005, p.283).

Averaging over all datasets, the AIC was lowest for all negative binomial models

followed by a close tie between the transformed OLS model and the negative binomial ZIP

model, which was followed by the ZIP model. All models were a drastic improvement over the

standard Poisson model. The implications are that, although the Poisson is rarely adequate when

the data is not equidispersed and/or is inflated or deflated, an intuitive climb up the ladder of

models may not be reasonable. There were features of these datasets including varying degrees

of zero-inflation and overall distributions that warrant further investigation toward appropriate

model selection. "If one were to fit a zero-inflated model, it would be advisable to present

quantitative evidence that the zero-inflation term was required. Based on the present results, it is

likely that a term for extra zeros is not needed, and a simpler model will usually suffice ..










special techniques are not generally necessary to account for the high frequency of zeros. The

negative binomial was found to be a good model for the number of zeros in counted abundance

datasets, suggesting that a good approach to analyzing such data will often be to use negative

binomial log-linear models" (Warton, 2005, p.287-288).

In regard to their problems with zero-inflated negative binomial convergence, Fanoye and

Singh (2006) developed an extension that improves convergence termed the zero-inflated

generalized Poisson regression (ZIGP) model. Their recent research revealed convergence in less

than 20 iterations for all trials. However parameter estimates and standard errors were often very

different than those produced by the ZIP model. They conclude, "Even though the ZIGP

regression model is a good competitor of ZINB regression model, we do not know under what

conditions, if any, which one will be better. The only observation we have in this regard at this

time is that in all of the datasets fitted to both models, we successfully fitted the ZIGP regression

model to all datasets. However, in a few cases, the iterative technique to estimate the parameters

of ZINB regression model did not converge" (p.128).

Greene (1994) used the Vuong statistic when comparing the Poisson, negative binomial

Poisson, and ZIP models. It was noted that the rank order for the Vuong statistics and the log-

likelihood estimates were in alignment. The conclusion suggested future research using the

Vuong statistic. "The use of Vuong' s statistic to test the specification seems not to have appeared

in the recent literature .. We conjecture that the Vuong testing procedure offers some real

potential for testing the distributional assumption in the discrete data context. In the cases

examined, it appears to perform well and in line with expectations" (Greene, 1994, p.30).

Shankar, Milton, and Mannering (1997) used the Vuong statistic to decide between the

negative binomial, ZIP, and negative binomial ZIP model for traffic accident data. They clarify









the interpretation of the statistic stating, "A value >1.96 (the 95% confidence level of the t-test)

for V favors the ZINB while a value < -1.96 favors the parent-NB (values in between 1.96 an -

1.96 mean that the test is indecisive) .. This test can also be applied for the ZIP(z) and ZIP

cases" (p.831).

Civettini and Hines (2005) explored the effects of misspecification on negative binomial

ZIP models. This included misspecification by leaving a variable out of the event stage that was

present in the event stage and misspecification by shifting a variable from the transition stage to

the event stage.

Applications

Lambert (1992), in formulating the ZIP model, applied it to the analysis of defects in

manufacturing. In terms of improperly soldered leads, 81% of circuit boards had zero defects

relative to the 71% to be expected under a Poisson distribution linked to a model with a three-

way interaction. This most complicated model had a log-likelihood of -638.20. This

dropped to -511.2 for the ZIP model. Although comparing to a different combination of

covariates, the negative binomial Poisson model fit better than the Poisson model but not as well

as the ZIP model.

Greene (1994) used credit-reporting data to investigate differences between the Poisson,

negative binomial Poisson, ZIP, negative binomial ZIP, as well as some of their aforementioned

variants and the specification of a probit link rather than the logit link. The data consisted of

1,023 people who had been approved for credit cards. The count variable of concern was the

number of maj or derogatory reports (MDR), which is the number of payment delinquencies in

the past 60 days.

For this sample, 89.4% had zero MDR. Given a mean of 0. 13, this frequency of 804 is

nearly double the 418 we might expect in a Poisson distribution. The skew of 4.018 is reduced to









2.77 when ignoring the zeros while the mean increases to 1.22. As expected, the negative

binomial Poisson resulted in improved fit (based on the Vuong test statistic), increased standard

errors and different parameter estimates. The ZIP model resulted in slightly worse fit than the

negative binomial Poisson while remaining much better compared to the Poisson model. If all of

the overdispersion was due to unobserved response heterogeneity then the results should be

similar for the negative binomial ZIP model. However, this model produced the best fit of all.

It is interesting to note that, again, the standard errors increase while the parameter

estimates are different relative to the ZIP model. In fact, of the 6 parameters, 4 estimates

decreased, 2 increased, and 1 switched in sign. Hence, there are two implications. First, the

negative binomial ZIP model was necessary to accommodate two sources of overdispersion to

adjust standard errors. Second, ignoring the negative binomial formulations would have led to

nonsensical parameter estimates driven by a sample mean of 0. 13.

Boihning, Dietz, Schlattmann, Mendonga, and Kirchner (1999) compared pre- and post-

intervention scores on the decayed, missing, and filled teeth index (DIVFT) for 797 children in

one of six randomly assigned treatment conditions. The results were not exhaustive; however,

the log-likelihood did decrease from -1473.20 to -1410.27 when going from the Poisson model to

the ZIP model. This study was somewhat unique in that all the covariates (sex, ethnicity, and

condition) were categorical, and that the conditions were dummy-coded represented as five

parameters. Also, this data had features that might suggest that zero-inflated models weren't

necessary. For pre-intervention, the proportion of zeros was 21.58%, which increased to only

28.99% at post-intervention. The means, with the zeros in the data, were 3.24 and 1.85,

respectively. Ignoring the zeros changed these means to 4.13 and 2.61, respectively. The skew,

with the zeros in the data, was 0.20 and 0.65, respectively. Ignoring zeros changed the skew to









0.08 and 0.63, respectively. In other words, many features of the data were consistent with what

would be expected of a Poisson, and possibly normal, distribution. Nonetheless, with these

means and frequencies, the Poisson distribution suggests overdispersion with 31 permissible

zeros for the pre-intervention and 125 permissible for the post-intervention whereas the data

revealed 173 zeros and 232 zeros, respectively. It then becomes a matter of whether the

overdispersion was due to the proportion of zeros in each condition or unobserved heterogeneity

in the event stage. The negative binomial ZIP model was not used to analyze this data.

Xie, He, and Goh (2001) analyzed the number of computer hard disk read-write errors.

Approximately 87% of the 208 cases were zeros. Given that the Poisson mean was 8.64, the

authors noted that the ZIP model is to be preferred over the Poisson model. However, this mean

is due to several values between 1 and 5, a few between 6 and 15, and 2 values of 75. These

latter two values appear to be so far from the others that they should have been treated as outliers

and addressed in some other manner.

Jang (2005) analyzed the number of non-home based trips per day from 4,416 households

in Jeonju City, Korea. The provided bar graph suggested that approximately 45% of the cases

were zeros. The Vuong statistic (Vuong, 1989) was used for model selection given that the

Poisson is not nested within the ZIP or negative binomial ZIP models.

The purpose of the article by Delucchi and Bostrom (2004) was to provide a brief

introduction to many possible methods for handling zero-inflation including standard t-tests,

bootstrapping,22 and nonparameteric methods. In doing so, they provided results from a study

involving 179 patients with opioid dependence assigned to either a methadone-maintenance or

methadone-assi sted-detoxification treatment. Five out of seven ways to segment the sample


22 See Jung, Jhun, and Lee (2005) for bootstrap procedures and simulation results for Type I and Type II errors.









resulted in zero-inflation ranging from 17% to approximately 66% zeros. The only two-part

model to be used was the ZIP model. The table of results revealed that, in terms of p-values, the

ZIP model performs either very similarly or very differently from the Pearson X2 test for the

proportion of zero values, the Mann-Whitney-Wilcoxon test of nonzero values, and/or the Mann-

Whitney-Wilcoxon test of difference in mean scores between treatment groups. It is possible that

these tests become more similar as the proportion of zeros declines but such conclusions are

based purely on the table of p-values.

Desouhant, Debouzie, and Menu (1998) researched the frequency of immature weevils in

chestnuts. One tree was measured over 16 years, another was measured over 11 years, and three

trees were measured on 1 year. The means ranged from .06 to .63. "None of the 30 distributions

fits a Poisson, X2 ValUeS being always very significant .. The ZIP distribution fits 25 out of 3 1

cases .. The NB distribution fits 20 out of the 31" (Desouhant, Debouzie, & Menu, 1998,

p.3 84). This led to the conclusion that researchers should consider both true zeros and

overdispersion (i.e., trees noted as 'contagious' and trees varying in random oviposition

behavior).

Shankar, Milton, and Mannering (1997) analyzed a 2-year summary of traffic accident

frequencies. For principal arterials, they chose a negative binomial model with data ranging from

0 to 84 (M~= 0.294, SD = 1.09). For minor arterials, they chose the negative binomial ZIP model

for data ranging from 0 to 7 (M~= 0.09, SD = 0.346). For collector arterials, they chose the ZIP

model for data ranging from 0 to 6 (M~= 0.61, SD = 0.279). Model selection was based on the

Vuong statistic. For example, they state, "As suspected previously, inherent overdispersion in the

data is due to the parent NB process and this was validated when the [negative binomial ZIP]










specification failed to provide a statistically better fit (the Vuong statistic <1.96, which

corresponds to the 95% confidence limit of the t-test" (p.833).

Slymen, Ayala, Arredondo, and Elder (2006) analyzed percent calories from fat and

number of days of vigorous physical activity from 357 females participating in a baseline

condition and one of three treatment conditions. Covariates included employment status,

education, martial status, cigarette smoking, and self-reported health. The zero-inflation was 294

out of 357 (82.4%). They compared models using likelihood ratio tests between the Poisson and

negative binomial Poisson and likewise between the ZIP and negative binomial ZIP. The AIC's

were inspected to compare the Poisson and ZIP models.

Not surprisingly, the negative binomial model fit better than the Poisson model. However,

the ZIP model did not fit better or worse than the negative binomial ZIP, and the parameter

estimates and standard errors were nearly identical. This suggests almost no overdispersion in the

data. Indeed, the nonzero percentages were as follows: 1 = 2.8%, 2 = 3.4%, 3 = 4.8%, 5 = 2.0%,

6 = 0.0%, and 7 = 2.0%. This suggests strong equidispersion leaning toward a uniform nonzero

distribution. The AIC's for both models were also nearly equal although both being considerably

smaller than the AIC for the Poisson model and somewhat smaller than the AIC for the negative

binomial Poisson model. Based on a 'smaller-is-better' heuristic, the authors favored the ZIP

model with an AIC of 562.5 over the zero-inflated ZIP model with an AIC of 565.

ZIP and Hurdle Model-Comparisons

The purpose of this section is to present a breadth of literature in which both the Hurdle

and ZIP models were either compared statistically and/or used to analyze real data. This also

includes extensions such as the negative binomial and tau formulations (e.g., ZIP(r)). Some

authors presented alternatives that seem to depart from the ZIP and Hurdle models too drastically

to be within the scope of this dissertation. For example, Lachenbruch (2001) used a two-part










model; however, the splitting formulation was not consistent with the literature. Further, the

model was compared to atypical formulations such as the Wilcoxon, Kolmogorov-Smirnov, and

z tests. As such, these types of articles are not included in the subsequent review. One exception

is Xie, He, and Goh (2001) who included a likelihood-ratio test for comparing the Poisson and

ZIP models.

Statistical

Greene (1994) proposed several 'zero-altered count models' for comparison. First, he took

Mullahy' s with-zero' s (WZ) adaptation of the Hurdle model and included a scalar estimate for

ease on computational burdens. Greene also presented an adaptation of Lambert' s ZIP known as

ZIP(r) and modified it for the negative binomial formulations terming them ZINB and

ZINB(r ). The intention was to identify a "procedure which will enable us to test the zero

inflated model against the simple Poisson model or against the negative binomial model. The

latter will allow us to make a statement as to whether the excess zeros are the consequence of the

splitting mechanism or are a symptom of unobserved heterogeneity" (Greene, 1994, p.10).

Greene developed a method for comparing the models; however, he noted that there was no a

priori reason to think that the Vuong statistic would be inferior.

Applications

Zorn (1996) examined the counts of actions taken by Congress addressing Supreme Court

decisions between 1953 and 1987. The zeros were seen to arise from two sources since many

cases will not be addressed unless there are lobbyists to pressure redress. Covariates included the

year of the decision, the political orientation of the decision, the presence of lower court

disagreement, the presence of precedence alteration, declaration of unconstitutionality, and

unanimous vote. The number of actions ranged from 0 to 11 (M~= 0. 11, SD = .64); however,

3,882 (95.8%) of the 4,052 counts were zeros. This contributed to an exceptionally high skew of









7.97. When ignoring the zeros, the skew was reduced to 1.86 (M~= 2.59, SD = 1.53). The

observed zeros were 107% of that which would be Poisson-expected.

Regardless, Poisson model results were in line with theory-driven expectations. However,

the test of overdispersion was significant when comparing the Poisson and negative binomial

Poisson resulting in fewer significant predictors than if ignoring overdispersion. The author also

fitted a generalized negative binomial model in which "the variance parameter is allowed to vary

as an exponential function of the same independent variables included in the model of the count"

(Zorn, 1996, p.9), which led to even better model Sit. However, due to zero-inflation, no model

provided reasonable estimate sizes given the low mean count.

Their analyses using the ZIP and Hurdle models yielded several findings. First, the

probability of remaining a zero in the transition stage was considerably lower for the Hurdle

model than for the ZIP model at lower levels of a predictor. This is a reflection of the asymmetry

of the Hurdle model. Second, parameter estimates and standard errors were similar between the

two models. They concluded that "at least in some circumstances the performance of ZIP and

hurdle Poisson models will be quite similar. This suggests that, as a practical matter and barring

any strong theoretical considerations favoring one over the other, the choice between them may

be made largely on the basis of convenience of estimation" (Zorn, 1996, p. 11).

Pardoe and Durham (2003) compared the Poisson, Hurdle, and ZIP models as well as

their negative binomial formulations using wine sales data. Of the 1,425 counts, 1,000 (70.2%)

were zeros. The authors noted that this is greater than the 67.8% that a Poisson distribution is

capable of predicting. Based on the AIC, the zero-inflated negative binomial ZIP performed best.

Surprisingly, the Hurdle model fit more poorly than did the Poisson model. It is possible, given

the unrealistically high AIC relative to the other models, that the value wasn't calculated









correctly. Alternatively, the distribution may not have been correctly specified since their

analysis included Bayesian estimates of the prior distributions. No discussion pertaining to the

Hurdle model was included. However, they did provide a novel procedure for comparing the

range of fit statistics across the zero-inflation models. This 'parallel coordinate plot for goodness

of fit measures' consists of an x-axis labeled Min on the left and Max on the right. The y-axis is

a series of horizontal lines each pertaining to a fit statistic (e.g., AIC, BIC). The ceiling x-axis

contains the labels for the models being compared. Then, points for each model are plotted on

the lines for the fit statistics at their relative location between Min and Max. A similar procedure

restricted to the AIC was used by Warton (2005). This technique was adapted to display the

coverage for simulated fit statistics.

Welsh, Cunningham, Donnelly, and Lindenmayer (1996) used zero-inflated rare species

count data to compare the Poisson, negative binomial Poisson, Hurdle, negative binomial Hurdle,

and ZIP models. Approximately 66% of the observations were zeros. They found little difference

between the Hurdle, negative binomial Hurdle, and ZIP model results. Since there was no

overdispersion, the authors recommended using the Poisson model. This is in line with Warton's

(2005) assertion that the more complex zero-inflation models may not always be necessary; at

least, this appears to be the case with 66% zero-inflation and equidispersion.

Discrepant Findings

What is exactly meant by zero-inflation? Min and Agresti (2005) define zero-inflated

count data as "data for which a generalized linear model has lack of fit due to disproportionately

many zero" (p. 1). This raises the question, "At what point does the frequency of zeros become

disproportionate to the frequency of non-zeros?" One statistical definition states that the

proportion of zeros is greater than that to be expected given the posited distribution (Zorn, 1996).

For example, for count data, the proportion of zeros should not be greater than that expected by a









Poisson distribution. However, there are three problems with this. First, there may be many

different proportions of zeros greater than that expected by a particular Poisson distribution.

Second, the definition assumes that the full model is distributed Poisson. Third, it ignores the two

potential sources of overdispersion for Poisson zero-inflated data.

The aforementioned AEWR example displayed a zero proportion of .7158 with a mean of

.54, a standard deviation of 1, and a skew of 1.971. Ignoring the zeros, although this event stage

distribution remains negative skewed, the mean increased to 1.91, and the level of skew dropped

to 0.96. The distribution for the categorical sex variable was binomial with approximately 43%

males and 47% females. The distribution for the age variable was roughly normal with a mean of

48.86, a standard deviation of 17.41, and a skew 0.87. Hence, with 71% zeros, a heavily skewed

distribution of 1.971, a moderately skewed nonzero distribution of 0.96, a normally distributed

continuous predictor, and a two-level categorical predictor led to the following findings:

1) the Hurdle model fit better than Poisson model; 2) the negative binomial Hurdle fit better than

negative binomial Poisson model; 3) the negative binomial Hurdle fit better than the Hurdle

model; 4) the negative binomial ZIP fit better than ZIP model; 5) the negative binomial ZIP

model descriptively fit better than all others; and, 6) the Hurdle and negative binomial Hurdle

model yielded nearly identical estimates and p-values. Hence, findings between the zero-

inflation models differed in terms of both fit and the significance of parameter estimates.

Although not all research presented sufficient information (e.g., data necessary to calculate

skew), there is clearly enough variation in results to warrant further research. Mullahy's (1986)

Hurdle model analyses were impacted by zero-inflation of .263 and not by zero-inflation of .616

or .404; however, this is in disagreement with findings that the Hurdle model adequately handles

zero-deflation (Min, 2003). Lambert' s ZIP analysis with 71.8% zeros favored the ZIP over the










negative binomial ZIP. Greene's (1994) ZIP analyses resulted in nonsensical results under .894

zeros and heavy skew (4.02); the negative binomial ZIP corrected this. Slymen, Ayala,

Arredondo, and Elder' s (2006) ZIP and negative binomial ZIP results were virtually identical;

however, their event stage distribution was uniform. This was confirmed by Warton's (2005)

finding that the negative binomial fits better than the ZIP only when zero-inflation and

overdispersion both are indeed present. Extending from this is Boihning, Dietz, Schlattmann,

Mendonga, and Kirchner' s (1999) findings that the ZIP actually fit better than the Poisson given

.216 and .289 zero-deflation. This is again in contrast to the suggestion that the Hurdle, and not

the ZIP, is appropriate for zero-deflated data. However, it could be argued that a normal

distribution should have been assumed given that the event stage distribution was relatively

normal .

When comparing the Hurdle model to the ZIP model, Zorn (1996) found similar results

given .958 zero-inflation, skew of 7.97, and a reduction of skew to 1.86 for the event stage.

These findings are in contrast to Zorn' s (1996) findings of greater zero-inflation and,

subsequently, greater skew. Welsh, Cunningham, Donnelly, and Lindenmayer (1997) also found

little difference between the Hurdle and ZIP models. Table 2-2 summarizes the findings from the

zero-inflation literature.

There are three factors that may have caused the anomalous findings. The first possibility

pertains to the effect of different types, quantities, and values for predictors. The second

possibility is the proportion of zeros for the outcome variable. The third possibility is the degree

of skew in the event stage for the outcome variable.

It has already been suggested that the Hurdle and ZIP models should be chosen given a

priori research about the source and nature of the zeros. Further, it has been established that the










negative binomial formulations are meant to handle additional overdispersion in the event stage.

However, the previous findings suggest that there are additional considerations such as the

proportion of zeros and the nature of the event stage distribution. The proportion of zeros in this

research ranged from as low as .20 (Delucchi & Bostrom, 1994) to .958 (Zorn, 1996).

Distributions for the event stage included those that were heavily positively skewed (Greene,

1994), moderately positively skewed (AEWR example), distributed normally (Boihning, Dietz,

Schlattmann, Mendonga, & Kirchner, 1999), and distributed uniformly (Slymen, Ayala,

Arredondo, & Elder, 2006).

The first possibility, pertaining to the covariates, can only be tested by varying an

incredibly large set of conditions ranging from small to large quantities of predictors as well as

their types (e.g., nominal, ordinal), and distributions. However, given a particular set of

covariates and corresponding values, the other two possibilities pertaining to zero-inflation and

skew can be explored. It is possible to vary the proportion of zeros for the outcome variable and

to simultaneously vary the degree of skew of the nonzeros for this outcome variable. Hence, the

following research questions are presented:

* Given one two-level categorical covariate with known values and one continuous covariate
with known values, what is the difference in the estimated log-likelihood between a) the
Negative binomial Poisson model vs. Poisson model; b) the Hurdle model vs. Poisson
model?; c) the Negative binomial Hurdle model vs. negative binomial Poisson model?; d) the
Negative binomial Hurdle model vs. Hurdle model; and, e) the Negative binomial ZIP model
vs. ZIP model?

* Given one two-level categorical covariate with known values and one continuous covariate
with known values, what is the difference in the estimated AIC between all models?

These questions were answered by establishing several levels of skew and zero-inflation.

The covariates and their values were Eixed as one continuous variable from a standard normal

distribution and one binary variable. Data for the outcome variable, for all levels of skew and









zero-inflation, were simulated to estimate log-likelihood values, AIC indices, covariate

coefficients, standard errors, and p-values. The details are delineated in the following chapter.

The obj ective is consistent with Zorn' s (1996) advice: "First and foremost, work should be

undertaken to better ascertain the statistical properties of the various estimators outlined here. It

is important that we determine the robustness of these techniques to skewness in the dependent

variable, model misspecification, and the host of other problems that all too frequently plague

political science researchers .. perhaps using Monte Carlo methods to assess under what

circumstances the results of the two may diverge" (Zorn, 1996, p. 12).









Table 2-1. Five pairs of nested models valid for statistical comparison
Valid Comparisons (Nested Models)
1 2 3 4 5
Poisson Simple Simple
NB Poisson Complex Simple
Hurdle Simple Complex
NB Hurdle Complex Complex
ZIP Simple
NB ZIP Complex















































NB ZIP over
OLS/NB ZIP
over Poisson
Equal


Zeros
Simulation

Simulation

.26, .62, .41
.718


.894


.824


.87


Superior Model Comments


Models Compared
Hurdle vs. ZIP

Hurdle vs. ZIP

Hurdle
ZIP vs. NB Poisson
vs. Poisson

ZIP vs. NB ZIP vs.
NB Poisson

Poisson vs. NB
Poisson; ZIP vs. NB
ZIP
ZIP vs. Poisson

Poisson vs. ZIP





Hurdle vs. ZIP
Poisson vs. Hurdle
vs. ZIP

ZIP vs. NB ZIP

ZIP vs. NB ZIP

NB ZIP vs. OLS vs.
NB ZIP vs. Poisson

Hurdle vs. ZIP


Table 2-2. Summary of literature on zero-inflation


Researchers)
Min & Agresti
(2004)
Min & Agresti
(2004)
Mullahy (1986)
Lambert (1992)


Greene (1994)


Slymen, Ayala,
Arredondo, and
Elder (2006)
Xie, He, and Goh
(2001)
Boihning, Dietz,
Schlattmann,
Mendonga, and
Kirchner' s
(1999)
Zorn (1996)
Pardoe and
Durham (2003)

Warton (2005)

Warton (2005)

Warton (2005)


Wel sh,
Cunningham,
Donnelly, and
Lindenmayer
(1997)


Hurdle


Zero-deflation

Zero-inflation


Equal


Hurdle .26
ZIP over NB
Poisson over
Poisson
NB ZIP over ZIP;
NB Poisson over
ZIP over Poisson
NB Poisson;
Equal

ZIP


Heavy skew;
Probit link

Uniform event
stage; Overall,
AIC's favor ZIP
Outliers

Zero-deflation;
normal event stage




Heavy skew
Based on AIC's


Overdi sp ersi on
favored NB ZIP
Rare convergence
for NB ZIP
Based on AIC's


No overdispersion


.216, .289 ZIP


Equal
NB ZIP over
Poisson over
Hurdle
ZIP or NB ZIP


.958
.702


Various


Very low ZIP


Various


.66









CHAPTER 3
METHODOLOGY

The previous chapter fulfilled three obj ectives. First, it described the statistical models and

methods for analyzing count data including that which is zero-inflated. Second, it presented

research, both technical and applied, pertaining to three models (Poisson, Hurdle, and ZIP) as

well as their negative binomial formulations. Third, it was concluded that there is considerable

divergence in findings between models and that such differences should be explored by

examining different levels of zero-inflation and skew for the count outcome variable. This led to

the following research questions:

Research Questions

* Given one two-level categorical covariate with known values and one continuous covariate
with known values, what is the difference in the estimated log-likelihood between a) the
Negative binomial Poisson model vs. Poisson model; b) the Hurdle model vs. Poisson
model?; c) the Negative binomial Hurdle model vs. negative binomial Poisson model?; d) the
Negative binomial Hurdle model vs. Hurdle model; and, e) the Negative binomial ZIP model
vs. ZIP model?

* Given one two-level categorical covariate with known values and one continuous covariate
with known values, what is the difference in the estimated AIC between all models?

As recommended by Zorn (1996), these questions were answered using a Monte Carlo

study in which the proportion of zeros and the skew for the distribution of the event stage counts

varied between simulations.

Monte Carlo Study Design

Monte Carlo studies begin in the same manner as other research methods. First, a problem

is identified as a research question. Second, the problem is made concrete in the form of a

hypothesis or set of hypotheses. Third, the hypotheses are tested using rigorous methods. Fourth,

conclusions are drawn from these results. Fifth, the implications and limitations are elucidated

for future researchers and practitioners.









For this study, the problem was zero-inflation for count outcomes. The problem was then

clarified in the form of the research questions stated above. It is at this point that Monte Carlo

studies differ from most other methods. For the Monte Carlo study, no real data is gathered.

Rather, a set of samples are generated based on given parameter specifications resulting in a

sampling distribution that is considered to be equivalent to that which would have been obtained

had this many real participants been available. This is a clear advantage over the true experiment

where statistics are used to infer from one sample to an entire population (Mooney, 1997). The

Monte Carlo study is not limited to a finite number of participants and is subsequently not prone

to violations of asymptotic theory (Paxton, Curran, Bollen, Kirby, & Chen, 2001). Paxton,

Curran, Bollen, Kirby, and Chen (2001, p.287) provide a succinct explanation as follows:

The researcher begins by creating a model with known population parameters (i.e., the
values are set by the researcher). The analyst then draws repeated samples of size Nfrom
that population and, for each sample, estimates the parameters of interest. Next, a sampling
distribution is estimated for each population parameter by collecting the parameter
estimates from all the samples. The properties of that sampling distribution, such as its
mean or variance, come from this estimated sampling distribution.

Similarly, Mooney (1997, p.2) explains,

Monte Carlo simulation offers an alternative to analytical mathematics for understanding a
statistic's sampling distribution and evaluating its behavior in random samples. Monte
Carlo simulation does this empirically using random samples fiom Imown populations of
simulated data to track a statistic' s behavior. The basic concept is straightforward: If a
statistic' s sampling distribution is the density function of the values it could take on in a
given population, then its estimate is the relative frequency distribution of the values of
that statistic that were actually observed in many samples drawn from that population.

The Monte Carlo simulations were performed using the R programming language (R

Development Core Team, 2006). R is an open-source language based on the commercially










available S-Plus program (Insightful, 2005) and is just one of many programs that can be used to

model zero-inflated data.23

Monte Carlo Sampling

Pseudo-Population

Mooney (1997) explains that defining the population parameters requires defining the

pseudo-population. In the pseudo-population, the values for a categorical factor, X1, with two

levels were constant at either 0 or 1; this is representative of a typical two-level categorical

factor such as sex. In the dataset, these values alternated. The Excel 2003 random number

generator was used to draw a random selection of 1,000 normally distributed values to represent

the continuous variable, X2~- N(0,1). This resulted in a pseudo-population of N = 1,000 where

the values for categorical X1 and continuous X2 were known. The values for X2 ranged from -

2.945 to 3.28 with a mean of 0, a median of 0.05, a standard deviation of 0.986, and skew of -

0. 127. These two sets of covariates and their distributions were chosen as a parsimonious

generalization to the basic ANCOVA general linear model that extends from analyses with either

quantitative or qualitative predictors.24 The simulations varied in terms of a) the amount of zero-

inflation present in the outcome variable scores; b) the amount of skew present in the event stage

outcome variable scores, and c) the generalized linear model.

The outcome variable, Y, was established as a deterministic variable in that it varied

systematically as a function of the specified distributions. As clarified by Mooney (1997),

"Deterministic variables are vectors of numbers that take on a range of values in a prespecified,

nonrandom manner" (p.6). The regression coefficients for X1 and X2 are random variables that

23 Others include Stata, LIMDEP, COUNT, MATLAB, and SAS. Preliminary analyses compared results between
SAS-code incorporated in research (Min & Agresti, 2004; Min, 2003) to publicly available R-code written by Simon
Jackman of the Stanford Political Science Computing Laboratory to verify the comparability of results.

24 The model is similar to that of the AEWR examples.










take on their realizations as a result of the relationship between deterministic Y and the two

random X covari ate s.

The Prespecified Zero Proportions

Justification for generating values with prespecified proportions of event counts with

frequencies also determined by prespecified proportions of zeros is justified by Mooney (1997).

"If we know (or are willing to make assumptions about) the components that make up a statistic,

then we can simulate these components, calculate the statistic, and explore the behavior of the

resulting estimates" (Mooney, 1997, p.67). In his case, the concern was bias determined by

calculating statistics and inspecting graphs as a result of Monte Carlo simulations. For this study,

the concern was goodness-of-fit for six models by comparing log-likelihoods and AIC's and

inspecting graphs as a result of Monte Carlo simulations over three levels of skew and five levels

of zero proportions.

Previous research displayed zero-inflation ranging from .20 (Mullahy, 1986) to .96 (Zorn,

1996). To reflect a range including both zero-deflation and zero-inflation, six pseudo-populations

were established differing in the proportion of zeros present in the count outcome variable. The

pseudo-populations contained either 0. 10, 0.25, 0.50, 0.75, or 0.90 proportions of zeros. 25

Pre-Specified Skew

To manipulate skew, the event stage distributions for the count outcome variable were

varied over three conditions. For each condition, proportions were specified and values were

drawn randomly from a multinomial distribution such that the frequencies of the values added to

the frequency for those already drawn to represent zero-inflation summed to 1,000. In other



25 Originally, a 0.00 proportion of zeros was included as a control condition. However, for the case of negative
skew, this is simply a count model truncated at one. And, for a normal distribution event-stage, this is semi-
continuous data often tested with different methods from those for zero-inflation (Min, 2002).










words, if .50 (or 500) observations were assigned a value of zero then the remaining .50 (or 500)

observations had values drawn from the prespecified multinomial distribution.

Event stage values ranging from one to five were sampled in order to represent a range

small enough to distinguish the distribution from one that might be analyzed as continuous given

a particular shape. The prespecified probabilities for each of the five counts were determined

primarily to achieve a particular level of skew and secondarily in order to achieve a convergent

and admissible solution. Hence, the proportions were not always exactly equal to .10, .25, .50,

.75, and .90; the approximates were selected to achieve convergence leading to more trust in

convergence for the Monte Carlo simulations. Table 3-1 displays the proportions of each count

as a function of the three levels of skew and five levels of zeros. Table 3-2 displays this same

information in terms of frequencies instead of proportions. Table 3-3 displays the descriptive

statistics for each distribution.

Random Number Generation

By definition, a random number is one in which there is no way possible to a priori

determine its value. Most statistical analysis software packages include random number

generators. However, these generated random numbers are not truly random. Usually, one

specifies a seed; when replications are performed using the same seed, the generated numbers are

identical to the first. Hence, the values are pseudo-random (Bonate, 2001). However, this

limitation is actually an advantage in that the researcher can check for errors in model

programming and run the analysis again with the same generated sample (Mooney, 1997).26





26 Technically, this is only true for most random number generators. The R programming language, by default, bases
its generation on the Wichman-Hill algorithm and the system clock resulting in a 626-integer seed (R Development
Core Team, 2006).









Another feature of Monte Carlo random sampling pertains to the desired distributions.

Typically, the random numbers are drawn from a uniform distribution, which is then followed by

a transformation to the desired distribution. As Mooney (1997) explains, "In its standard form,

U(0, 1), the uniform distribution is the building block of all Monte Carlo simulation work in that

from it, in one way or another, variables with all other distribution functions are derived. This is

because the U(0, 1) distribution with its 0 < x < 1 range, can be used to simulate a set of random

probabilities, which are used to generate other distribution functions through the inverse

transformation and acceptance-rejection methods" (p.10).

The random number generation was performed using R 2.3.1 (R Development Core Team,

2006). The procedure requires the generic sample command in which the following were

specified: 1) a range of counts, which in this study, was from one to five (not zero to five since

proportions of zeros were already drawn from the pseudo-population), 2) the number of values to

draw, which in this study was one minus the prespecified proportion of zeros, 3) the proportions

for each value in the range, which in this case was one of three possibilities determining skew,

and 4) the specification to sample with replacement. The seed was arbitrarily set at 6770.

Sample Size

Determining the appropriate sample size for each simulate is an important concern. This

could range from zero to infinity. However, if the sample size is too small then it is not safe to

assume that estimates are asymptotically normal. On the other hand, computer time and burden

increases as sample size increases. The sample size was based on the highest found in the

literature pertaining to zero-inflated count data, which was n = 1,000 (Civentti & Hines, 2005).

Simulation Size

Determining the number of simulations is also an important concern since too few

replications may result in inaccurate estimates and too many replications may unnecessarily









overburden computer time and performance (Bonate, 2001). Hur' s (1999) research pertaining to

the ZIP model with random effects set the number of simulations at 200. Min and Agresti (2005)

were able to sufficiently compare the goodness of fit for several competing models using 1,000

simulations. Likewise, Civettini and Hines (2005) selected 1,000 simulations when researching

misspecification in negative binomial ZIP models. Lambert (1989) set the number of simulations

at 2,000 when researching the asymptotic properties of the ZIP model.

Mooney (1997) states that "The best practical advice on how many trials are needed for a

given experiment is "lots"!i Most simulations published recently report upward from 1,000 trials,

and simulations of 10,000 and 25,000 trials are common" (p.58). Given the previously noted

problems with convergence for the negative binomial ZIP model, it seems prudent to minimize

the number of simulations as much as possible. However, it is also important to simulate under

conditions already found to produce asymptotic results. Hence, similar to Lambert' s (1989)

seminal study and equal to the maximum found in the literature, the number of simulations was

set at 2,000 for each condition (S = 2,000).

Iteration Size

Iteration size is not much of a concern for the basic Poisson and negative binomial

Poisson model. However, obtaining valid estimates for the Hurdle model, ZIP model, and their

negative binomial counterparts requires selecting an appropriate maximum number of iterations.

Too few iterations can lead to incorrect estimates or, even worse, premature declaration of

nonconvergence. Too many iterations results in unnecessary computer time and burden. Various

procedures for analyzing these models in R have maximum iterations of 500, 5,000, and 50,000.

It was important to determine an iteration size that would be equal across analyses and large

enough given some models' potential for nonconvergence. These concerns were deemed more










important than excessive computational burden. Hence, the procedure with the largest iteration

size was selected for all procedures leading to 50,000 iterations per analysis.

Distribution Generation

The following describes the procedure for generating the distributions for each

simulation.

* The total proportion of counts out of 1,000 to be sampled was reduced by the prespecified
proportion of zeros. Hence, if the proportion of zeros was .50 then the proportion of event
stage counts was 1.00-0.50 = .50. Translated into frequencies, this is 1,000-(0.50 1,000) =
500

* The generic R 'sample' procedure was used to sample with replacement from the event stage
counts according the specified proportions depending on the skew condition. The seed was
set at 6770.

* The values were sampled over N = 1,000.

* Each sample was simulated S= 2,000 times.

* The data over all S = 2,000 at N= 1,000 were then stored in separate files as they were
created. The filenames conformed to the labeling format where the model was replaced by an
underscore (e.g., _25Pos was the filename for the S = 1,000 datasets at N=2,000 where the
proportion of zeros was .25 and the skew for the event stage was positive).

Monte Carlo Models

As previously discussed, generalized linear models include a random component, a

systematic component, and a link function. The X1 and X2 COnstants form the systematic

component in the pseudo-population's generalized linear model. The random component

specification for the distribution of the outcome mean varied from pseudo-population to pseudo-

population. The base level generalized linear model assuming a normally distributed outcome is

given by

Y = Po + fl,(X,l )+ A(X23> + (3-1)









Subsequent models extended from this base model to form the six distributions for

deterministic Y The first model, which was the Poisson generalized linear model with a log

link, is given by

log(il) = o, + (XI)+ 2,(X27) (3 -2)

Table 3-4 displays the parameters for this model over all conditions of skew and zero-

inflation. For example, the analysis with a .10 proportion of zeros and a positively skewed event

stage distribution yielded

log(il,) = .450 +.004(X,) -.037(X2) (3-3)

For both predictors, the coefficient near zero transforms to an exponentiation near one. The

coefficient for X1 is lowest (.001) for the negatively distributed data with a .25 proportion of

zeros and highest (.007) for the positively distributed data with a .50 proportion of zeros. The

coefficient for X2 is lowest (-.007) for the negatively distributed data with a .10 proportion of

zeros and highest (-. 165) for the normally distributed data with a .90 proportion of zeros. Hence,

for the two-level categorical X1 variable, changing from zero to one multiplies the mean of the

outcome variable by approximately exp(0.00), which equals one. For the simulated data, the test

for the coefficient estimates corresponding to these pseudo-population parameter values is

approximately Ho: P = 0.

The second model, which was the negative binomial formulation of the Poisson model, is

the same as the Poisson model with the addition of a dispersion parameter. Table 3-5 displays the

parameters for this model over all conditions of skew and zero-inflation. Like the Poisson model,

the test for the simulated data is Ho: p = 0. The dispersion parameter is also included in Table 3-

5. For the simulated data, the test that this parameter equals zero has equivalent results to the

model comparison tests conducted.









The third and fifth models were the general formulations for the Hurdle and ZIP models

given by


logit (p, )= log( )= P, + 7,(Z,I)+ P(Z,I) (3 -4)
1 p,
log(l,) = P,+ P(X,I)+ P(X,I) (3-5)

while the fourth and six models were their negative binomial formulations. Tables 3-6 through 3-

13 display the pseudo-population coefficients, standard errors, log-likelihood values, and AIC

values over all conditions of proportions of zero and skew. Although there are four models (i.e.,

Hurdle, negative binomial Hurdle, ZIP, and negative binomial ZIP), there are eight tables since,

for each model, there are separate results for the transition (zeros) stage and events (nonzero

counts) stage.

Monte Carlo Analysis Procedures

A generic looping procedure was written in the R programming language to retrieve each

of the 15 sets of simulated data. Each dataset was analyzed with each of the six models. The

Poisson model used the generic glm procedure in R, which requires specification of the model,

Poisson distribution, and log link. The negative binomial Poisson model used the glm.nb

procedure in R from the generic M4rSS library. This procedure requires only the specification of

the model. The R Core Development Team (2006) describes the procedure as follows: "An

alternating iteration process is used. For given 'theta' [dispersion] the GLM is fitted using the

same process as used by 'glm()'. For fixed means the 'theta' parameter is estimated using score

and information iterations. The two are alternated until convergence of both is achieved" (R

Core Development Team, 2006).

The Hurdle model used the hurdle procedure in the pscl library authored by Simon

Jackman, PhD of the Political Science Computing Laboratory at Stanford University. The










procedure requires specification of the count response variable (Y), the Poisson distribution for

the event stage, the logit link function for the transition stage, and the models for both the

transition stage and event stage, which for this study were X1 and X2. The negative binomial

Hurdle model also used the hurdle procedure but specified a negative binomial rather than a

Poisson distribution for the event stage. This hurdle model procedure maximizes the log-

likelihood using either the Broyden-Fletcher-Goldfarb-Shanno (BFGS) or Nelder-Mead

methods. The Nelder-Mead (default) was selected for solution optimization.

The ZIP model used the zicounts procedure in the zicounts library authored by Samuel

Mwalili, doctoral student in biostatistics at Katholieke Universeteit Leuven (Netherlands).

Similar to the hurdle procedure, zicounts requires specification of the count outcome variable and

models for both the transition stage and event stage. The distribution is specified as "ZIP".

Optimization procedures include the BFGS, Nelder-Mead, and conj oint gradient (CG) methods;

for consistency, the Nelder-Mead was chosen. The negative binomial ZIP used the same

procedure with the specification of the "ZINB" distribution. The specific procedure for analyzing

the data with the models was as follows:

* Three separate loops were established for the negatively skewed, normal, and positively
skewed distributions.

* Arrays were created to store the log-likelihood, AIC, coefficient estimates, standard errors,
and p-values.

* The outermost loop for each distribution pertained to the five conditions of varying zero
proportions; at each loop, the data corresponding to this zero-condition and distribution was
loaded.

* Within these loops, another looping procedure pertained to the six models that analyzed the
data.

* Within these loops another looping procedure was defined for the number of simulations in
the data set.










* It was at this point that the data corresponding to a particular distribution, proportion of
zeros, model, and simulation were analyzed with the calculated AIC, log-likelihood,
coefficient estimates, standard errors, and p-values transferred to the appropriate array.
Models that failed to converge were automatically coded as NA. 27

* The results for each statistic over all simulations for a particular distribution, model, and
proportion of zeros were then exported in comma-separated format for subsequent analyses.
Hence, the three distributions by fiye proportion of zeros conditions by six models yielded 90
data Hiles each containing columns pertaining to a particular statistic or set of statistics and
2,000 rows of simulated results.
Analysis Design

A series of tests were conducted using the simulated data. The results and graphical

output were created using R 2.01 (R Development Core Team, 2006) and SPSS 14.0. The design

for a particular analysis depended on the research question. These questions were as follows:

* Given one two-level categorical covariate with known values and one continuous covariate
with known values, what is the difference in the estimated log-likelihood between a) the
Negative binomial Poisson model vs. Poisson model; b) the Hurdle model vs. Poisson
model?; c) the Negative binomial Hurdle model vs. negative binomial Poisson model?; d) the
Negative binomial Hurdle model vs. Hurdle model; and, e) the Negative binomial ZIP model
vs. ZIP model?

* Given one two-level categorical covariate with known values and one continuous covariate
with known values, what is the difference in the estimated AIC between all models?

The first question was answered by calculating log-likelihood values for the six models

over the six zero proportion conditions and the three skew conditions. The deviance statistic was

then calculated as -2(LLs-LLc) where LLs is the model with less parameters (i.e., the simple

model) than the LLc (i.e., the more complex model). Since there were 5 conditions for the

proportions of zeros and 3 conditions for skew, this led to a) a total of 15 separate analyses

comparing the fit of the negative binomial Poisson model and the Poisson model; b) 15 separate

analyses comparing the fit of the Poisson model and the Hurdle model; c) 15 separate analyses

comparing the fit of the negative binomial Poisson model and the negative binomial Hurdle

27 This number was chosen to reduce the probability of obtaining a valid result that would be identified as a
nonconvergant solution










model; and d) 15 separate analyses comparing the fit of ZIP model and the negative binomial

ZIP model.

Test statistics for the deviance were assumed asymptotically chi-square with degrees of

freedom equal to the difference in the number of parameters between the two models. Some

models differed only by the dispersion parameter; these were a) the comparison of the Poisson

model and negative binomial Poisson model; b) the comparison of the Hurdle model and

negative binomial Hurdle model; and c) the comparison of the ZIP model and negative binomial

ZIP model. The log-likelihood statistic for the Hurdle model is based on the log-likelihood

statistics from each of its two parts. Given three parameters in the Poisson model (i.e., the

intercept and the two predictor coefficients), there are six parameters in the Hurdle model.

Hence, the degrees of freedom for the model comparison test are the difference, which are three.

The same is true for comparing the negative binomial Poisson model to the negative binomial

Hurdle model where including the dispersion parameter leads to subtracting 4 parameters from 7

parameters .

Each of the 2,000 goodness-of-fit statistics for the simpler model was subtracted from each

of the 2,000 goodness-of-fit statistics for the more complex model. These results were then

multiplied by -2. Each of these values was then compared to the chi-square distribution with the

appropriate degrees of freedom. This yielded a p-value representing the probability of obtaining

a statistic this high or higher given that the simpler model adequately fits the data. Values

exceeding this critical chi-square statistic based on a Type I error rate of a = .05 were coded "1"

with these results suggesting that the more complex model fits the data better than the Poisson

model. Values failing to exceed the critical value were coded "O" with these results suggesting

adequate fit for the simpler model.









The results were thus the proportion of simulated datasets favoring the more complex

model over the simpler model for a particular proportion of zeros and level of skew. Output

included 1.) descriptive statistics for the goodness-of-fit for each model and 2.) boxplots for the

difference in goodness-of-fit between the two models.

Answering the second question was done in a similar fashion to answering the first

question. This is due to the fact that the AIC is a linear transformation of the log-likelihood

statistic with a result that is positive in sign and is interpreted in a lower-is-better fashion.

However, these analyses did not involve comparisons to the chi-square distribution. As

previously explained, the AIC should not be used in this manner. However, the advantage is that

the AIC can be used to descriptively compare all models regardless of whether one is nested or

not within another. Boxplots were created to display the range of AIC's produced over

simulations for each analysis.










Table 3-1. Proportions of counts as a function of zeros and skew
Study Proportion Proportion of Remaining Nonzero Values
ofZeros
Ones Twos Threes Fours Fives
Posl0 0.10 0.504 0.227 0.088 0.054 0.027
Norml10 0.099 0.198 0.306 0.198 0.099
Negl10 0.027 0.054 0.091 0.227 0.501
Pos25 0.25 0.418 0.190 0.075 0.046 0.021
Norm25 0.083 0.166 0.254 0.166 0.081
Neg25 0.022 0.045 0.075 0.188 0.420
Pos50 .50 0.280 0.125 0.050 0.030 0.015
Norm50 0.053 0.107 0.175 0.110 0.055
Neg50 0.015 0.030 0.050 0.125 0.280
Pos75 0.75 0.140 0.062 0.025 0.015 0.008
Norm75 0.028 0.052 0.089 0.053 0.028
Neg75 0.008 0.015 0.025 0.062 0.140
Pos90 0.90 0.056 0.025 0.010 0.006 0.003
Norm90 0.011 0.021 0.035 0.024 0.009
Neg90 0.004 0.005 0.010 0.025 0.056

Table 3-2. Frequencies of counts as a function of zeros and skew
Study Frequency Frequency of Individual Nonzero Values
of Zeros
Ones Twos Threes Fours Fives
Posl0 100 504 227 88 54 27
Norml0 99 198 306 198 99
Negl0 27 54 91 227 501
Pos25 250 418 190 75 46 21
Norm25 83 166 254 166 81
Neg25 22 45 75 188 420
Pos50 500 280 125 50 30 15
Norm50 53 107 175 110 55
Neg50 15 30 50 125 280
Pos75 750 140 62 25 15 8
Norm75 28 52 89 53 28
Neg75 8 15 25 62 140
Pos90 900 56 25 10 6 3
Norm90 11 21 35 24 9
Neg90 4 5 10 25 56











Table 3-3. Descriptive statistics for each distribution
Range = 0 to 5


Skew


Mean


Range = 1 to 5
Mean Std.Dev.


Std.Dev.
1.127
1.414
1.619
1.181
1.635
2.054
1.149
1.710
2.253
0.928
1.422
1.914
0.622
0.965
1.318


Skew
1.45
0.08
-1.43
1.42
0.08
-1.86
1.44
-0.01
-1.44
1.44
-0.00
-1.44
1.46
-0.06
-1.51


Posl0
Norm10
Negl0
Pos25
Norm25
Neg25
Pos50
Norm50
Neg50
Pos75
Norm75
Neg75
Pos90
Norm90
Neg90


1.570
2.700
3.820
1.310
2.250
3.190
0.880
1.510
2.130
0.440
0.750
1.060
0.180
0.300
0.420


1.155
-0.310
-1.356
1.092
-0.069
-0.673
1.514
0.587
0.236
2.577
1.660
1.349
4.490
3.268
2.908


1.750
3.000
4.250
1.750
2.990
4.840
1.750
3.010
4.250
1.760
3.000
4.240
1.750
2.990
4.240


1.051
1.150
1.053
1.046
1.149
0.367
1.053
1.140
1.053
1.064
1.149
1.064
1.058
1.124
1.084


Table 3-4. Poisson model:


pseudo-population parameters


Pos10
Nonn10
Neg10
Pos25
Nonn25
Neg25
Pos50
Nonn50
Neg50
Pos75
Nonn75
Neg75
Pos90
Nonn90
Neg90


.450
.992
1.340
0.270
.807
1.159
-138
.407
.752
-829
-.291


-1475.3
-1806.7
-2019.2
-1468.1
-1913.9
-2265.35
-1334.3
-1888.5
-2370.5


2956.6
3619.3
4044.5
2942.3
3833.8
4534.7
2674.5
3783.0
4746.9
1951.7
2929.6
3845.3
1099.0
1711.1
2310.2


.068 .005 .095 -.077
.052 .003 .073 -.079


.048 -972.9
.037 -1461.8
.031 -1919.7


.057 .043 .002 .061 -.044


-546.5
-852.6
-1152.1














Table 3-5. Negative Binomial Poisson model: pseudo-population parameters

Po sw P s, p, .s LL


Pos10

Norml0

Neg10

Pos25

Norm25

Neg25

Pos50

Norm50

Neg50

Pos75

Norm75

Neg75

Pos90

Norm90

Neg90


-1594.5

-1998.9

-2277.9

-1509.4

-1931.5

-2235.5

-1290.1

-1723.0

-2041.3

-910.4

-1275.6

-1567.4

-511.6

-751.1

-957.3


(zeros): pseudo-population parameters

sP, A1 saq A


AIC


Pos10

Norml0

Neg10

Pos25

Norm25

Neg25

Pos50

Norm50

Neg50

Pos75

Norm75

Neg75

Pos90

Norm90

Neg90


-2.198

-2.198

-2.198

-1.099

-1.098

-1.099

.000

-.005

.000

1.099

1.100

1.099

2.205

2.206

2.204


-1409.5

-1797.1

-1928.3

-1465.5

-1787.6

-1897.9

-1296.6

-1510.1

-1583.9

-865.2

-970.3

-1008.1

-444.9

-486.4

-503.0


2830.9

3606.1

3868.7

2943.0

3587.2

3807.9

2605.1

3023.2

3179.7

1742.4

1952.6

2028.3

901.8

984.8

1018.1


Table 3-6. Hurdle model

A























































Pos10

Norml0

~Neg10

Pos25

Norm25

~Neg25

Pos50

Norm50

~Neg50

Pos75

Norm75

~Neg75

Pos90

Norm90

Neg90


-2.197

-2.197

-2.205

-1.099

-1.099

-1.099

.000

.000

.001

1.099

1.099

1.099

2.206

2.244

2.205


-1395.7

-1800.2

-1936.1

-1455.3

-1790.2

-1904.4

-1288.9

-1511.9

-1588.1

-861.1

-971.2

-1010.3

-443.3

-487.3

-503.9


Table 3-7. Hurdle model (events): pseudo-population parameters


SP, 1


Sa P2


AIC


Pos10

Norml0

Neg10

Pos25

Norm25

Neg25

Pos50

Norm50

Neg50

Pos75

Norm75

Neg75

Pos90

Norm90

Neg90


.212

1.035

1.430

.214

1.033

1.431

.208

1.035

1.430

.213

1.032

1.428

.179

1.106

1.420


-1409.5

-1797.1

-1928.3

-1465.5

-1787.6

-1897.9

-1296.6

-1510.1

-1583.9

-865.2

-970.3

-1008.2

-444.9

-486.4

-503.0


2830.9

3606.1

3868.7

2943

3587.2

3807.9

2605.1

3023.2

3179.7

1742.4

1952.6

2028.3

901.8

984.8

1018.1


Table 3-8. Negative Binomial Hurdle

0o Sp, 1


model (zeros): pseudo-population

S4 P2 S ,


parameters
LT,


AIC


2803.4

3612.3

3884.1

2922.5

3592.3

3820.8

2589.8

3035.8

3188.2

1734.3

1954.4

2032.5

898.7

986.6

1019.8














Table 3-9. Negative Binomial Hurdle model (events): pseudo-population parameters

P S~G S P .S Theta LL AIC


Pos10

Nonn10

Neg10

Pos25

Nonn25

Neg25

Pos50

Nonn50

Neg50

Pos75

Nonn75

Neg75

Pos90

Nonn90

Neg90


-.067 .108

1.033 .030

1.429 .024

-.042 .114

1.030 .033

1.431 .026

-.072 .146

1.037 .041

1.429 .032

-.075 .208

1.031 .058

1.427 .046

-.109 .332

1.051 .091

1.421 .072


.009 .091

.004 .043

.002 .034

.009 .098

.005 .047

.001 .037

.020 .122

.006 .057

.004 .046

.012 .173

.005 .081

.002 .064

.078 .273

-051 .130

.007 .102


-.071 .046

-012 .022

-002 .017

-.067 .051

-027 .025

-015 .019

-.065 .062

-025 .029

-012 .023

-.097 .088

-054 .041

-011 .032

-.081 .140

-048 .066

-028 .052


2.039

165.718

164.970

2.234

166.348

166.006

2.030

166.574

166.526

1.950

166.445

166.377

2.047

91.979

166.606


-1395.7

-1800.2

-1936.1

-1455.3

-1790.2

-1904.4

-1288.9

-1511.9

-1588.1

-861.1

-971.2

-1010.3

-443.3

-487.3

-503.9


2803.4

3612.3

3884.1

2922.5

3592.3

3820.8

2589.8

3035.8

3188.2

1734.3

1954.4

2032.5

898.7

986.6

1019.8


Table 3-10. ZIP model (zeros): pseudo-population parameters

Ao sp, A1 sa4 A


AIC


Pos10

Nonn10

Neg10

Pos25

Nonn25

Neg25

Pos50

Nonn50

Neg50

Pos75

Nonn75

Neg75

Pos90

Nonn90

Neg90


-27.010

-3.114

-2.365

-15.460

-1.379

-1.162

-.890

-.126

-031

.610

1.032

1.079

1.797

2.133

2.188


4434.000

.038

.175

3834.000

.132

.109

.198

.097

.091

.142

.057

.110

.183

.152

.071


11.490

.002

.002

-3.418

.008

.001

.051

.002

.000

.010

.005

-001

.032

.002

.007


4435.000 -.363

.527 .050

.246 .049

.002 1.213

.186 -.056

.154 -.025

.270 -.050

.137 .030

.129 .035

.198 -.013

.081 -.054

.147 .044

.255 .104

.214 .128

.101 -.028


150.700

.253

.125

.001

.098

.078

.141

.070

.065

.010

.040

.075

.129

.108

.051


-1478.3

-1800.1

-1931.3

-1471.1

-1790.6

-1900.9

-1299.5

-1513.1

-1586.9

-868.2

-973.3

-1011.2

-447.9

-489.4

-506.0


2968.6

3612.1

3874.6

2954.3

3593.1

3813.9

2611.1

3038.3

3185.7

1748.5

1958.5

2034.3

907.7

990.8

1024.1
















sP 4


sa P


AIC


Pos10

Nonn10

Neg10

Pos25

Nonn25

Neg25

Pos50

Nonn50

Neg50

Pos75

Nonn75

Neg75

Pos90

Nonn90

Neg90


.450

1.035

1.430

.270

1.032

1.431

.205

1.039

1.429

.213

1.015

1.428

.180

1.017

1.422


-1478.3

-1800.1

-1931.3

-1471.1

-1790.6

-1900.9

-1299.5

-1513.1

-1586.9

-868.2

-973.3

-1011.2

-447.9

-489.4

-506.0


2968.6

3612.1

3874.6

2954.3

3593.1

3813.9

2611.1

3038.3

3185.7

1748.5

1958.5

2034.3

907.7

990.8

1024.1


Table 3-12.


Negative



-15.280

-3.110

-2.370

-17.090

-1.380

-1.162

-4.895

-.131

-031

.130

1.015

1.080

1.462

2.133

2.187


Binomial ZIP model (zeros):

sp P1 s,


pseudo-population parameters

2 Sp LL


AIC


Pos10

Nonn10

Neg10

Pos25

Nonn25

Neg25

Pos50

Nonn50

Neg50

Pos75

Nonn75

Neg75

Pos90

Nonn90

Neg90


142.300

.375

.175

501.400

.132

.109

3.373

.097

.091

.364

.107

.104

.386

.152


-3.880

.0165

-001

2.390

-.002

.000

1.352

.008

.000

.015

.001

-002

.047

.002


830.400

.527

.247

473.600

.186

.037

2.211

.014

.129

.271

.151

.147

.298

.214


.719

.0510

.047

1.700

-.059

-016

-1.086

.031

.035

-.040

.033

.044

.098

.128


102.100

.253

.125

133.500

.983

.019

1.304

.069

.065

.136

.076

.075

.151

.108


-1479.3

-1801.1

-1932.3

-1471.4

-1791.6

-1901.9

-1292.4

-1514.1

-1587.9

-865.1

-974.3

-1012.2

-447.3

-490.4


2972.6

3616.1

3878.7

2956.7

3597.1

3817.9

2598.8

3042.3

3189.7

1744.3

1962.6

2038.3

908.7

994.8


.150 -.001


.212 .136


.107 -507.0 1028.1


Table 3-11. ZIP Model (events): pseudo-population parameters














Table 3-13. Negative Binomial ZIP model (events): pseudo-population parameters

o S, P1 Sr P Sp Theta LL AIC


Pos10

Nonn10

Neg10

Pos25

Nonn25

Neg25

Pos50

Nonn50

Neg50

Pos75

Nonn75

Neg75

Pos90

Nonn90

Neg90


.036 .004

.030 .004

.024 .001

.040 .003

.033 .005

.026 .001

.067 .045

.040 .007

.032 .004

.204 .010

.057 .005

.045 .000

.327 .072

.091 .017

.071 .006


-.037 .026

-013 .021

-003 .017

-.024 .029

-027 .025

-016 .019

-.080 .052

-026 .029

-012 .023

-.099 .087

-054 .040

-010 .032

-.080 .140

-052 .065

-028 .051


14.69

12.64

13.02

3.16

10.88

15.94

.57

11.70

14.19

.69

12.02

12.99

.73.

12.76

14.73


-1479.3

-1801.1

-1932.3

-1471.4

-1791.6

-1901.9

-1292.4

-1514.1

-1587.9

-865.1

-974.3

-1012.2

-447.3

-490.4

-507.0


2972.6

3616.1

3878.7

2956.7

3597.1

3817.9

2598.8

3042.3

3189.7

1744.3

1962.6

2038.3

908.7

994.8

1028.1









CHAPTER 4
RESULTS

This chapter presents the results based on the previously discussed methods and

procedures for analyzing data with varying proportions of zeros and varying event stage

distributions. First, the results are presented using the data for the pseudo-population in which

the proportions for each count level are exactly that which was rando mly sampled from in the

simulations. Tables and figures are included to support interpretation. Second, the results are

presented outlined by the skew level (i.e., positive, normal, and negative) and by the proportion

of zeros within that skew level (i.e., .10, .25, .50, .75, and .90). For each combination of

conditions, the results are presented for the five model comparisons. Third, the results are

summarized separately for the negative, normal, and positive event count distributions. Tables

and figures are included to support interpretation. The primary purpose of the results was to

assist in answering the research questions, which were as follows:

* Given one two-level categorical covariate with known values and one continuous covariate
with known values, what is the difference in the estimated log-likelihood between a) the
Negative binomial Poisson model vs. Poisson model; b) the Hurdle model vs. Poisson
model?; c) the Negative binomial Hurdle model vs. negative binomial Poisson model?; d) the
Negative binomial Hurdle model vs. Hurdle model; and, e) the Negative binomial ZIP model
vs. ZIP model?

* Given one two-level categorical covariate with known values and one continuous covariate
with known values, what is the difference in the estimated AIC between all models?

Pseudo-Population Results

Before addressing the results of the Monte Carlo simulations, it is necessary to discuss

the results when each model was analyzed with the pseudo-population data. The prespecified

proportions for each count were displayed in Table 3-1. Table 3-2 displayed this same

information in terms of frequencies instead of proportions. Table 3-3 displayed the descriptive

statistics for each distribution. Tables 3-4 through 3-13 displayed the coefficients, standard









errors, log-likelihood values, and AIC values with each table pertaining to either a specific model

or one of the two stages for a specific model.

The following sections provide the results when comparing models using the pseudo-

population data. First, the Poisson model results were compared to the negative binomial Poisson

model results, both descriptively via AIC's and inferentially via the deviance model comparison

test. Second, the results are presented in a likewise manner for the three Hurdle model

comparisons. These comparisons were a) the Hurdle model vs. the negative binomial Hurdle

model, b) the Poisson model vs. the Hurdle model, and c) the negative binomial Poisson model

vs. the negative binomial Hurdle model. Third, the results are presented for the comparison of

the ZIP model and the negative binomial ZIP model. Fourth, descriptive results are presented

comparing all models for each of the five proportions of zeros.

Pseudo-Population Poisson Models

Based on the AIC, for all proportions of zeros the data fit the Poisson model better when

the distribution was positively skewed than when normally distributed. In addition, this data fit

the Poisson model better when the distribution was normally distributed than when it was

negatively skewed. For example, for the negatively skewed distribution with a .10 proportion of

zeros, the AIC was 4,044.5. As the curve shifted left to a normal distribution, the AIC dropped to

3,619.3, and as the curve shifted further left to a positively skewed distribution, the AIC dropped

to 2,956.3.

The same pattern emerged for the negative binomial Poisson models. For example, for the

negatively skewed distribution with a .10 proportion of zeros, the AIC was 4,563.8. As the curve

shifted left to a normal distribution, the AIC dropped to 4,005.9, and as the curve shifted further

left to a positively skewed distribution, the AIC dropped to 3,196.9.










For the .10 proportion of zeros condition, the Poisson models had a lower AIC than that

calculated for the negative binomial Poisson model. For the .25 proportion of zeros condition,

the AIC's were approximately equal between the two models. For the .50, .75, and .90

proportions of zeros conditions, the AIC was lower for the negative binomial Poisson model than

for the Poisson model.

The deviance statistic was calculated comparing the Poisson log-likelihood and negative

binomial Poisson log-likelihood for each skew condition and zero proportion condition. These

statistics are displayed in Table 4-1. For all analyses, the Poisson model is nested within its

negative binomial Poisson formulation differing by one degree of freedom (i.e., the dispersion

parameter in the negative binomial Poisson model). Hence, at the .05 Type I error rate and

assuming deviance statistics asymptotically distributed chi-square, a deviance exceeding 3.84

suggests better fit for the more complex negative binomial Poisson model. This was the result for

all analyses in which the proportion of zeros was .50 or greater. However, for all analyses with

.10 and .25 proportions of zeros, the statistic was significant in favor of the Poisson model.

Pseudo-Population Hurdle Models

Hurdle vs. Negative Binomial Hurdle

For the Hurdle models, regardless of the proportion of zeros, the positively skewed

distributions had a lower AIC than did the normal distributions. These, in turn, had a lower AIC

than the negatively skewed distribution. The same was true for the negative binomial Hurdle

models. However, for both models, the difference between AIC's for the three skew conditions

decreased as the proportion of zeros decreased (i.e., the models became more similar in fit).

When comparing the Hurdle models and negative binomial Hurdle models, the AIC's appear to

be similar regardless of the proportion of zeros and regardless of skew.










The deviance statistic was calculated comparing the Hurdle log-likelihood and negative

binomial Hurdle log-likelihood for each skew condition and zero proportion condition. These

statistics are displayed in Table 4-2. For all analyses, the Hurdle model is nested within its

negative binomial Hurdle formulation differing by one degree of freedom (i.e., the dispersion

parameter in the negative binomial Hurdle model). Hence, at the .05 Type I error rate and

assuming deviance statistics asymptotically distributed chi-square, a deviance exceeding 3.84

suggests better fit for the more complex negative binomial Hurdle model.

For the positively skewed distributions, the negative binomial Hurdle model fit

significantly better than the Hurdle model, except when the proportion of zeros was .90. In this

case the deviance statistic did not exceed 3.84. For the normal distributions, the Hurdle model fit

significantly better than the negative binomial Hurdle model when the proportion of zeros was

.10 or .25. When the proportion of zeros was .50, .75, or .90, the deviance did not exceed 3.84.

For the negatively skewed distributions, the Hurdle model fit significantly better than the

negative binomial Hurdle model when the proportion of zeros was .10, .25, .50, or .75. When the

proportion of zeros of .90, the deviance statistic did not exceed 3.84.

Poisson vs. Hurdle

Deviance statistics were also calculated to compare the Poisson log-likelihood and Hurdle

log-likelihood for each skew condition and zero proportion condition. These statistics are

displayed in Table 4-3. For all analyses, the Poisson model is nested within the Hurdle model

differing by three degrees of freedom. Hence, at the .05 Type I error rate and assuming deviance

statistics asymptotically distributed chi-square, a deviance exceeding 7.82 suggests better fit for

the more complex Hurdle model.

The deviances were large supporting the Hurdle model fit over the Poisson model fit. In

fact, several of the deviances were over 1,000. There was only one analysis that did not favor the









Hurdle model fit over the Poisson model fit. This was the deviance calculated for the positively

skewed distribution with a .25 proportion of zeros.

Negative Binomial Poisson vs. Negative Binomial Hurdle

Deviance statistics were also calculated to compare the negative binomial Poisson log-

likelihood and the negative binomial Hurdle log-likelihood for each skew condition and zero

proportion condition. These statistics are displayed in Table 4-4. For all analyses, the negative

binomial Poisson model is nested within the negative binomial Hurdle model differing by three

degrees of freedom (i.e., the duplication of parameters in the negative binomial Hurdle model to

represent both a transition stage and an event stage).28 Hence, at the .05 Type I error rate and

assuming deviance statistics asymptotically distributed chi-square, a deviance exceeding 7.82

suggests better fit for the more complex negative binomial Hurdle model.

As was the case when comparing the Poisson model and the Hurdle model, the deviances

were large. Likewise, there was one deviance that did not exceed the critical value. However, it

was for the positively skewed distribution with .50 zeros rather than the positively skewed

distribution with .25 zeros.

Pseudo-Population ZIP Models

The pseudo-population results for the ZIP models were rather similar to those obtained for

the Hurdle models. Graphically, regardless of skew condition and proportions of zeros, the AIC's

for the ZIP models and negative binomial ZIP models were very similar. Additionally, the AIC's

appeared to be equal between the .10 and .25 proportions of zeros conditions.

The deviance statistic was calculated comparing the ZIP log-likelihood and negative

binomial ZIP log-likelihood for each skew condition and zero proportion condition. These



28The presence of a dispersion parameter is now redundant between models.









statistics are displayed in Table 4-5. For all analyses, the ZIP model is nested within its negative

binomial ZIP formulation differing by one degree of freedom (i.e., the dispersion parameter in

the negative binomial ZIP model). Hence, at the .05 Type I error rate and assuming deviance

statistics asymptotically distributed chi-square, a deviance exceeding 3.84 suggests better fit for

the more complex negative binomial ZIP model.

The deviance was significant for only two comparisons. When the distribution was

positively skewed, the fit was significantly better for the negative binomial ZIP model than for

the ZIP model when the proportion of zeros was either .50 or .75. All other results suggested that

the ZIP model fit was adequate.

Comparing AIC's For All Models

For the .10 proportion of zeros condition with negative skew, the AIC's were

approximately equal for all models except for the Poisson models. The AIC for the Poisson

model was higher than for the other models, and the AIC for the negative binomial Poisson

model was highest of all. When the distribution was normal, the only model to have a noticeably

different AIC was the negative binomial Hurdle, which was again highest of all. For the

positively skewed distribution, there was some separation between the ZIP and Hurdle models,

with the Hurdle models having the lower AIC. The differences between these models and their

negative binomial formulations appeared to be trivial. The Poisson model appeared to have the

same AIC as those displayed for the ZIP models. The AIC for the negative binomial was again

highest of all. Between the three distributions, the AIC declined from the negatively skewed

distribution to the normal distribution to the positively skewed distribution.

For the .25 proportion of zeros, there was little distinction between the Poisson models

and negative binomial Poisson models for all distributions. Further, there was no distribution

displaying a nontrivial distinction between the Hurdle models and the ZIP models. For the









positively skewed distribution, all six distributions appeared approximately equal with a slightly

higher AIC apparent for the negative binomial Poisson model. Between the three distributions,

the AIC's appeared equal for the negatively skewed distribution and the normal distribution;

However, the AIC's for the normal distribution were considerably smaller than the AIC's for the

other two distributions.

For the .50 proportion of zeros condition with negative skew, the results for the Poisson

model and the negative binomial Poisson model reversed. For the negatively skewed distribution,

the AIC for the negative binomial Poisson was higher than those for the ZIP and Hurdle models,

while the AIC for the Poisson model was higher yet. As in the .50 proportion of zeros condition,

the AIC appeared equal for the ZIP and Hurdle models. Also, as in the .50 proportion of zeros

condition, there appeared to be no difference between any of the models for the positively

skewed distribution. Between the three distributions, the AIC's declined from the negatively

skewed distribution to the normal distribution to the positively skewed distribution with declines

being most rapid for the Poisson and negative binomial Poisson models.

Beyond the fact that the overall AIC was lower for the .90 proportion of zeros condition

than for the .75 condition, these last two conditions displayed similar results. For each

distribution, the Hurdle and ZIP model AIC's were approximately equal. They declined slightly

from the negatively skewed distribution to the normal distribution and declined a bit more from

this distribution to the positively skewed distribution. For the negatively skewed distribution, the

negative binomial Poisson AIC was considerably higher than the AIC's for the ZIP and Hurdle

models; the AIC for the Poisson model was highest of all. For the normal distribution, both the

Poisson and negative binomial Poisson model AIC's declined by approximately 50% of their

value for the negatively skewed distribution. For the positively skewed distribution, they