1 ESTIMATING THE SUPPL Y OF FOREST CARBON O FFSETS: A COMPARISON OF BEST WORST AND DISCRETE C HOICE VALUATION METH ODS By JOS R. SOTO A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FU LFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 201 3
2 201 3 Jos R. Soto
3 Teresa Leal (Mom) & Jos A. Soto (Dad)
4 ACKNOWLEDGMENTS My deepest gratitude toward s my dissertation committee chair, Dr. Damian Adams for his direction and unyielding support during these last two years of my doctoral studies. I decided to pursue a PhD in Food and Resource Economics to explore my curiosity for the field of Natural Resou rce Economics, but after finishing the field requirements of my degree, I lacked the focus to develop my dissertation. Dr. Adams provided the guidance and support needed to apply the skills of my field, and develop the discipline of independent doctoral re search. I strongly feel that he provided the necessary balance of guidance and independence needed to develop a professional research background. I would also like to thank my dissertation co chair, Dr. Michael Olexa for his commitment to support my gradua te studies at an early stage of my academic career. Dr. Olexa inspired me to explore the area of forest carbon markets, as an integral part of ecosystem services in Florida. His dedication and legal expertise provided the guidance needed to narrow down my research focus. I also thank Dr. Francisco Escobedo, who is an amazing mentor and a great and conservation, and for having the patience to guide an agricultural economist seeking to explore policies in his field. He had the grace to guide someone who, at the beginning of his dissertation, knew very little about forestry. I would also thank Dr. Sherry Larkin for providing shrewd and prompt feedback on my research. Dr. Larkin shared her vast experience in survey research, which helped shape my entire dissertation, from developing the survey design, to guiding the final elements of analysis in my study.
5 I am also thankful of the Florida Forest Stewardship Coordinator, Chris Dem ers for proving the network of landowners needed for my research, and for assisting the development of my survey design. Chris provided a very sharp and experienced perspective of the Florida forest landowners, which help refine the language needed to bett er communicate with the subject of my research. I would also like to thank Drs. Lisa House and Zhifeng Gao, as well as my doctoral colleague Sergio Alvarez for providing important feedback on the design of my research survey tool. I also want to t h ank the Florida Forest Service Stewardship Ecosystem Services and PINEMAP projects for funding this research. Finally, my deepest gratitude towards my dear friend, Geeti Shirazi Mahajan for her invaluable editorial feedback. Her countless hours correcting the gram matical mistakes of my dissertation are greatly appreciated.
6 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ .. 4 LIST OF TABLES ................................ ................................ ................................ ............ 9 LIST OF FIGURES ................................ ................................ ................................ ........ 11 ABSTRACT ................................ ................................ ................................ ................... 15 CHAPTER 1 INTRODUCTION ................................ ................................ ................................ .... 17 Existing Carbon Markets ................................ ................................ ......................... 20 Attitudes and Willingness to Accept ................................ ................................ ........ 21 Best Worst Choice vs. Discrete Choice Experimentation ................................ ....... 23 The Supply of Carbon Offsets ................................ ................................ ................. 24 Summary ................................ ................................ ................................ ................ 25 2 CARBON OFFSET OPPORT UNITIES FOR FLORIDA LANDHOLDERS .............. 26 A Brief History of US GHG Markets ................................ ................................ ........ 27 Supplying and Certifying Offsets ................................ ................................ ............. 30 Contract Length and Gen eral Description ................................ ........................ 33 Afforestation/Reforestation ................................ ................................ ............... 34 Improved Forest Management ................................ ................................ ......... 35 REDD and Avoided Conversion ................................ ................................ ....... 35 Permanence ................................ ................................ ................................ ..... 36 Market Demand ................................ ................................ ................................ ...... 37 Potential Barriers to Entry ................................ ................................ ....................... 39 Improve Forest Management, Afforestation and Reforestation, and Reduce Emissions from Deforestation and Degradation ................................ ............ 40 Contract Duration and Penalty for Early Withdrawal ................................ ........ 41 Additionality ................................ ................................ ................................ ...... 41 Formal Management Pl an ................................ ................................ ................ 41 Risk and Institutional Trust ................................ ................................ ............... 42 Perceptions and Misinformation ................................ ................................ ....... 42 Returns to scale ................................ ................................ ............................... 43 Special Requirements ................................ ................................ ...................... 43 Discussion ................................ ................................ ................................ .............. 43 Summary ................................ ................................ ................................ ................ 45
7 3 ATTITUDES AND WILLIN GNESS TO ACCEPT COMP ENSATION FOR CARBON OFFSET PRODUC TION IN FLORIDA: APP LICATION OF BEST WORST CHOICE MODELIN G AND DISCRETE CHOIC E EXPERIMENTATION .. 46 Background ................................ ................................ ................................ ............. 46 Statistical Models ................................ ................................ ................................ .... 52 Best Worst Choice ................................ ................................ ........................... 54 Discrete Choice Experimentation ................................ ................................ ..... 59 Methods ................................ ................................ ................................ .................. 60 Survey Instruments ................................ ................................ .......................... 60 Coding ................................ ................................ ................................ .............. 61 Estimation ................................ ................................ ................................ ......... 62 Choices and Variables ................................ ................................ ............................ 65 Results and Discussion ................................ ................................ ........................... 65 Alternative Model Specifications ................................ ................................ ............. 78 Summary ................................ ................................ ................................ ................ 79 4 C OMPARISON OF BEST WORST SCALING AND DI SCRETE CHOICE EXPERIMENTATION ................................ ................................ .............................. 81 Background ................................ ................................ ................................ ............. 82 Research Objective and Methodology ................................ ................................ .... 85 BWS and The Effects of Ambiguity for Empirical Researchers ........................ 85 External Validity ................................ ................................ ................................ 85 Estimation and Research Design ................................ ................................ ............ 86 Best Worst Scaling ................................ ................................ ........................... 87 Paire d Conditional Logit ................................ ................................ ............. 87 Paired Random Parameters Logit ................................ .............................. 88 Frequency Counts ................................ ................................ ...................... 88 Discrete Choice Experimentation ................................ ................................ ..... 89 Results ................................ ................................ ................................ .................... 91 Comparisons of Importance Measures ................................ ................................ 106 Limitations ................................ ................................ ................................ ............. 113 Discussion and Summary ................................ ................................ ..................... 114 5 E STIMATING THE SUPPLY OF FOREST CARBON OFF SETS : S TATED PREFERENCE MEASUREMENTS OF CARBON SEQUESTRATION IN FLORIDA ................................ ................................ ................................ .............. 116 Conceptual Framework ................................ ................................ ......................... 118 Estimating WTA Compensation for Pr oducing Carbon Offsets ...................... 119 Estimating the Probability of Enrollment ................................ ......................... 123 Combining BWC and DCE Survey Data with FIA Sample Plots ........................... 128 Carbon Sequestration Estimates ................................ ................................ .......... 134 Additionality ................................ ................................ ................................ .... 138 IFM Add itionality Estimation ................................ ................................ ..... 139 REDD Additionality Estimation ................................ ................................ 141
8 Results ................................ ................................ ................................ .................. 143 Discussion and Summary ................................ ................................ ..................... 147 6 CONCLUSIONS ................................ ................................ ................................ ... 148 APPENDIX A THE SURVEYS ................................ ................................ ................................ .... 150 Introduction and Filter Questions Survey ................................ ........................ 150 Instructions for Discrete Choice Experimentation ................................ ........... 151 Instructions f or Best Worst Choice ................................ ................................ 152 Demographic Questions ................................ ................................ ................. 153 B NON LINEAR CONSIDERATIONS ................................ ................................ ....... 155 LIST OF REFERENCES ................................ ................................ ............................. 157 BIOGRAPHICAL SKETCH ................................ ................................ .......................... 162
9 LIST OF TABLES Table page 2 1 Market share and prices of over the counter transactions for 2011. ................... 39 3 1 Empirical studies on willingness to participate in hypothetical carbon offset markets in North Amer ica ................................ ................................ ................... 48 3 2 Attributes and attribute variable levels in discrete choice experimentation and best worst choice ................................ ................................ ................................ 58 3 3 Effect s coding ................................ ................................ ................................ ..... 61 3 4 Characteristics of discrete choice experimentation, best worst choice, and national woodland ownership survey respondents ................................ ............. 63 3 5 Results from discrete choice experimentation: conditional logit model estimations ................................ ................................ ................................ ......... 68 3 6 Differences in marginal willingness to accept ($/choice) for discrete choice experimentation (Model 2) estimates ................................ ................................ .. 69 3 7 Results from binary choice model: random effects model estimations ............... 70 3 8 Diffe rences in marginal willingness to accept ($/choice) for binary (Model 2) estimates ................................ ................................ ................................ ............ 71 3 9 Results from best worst scaling: clogit estimates adjusting for covariates ......... 74 4 1 Results from best worst scaling: clogit estimates ................................ ............... 91 4 2 Results from best worst scaling: random parameters logit model estimations ... 93 4 3 Results from best worst scaling: random parameters logit model estimations excluding impact variables ................................ ................................ .................. 95 4 4 Results from best worst scaling: frequency counts ................................ ............. 99 4 5 Results from discrete choice experimentation model: random parameters logit model estimations ................................ ................................ ..................... 102 4 6 Mean willingness to accept ($/choice) estimates for discrete choice experimentation: random parameters logit (Model 1) ................................ ....... 103 4 7 Mean willingness to accept ($/choice) estimates for discrete choice experimentation: random parameters logit (Model 2) ................................ ....... 104
10 4 8 Differences in mean willingness to accept ($/choice) for discrete choice experimentation: random par ameters logit (Model 1) estimates ....................... 10 4 4 9 Order of differences between best worst scaling level scale attributes for BWS_clogit, BWS_RPL, and BWS_Freq (1 having the lowest distance) ......... 108 4 10 Ordering of relative impact of attribute/scale level ranges across best worst scaling models estimated using random parameter logit, conditional logit, adjusted mean frequency counts (1 h aving the least impact) ........................... 109 4 11 Comparison between rank and ordering of relative impact of attribute/scale level ranges across all methods used in Chapter 3 and Chapter 4 (1 having the lea st impact) ................................ ................................ ............................... 111 4 12 Comparison of order of differences between levels across all methods used in Chapter 3 and Chapter 4 (1 having the least impact) ................................ ... 111 5 1 Results from binary choice model: logit model estimations for survey respondents who reported a correct with address and/or zip code ................... 121 5 2 Summary stati stic of variables from carbon program survey: mean and ................................ ................................ ...................... 129 5 3 ................................ 132 5 4 Summary statistics of variables from Forest Inventory Analysis. Means and ................................ ................................ ..................... 133 5 5 Carbon pool in commerc ial pine plantations in Florid a ................................ ..... 140
11 LIST OF FIGURES Figure page 2 1 Map of Voluntar y Carbon Standard project sites ................................ ................ 31 2 2 2012 map of Climate Action Reserve projects s ites ................................ ........... 32 2 3 Map of GreenTrees Forest Carbon Project identifying GHG sources. The dots in this figure represent sources and sinks of carbon. A) Arkansa s. B) Louisiana. C) Mississippi ................................ ................................ .................... 33 2 4 Historic Voluntary C arbon Market transaction volume ................................ ........ 38 3 1 Ex ample of discrete choice experimentation question presented to survey respondents ................................ ................................ ................................ ........ 54 3 2 Example of best worst choice question, used to estimate best worst scaling and binary models ................................ ................................ .............................. 54 3 3 Willingness to accept ($/choice): discrete choice experimentation (Model2) vs. binary (Model2) ................................ ................................ ............................. 71 3 4 Willingness to acce pt ($/choice): discrete choice experimentation (Model3) vs. binary (Model3) ................................ ................................ ............................. 72 3 5 were available for selection ................................ ................................ ................ 72 3 6 were available for selection ................................ ................................ ................ 73 4 1 Rel ative desirability of carbon offset institutional elements: estimated with BWS_RPL excluding impact variables. ................................ .............................. 97 4 2 Relative desirability of carbon offset institutional elements: estimate d with BWS_RPL ................................ ................................ ................................ .......... 97 4 3 Relative desirability of carbon offset institutional attributes: estimated with BWS_RPL ................................ ................................ ................................ .......... 98 4 4 Best levels of carbon program institutional components ................................ ........... 100 4 5 Best tribute levels of carbon program institutional components ................................ ......................... 100
12 4 6 Best carbon program institutional components, di vided by the number of times they were available ................................ ................................ ........................... 101 4 7 Willingness to quantitatively coded ................................ ................................ .......................... 105 4 8 Willingness to ................................ ............................. 106 4 9 Best worst scaling model comparison: estimated using random parameter logit, conditional logit model, adjusted mean frequency counts (the upper horizontal axis corresponds to the adjusted mean of frequency counts, and the lower to the paired BWS estimations of clogit and RPL) ............................ 107 4 10 Best worst scaling attribute impact estimate comparison of random parameter logit and conditional logit models ................................ .................... 108 4 11 Best wors t scaling order of difference between attribute levels of random parameter logit and conditional logit models ................................ .................... 109 4 12 Best worst scaling order of relative impact of attribute/scale level r anges across methods ................................ ................................ ................................ 110 4 13 Order of attribute impact ................................ ................................ ................... 112 4 14 Order of difference between levels ................................ ................................ ... 113 5 1 Map of biggest forestland plot owned by BWC survey participants. ................. 120 5 2 Willingness to accept ($/attribute level) comparison of various models from ................................ ......................... 122 5 3 Supply response with 95% confidence interval. Supply response by landowner using Binary_logit for a program with 5 years contract duration, ................................ .. 123 5 4 Scenario 1 supply response. Supply response by landowner using Binary_logit for a program with 5 ................................ ...... 124 5 5 Scenario 2 supply response. Supply response by landowner using Binary_logit for a program with 40 ................................ ...... 125
13 5 6 Scenario 3 supply response. Supply response by landowner using Binary_logit for a program with 100 withdrawal. ................................ ...... 126 5 7 Multi cenario supply comparison. Supply response by landowner using Binary_logit for programs with several levels of contract duration, the use of ................................ ................... 127 5 8 Multi scenario supply comparison. Supply response by landowner using Binary_logit for programs with several levels of contract duration, the use of ................................ ......................... 127 5 9 Map of the biggest reported plot of forestland from Combined data ................. 131 5 10 Exam ple of Question 17 of survey. ................................ ................................ ... 134 5 11 Answers to Question 17 of the BWC and DCE surveys: the horizontal axis displays the percent of their Florida forestland willing to enroll in a carbon program described in Figure 5 6. ................................ ................................ ...... 135 5 12 Interaction between forest acreage reported and the percent of forestland willing to enroll in Scenario 1 ................................ ................................ ............ 136 5 13 Supply response with 95% confidence interval. Supply response by landowner using Binary_logit for a program with 5 years contract duration, ................................ .. 137 5 14 Total carbon sequestration estimates for respondents in FIACombined. The horizontal axis is the survey ID and the units of the vertical axis are in metric tons per year. ................................ ................................ ................................ .... 138 5 15 Map of forestland assumed to be under IFM ................................ .................... 142 5 16 Map of forestland assumed to be under REDD ................................ ................ 142 5 17 Carbon sequestration and additionality for IFM and REDD at 10% .................. 143 5 18 Carbon sequestration and additionality for IFM and REDD at 20% .................. 144 5 19 Carbon sequestration and additionality for IFM and REDD at 50% .................. 144 5 20 Supply of carbon additionality for Scenario 1 using IFM and REDD at 10% ..... 145 5 21 Supply of carbon additionality for Scenario 1 using IFM and REDD at 10% ..... 146 5 22 Supply of carbon additionality for Scenario 1 using IFM and REDD at 10% ..... 146 B 1 Discrete choice experimentation estimates of Mode 1 vs. attribute levels ........ 155
14 B 2 Binary estimates of Mode 1 vs. attribute levels ................................ ................. 155 B 3 O rthogonal main effects design. ................................ ................................ ....... 156
15 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 ESTIMATING THE SUPPL Y OF FOREST CARBON O FFSETS: A COMPARISON OF BEST WORST AND DISCRETE C HOICE VALUATION METH ODS By Jos R. Soto May 201 3 Chair: Damian C. Adams Cochair: Michael T. Olexa Major: Food and Resource Economics The use of forests to offset greenhouse gas (GHG) emissions has been promoted as a cost effective policy to deal with global warming. Carbon markets often e ncourage forest landowners to sequester carbon in exchange for compensation. This dissertation uses one of the most comprehensive lists of Florida non industrial private forest landowners to examine the role of forests within a carbon market framework. Fir st, it discusses the opportunities of Florida landowners to particip ate in existing carbon markets, and provides a description of the St sequestration potential. Next, it describes the implementation of two different conjoint choice tasks (best worst choice and discrete choice experimentation), which offer multiple options to assess attitudes of landowners towards different carbon offset programs, as well as various avenues to estimate willingn ess to accept compensation to part icipate in these programs This is followed by a comparison of these two methods of choice elicitation, and a description of the implications to the field of limited dependent variables. Lastly, it predicts the potential for enrollment in various carbon
16 of fset programs, and estimates the supply of carbon from sample plots located in the Northeast, Northwest, and Central areas of Florida. C arbon markets often use different platforms that vary in terms of contract length, penalties for withdrawal, etc. These differences in available carbon programs send signals to both consumers and potential producers of carbon credits, which often cause s confusion, price variations, and potential barriers to participation. This study examines these barriers and attitudes, an d identifies the carbon market institutional hus, t hese lands can play a key role in reducing GHGs by sequestering a pproximately 9.5 million metric tons of carbon per year 1 Despite the fact that the majority (65%) of forests are owned by non industrial private forest (NIPF) landowners 2 no previous research has been done to assess their institutional preferences regard ing carbon offset markets. This dissertation therefore investigates barriers to participate in forest carbon markets in Florida, and estimates the supply of offsets for the most promising institutional arrangements. The results of this study indicate that landowners would need between $20 to $30 acre per year to be positively affected by revenue, while the inclusion of a penalty for early withdrawal from a program increases the cost of participation by approximately $4.45 to $10.41 acre per year. 1 See Mulkey S., J. Alavalapati, A. Hodges, A.C. Wilkie and S. Grun wald (2008), Opportunities for greenhouse gas reduction by agriculture and forestry in Florida. University of Florida, School of Natural Resources and Environment Department of Environmental Defense, Washington D.C. 2 Florida 2010 Forest Inventory & Ana lysis Factsheet http://www.srs.fs.usda.gov/pubs/su/su_srs043.pdf
17 CHAPTER 1 INTRODUCTION Climate change is likely to impact the environmental and economic stability of the World Global warming has been associated with the anomalous heat waves and droughts in Texas, Oklahoma and Mexico in 2011 (Hansen et al., 2012), which cost the State of Texas an estimated $5 billion 1 Florida, as a peninsula, is particularly vulnerable to this threat. Sixty year projections of current climate trends have placed major port cities like Miami at risk of coastal flooding from storm surge s and hig h winds, with massive population e xposure (4.8 million people) and close to $3.5 trillion assets at risk (Nicholls et al., 2008). When policy makers seek to develop instruments to ameliorate the threat of global climate change by reducing greenhouse gas em issions they have several policy options to consider, namely, prohibitions, pollution permits, taxes, etc. I t is often important to under stand regional differences (Kaetzel et al., 2012) and cultural barriers (see Fisher and Charnley, 2010), in order to d esign the most cost effective program with optimal enrollment Fores t carbon markets currently exist, but continue to emerge in the US (Charnley et al., 2010). Understanding institutional barriers to these potential environmental markets (e.g., contract l ength, institutional trust, compensation, etc.) can lead to increased pollution reductions at optimal abatement costs. Lack of available knowledge on these issues has been cited in the literature as barriers to participate in similar programs (e.g., Butter field et al., 2005). 1 Time Magazine article retrieved on Aug. 31, 2011: http://www.time.com/time/nation/article/0,8599,20911 92,00. html
18 Greenhouse gasse s have been identified as drivers affecting the warming and coolin g of global climate They include carbon dioxide (CO 2 ), methane, and nitrous oxide, which have significantly increased as a result of human activi ties si nce 1750 (IPCC, 2007). Carbon dioxide is the second most abundant of all GHG s but has a high capacity to influence incoming and outgoing energy in the Earth atmospheric system via its radiative force the long run atmospheric accumulation of these gas es is global, which allows for their removal from any location to impact long term climate trends It is widely recognized that forests have the capacity to sequester and store CO 2 from the atmosphere (e.g., U .S. EPA, 2005). Several regional and national studies have found that using forests for this purpose merits consideration from both policy makers and landowne rs (e.g., Lubowski et al., 2005; St ainback and Alavalapati, 2002). Anthropomorphic GHGs emissions resulting from economic activities, which have with an output of production suc pollution (Jaffe et al., 2005). Coase (1960) resolved that these types of problems stem from ambiguous specificatio n of property rights, namely, whether GHG emitters have the right to release GHGs, or non are clearly specified (and market transaction costs are sufficiently low), then there will be an incentive to have an arrangement that will either compensate for the costs of pollution, or compensate for the economic loss of pollution abatement (Coase, 1960).
19 Another potential solution would be to incentivize forest landowners to sequester GHGs ( Raymond and Shi evely, 2008 ). Emissions trading or c ap and trade (CAT) programs create carbon markets, by allowing polluters to emit more GHGs than they are allowed, while paying others to stop polluting or capture/offset them elsewhere (for an extensive definition of CAT please see Raymond and Shievely, 2008). These frameworks typically include CO 2 ffsets provided the second largest sources of both supply and demand in the (global volunt ary carbon ) market, with companies taking on 5.6 MtCO 2 e (m etric ton carbon dioxide equivalent), just over the 4.9 MtCO 2 2011). The US does not have a national CAT program, but a recent national survey re vealed that 75% of U.S. voters favor regulating CO 2 as a greenhouse gas pollutant 2 The State of Florida has yet to create a carbon offset market or take part in any regional cap and trade agreement (e.g., Regional Green House Gas Initiative), but there is certainly a framework in place in multiple voluntary national programs (e.g., Climate Action Reserve, Voluntary Carbon Standard) to potentially take advantage of its vast public and private forestlands (over 17 million acres) to not only protect the state from climate change, but to create wealth by carbon offsets. Mulkey et al. (2008) estimate that the afforestation of 5% of Florida range and pasture lands, along with increased management intensity on pine plantations can potentially yield $139.6 million 2 Zabanko, Deborah (2012) U.S. voters favor regulating carbon dioxide: sur vey. Chicago Tribune, April 26, 2012
20 to producers 3 Yet, despite the economic benefits of engaging these markets, no empirical study has been done to assess the attitudes of Florida private forest landowners towards these programs in such a way that inform s poli cymakers and resource manager s. The overall purpose of this research is to examine barriers to participate in forest carbon offset markets in Florida. The specific goals are 1) to identify optimal institutio nal structures of carbon market programs that increase participation at effici ent cost s ; 2) to estimate the supply of carb on in Florida; and 3) to make a contribution to the measurement of bes t worst scaling by assessing the p otential of best worst choice to produce measurements of traditional discrete choice experimentation, and best worst scaling. Existing Carbon Markets C hapter 2 discusses current and emerging forest carbon market opportunities for Florida landowners The use of forest carbon markets that pay landowners to capture GHG emissions for example by planting trees, p reventing forest degradation, or improving forest management practice s, are currently being considered by 22 U.S. states, two Canadian provinces, and six Mexican observant regions 4 number of environmental, economic, and social constraints curr ently limit carbon market participation by forest owners. Key issues in clude: the low price of carbon and high cost of market entry; whether small landowners can gain market access; how to meet 3 These estimates were done using $20 per metric ton CO2 equivalent and do not reflect costs of creation and maintenance of mitigation projects. 4 See Regional Greenhouse Gas Initiative, Western Climate Initiative, and Midwestern Accord
21 requirements such as management plans and certification; and whether managing for Landowners in the Southeast are currently able to participate in three major carbon offset certification programs: American Carbon Registry (ACR), Clim ate Action Reserve (CAR), and Voluntary Carbon Standard (VSC). Georgia, Alabama, Mississippi, South Carolina, and Louisiana are currently participating in these programs with more than 25 forest projects with thousands of acres, but Florida only makes use of these opportunities with six landfill projects and one energy program 5 6 7 This chapter qualitatively analyzes the major requirements of ACR, VCS, and CAR, and reviews existing literature related to Florida non industrial private forest landowners to identify potential barriers to participate. Three major categories were identified as potential barriers to participation: compensation, contractual commitment (length and penalties for early withdrawal), and risk. Attitudes and Willingness to Accept Chap ter 3 characterizes the guiding structure of car bon production in Florida, by i dentifying barriers to participate in a hypothetical carbon of fset program, and estimating willingness to participate of non industrial private forest landowners. To explore opportunities in new markets, it is not only appropriate to elicit willingness to accept (WTA) compensation for landowners to produce offsets (e.g., Shaikh et al., 2007), but also to research institutional factors that may influence the likelihood of land managers 5 See project listings for CAR: http://www.climateactionreserve.org/how/projects/ 6 See project listings for ACR: http://americancarbonregistry.org/carbon registry/projects/project list by type/acr_atct_topic_view?b_start:int=0& C= 7 See project listings for VCS: http://www.vcsprojectdatabase.org/
22 to participate in such programs (e.g ., Fletcher et al., 2009) To address this, a conjoint choice survey was developed and electronically administered in December 2011 to 920 Florida Forest Stewardship Program participants, with a response rate o f 34%, of which 189 completed the entire survey. Four program attributes were included: contract length (5 to 100 years), annual compensation ($5 to $30 per acre), penalty for early withdrawal (penalty, or no penalty), and the type of risk tool ( i nsurance or risk pool), along with demographic and institutional trust questions. Two methods of choice elicitation were used (discrete choice experimentation and best worst choice), which only varied on the type of conjoint method. Respondents were randomly presen ted with only one survey type, which resulted in equal response rates. One that required respondents to choose among three single profile of attributes and consisted of t wo separate instructions: 1) to select a most preferred and least preferred attribute level, and 2) to consider the profile as a single carbon program and choose whether to enroll or not. The three different sets of instructions for these two conjoint task s resulted in three valuation models: Best Worst Scaling (BWS), Binary Logit (Binary) and Discrete Choice Experimentation (DCE), and their elicitation formats in the context of hypothetical carbon sequ estration markets in Florida. These models were used t o test the hypotheses that several factors do not affect willingness to accept payments for carbon sequestration: (1) carbon contract revenue (2) institutional factors (risk tool, contract length ), and (3) respondent characteristics.
23 The results from thi s chapter indicate that non industrial private forest landowners in this study are more influenc ed by carbon offset revenue, tha n penalty for early withdrawal, and contract duration. Carbon market programs that offer compensations $20 or $30 acre per year have a positive impact on participation, while $5 and $10 acre per year are less desirable The least preferred component of this study seems to be a contract commitment of 100 years. A program with this duration would elicit an incre ase in cost of partici pation of $28.53 to $37.78 acre per year, while a 10 year commitment woul d lower cost by $12.48 or $15.07 acre per year. Best Worst Choice vs. Discrete Choice Experimentation C hapter 4 makes a contribution to the measurement of BWS, by applying and comp aring best worst choice (BWC), an innovation proposed by Flynn et al. (2007) and applied by Coast et al. (2006), with the conventional discrete choice experiment (DCE) method (see Louviere et al. 2000). The limitation of BWS with regard to the estimation o f willingness to pay (see Louviere and Islam, 2008) has been somewhat circumvented in the field of applied economics by implementing both BWS and DCE tasks in a single survey (e.g., Lusk and Parker, 2009). This approach is very likely to increase both choi ce task complexity and survey length. But, BWC may solve this problem, by asking respondents to perform two tasks: 1) select a best and a worst attribute from a profile, and 2) to accept or reject the sc enario as a whole Thus, the latter instruction resem bles the Binary model, which allows for the estimation of WTA. This chapter builds on the work done by Louviere and Islam (2008) that compares BWS with Binary estimates, and Potoglou et al. (2011), which analyzes differences between BWS and DCE measurement s. By comparing BWC with DCE, we provide the first assessment of measurements of BWC, compared to the su fficiently
24 well documented DCE. T he results of this study will provide applied economists evidence of the performance of a conjoint tool that take s adva ntage of BWS, while estimating WTA, without having to resort to two choice tasks in a single survey thereby reducing choice task complexity. The Supply of Carbon Offsets Finally, Chapter 5 combines carbon sequestration data from the Forest Inventory Analysis of the USDA Forest Service 8 to simulate an improved forest management practice carbon program, and trace a supply curve of forest carbon offsets of Florida Stewardship Program participants. Using the findings of previous chap ters, a Linear logit model (see Louviere et al., 2000) is used to predicted the probability of participating in various carbon offset programs scenarios of low, medium, and high preference (e.g. Markowski Lindsay et al., 2011; Kline et al., 2000). Estimate s of carbon additionality were estimated using the findings of Mulkey et al. (2008) that estimate average per acre forest carbon sequestration rates in Florida via simulation tools. This chapter finds a higher probability of participation in programs that offer more than $30 per acre per year, and commitment periods of 30 years or less, along with significant supply shifts stemming from the inclusion of characteristics such as penalty for early withdrawal from a carbon offset contract. The respondents in th is survey are well represented in areas of the Northeast, Northwest, and Central portions of Florida. 8 http://www.fia.fs.fed.us/
25 Summary This study uses Best Worst Choice modeling and Discrete Choice Experiment to estimate WTA compensation for producin g carbon offsets in Florida, an d identifying potential barriers to participation. The data from these models was enriched with carbon sequestration data from the Forest Inventory Analysis, in order to estimate the supply of carbon offsets of representative plots in the Central and North ern portions of Florida. The surveys had a response rate of 34%, and the results show a preference for hypothetical programs with high revenues and no penalty for withdrawal; and a positive of WTA from DCE. Florida has great potential for producing carbon offsets, and the results of this study will generate interest the forest landowner community, private sector consumers of carbon offsets, and policy makers. In addition, this dissertation will also be a contribution to the applied economics community, by exploring the potential of Best A ccept.
26 CHAPTER 2 CARBON OFFSET OPPORT UNITIES FOR FLORID A LANDHOLDERS In recent years, several institutions and entrepreneurs in Florida have expressed interest in engaging carbon markets or projects that aim to reduce greenhouse gas emissions Since 2008, the University of Florida has purchased nearly 8,000 to ns of carbon offsets from Earth Givers Inc. (a local non profit), to spearhead the first carbon neutral football season in NCAA history, and to have its first carbon neutral commencement ceremony ( Cravey [personal communication], 2011). Also, i n 2007, the City of Miami entered into a contract with the Chicago Climate Exchange to further their goal of reducing 6% of GHG emissions by 2010, while allocating $500,000 of their 2007 08 budget to the Office of Sustaina bility (Acosta, 2009). Two years ago the stat e governments to levy non ad valorem (no value added for taxation purposes) assessments to fund qualifying improvements in energy conservation, and renewable energy as well as allowing them to adopt ordinances or resolutions that provide upfront funds to cover the financial costs of these environmental improvements (Friedman and Glinn, 2010). As of this writing, Florida has yet to register any current forest carbon of fset projects using the major available certification platforms (see ACR, CAR, and VCS project listings 1 2 3 ). The long term negative effects associated with climate change (see IPCC, 2007) highlight an additional role for forestlands to assist the economi c and human wellbeing 1 See project listings for CAR: http://www.climateactionreserve.org/how/projects/ 2 See project listings for ACR: http://americancarbonregistry.org/carbon registry/pro jects/project list by type/acr_atct_topic_view?b_start:int=0& C= 3 See project listings for VCS: http://www.vcsprojectdatabase.org/
27 of the State of Florida. In spite of having almost 50% of its land area covered by forests, CO 2 equivalent ( CO 2 e) units (Center for Climate Strategies, 2008). Converting 1.85 million acres of forestland from lower to higher forest management intensity, and 2.9 million from medium to high intensity, along with 5% afforestation of range and pasture lands could result in 2.7 MMt of additional carbon sequestr ation per year (Mulkey et al., 2008). Using a $20 per metric ton CO 2 e, this change can potential yield $23 million for afforestation and $116.8 million for changes in management intensity. The proposed additional sequestration will not make Florida carbon neutral, but it will certainly contribute to efforts that mitigate the negative effects of climate change. This second chapter provides an overview of the current opportunities available for Florida forest landowners to seek compensation for sequestering c arbon. The chapter first provides a brief historical overview of GHG markets in the US, then elaborates on existing national frameworks to certify carbon offsets, followed by an analysis of potential barriers to participate. A Brief History of U S GHG Marke ts Over the past 20 years, there have been several major efforts to institute federal policies that regulate GHG emissions in the US In 1992, the US ratified the United Nations Framework Convention on Climate Change, which became the negotiating framework of GHG reductions under the Kyoto Protocol (Charnley et al., 2010). This protocol required countries to reduce GHG emissions to an average of 5% below 1990 levels by 2012 (the requirements differed by schedule), with developed countries facing more obliga tions than their less developed counterparts. The widely recognized success (by business and environmental groups) of the 1990s Acid Rain Program paved the way
28 for the European Union to initiate a cap and trade program in 2005, in order to meet their Kyoto Protocol obligations (Raymond and Shively, 2008). Yet all the excitement of the 1990s was not enough to overcome bipartisan disagreement on exempting developing countries from emissions reduction requirements under the Kyoto Protocol, and the US failed to ratify the 1997 protocol (see Senate Resolution 98, 105 th Congress, 1 st Session ) However, the idea of a national CAT program weathered through and resurfaced in the late 2000s, with a Supreme Court ruling (Massachusetts v. EPA) requiring the Environmenta l Protection Agency to regulate GHG s as air pollution Vice President Al highly influenced public opinion, a s did presidential candidates showing support for a national CAT (Adams, 2009) This momentum culminated with the pa ssing of US House of Representatives Bill 2454 (t he American Clean Energy Sec urity Act of 2009), which included the possibility of utilizing forest carbon offsets to mitigate GHG emissions The bill lost viability with the recent economic down turn and it did not make out of the US Senate (Broder and Krauss, 2010). In 2011 the system) lost its CAT component (Gronewold, 2011) leaving potential carbon producers less than satisfied (Sager [personal communication], 2011). The res iliency o f CAT continues to merit consideration by 22 state s under three major regional blocks: Western Climate Initiative (WCI), Regional Greenhouse Gas Initiative (RGGI), and the Midwest ern Greenhouse Gas Accord (MGGA). As of this writing, RGGI is the on ly active CAT regional program in the US, and is currently facing
29 the withdrawal of one of its 10 regional members (New Jersey), 4 and MGGA negotiations appear to be idle 5 The majority of WCI member states are not planning to implement a CAT prog ram, excep t for New Mexico (although the current governor ), California and three Canadian Provinces ( Alter [personal communication], 2011). The demand for RGGI CO 2 allowances has fallen dramatically and the prices have reached their virtual price floor of $1.86 per ton 6 This CAT program in the Northeast regulates power plants, and only allows for 3.3% of their allowances to come from CO 2 offsets. Forest projects are included in this system, but only wit hin the 10 participating member states ( Connecticut, Maine, Maryland, Massachusetts, New Hampshire, New Jersey, New York, Rhode Island, Vermont, and New Jersey ). of 2006 (AB32). The goals are to reduce GHG emissions to 1990 levels by the year 2020 7 This initiative aims to create a very ambitious statewide CAT program on all aggregate GHG emitters, and it plans to allow forest carbon offsets. The California Department of Forestry and Fire Protection estimates that 5 million metric tons of CO 2 8 Stanley 4 WJS Article, May 26 th 2011: http://online.wsj.com/article/SB10001424052702304520804576347560078618994.html 5 Point Carbon Fed 25 th 2011: http://www.pointcarbon.com/pages/shop/1.1510367 6 th 2012: http://www.huffingtonpost.com/robert stavins/low prices a problem maki_b_1461501.html 7 CA EPA website: http://www.arb.ca.gov/cc/ab3 2/ab32.htm 8 th 2012: http://www.lao.ca.gov/analysis/2012/resources/cap and trade auction revenues 021612.aspx
30 and Hamilton, 2012), a nd is likely to be linked to other networks under the Western Climate Initiative 9 Supplying and Certifying Offsets Florida is not currently participating in RGGI or WCI, but landho lde rs have three major certification options to engage carbon offset markets: CAR, ACR, and VCS These are voluntary non profit carbon offset certification programs that slightly differ in protocol requirements, but encompass similar types of forest offset ac tivities: afforestation/reforestation (AR), improve d forest management (IFM), and reducing emissions from deforestation and degradation (REDD). They play the role of independent third party standards for carbon credits that are sold in domestic and foreign markets. In 2011, these certifiers guided the development of 76% of all (independent third party standards) transacted credits in voluntary markets (Peter Stanly and Hamilton, 2012). As of this writing, no Florida forest projects have been registered unde r these three programs, but there are six landfills certified to receive CO 2 equivalent credits for trucks with electricity to avoid using diesel under ACR, one VCS pr oject in Lee County receiving credits for incinerating municipal waste to generate electricity, 10 11 12 and 39 9 th 2010: http://www.huffingtonpost.com/robert stavin s/ab 32 rggi and climate ch_b_749791.html 10 See project listings for CAR: http://www.climateactionreserve.org/how/projects/ 11 See project listings for ACR: http://americancarbonregistry.org/carbon registry/projects/project list by type/acr_atct_topic_view? b_start:int=0& C= 12 See project listings for VCS: http://www.vcsprojectdatabase.org/
31 certification program) 13 The southeastern states of Georgia Alabama Louisiana Mississippi and South Carolina are however currently participating in these programs with: 25 forest projects, projects. (F igure 2 1 & Figure 2 2) One of the most ambitious forest projects in ACR (GreenTrees Forest Carbon million acres in the Lower Mis sissippi Alluvial Valley roximately 25 million acres in Louisiana, Mississippi, Arkansas, Kentucky, Tennessee, Missouri and Illinois 14 (Figure 2 3) Figure 2 1. Map of V oluntary Carbon Standard project sites. [Figure ad a pted from VCS project listings website: http://www.vcsprojec tdatabase.org ] 13 The Gold Standard project website: https://gs2.apx.com/mymodule/mypage.asp 14 Project website: http://americancarbonregistry.org/carbon registry/projects/greentrees fores t carbon project
32 Figure 2 2. 2012 map of Climate Action Reserve p rojects sites. [Figure ad a pted from CAR project listings website: http://www.climateactionreserve.org/how/proj ects/ ]
33 Figure 2 3. Map of GreenTrees Forest Carbon Project identifying G HG sources. The dots in this figure represent sources and sinks of carbon. A) Arkansas. B) Louisian a. C) Mississippi. [Figure ad a pted from American Carbon Registry project listings website: http://americancarbonregistry.org/carbon registry/projects/project l ist by type/acr_atct_topic_view?b_start:int=0& C= ] Contract Length and General Description These three programs have commitment periods that range from 40 to 100 years (ACR 40, VCS 20 100, CAR 100) Recent (e.g., Markowski Lindsay et al., 2011) studies hav e identified contract duration as a potential barrier to participation in similar programs. Landownership can be both private and public, and they typically retain their ownership status during their production period, except for CAR, which requires partic ipants of REDD to reclassify their property as conservation easements or to transfer them to public ownership. Most programs allow for the use of the clean development mechanism methodologies defined by the Kyoto Protocol, and for the implementation of mul tiple A B C
34 activities, that is, a mix of AR, REDD, and IFM. The registration processes and procedures vary within each project type but all involve a project manager who submits a proposal to an offset certifi cation program (ACR, VCS, CAR) that screens applica tions, and determines who is eligible to then be inspected by an i ndependent third party verifier who confirm s the validity of the project. If approved, the project is registered and carbon credits are issued in accordance with project type and production capacity Afforestation/Reforestation AR involves the restoration of forest covered lands that are not at optimal stocking levels rograms certification is consider ed for areas that were not converted to non forests within the last 1 0 years prior t o the beginning of a project, or for areas that suffer ed a significant natural disturbance such as hurricane or fire (ACR, VCS, CAR). Some programs have special requirements, such as ACR, that normally considers projects with start dates after 1997, or CAR, wh ich does not allow lands that have been previously registered as Forest Projects (unless the project was terminated due to unavoidable reversal). The starting date for this project type is at the beginning of the planting season (ACR, VCS), or at the remov al of obstacles to growth (CAR). After the project is approved, it receives a crediting period of 40 to 100 years (40 ACR, 20 100 VCS, 100 uction characteristics. Carbon offsets are mea sured by comparing a baseline estimation of all GHG sinks with the subsequent sequestration rates implemented by project. The three programs offer slightly different ways to calculate baselines, but in essence, all of them
35 take an inventory of current GH G stocks and leaks, which are then compare d with the Improved Forest Management The IFM activities include: making managerial adjustments of conventional logging, no logging, extending rotati on periods going from low to high productive forests, thinning diseased or suppressed trees, managing competing brush and short lived species, and increa sing tree density. This project is considered to have started after the application of the new management regime. Their crediting pe riod ranges from 20 100 years (20 ACR, 20 100 VCS, 100 CAR) The VCS considers lands that have not been converted to non forests for at least 10 years prior to th e beginning of the project, and CAR requires lands that have not been previously registered as Forest Projects ( unless project was terminated due to unavoidable reversal ), and to have les s than 10% tree canopy cover The baseline estimation used to compare improvements of carbon stock requires land managers to identify a credible alternative forest management scenario (ACR), or to provide 5 al historical a qualitative characterization of likely vegetative conditions and activities that would have occurred without the project (including laws, statutes, regulations or other legal mandates), along with 20 samples plots to perform a compute r simulation for 100 years REDD and Avoided Conversion REDD is a n approach to avoid planned, unplanned, and/or illegal deforestation of lands that are threatened by urban development, industrial tree production, and/or changes in legislation. The starting dates for ACR and VCS are at the implementation of
36 and CAR initiates after the recording of a conservation eas ement, or tr ansfer to public ownership REDD projects are credited for 10 100 years (10 ACR, 20 100 VCS, 100 CAR), and CAR does not allow projects to tak e place on previou sly registered Forest Projects, unless the project was terminated due to unavoidable reve rsal The additional sequestration of this protocol is measured by comparing existing carbon stocks, with the expected stocks of the threatening contingency (i.e. roa ds, buildings, timber harvest), namely, the carbon stocks of the treating contingency become the baseline of the project. Also, VCS demands a reassessment every 10 years, while CAR requires the same type of qualitative and forecasting methods of IFM. Permanence The project s boundaries (standard sources, sinks, and reservoirs) are defined within the property, but managers are responsible for leaks that may occur outside its bounds, due to the project s own activities. Hence, if a project displaces cattle grazing, urban development, transportation, or tree farming, then those activities wo uld be considered leaks. This caveat specially applies to REDD, which requires regular monitoring of the owner who was originally planning to degrade or deforest a land. Leaks from intentional or unintentional (i.e. natural disaster) reversals are managed by instituting a s eries of risk management tools, such as: allowing participants to propose insurance products (ACR), carbon banking pools, a.k.a. buffer pools (ACR, VCS, CAR), and in some cases a buy out option (ACR). Buffer tools are used by programs to all registered pr ojects They work by allowing project managers to deposit a percentage of offsets (similar to insurance premiums) into an account controlled and managed by
37 the program. The pool of offsets is u sed to cover carbon losses from unexpected reversals (e.g., wildfires hurricanes, etc. ). The amount deposited and refunded varies with each progr am, for example, ACR refunds 10% of offsets to producers every 5 years of non reversals, whereas VCS and CAR a sk for a certain percentage of offset considered lower in cases of easements or deed commitments. Each buffer or insurance is required for entire dur ation of the commitment period. Market Demand The only existing CAT market in the US is RGGI, but as previously mentioned, Assembly Bill 32 will create a CAT market in the State of California, but the participation of o ut of state forest projects has yet to be determined. In 2011, the California Air Resources Board approved CAR protocols for early action compliance of CAR credits (Peter Stanly and Hamilton, 2012). ACR, CAR, and VCS can only be traded in voluntary carbon markets (VCM). In 2010, the only available VCMs in the US were over the counter (OTC) transactions and CCX (Charnley et al., 2010), but the latter no longer exists. (Figure 2 3) Florida landowners attempting to sell their certified credits would have to ma ke use of OTC markets. These involve the use of private contracts between offset provider and buyer (Charnley et al., 2010). The transaction occurs via a private exchange (e.g., Climax) or directly thought a broker or online retail (Peter Stanly and Hamilt on, 2012). Most transactions occur directly. (Figure 2 4)
38 Figure 2 4. Historic Voluntary Carbon Market t ransaction volume. [Figur e ad a pted from Peters Stanley M, Hamilton K (2012) State of the Voluntary Carbon Mar kets 2012: Developing Dimension (Page 9, Figure 7), Ecosystem Marketplace and Bloomberg New Energy Finance, Washington, DC .] The rest of this section summarizes the findings of the State of the Voluntary Carbon Market 2012 by Peter Stanly and Hamilton (2012). Since the vast majority of OTC transa ctions occur via private brokers and individuals, prices and transacted volumes are not centralized in a market type dataset. Peter Stanly and Hamilton report the findings of a survey collected from 312 offset suppliers, seven exchanges, and all major regi stries (including ACR, VCS, and CAR). In 2010 2011, the market share of over the counter (OTC) transactions grew by 20%, while forest projects decreased by 38%. Respondents of the survey attributed this decrease to the recent financial crisis. The OTC vol ume in 2011 constituted 1% (95 Mt CO 2 e, valued at $572 million) of the global carbon market. AR and IFM projects transacted 7.6 Mt CO 2 e and REDD 7.3 Mt CO 2 e.
39 purchased 28 Mt CO 2 e at a cost of $1 51 million. North America supplied 30Mt CO 2 e (23% forest) worth $178 million. Of this volume, 92% were sold domestically, and the rest in Europe. In 2011, the US purchased a total of $159 million worth of OTC offsets, at an average price of $6 per Mt CO 2 e. ACR, VCS, and CAR had a price range of $0.10 to $30 per Mt CO 2 e. The average prices for VCS, ACR, and CAR during 2011 were $3.70, $5.8, $7.30 per Mt CO 2 e, respectively. CAR appeared to be fetching higher prices, given the pre compliance approval of their pro tocols under the California CAT. ACR had the highest forest OTC transactions in North America with 2.56 Mt CO 2 e, followed by 1.63 of VSC, and 1.17 with CAR. Most transactions for these programs occurred in North America, with the exception of 25% of VCS. (T able 2 1) Table 2 1. Market share and prices of over the counter transactions for 2011. Market Share World Volume (Mt CO 2 e) North American Volume (Mt CO 2 e) Forest Share Price range ($/Mt CO 2 e) Avg. prices ($/Mt CO 2 e) VCS 58% 41 10.2 16% 0.12 30.00 3.70 ACR 6% 4 4 64% 0.10 10.00 5.80 CAR 12% 9 9 1 3% 0.35 15.35 7.30 US 41% 30 30 23% 6 [Data adapted from Peters Stanley M, Hamilton K (2012) State of the Voluntary Carbon Mar kets 2012: Developing Dimension, Ecosystem Marketplace and Bloomberg New Energ y Finance, Washington, DC .] Potential Barriers to Entry Specific barriers to participate in forest carbon offset markets have yet to be examined in the State of Florida, but similar programs have been explored in Massachusetts, Texas, and Western Canada (e.g., Markowski Lindsay et al., 2011; Lee
40 2010; Shaikh et al. 2007). Fischer and Charnley (2010) have also explored the literature of non industrial private forest landowners to identify the cultural elements that may influence participatio n in sequestration programs. Other land use studies of NIPF landowners in Florida have explored willingness to accept compensation to convert forestland to biofuel crops area, which offer interesting information on reservation prices and forest attitudes ( Pancholy et al., 2011). A Massachusetts study of NIPFs from 2010 by Markowski Lindsay et al. (2011), presented participants with different carbon sequestration programs and elicited their ratings. The programs included the following attributes: management plan (required/not required), contract length (15 or 30 years), percent of land required to enroll (50 or 100%), revenue ($10, $100, $1000), additionality, penalty for early withdrawal (no penalty or repay earnings plus 20% fee), and institutional trust ( implemented by public or private sector). With a response rate of 43%, the study found significant preferences for programs with higher net revenue, no withdrawal penalty, shorter contract lengths, at forests must be managed The following subsections discuss some of the key barriers identified by Markowski Lindsay (2011) and Fischer and Charnley (2010), which may be applicable to Florida NIPF landow ners in ACR, CAR, and VCS. Improve Forest Management, Afforestation and Reforestation, and Reduce Emissions from Deforestation and Degradation Fischer and Charnley (2010) identify REDD as being compatible with the legacy values and their reluctance to harvest
41 ecological values of family landowners. They cite literature evidencing a potential reluctance to engage in AR, on the grounds that planting new trees commits them to long term financial investments, but the fact that the economics of afforestation values are highly regional, and the high stumpage values of the South might make it more appealing. Contract Duration and Penalty for Early Withdrawal Fischer and Charnley (2010) find evidence that family forest owners are concerned about losing property rights, as well as reluctant to make large financial investments for long term benefits. Markowski Lindsay et al. (2011) also identifie d preferences for shorter durations and no penalty for withdrawal. H aving commitment periods that range from 40 to 100 years ( ACR 40, VCS 20 100, CAR 100) might create rigidity in the property rights of landholders who might be considering developing their Florida l ands in the not so distant future. Additionality Markowski Lindsay et al. (2011) found evidence that NIPFs prefer programs that This attribute is present in most carbon certification programs, and it represents the method by which they justify compensation for implementing actions that increase (or protect) carbon sequestration. It is very unlikely that this component will disappea r from ACR, CAR, or VCS. Formal Management Plan T he above Massachusetts study found that eligibility r equiring management plans had a no significant influence in willingness to participate. Fischer and Charnley (2010) also explain that NIPFs already use ma nagement practices that promote carbon
42 sequestration, but landowners management programs may be attributable to their values for privacy and autonomy and skep Risk and Institu tional Trust I n spite of risk pools and insurance services, the fact that most programs see no difference between intentional and unintentional revers als might be a cause for est ownership compounded by concerns about the risks and burdens of entering into Markowski Linds ay et al. (2011) was found to be insignificant to a .1 level. Perceptions and Misinformation In a personal interview in 2011, former third party verifier Scott Sager o f Environmental Services, Inc., mentioned that one of the biggest obstacles to participat ion was bad information Sager gave an example o f aggregators who told potential Florida clients that CCX was going to provide ample demand, b The Massachusetts study found that some knowledge about the role of forest in to a .01 level of significance. The survey Fischer and Charnley (2010) found that landowners would be more willing to take action if they view themselves more vulnera ble to climate change.
43 Several efforts of key states are helping to reduce information insecurities. The State of Oklahoma provides a state run carbon program with authority to verify carbon developing verification protocols, and supporting research 15 Returns to scale The Fischer and Charnley (2010) study found that lack of overriding financial motivations and contract risks will likely limit participation. Given the cos ts of ver ification and maintenance (these c osts depend on project type and land characteristics), these programs may not seem profitable f or small scale landholders. Sager [personal interview] (2011) res, given current Special Requirements S ome programs have specific forest management requirements depending on the methodology and project type, as is the case of VCS, but CAR has a diversity of native species requirement that demands a 95% or higher level of native species within the first 50 years of their commitment period. Fischer and Charnley (2010) find evidence of family forest values that include: enjoying beauty and scenery, protecting nature and biodi versity; both of which correlate with some of these program requirements, but there might be some hidden costs associated with invasive species or planting native species. Discussion As seen, ACR, CAR, and VCS offer several differences in program requirem ents that may attract or dissuade the participation of landowners in carbon offsets. For 15 Oklahoma Carbon Program website: open http://www.ok.gov/conservation/Agency_Divisions/Water_Quality_Division/WQ_Carbon_Sequestration/Ab out_the_Program/
44 example, landowners interested in shorter time commitments might consider a REDD or IFM protocol, which does not require engaging in a long term financial commitment. This group might also prefer to avoid the 100 year obligations of CAR, by using the more flexible commitment periods of VCS (20 100 years) and ACR (40 years). But landowners concerned with property rights may not be willing to participate in CAR, which re quires land to be transferred to an easement or to public property for the duration of some protocols (REDD). Owners with lands invaded by non native species would also experience additional costs associated with the diversity of native species requirement of CAR, and available in VCS. Potential participants concerned with revenue and market stability, might prefer to participate with CAR. This certification currently fetches higher prices, and in lieu of the California Air Resources Board approval of CAR p rotocols for early action compliance, this is likely to continue. CAR credits might be grandfathered into the new California CAT program, which might be a way to access this market from a non WCI member state, like Florida. For the most part, VCS is the m ore flexible platform in terms of time commitment, but it also attracts the lowest prices. An advantage of this certification is that it can be traded in more countries and exchanges than ACR or CAR (Peter Stanly and Hamilton, 2012). Landowners not intere sted in ACR, CAR, or VCS might also want to consider other options in the OTC market. If they can find a willing buyer, they would be able to craft their own protocol tailored to their own preferences. This opportunity might not be
45 so viable, given the exi stence of reputable platforms such as the ones described in this chapter. Summary Carbon markets offer Florida landowners the opportunity to take advantage of their vast forest resources to create wealth, decrease long term vulnerability from climate ch an ge, and strive to maintain the natural florid character of the state. This chapter examined existing options to participate in carbon markets, and identified differences in available certification programs. In addition, a comparison of these programs was presented, highlighting potential barriers to entry A ny national, regional, or state C AT program in the US is likely to accept forest carbon of fsets to make up for pollution using protocols similar to the ones explained in this chapter. Therefore, policy makers in Florida interested in mitigating GHG or increasing options for forest landowners, are going to find useful information in this chapter regarding existing carbon offset certification options. Creating a statewide offset market or joining a reg iona l cap and trade program is going to require the cooperation of all s ectors of society, but a first good step would be to inform landowners about existing opportunities to engage in forest carbon offsets.
46 CHAPTER 3 A TTITUDES AND WILLING NESS TO ACCEPT COMPE NSATION FOR CARBON OFFSET PRODUCTION IN FLORIDA: APPLICATION OF BEST WORST CHOICE MODELING AND DI SCRETE CHOICE EXPERI MEN T ATION The study of institutional, cultural, or regional preferences for forest offset markets can shed light on potential barriers to p articipate in valuable efforts to mitigate the threat of climate change. This chapter characterizes the guiding structure of car bon production in Florida by i dentifying barriers to participate in a hypothetical carbon of fset program, and by estimating land owner willingness to accept (WTA) compensation for enrolment. In December 2011, 310 Florida Forest Stewardship Program (FSP) affiliates (FSP participants and Tree Farm members) responded to an electronically administered conjoint choice survey of hypotheti cal carbon offset programs. This study finds evidence that landowners in this sample are willing to engage in carbon offset markets. The survey also exhibits important preferences for various institutional components. These results can be useful to policy makers interested in regional landowner preferences of carbon offset markets in the Southeast. Background Florida landholders have three major national options to engage forest carbon marke ts: Climate Action Reserve (CAR), American Carbon Registry (ACR) a nd Voluntary Carbon Standard (VCS) (Chapter 2) These are non profit carbon offset certification programs that slightly differ in protocol requirements, but encompass similar types of forest offset activities. The programs have commitme nt periods that rang e from 2 0 to 100 years (ACR 40, VCS 20 100, CAR 100), and compensations range from $2.50 to $30 per ton of carbon dioxide equivalent (see Charnley et al., 2010). Risk from intentional or unintentional (i.e. natural disaster) reversals is managed by institu ting a
47 series of accountability measures, such as, allowing participants to propose insurance products (ACR), carbon buffer pools (ACR, VCS, CAR), and in some cases a buy out option (ACR). eversals among all registered producers, similar to insurance. They work by allowing project managers to deposit a percentage of offsets (similar to insurance premiums) into an account controlled and managed by the program. The pool of offsets is used to c over carbon losses from unex pected events, such as wildfires, or hurricanes (American Carbon Registry, 2010). A number of recent empirical studies have explored some of the institutional aspects of carbo n markets in North America ( Table 3 1). In the abse nce of a national (or regionally present) carbon offset market that provides analysts with observations of indirect market transactions, the studies seen in Table 3 1 have used hedonic analysis to look for evidence of landowner WTA compensation to produce forest carbon offsets. The State of the Voluntary Market report ( Peters Stanley and Hamilton, 2012) is the closest publication of voluntary carbon offset market data in North America. The report consists of a survey to carbon offset brokers, and certificat ion programs to assess market demand and prices. This annual report provides valuable aggregate secondary price data, but it lacks the specific (and regional) focus needed for regional carbon market studies.
48 Table 3 1. E mpirical studies on willingness t o participate in hypothetical carbon o ffset m arkets in North America Reference Data Attribute Levels Markowski Lindsay et al., 2011 (Ratings: 1 5) Li, 2010 (Random Utility Model) Fletcher et al., 2009 (Ratings: 1 10) Shaikh et al., 2007 (RUM) van Kooten et al., 2002 (Ratings: 1 3) Massachusetts family forest owners (n=930) Texas non industrial private forest landowners (n=1,032) Massachusetts non industrial private forest landowners (n=17) Canadian landowners (n=260) Canadian farmers (n=182) Revenue Management Plan Enrolled acreage Time Additionality Implementer Withdrawal Penalty Time Revenue Eligibility Revenue Withdrawal Penalty Revenue Time Revenue $10, $100, $1000 Require/ not required 50 or 100% 15 or 30 years Require d/not required Private/Public sector No penalty or earnings plus 20% fee 1, 5years, or conservation easement status $2 to $42 acre/ year Formal Plan, No Plan $5, $15, or $30 acre/ year No Penalty, $10 per acre Bids ( $1 60 acre per year ) 10 years In a 2000 survey of 2,000 randomly selected farmers of northeastern British Colombia, Alberta, Saskatchewan, and Manitoba, Shaikh et al. (2007), used a random utility model to elicit WTA bids of a hypothetical Western Canadian car bon program to afforest marginal agricultural land. The bids offered participants a tree planting program with a 10 year duration, no monitoring, establishment, or management costs, annual compensations that ranged from $1 to $ 60 acre per year (bid levels were selected from a pilot study), and the option for ownership of all trees after the end of the program.
49 accepting the bid, with 13% of surveys fully completed (260 observati ons). Price (bid) was the only varying factor in their surveys, which were randomly sent to different participants. The study illustrated various demographic, cultural, and social factors (soil, education, visual landscape appeal, etc.) influencing partici pation, and estimated WTA bids, but only within confine s of the particular institutional structure mentioned above (to see the exact wording of this hypothetical question, please see the Appendix section of Shaikh et al. (2007)). The average WTA estimates to get farmers to plant blocks of trees was $33.59/acre. V an Kooten et al. (2002) also sampled Canadian farmers in 2000 to elicit ratings (1 3) with a question asking if respondents have ever considered large scale tree planning for level observations such as age, experience with contracts that restrict land use, etc. The Fletcher et al. (2009) did a similar study in 2007, usi ng a pilot survey of 17 private landowners from Massachusetts (randomly selected from a list of landowners who owned 3 or more parcels, each participant was compensated with $50) to also elicit the likelihood of producing carbon offsets. Participants were surveyed on socioeconomic questions management activities, reason s for owning land, but also asked to rate (1 10, 10 being the better option) six alternative carbon credit programs with four varying institutional attributes: eligibility (formal managemen t plan or no plan), time commitment (5 or 10 years), expected payment ($5, $15, or $30) acre per year, and penalty for withdrawal (none or $10 per acre). All options required project verification by a
50 professional forester. Their results using a Tobit mode l indicate that ratings increase with expected payment and commitment length, but decrease with penalty for withdrawal. Logit estimates of WTA were about 5% with $15, 13% at $30, and 33% at $50. While this study was limited by its pilot study nature, it in novated carbon market research by exploring WTA in the context of different institutional arrangements. This work by Fletcher et al. (2009) was followed by Markowski Lindsay et al. (2011) with a 2010 Massachusetts family forest owners survey administered t o 930 participants. The results yielded 402 observations of attribute level ratings (1 5) of carbon sequestration programs. The program attributes included were: management plan (required/not required), contract length (15 or 30 years), percent of land req uired to enroll (50 or 100%), revenue ($10, $100, $1000), additionality, penalty for early withdrawal (no penalty or repay earnings plus 20% fee), and institutional trust (implemented by public or private sector). The results from a random effects order pr obit found significant preferences for programs with higher net revenue, no withdrawal requirement that forests must be managed to sequester more carbon than if nothing was done for a similar analysis) The Texas Forest Service conducted a similar survey in 2009 (Li, 2010), of non industrial Texas landowners, exploring WTA at different levels of contractual duration. Participants of this study (20% survey response rate, which resulted in 1,032 onsider
51 hypothetical carbon program was presented, otherwise, they skipped this part, and were taken to a section where they rated factors that would prevent them from s elling environmental credits. The hypothetical program consisted of a contract with three different time commitment levels, each with a different annual per acre compensation (annual at $8, 5 years at $9, and conservation easement status for $10), to sell environmental credits, with an option for timber harvesting, as long as it generated additi onal credits ( Appendix A of Li (2009) for exact survey form). Factors affecting participation wer e analyzed using a Logit model. Additional questions in this study i ncluded awareness of carbon credits, size of forest landownership, current cost share participation, and importance of managing forestland for producing income WTA was estimated using the Contingent Valuation (CV) method. Methodologically, Shaikh et al. (2007) and Li (2010) are the closest related to this study, by postulating a hypothetical carbon market scenario, and asking respondents to either accept or reject it ( Figure 2 2). The ratings method used by Fletc her et al. (2009) is similar to the use of BWC in this study in the sense that both are ordinal in nature, but differ in terms of choice tas k ( Figure 2 2). In addition to using BWC, I used a discrete ch oice experimentation (DCE) conjoint task to of WTA used in applied economics, and to increase robustness by assessing attitudinal measurements with multiple statistical models. The m ajority of attributes in Table 3 1 w ere explored in this paper (Table 3 2 ), but using levels that closely resemble the requirements of curre nt available carbon cer tification programs for Florida landowners ( i.e. CAR, VCS, ACR)
52 are landowners who already have something akin to this requiremen t via the participation in Florida Stewardship or Tree Farm Programs. Other contractual were not included in this study to emulate existing certification programs (CAR, VCS, ACR do not require 100% enrolled acres), or to avoid survey complexity, respectively. governmental carbon offset program) was explores in this study, but not included in the conjoint choice questions. This study contributes to the literature (Table 3 1) by examining a new attribute, were asked to choose isk pool or insurance. We also extend the range of commitment periods explored in Table 1 1, by using a 5 to 100 years, which also simulates market realism. Statistical Models To choose the most appropriate discrete choic e experiment model to estimate the traditional marginal WTA values, Louviere e t al. (2000) advise researchers to consider the following design objectives: identification, precision, cognitive complexity, and market realism. Accordingly, this study simula te s market realism by using BWC (Figure 2 2), which uses a combination of BWS and discrete choice experimentation ( e.g. Coas t et al., 2006), and a DCE ( Figure 2 1) with multiple profile options (e.g. Lusk and Parker, 2009). Also BWC produces binary choice da ta (Binary hereafter) that I interpret as discrete choice e xperimentation type to estimate WTA
53 The BWS was recently applied to examine preferences in forest management stakeh As noted in Ohler et al. (2000), are binary (e.g., category or brand consideration, buy now or wait, etc), and binary tasks are Yet in regards to market realism, the current situation of carbon markets in the US have multiple certification options ( Chapter 2 ) that vary in terms of duratio n, risk options, etc. So, existing options might be better simulated by a DCE which offers multiple options for direct tradeoffs of entire attribute profiles For example, in the case of time commitments (up to 100 years), a landowner in Florida seeking carbon of fset certification via one of the currently available platforms would likely research or hire a consultant that will offer a cross comparison analysis of ACR, CAR, VCS and/or other program protocols, offering its customer multiple options with varying con tractual length Landowners would either choose one or none (maintain the status quo). DCE closely mimics this task by offering two or more options to participants, who are then instructed to select one or none; whereas B inary presents conjoint options, wh ich elicits an acceptance or rejection option It may be argued that a national carbon market program would homogenize all options by offering a single protocol, but even in that scenario, this analysis might still hold market realism, given that there may be other options to certify with international (e.g. Clean Development Mechanism), state (e.g. California AB 32), or private programs (e.g. Chicago Climate Exchange). In any case, the application
54 of both choice elicitation methods, BWC and DCE, will add r obustness to this study by exploring attribute tradeoffs in the context of multiple realistic market scenarios. Figure 3 1. Example of d iscrete c hoice e xperimentation q uestion p resented to survey r espondents Figure 3 2. Example of best worst choice q uestion, used to estimate best w orst s caling and b inary m odels Best Worst Choice For the past two decades, economists working on environmental issues have been using methods known as stated preferences, conjoint analysis, and attribute based methods. These non market valuations techniques typically require participants
55 to rank, choose, or rate a particular scenarios of attributes on a given scale (e.g. Foster and Mourato, 2002; Elrod et al., 1992; Fletcher et al., 2009). A relatively new innovation in scali ng methods (best worst) introduced by Finn and Louviere (1992), is currently gaining popularity in the fields of marketing business, health, and applied economics (e.g. Marley et al., 2008; Flynn et al., 2008; Lusk and Briggeman, 2009). The approach consis ts of creating profiles of different attribute levels, and asking participants to tool measures the maximum difference (maxdiff) between attribute levels under a common utility scale while offering an alternative over some of the shortcomings of the previously mentioned methods. Best Worst Choice was first implemented by Coast et al. (2006), and is one of the most recent innovations in the field of BWS. This model can be interpreted as a single profile choice model, and it works by constraining all attributes to have a represented level in each profile, and to include a second instruction, asking participants .g. Coast et al., 2006). The characterization of properties for BWC were derived by Marley et al. (2008), where they also propose an empirical design that may allow for the separation of Equation 3 1 is from Marley et al. (2 008), and represents a paired estimation of Z i and Z j are the chosen best worst pair, and Z k and can by any other pair within a set of pairs M. The influence of the judgment can show up in the utilit y value (this is the typical measurement estimated in DCE and Binary models), the weight or both.
56 BW z (Z i ,Z j ) = (3 1) Marley et al. (2008) propose the use of alternating BWC instructions, which alter BWC answers and allow for the Binary and BWS components of BWC to estimate various combinations of and that are used to effectively separates utility and importance. For example, if a survey is eliciting responses for trips to Mexico, in question 1, the participant may be asked to answer BWC instructions (Figure 3 2) assuming that the trip is for business, and in question 2 to answer assuming the trip is for personal pleasure. This allows for BWC answers to depend on instructions, and have two sets of BWS and Binary measurements that are used to solve for which is the overall estimate of a typical BW S (Equation 3 2). BW z (Z i ,Z j ) = (3 2) This study presents a BWC with consistent instructions, but the separation of these measurements is the subject of a follow up study using the estimates from this research, along with a subsequent survey to a randomly selected subpopulation of this study. version of the BWC conjoint cho ice task seen Figure 3 2, but with alternating instructions for each of the conjoint choice questions. Namely, question one will read,
57 non governmental gover nmental subsequent questions will alternate between these two instructions. Each question will followed by a question of enrolling or not enrolling in the carbon credit program (Figure 3 2). This process will assume that their responses will depend on program provider (non governmental vs. governmental), which allows for the separation of measurements (see Marley et al., 2008). The BW S estimations of this study are based on Equation 3 2, which confounds weights of importance and utility measurements in which is used to estimate the parameters in the equation below: (3 3) ( lays on latent utility scale) of a best and worst pair in choice question i (i=1,2,..,12), and the 16 (4 attribute impact variable a nd 12 level scale values) independent variables (Table 3 2). Each attribute has one impact variable and one level scale value for each level. BWS allows for a separate parameter estimation of the over all average impact of an attribute ( ) and the attribute level ( ) scale value (see Flynn et al., 2007). Hence, for choice i, the attribute chosen as best had its impact variable ( ) taking a value of 1, and the worst choice taking a 1, with th e rest of the impact variables taking values of 0. The rest of the level scale values are effects coded. (See guidelines for paired estimations in Flynn et al., 2007)
58 Table 3 2. Attributes and attribute variable levels in discrete choice experimentation a nd best worst c hoice Attribute Definition Levels Risk Tool Penalty Time Revenue Options for risk reduction in forest project Fines for leaving the program early Commitment period Carbon credit payments of acre per year, after costs Insurance Ris k Pool No Penalty Penalty 5 years 10 years 40 years 100 years $5 $10 $20 $30 As seen in Figure 3 2, BWC attempts to capture BWS behavior as well as binary choice data of an entire profile of attributes; we can observe utility values from the latter, and BWS weights from the former. The utility component of this model is of particular importance to applied economists, given that it allows for the conventional estimation of one the most widely used measurements, WTP or WTA. This method is estimated with two models, namely, the first task (BWS from now the tools of BWS, and the second task (Binary hereafter in the carbon credit program can be estimated using a binary logit, or random effects logit (e.g. Coast et al., 2006). There are multiple ways to estimate BWS, but these worst pairs, and se attribute level observations (see Flynn et al., 2006). The latter is an approximation of the former, and it may lead to larger standard errors (see Flynn et al., 2006). This study uses Paired estimation for BWS and random effects logit for Binary. An or thogonal main effects plan (OMEP) taken from Table 9 (Figure B 3) of
59 Street et al. (2005) was used to construct this survey, which resulted in 16 BWC scenarios (see Flynn et al., 2006). Discrete Choice Experimentation Discrete Choice Experimentation was f irst introduced by Mc Fadden (1974), and it is among the most widely used methods of stated preference elicitation (e.g., Lusk and Parker, 2009; Lancsar et al., 2007). Multiple fields, and a significant amount of studies have vetted this model, and when com pared with real choices, the model has been found to estimate results with a high degree of preference regularity and accuracy (see Louviere et al. (2000) for a literature review of comparative studies). A number of comparative studies of conjoint models have used it for external validation (e.g., Louviere and Islam 2008; Elrod et al., 1992). DCE models consist of several choice sets, each containing two or more options, where participants are asked to choose one (see Louviere et al., 2000). This study a 1 00% D efficient design to construct choice sets of size thre option. F ollowing the guidelines of Strategy software package SAS to generate a starting OMEP, used to create the first choice of eac h question set, followed by systematic changes to the levels of attributes to create the remaining choices. Given that Street et al. (2005) had created an optimal design with the attributes and levels of desired for this study, the choice sets were created using Table 9 from their study (Figure B 3). This process resulted in 16 questions, each with four choices This objectives of this chapter are to 1) use best worst choice and discrete choice experimentation to test the hypotheses that several factors do not affect willingness to accept payme nts for carbon sequestration: (a ) carbon contract revenue
60 (b) institutional factors (risk tool, contract length, penalty for early withdrawal), and (c) respondent characteristics; 2) to contribute to the literature of forest carbon market research (Table 4 1), by implementing the first regional study of forest carbon markets in Florida using hedonic analysis of non industrial private forest landowners; and 3) to contribute to the literature stated preference methods in applied economics and forestry, by implementing/introducing best worst choice in the context of natural resource economics. Methods S urvey Instruments The electronic surve ys were administered to participants of the Florida Stewardship Program 1 (FSP). On e of the most comprehensive non industrial private landowner lists in the state (see Pancholy et al., 2009). This subpopulation (over 900 members) of Florida landowners is an ideal representation of private managers, for they are highly motivated, versed i n forest management practices and barriers, organized into a reliable exte nsion network, and likely be the group who would seriously consider participating in a forest carbon offset market Following the guidelines of Dillman et al. (2009), some key questi ons from the National Woodland Owners Survey 2 (NWOS) were ad a pted, to better compare the characteristics of the subpopulation of forest landowners in this study to those of the most representative statewide forest landowners survey. The NWOS is carried ou t as part of the USDA, with the goal of characterizing the private forest landowners of the U.S. They also follow ed Dillman (2001), and randomly select a portion (10 20%) of the 1 http://www.sfrc.ufl.edu/Extension/florida_forestry_information/additional_pages/for est_stewardship_program .html 2 http://www.fia.fs.fed.us/nwos/quest/
61 full sample of private owners in ea ch state, i ncluding forest industry compan ies, partnerships, tr ibes, families, and individuals (Butler and Leatherberry, 2004) Florida is currently being surveyed by NWOS. See Table 2 for a comparison of survey respondent demographics of this study with those of 2002 2006 NWOS survey. Coding The D CE and Binary models were coded using ef fects coding for all independent variables. The dependent variable for DCE took 1 if chosen, and 0 otherwise, and Binary took 1 if profile was accepted and 0 if not. Effects coding works by coding all attribute level s except one (base level), with a 0 if absent, 1 present, and 1 if the base level is present. The base level then corresponds to the negative sum of the estimates of all other levels of the given attribute. Following Flynn et al. (2006), paired estimatio n instructions, all attributes of BWS were also coded using effects coding. (Table 3 3 ) Table 3 3. Effects c oding Attribute Effects c oding Effects c oding Effects c oding Contract length 10 years 40 years 100 years 5 years 1 1 1 10 years 0 0 1 40 yea rs 0 1 0 100 years 1 0 0 Revenue $10 $20 $30 $5 acre per year 1 1 1 $10 acre per year 0 0 1 $20 acre per year 0 1 0 $30 acre per year 1 0 0 Pe nalty for early withdrawal No Penalty 1 Penalty 1 Risk t ool type In surance 1 Risk Pool 1
62 Estimation BWC was estimated with two models, BWS and Binary. T he data of the binary ical carbon offset program ( Figure 3 2) was estimated using ra ndom effects logit (REL) to The statistical tions that are assumed to be independently Bernoulli distributed (Rodriguez and Elo, 2003). e task of BWC (Figure 3 2), BWS, were estimated using paired estimation, which fits the data for use in statistical software STATA, under the conditional logit command, which allows for interaction of covariates (see Flynn et al., 2006). Given that the OMEP used for BWS was unbalanced namely, not all best worst pairs were equally available for selection (Figure B 3), th is study adjusted for this bias using the guidelines provid ed by Flynn et al. (2006). The adjustment is performed with frequency weights via a two step process: 1) divide the number of times each best worst pair was chosen by the number of times it was ava ilable to be in across all scenarios and individuals (availability total); 2) multiplying this string of numbers by o ne of the availability totals. The DCE model was estimated using a conditional logit model (Louviere et al., 2000), and adjusted to represe nt the population of landowners in Florida (e.g., Lusk and Parker, 2009). Namely, population weights were created using all the demographic variables in Table 3 4 by implementing iterative proportional fitting techniques. This practice is common in survey research and it works by forcing the sample proportions to match those of the population (N ational Woodland Ownership Survey in this case). The
63 NWOS is the most comprehensive survey of non industrial private landowners in Florida. An alternative specific constant (ASC) was included in the model, which is a Adamowicz et al., 1998). Table 3 4. Characteristics of discrete choice experimentation, best worst choice, and nation al woodland ownership survey r espondents Category NWOS b (n=4900) BWC (n=93 ) DCE (n=85 ) Under 35 years 35 to 44 years 45 to 54 years 55 to 64 years 65 to 74 years 75+ years Less than 12th grade High school graduated or GED Some College Associate or techni cal degree Bachelor's degree Graduate degree Female Annual HH income less than $25,000 Annual HH income $25,000 to $49,999 Annual HH income $50,000 to $99,999 Annual HH income $100,000 to $199,999 Annual HH income $200,000 or more 1 9 acres 10 49 acres 50 99 acres 100 499 acres 500 999 acres 1000 4999 acres 5000+ acres Home <1 mile from their forestland NGO ( government carbon Scale of 1 to 5, 1 having the least importance) 1.65% a 7.10% 18.04% 28.69% 21.14% 19.29% 7.76% 16.08% 16.61% 9.76% 19.96% 23.18% 14.47% 10.90% 17.84% 22.84% 14.45% 18.71% 18.49% 22.71% 6.71% 33.53% 10.80% 3.94% 3.80% 47.14% 0.00% 3.37% 31.46% 31.46% 20.22% 8.99% 0.00% 2.25% 10.11% 12.36% 35.96% 35.96% 17.98% 2. 25% 11.24% 31.46% 20.22% 7.87% 3.82% 18.32% 17.56% 31.30% 8.40% 8.40% 12.21% 37.50% 3.13 c (1.37) d 1.11% 6.67% 21.11% 40.00% 24.44% 5.56% 2.22% 6.67% 6.67% 13.33% 40.00% 30.00% 15.91% 2.22% 10.00% 31.11% 24.44% 7.78% 3.85% 18.46% 18.46% 29.23% 10.77% 3.08 % 16.15% 42.53% a Percent of respondents falling in the respective category b National Woodland Owners Survey results for Florida c Mean d Standard deviation
64 Both DCE and Binary models were estimated using multiple specifications of both quantitativ and were quantitatively coded for some models to attain multiple estimations of marginal WTA. BWS Impact v alues are the average utility for the given attribute across all its levels, and scale values are the estimations of attribute level importance (relative importance of an attribute level as compared to other attribute levels on a common utility scale). In accordance with common practice, a full model was estimated to identify the attribute with the lowest impact (Time), in order to omit it from the final model, and use it as the the reference point. I intuitive understanding of the estimates. A negative sign on a coefficient does not imply a negative relationship with the dependent variable, but that it lays to the negative side of the reference case, under a common underlying scale. For example, the impact attribu which means that the average utility across all levels of this attribute is higher than the average utility across al Given that respondent level data does not vary for potential best worst pairs, these covariates were interacted with choice outcomes, in order to provide variation to individual characteristics (e.g. Flynn et al., 2008). The use of cova riates still allows for impact and scale values to be interpreted as averages across the entire sample. The interactions represent the additional utility that the particular demographic experiences for the given attribute or level. The model in Table 3 9 w as evaluated using other
65 insignificant in most interactions with choice outcomes, and thus excluded from the final represents an i nstitutional trust proxy have a non governmental carbon credit program rather than a governmental carbon = very unimportant, 2 = somew very important Choices and Variables The general guideline s in Chapter 9 of Louviere et al. (2000) for stated preference choice modeling were used to develop this choice tasks. The attributes and levels for this study were selected from qualitative research on features in currently available carbon offset programs (e.g., Climate Action Reserve, American Carbon Registry, etc.), similar studies ( e.g., Fletcher et al., 2009), and nine phone interviews with FSP members during the Summer of 2011 A similar approach was taken to select demographic and personal questions. These efforts were followed by the implementation of a pilot data collection inst rument that was tested for performance and accuracy with 24 participants with forestry, policy, and survey design backgrounds The final survey was electronically implemented following the format and procedural recommendations of Dillman et al. (2009). R esults and Discussion A total of 920 surveys were administered and 310 responded, for a 3 4% response rate which is fairly high according to the studies from Table 3 1 (13% of surveys were completed in Shaikh et al., 2007; 20% in Li, 2010; and 43% in Marko wski
66 Lindsay et al., 2011) After accounting for filter questions, and participants who did not provide full answers to all demographic questions (with the exception of location questions, such as zip code and address of largest forest plot) the sample si ze was reduced to 178 (93 with a BWC task, and 85 with DCE). All WTA estimates for DCE and coefficient to the price coefficient, with units of US dollars per choice. Table 3 5 presents the results of three estimations of DCE u sing a statistical logit model. Model 1 uses effects coding for all variables, The last two columns disp lay marginal WTA estimates for M odels 2 and 3. The results in Table 3 5 are very consistent across models in terms of significa nce, expected sign, and magnitude insignificant which implies that respondents had no significant preference for either risk pool or insurance This attributes in these hypot hetical programs significant at a 1% level for all models, and elicits the second highest increase in WTA for Model 2 and the highest for Model 3. This indicates that the would require an increase of $6.46 acre per year in compensation for Model 2 and $ 6. 4 4 in Model 3. Model 1 shows an in significant relationship of the lower two levels of compensation ($5, $20, and $10), and pos itive and significant for the $30 acr e per year. Models 1 and 2 reveal a significant (at a 5% level of significance) and positive (except for 5 year contract, which is .09) preference
67 for the lowest three levels of ), and negative for a ract years, then 40. The WTA estimates of Model 2 show the same relationship in terms of sign and magnitude. The inclusion of a 100 year commitment in a program is estimate d to increase $38.2 3 acre per year in compensation cost s. The inclusion of other time commitments would elicit a decrease in compensation costs of $17.95, $15.07, and $5.2 acre per year, for 5, 10, and 40 years contract respectively. Table 3 6 shows landowner WTA compensati on to switch between attribute Lancsar et al. (2007) to compare attribute impacts of various methods of estimating discrete choice experimentation model s The order is determ ined by the absolute value of column 2, which means that the higher the order of impact, the higher the absolute WTA difference between attribute levels. This table s hows that it would require a $43.43 acre per year compensation to have a landowner switch from a pro $9.87 acre per year it a decrea se in compensation cost of $12. 9 2 acre per year.
68 Table 3 5. Results from discrete c ho ice experimentation: conditional logit model e stimations Attribute Model 1 All effects coded Model 2 Revenue quantitative Model 3 Revenue & Time quantitat ive WTA Model 2 WTA Model 3 Insurance 0.09 (0.06) b 0.08 (0.06) 0.08 (0.06) $ 1.26 $ 1.17 Risk Pool 0.09 e 0.08 e 0.08 e $ 1.26 $ 1.17 No Penalty 0.35*** a (0.07) 0.42* (0.07) 0.42* (0.07) $ 6.46 $ 6.44 Penalty 0.35 e 0.42 e 0.42 e $ 6.46 $ 6.44 Revenue Q uantitative 0.06* (0.01) 0.06* (0.01) Time Quantitative 0.03* (0.00) $ 0.47 $5 acre per year 0.09 e $10 acre per year 0.05 (0.13) $20 acre per year 0.13 (0.13) $30 acre per year 1* (0.1) 5 year contract 0.09 e 1.17 e $ 17.95 10 year contract 0.94* (0.17) 0.98* (0.17) $ 15.07 40 year contract 0.36** (0.18) 0.34* (0.18) $ 5.2 100 year contract 2.44* (0.42) 2.48* (0.42) $ 38.23 ASC 3.33* (0.16) 4.35* (0.21) 3.02* (0.17) Number of Respondents 85 85 85 Number of Choices 5440 5440 5440 Log Likelihood 873.74 8788.06 880.66 Chi Square Statistic c 2156.33 2147.69 2142.48 a One (*), two (**), and (***) asterisk represent 0.01, 0.05 0.10 level of statistical significance, respectively. b Number i n parentheses are standard errors. c Chi square statistic associated with a test of the hypothesis that all model parameters are zero. e Effects coding: negative sum of the above level scale values corresponding to this attribute.
69 Table 3 6. D ifferences in marginal willingness to a ccep t ($/choice) for discrete choice experimentation (Model 2) e stimates Attribute Difference in WTA Absolute value Order of impact WTA to go from Insurance to Risk Pool $2.52 $2.52 1 WTA to go from No Penalty Penalty $12. 92 $12.92 4 WTA to go from a 5 to 10 year contract $2.88 $2.88 2 WTA to go from a 10 to 40 year contract $9.87 $9.87 3 WTA to go from a 40 to 100 year contract $43.43 $43.43 5 Table 3 7 presents the results of three Binary mode ls with the same effe cts coding used in DCE ( Table 3 3). These estimates were similar in terms of sign and significance to the results of DCE from Table 3 5, except for the variables $20 and $10 acres per ar The magnitudes of most coefficients in this table are slightly same interpreta tion given for Table 3.6 but the WTA was estimated higher for m odels 2 and 3, at $10.14 acre per year, and $9.00 acre per year respectively. The coefficient for Mo del 2, but slightly lower at $28.53 acre per year. The most preferred, or lowest WTA estimate ( $31.47 years of contract) whereas DCE estimated this attribute level to be the second lowest in terms of WTA.
70 Table 3 7. Results from binary choice model: r andom effects model e stimations Attribute Model 1 All effects coded Model 2 Revenue quantitative Model 3 Revenue & Time quantitative WTA Model 2 WTA Model 3 Insurance 0.00 (0.08) 0.00 (0.08) 0.05 (0.08 ) $0.02 $0.83 Risk Pool 0.00 e 0.00 e 0.05 e $0.02 $0.83 No Penalty 0.72* a (0.09) b 0.61* (0.08) 0.54* (0.08) $10.14 $9.00 Penalty 0.72 e 0.61 e 0.54 e $10.14 $0.50 Revenue Quantitative 0.06* (0.01) 0.06* (0.01) Time Quantitative 0.03* (0.00) $0.83 $5 acre per year 0.41 e $10 acre per year 1.3* (0.18) $20 acre per year 1.15* (0.15) $30 acre per year 0.56* (0.16) 5 year contract 2.05 e 1.89 e $31.47 10 year contract 0.89* (0.15) 0.75* (0.14) $12.48 40 year contract 1.02* (0.15) 0.93* (0.14) $15.42 100 year contract 1.92* (0.18) 1.71* (0.16) $28.53 Constant 0.23*** (0.39) 1.2* (0.38) 0.2 (0.36) Number of Respondents 97 97 97 Number of Choices 1552 1552 1552 Log Likelihood 609.24 639.29 669.39 Chi Square Statistic c 225.68 226.43 199.82 a One (*), two (**), and (***) asterisk represent 0.01, 0.05 0.10 level of statistical significance, respectively. b Number in parentheses are standard errors. c Chi square statistic associated with a test of the hypothesis that all model parameters are zero. e Effects coding: negative sum of the above level scale values corresponding to this attribute
71 Table 3 8. Differences in marginal willingness to a ccept ( $/choice) for binary (Model 2) e stimates Attrib ute Difference in WTA Absolute value Order of impact WTA to go from Insurance to Risk Pool $0.04 $0.04 1 WTA to go from No Penalty to Penalty $20.28 $20.28 4 WTA to go from a 5 to 10 year contract $18.99 $18.99 3 WTA to go from a 10 to 40 year contrac t $27.90 $27.90 5 WTA to go from a 40 to 100 year contract $13.11 $13.11 2 Table 3 8 presents Binary Model 2, seen in Table 3 5. These estimates significantly different than those of DCE. The magnitudes are higher, and the order of impact is different. For these esti mates, it would require a $13.11 acre per year compensation to have a landowner eas DCE estimates thi s to be $43.43 acre per year. The most noticeable change came from ear $9.87 acre per year in DCE, to $27.9 acre per year. This model indicates that elicit a decrease in compensation cost of $20.82 acre per year. Figure 3 3. Willingness to accept ($/choice): discrete choice experimentation (Model2) vs. binary (Model2)
72 Figure 3 4 Willingness to a ccept ($/choice): d iscrete c hoice e xperimentation (Model3) vs. b inary (Model3) Figure 3 were available for selection
73 Figure 3 attribute choices divided by the number of times they were available for selection All o f the choice outcomes in Table 3 9 are significant to a 1% level. Ten covariate interactions were insignificant to a 10% level and five were significant at a 5% level The type of risk tool is significant to a 1% level, and it indicates a preference for important attribute, followed by less important for Withdrawal was the least important The level scale values (average effect of an attribute level) indicate preferences similar to those found in the previous models (DCE and Binary) That is, (the level scale value) is less important and the higher per per are more important than the lower two more important than the higher two (40 and 100 years).
74 Table 3 9. Re sults from best worst scaling: clogit estimates adjusting for c ovariates Attribute Impacts Coefficient Rank Time 0 (reference) 1 Risk Tool 0.78* (0.04) 3 Penalty 0.38* (0.04) 2 Revenue 3.62* (0.05) 4 Age Risk Tool 0.12* (0.01) Age Penalty 0 .1* (0.01) Age Revenue 0.08* (0.01) Education Risk Tool 0.34* (0.01) Education Penalty 0.14* (0.01) Education Revenue 0.36* (0.01) Male Risk Tool 0.26* (0.01) Male Penalty 0.29* (0.01) Male Revenue 0.2* (0.01) Income R isk Tool 0.05* (0 .00 ) Income Penalty 0.04* (0 .00 ) Income Revenue 0.08* (0 .00 ) Acres Risk Tool 0.0003358* (0 .00 ) Acres Penalty 0.0002792* (0 .00 ) Acres Revenue 0.0000956* (0 .00 ) NGO Risk Tool 0.01* (0.01) NGO Penal ty 0.02* (0.01) NGO Revenue 0.25* (0.01) Level Scale Values Coefficient Insurance 0.19* (0.03) Risk Pool 0.19* (0 .00 ) No Penalty 1.31* (0.03) Penalty 1.31* (0 .00 ) $5 acre per year 0.04* (0 .00 ) $10 acre per year 0.53* (0.0 7) $20 acre per year 0.32* (0.07) $30 acre per year 0.25* (0.07) 5 year contract 0.94* (0 .00 ) 10 year contract 0.92* (0.06) 40 year contract 0.96* (0.06) 100 year contract 0.19* (0.03) a One (*), two (**), and (***) asterisk represe nt 0.01, 0.05, 0.10 level of statistical significance, respectively. b Number in parentheses are standard errors. c Chi square statistic associated with a test of the hypothesis that all model parameters are zero. e Effects coding: negative sum of the abo ve level scale values corresponding to this attribute.
75 Table 3 9. Continued Attribute Impacts Coefficient Age Insurance 0.06* (0 .00 ) No Penalty 0.08* (0 .00 ) Age $10 acre per year 0.01* (0.01) Age $20 acre per year 0.05* (0.01) Age $30 acre per year 0.02* (0.01) Age 10 year contract 0.06 (0.01) Age 40 year contract 0.04* (0.01) Age 100 year contract 0.11* (0.01) Education Insurance 0.07* (0.01) Education No Penalty 0.03* (0.01) Education $10 acre p er year 0.0033933* (0.01) Education $20 acre per year 0.04 (0.01) Education $30 acre per year 0.08* (0.01) Education 10 year contract 0.14* (0.01) Education 40 year contract 0.05* (0.01) Education 100 year contract 0.23* (0.01) Male Insurance 0.09* (0.01) Male No Penalty 0.07* (0.01) Male $10 acre per year 0.08* (0.02) Male $20 acre per year 0.05* (0.02) Male $30 acre per year 0.07* (0.02) Male 10 year contract 0.15* (0.02) Male 40 year cont ract 0.1* (0.02) Male 100 year contract 0.4* (0.02) Income Insurance 0.0001787 (0 .00 ) Income No Penalty 0.03* (0 .00 ) Income $10 acre per year 0.01 (0.01) Income $20 acre per year 0.02 (0.01) Income $30 acre per year 0.03* (0.01) Income 10 year contract 0.08* (0.01) Income 40 year contract 0.01 (0.01) Income 100 year contract 0.05* (0.01) Acres Insurance 0.00000583* (0 .00 ) Acres No Penalty 0.00000829* (0 .00 ) Acres $10 acre per year 0.0000217 (0 .00 ) Acres $20 acre per year 0.0000375* (0 .00 ) Acres $30 acre per year 0.00000699 (0 .00 ) a One (*), two (**), and (***) asterisk represent 0.01, 0. 0 5, 0.10 level of statistical significance, respectively. b Number in parentheses are stan dard errors. c Chi square statistic associated with a test of the hypothesis that all model parameters are zero. e Effects coding: negative sum of the above level scale values corresponding to this attribut e
76 Table 3 9. Continued Attribute Impacts Coeffic ient Acres 10 year contract 0.0000266* (0 .00 ) Acres 40 year contract 0.0001* ( 0.00 ) Acres 100 year contract 0.0001* (0 .00 ) NGO Insurance 0.001289 6 (0 .00 ) NGO No Penalty 0.07* (0 .00 ) NGO $10 acre per year 0.05* (0.01) NGO $20 acre per year 0.02 (0.01) NGO $30 acre per year 0.08* (0.01) NGO 10 year contract 0.01 (0.01) NGO 40 year contract 0.18* (0.01) NGO 100 year contract 0.01 (0.01) Number of Respondents 93 Number of Choices 17856 Log Likel ihood 197552.99 Chi Square Statistic c 156458.72 a One (*), two (**), and (***) asterisk represent 0.01, 0. 0 5, 0.10 level of statistical significance, respectively. b Number in parentheses are standard errors. c Chi square statistic associated with a test of the hypothesis that all model parameters are zero. e Effects coding: negative sum of the above level scale values corresponding to this attribute Table 3 highly significant (at per per whereas the relative difference between $20 and $10 is .85 units. This implies that upper two level the results of the previous other two models. The three models agree that the lower two levels of compensation are the least preferred within this attribute. The highest importance was
77 The results of this study are similar to the general findings of attribute influence from the Massachusetts study b y Markowski Lindsay et al. (2011). Their study indicates that participants prefer higher net revenue, no penalty for withdrawal, and shorter contractual commitments. The range of the compensation values in their study ($10, $100, $1000) and its emphasis on supply analysis (willingness to participate) of specific carbon offset programs (amount of required enrolled acreage and whether the program was administered by the public or private sector) did not produce comparable WTA estimates. Their study analyses t he probability of participation of three carbon offset for withdrawal, and $10 acre per year compensation, is the only scenario that includes attribute levels similar this study (the other two scenarios included compensations of $100 and $1000 acres per year). The participation rates for this scenario are low (between 2% and 5%), which is relatively similar to the results from this study that finds a negative and significan t influence of a $10 acre per year compensation, but a positive and low WTA ( $5.2 acre per year) for a 40 year contract. Another Massachusetts study by Fletcher et al. (2009), finds average participation percentage estimates of 5% for $15 acre per year c ompensation, 13% for $30, and 33% for $50 which generally agree with the results of this study. Namely, this study finds negative and significant estimates for compensations that are less than $20 acre per year, and positive and significant for $20 or $30 The Texas study by Li (2 010 ) considered contract durations of 1 year, 5 years, conservation easement status WTA of $15.15 acre per year for 1 year, $19.92 acre per year for 5 years, and $27.36
78 acre pe r year for conservation easement. His estimates of WTA for a 5 year contract fall within range of WTA estimates ($17.95 to $31.47 acre per year) from my study. Shaikh et al. (2007) surveyed Canadian landowners and estimated an average WTA of $33.59 acre pe r year to get farme study finds lower WTA estimates ($12.48 or $15.07 acre per year) for a 10 year contractual commitment The difference in WTA estimates between my study and Shaikh et al. (2 007) may be due to the fact that my study did not include specific land management requirements for carbon market programs, or attributed to regional and cultural differences between Florida and Canadian landowner preferences. Alternative Model S pecificat ions The estimated parameter s of Binary ( Model 1 ) and DCE ( Model 1 ) were graphed against their attribute levels, and a non linear relationship was visually detected for the A quadratic specification of the quanti tatively 10% level in Binary, and in the case of DCE, the linear version of the variable (which was estimated along with the quadratic specification) lost significanc e at a 10% level. Non linear specifications are the source of future work, but I suspect that the problem 3 2) An effort was made to adjust the random effects logit estimation of Binary to represent the population of landowners in Florida, by using the same approach applied to the conditional logit estimation of DCE but the parameter estimations were drast ically changing in sign and significance. This is the subject of future research, but I suspect that using a random effects model logit, which adjusts for the clustering of individual
79 marginal totals of sample proportions to conform with population totals. As seen in Figure 3 3 and Figure 3 4 WTA estimates for both Binary and DCE models seem to produce similar results but the most drastic difference in WTA came 40 year contract seen in Figure 3 4, which has a relatively similar magnitude, but with opposite sign. The sign difference may be due to the tipping point in which Sum mary In this study I have used three different models to analyze the preferences of a subpopulation of Florida landowners towards different aspects of contemporary carbo n markets in North America. This study indicates that landowners are not very affected by use of risk pooling, where they deposit carbon credits to address project uncertainty, or insurance type th at is typically used in crops. Using measurements of marginal willingness to accept, this study estimates that including a penalty for early withd rawal in this type of program would increase the cost of participation by approximately $6.46 to $10. 1 4 acre per year. The effect of a program that offers compensations of $5 or $10 acre per year seems to have a ne gative or less desirable effect than a pro gram that offers $20 or $30. Landowners seem to prefer contract durations of 5 to 40 years, while strongly disfavoring a 100 year commitment. A program with a 100 year contract would elicit an increase in cost of participation of $28.53 to $37.78 acre per year, while a 10 year commitment woul d lower cost by $12.48 or $15.07 acre per year Overall, revenue appears to be the most highly valued aspect of the components of this study, followed by the type of risk tool used to manage uncertainty, which is slight ly more valued than having a penalty for early withdrawal. Contract length was the least valued aspect.
80 In addition to exploring the attitudes of landowners, this paper compared a relatively new innovation in best worst scaling that includes an additional task per question. It asks respondents to evaluate the options presented in each question as a single profile, and to reject or accept it. There is no publication to my knowledge in the field of applied economics using this model, and the potential of thi s method to accurately est imate wiliness to accept (or WTP ), as well as measuring attribute impact, makes it appealing to applied researchers interested in binary discrete choice experimentation. The r esults seem to indicate that the estimates of WTA from this model are higher in magnitude than those of the traditional discrete choice experimentation (with more choice options) but they generally correlate in sign and relative magnitude.
81 CHAPTER 4 C OMPARISON OF BEST WORST SCALING AND DISCRETE CHOICE EXPERIM ENTATIO N There is little consensus in definition or measurement of product attribute importance (Louviere and Islam, 2008). The increase in use of best worst scaling methodologies in the fields of Forestry, Healthcare, and Applied Economics (e.g. Loureiro and Arcos, 2012; Lusk and Parker, 2009) using various estimation methods (e.g. multinomial logit, random parameters logit, frequency scores), raises an empirical question of what is being measured and how it compares to well established or traditional meas urements of utility. This study uses discrete choice experimentation for external validity, and performs a cross validation/comparison of best worst choice. The latter model produces binary observations akin to DCE as well as BWS observations. Each of the se models measures utility in a different manner, namely BWC measures direct tradeoffs of attribute level differences, whereas DCE measures indirectly via the choosing of profiles composed of attribute levels. Yet, b oth models eli cit tradeoffs between attr ibute /features, which reflect real market decisions, but BWS is easier to implement (Louviere and Islam, 2008), and the same may be true of BWC. To the best of my knowledge, this study is the first comparative study of BWC and DCE. A recent study by Potogl ou et al. (2011) implemented an empirical comparison of BWS and DCE, and found evidence of estimates that reveal similar patterns in preferences. Their experiment did not include price (or cost), which facilitates estimation of WTA/WTP, thereby potentially cancelling unknown confounding factors of this latent class models (see Flynn et al., 2007). Louviere and Islam (2008) did a similar comparative study of BWS importance weights, and DCE willingness to pay (WTP)
82 BWS importance, and a binary response DCE model for external validity. Their results indicated low correlations between BWS estimates and marginal WTP. This chapter uses a different DCE model for external validity, which includes more choices, and is more common in the field of applied economics. The BWS estimation of BWC is also applied differently, using paired model s of conditional logit and random parameters, as well as the aggregate measure of frequency scores. However, the former allows for inferences about the effects of respondent level covariates (see Flynn et al., 2008). This chapter also compares the marginal WTA estimates of BWC Binary and DCE. The ter 3 is not address in this chapter but is the subject of upcoming research (as explained in Chapter 3, the additional survey with alternating instructions is required for this separation, which I am currently designing and exploring its feasibility) Ba ckground For the past two decades, economists working on environmental issues h ave been using conjoint analysis, and attribute based methods preferences Non market valuations are typically examined with these tools, which require participants to rank, choose or rate a particular scenario of attributes on a given scale (e.g. Foster and Mourato, 2002; Elrod et al., 1992; Fletcher et al., 2009). A relatively new innovation in scaling methods (best worst) introduced by Finn and Louvie re (1992), is currently gaining popularity in the fields of marketing business, forestry, health, and applied economics (e.g. Marley et al., 2008; Flynn et al., 2008; Lusk and Briggeman, 2009). The approach measures the maximum difference between attribute levels, while offering an alternative over some of the shortcomings of the previously mentioned methods.
83 BWS models offer participants a set of items or attribute levels and instruct them to choose the two items with the largest perceptual difference on a n underlying scale, with varying sets of three or more choices, until an experimentally designed number of choice sets are complete (for a complete background of BWS see Flynn et al., 2007). While comparing the advantages of BWS and ratings, Lusk and Briggeman (2009) explain that in rating scales, participants are not forced to make trade offs of relative importance, while in the real world, trade offs between choi ces are made on a daily basis. A nother deficiency of ratings comes from the notion that different people are likely to use the same scale differently, leaving the interpretation of ratings to an ordinal scale at best (Lusk and Bgriggeman, 2009) BWS minimizes th e chances of introducing assumptions about human decision making, by forcing respondents to consider only the extremes of the utility space ( Flynn et al., 2007 ). This chapter takes advantage of the well documented empirical record of DCEs, to "reasonably a ssume that DCE derived measures accurately reflect importance in assumptions and the nature of their estimates represent conditional demands, but BWS rage utility across levels) estimation of all attributes (except one) traditional accept/reject question along side BWS, the strengths of both can be drawn upon ( see Flynn et al., 2007). This approach is commonly used in the field of applied economics, given that it provides an option to estimate one of the most important measurements in their field, willingness to pay (e.g. Lusk and Parker, 2009). Another
84 method includes an additional instruction commonly provided by BWS, which asks Chapter 3 ) the particular profile of attribute levels (e.g. Coast et al., 2006, Marley et al., 2008). Binary answers to non market goo ds are common in the field of environmental economics, which allow for the estimation of random utility models (RUM) with referendum type questions (e.g. Hanemann, 1984; Shaikh et al., 2007). BWC elicits such responses, along side direct estimates of BWS. The DCE used for external validity in this study has also been shown to approximate RUM estimates (McFadden and Train, 2000) using mixed logit with approximate choice of variables and mixing distribution. This chapter also compares estimates DCE mixed logi t (RPL from now on) with measurements of BWS and Binary estimates contained in BWC. There is increase d importance in weights and WTA measures in the field of applied economics, as well as sizes for BW and evidence that increasing sample sizes of choice based conjoint experiments yield more accurate estimates o f WTA (Lusk and Norwood, 2005). So, this study aims to provide cross validation/comparison to the results of Louviere and Islam (2008), which suggest DCE measures can be systematically impacted by what is included/excluded, thus inference measures from this model will appear sensitive to context effects. By adding a price attribute, and controlling for other survey design ele ments, I also aim to augment the analysis of Potoglou et al. (2011). The inclusion of an accept/reject question to the BWS profiles will be the first empirical comparison of BWC to DCE.
85 Research Objective and Methodology BWS and The Effects of Ambiguity fo r Empirical Researchers As explained the previous chapter, BWS estimates of BWC produce measurements of using the following equation: BW z (Z i ,Z j ) = (4 1) Thus, estimates of the utility values of which may elicit an erroneous or ambiguous interpretation of the impa ct that a variable has on judgment. This entanglement cannot be separately identifiable from a single design involving a finite set of profiles (see Marley et al., 2008), and the issue of BWS estimation and interpretation becomes an empirical question in t he applied field. The aim of this chapter is to estimate BWS using some of the most widely used procedures, and to empirically compare these measurements with two types of discrete choice experimentation tools (DCE and Binary) that are common in the field of applied economics. A secondary goal is to compare the welfare estimates of BWC derived using Binary (Chapter 3) to examine the reliability of this technique in estimating WTA measurements. External Validity The use of DCE data to compare other limited d ependent variable estimations has several advantages, and it is the tool of choice for most empirical studies examining the performance of BWS (e.g. Potoglou et al., 2011; Louviere and Islam, 2008; Lancsar et al., 2007). According to Louviere and Islam (20 established
86 record of high correspondence of experimental measures with measures derived from hypothetical carbon markets in Florida, which is the subje ct of this study. Estimation and Research Design The attributes in Table 3 2 were arranged using an OMEP for both BWC and DCE survey designs (e.g. Coast et al., 2006), to control for potential structural design factors that may affect this empirical compa rison. The design from Figure B 3 was level attributes and two 4 level attributes. Options 2 and 3 of the OMEP were design by chapter. Please refer to Chapter 2 and Chapter 3 for a detailed description of the qualitative methods used to select the current survey attributes and participants of this chapter. The attribute levels of both sur veys use effects coding, which is recommended for BWS by Flynn et al. (2007), and optional for DCE (see Louviere et al., 2002). Other aspects of design, such as context dependent issues discussed in Louviere and Islam (2008) were addressed in equal manner for both surveys to control for the effects of ambiguity. (Appendix A includes examples of both surveys) All models were constructed using a 100% D efficient design. F ollowing the guidelines of Strategy kage SAS to generate a starting OMEP, used to create the first choice of each question set, followed by systematic changes to the levels of attributes to create the remaining choices. Given that Street et al. (2005) had created an optimal design with the a ttributes and levels of
87 desired for this study, the choice sets were created using Table 9 from their study (Figure B 3). This process resulted in 16 questions, each with four choices The design was not balanced, but the appropriate balancing techniques recommended for BWS (see Flynn et al., 2007) were applied to the appropriate designs to make up for the fact that attributes 1 and 2 (risk tool and penalty for withdrawal) from Option 1 in appeared with more times in attributes 2 and 3 (contract length and revenue). Best Worst Scaling The observations of direct attribute level tradeoffs from BWS can be estimated using a number of methods (see Flynn et al., 2007 for a detailed a description which, as the name suggests, use pairs of best and worst observations to make frequencies or individual selections to approximate the paired estima tions. I estimated two paired models (Conditional Logit and Random Parameters Logit) and one marginal method using sample level frequency scores of best and worst choices. The former approach is theoretically consistent with maximum distance choice models, but marginal estimations have been found to be in agreement with such paired models (see Flynn et al., 2008). Paired Conditional Logit For this model (BWS_clogit from now on) I follow the estimation specifications of Flynn et al. (2008), which allow for t he inclusion of covariates (Chapter 3) representing additional impact a given attribute had upon the particular landowner sociodemographic group. Given that marginal approaches are an approximation to the paired models, only
88 the latter was considered for c ovariate interactions. The results from the previous chapter were incorporated for this comparison, along with another specification excluding covariates (Table 4 2). Paired Random Parameters Logit The estimation of BWS from Chapter 3 assume that all indiv iduals in the sample place the same level of importance on each value, this assumption is relaxed by allowing importance parameter of best worst choice i of individual j to be specified as where and are the mean and standard deviations of in the population, and is a normally distributed random component with mean zero and unit standard deviation (Lusk and Briggeman, 2009). The use of a reference case (omitting one attribute level or impact variable) to be normalized to zero is also used in this model to both avoid colinearity in effects coding, and to serve as a benchmark for comparing all other estimates. All attribute lev els and impact variables were assumed to be independently normally distributed in the population. This approach will be referred to as BWS_RPL from now on. Frequency Counts Frequency Counts (BWS_Freq) is an approximation estimate of BWS, and it was used in a recent application of BWS to the field of forestry (see Loureiro and Arcos, 2012). Following the analysis of Louviere and Flynn (2010), these estimates were each of the 16 profiles in our design, and aggregating their frequencies. These least) and adjusted for the
89 by the for Discrete Choice Experimentation The results from Chapter 3 using conditional logit estimates of DCE were incorporated in this analysis, along with the random utility estimations using RPL (see Tr ain, 2003). All non price attributes in DCE using RPL (DCE_RPL) were assumed to be independently normally distributed, and the price coefficient was not allowed to vary in the population to make sure that WTA estimates are normally distributed (e.g. Lusk e t al., 2003; Train, 2003). The confidence intervals for WTA estimates of DCE_RPL were estimated using 10,000 repetitions of the Krinsky Robb method (Krinsky Robb, 1986). Other specifications such as adjusting the population proportions were similar to the ones applied to conditional logit estimations of DCE (DCE_clogit) from Chapter 3. WTA measurements were estimated for DCE_clogit, DCE_RPL, and Binary (Chapter 3) models. Louviere and Islam (2008) elaborate on the significant advantages using WTA for model comparison purposes. Attribute utilities in discrete choice models are measured on interval scales unique to each attribute and individual. These estimates of utilities are confounded by the magnitudes of estimated attribute parameters with error variances of individuals. WTA estimates cancel these effects by offering a common utility scale in terms of price/choice, which allows for the appropriate comparison of attributes. All DCE measurements of WTA (see Louviere et al., 2000, Chapter 12) were assumed to come from a linear indirect utility function. For example, a 2 level attribute, such as penalty for withdrawal would have a WTA given by the ratio of
90 where the units of are utility per level, and are utility per dollar. BWS has the advantage of producing level scale parameter estimates with a common utility scale. The purpose of estimating DCE using random parameters logit i s to make use of the WTA estimates to evaluate the performance of BWS, and to compare them to the Binary WTA estimates from Chapter 3. The same method used in Chapter 3 to adjust population proportions to match those of the National Woodland Ownership Surv ey results from Florida were used for these estimations. An alternative specific constant (ASC) was included in the model, which is a controls for (or represents) the util option. Adamowicz et al. (1998) utilized a similar ASC for a similar orthogonal desing using multiple conjoint estimations, and explain that a lack of inclusion of ASC may their study, they elaborate on the ambiguity surrounding the interpretation of this parameter, namely, if negative and significant, this may be insitutional trus t or understanding of carbon market programs), but it may also indicate a (although survey ins tructions explicitly asked respondents to consider each choice set in isolatin from the following or previous questions sets) behavioral interpretation.
91 Results The results from Table 4 1 were estimated using the same procedure as Chapter 3, but omitted t he interaction of covariates. All attribute impacts and scale parameters 5% level. Risk Tool was used as the reference case, and the most valued attribute was Reven ue, while Penalty the least valued. The attribute impacts produced a different order of rank than the one found in the previous chapter (the BWS clogit estimation with individual covariates from Chapter 3 will be referred to as BWS_clogit_demo from now on) adjusting for covariates, namely, most of the interaction of contract length wi th individual covariates found in Table 3 10 were significant, which suggests that the effects of this attribute on utility varies within subpopulations of participants. The rest of the estimates produce similar results to Chapter 3, in terms of a preferen ce for no penalty for withdrawal, the two higher levels of Revenue. The attribute level scale parameters of Risk Tool were significant in this model, and show a preference for the use of Risk Pool. Table 4 1. Results from best worst s ca ling: clogit estima tes Attribute Impacts Coefficient Rank Risk Tool 0 2 Time 0.44* (0.01) 3 Penalty 0.01** (0.01) 1 Revenue 1.89* (0.01) 4 a One (*), two (**), and (***) asterisk represent 0.01, 0.05, 0.10 level of statistical significance, respectively. b Number in parentheses are standard errors. c Chi square statistic associated with a test of the hypothesis that all model parameters are zero. e Effects coding: negative sum of the above level scale values corresponding to this attribute
92 Table 4 1. Continued L evel Scale Values Coefficient Insurance 0.05* (0.01) Risk Pool 0.05 No Penalty 0.32* (0.01) Penalty 0.32 $5 acre per year 0.39 $10 acre per year 0.16* (0.01) $20 acre per year 0.3* (0.01) $30 acre per year 0.24* (0.01) 5 year contract 0.11 10 year contract 0.04* (0.01) 40 year contract 0.21* (0.01) 100 year contract 0.36* (0.01) Number of Respondents 93 Number of Choices 17856 Log Likelihood 207100.15 Chi Square Statistic c 137364.39 a One (*), t wo (**), and (***) asterisk represent 0.01, 0.05, 0.10 level of statistical significance, respectively. b Number in parentheses are standard errors. c Chi square statistic associated with a test of the hypothesis that all model parameters are zero. e Effe cts coding: negative sum of the above level scale values corresponding to this attribute The estimation results of BWS_RPL are displayed in Table 4 2, which produce measurements of importance for level scale values and attribute impacts. These estimates a re relative to Risk Tool, and most mean estimates were significant at a 1% per impacts we re significant at a 1%, while none of the level scale values had a significant rank order was also differs in this model from the results of BWS_clogit and BWS_clogit_de mo. These results will be further analyzed in the subsequent comparison analysis section.
93 Table 4 2. Results from best worst scaling: r andom parameters logit model e stimations Attribute Impacts Coefficient Rank Share of preference Risk Tool Reference Case 0 1 9.07% Contract Length Mean St. dev. 0.61* (0.11) 0.97* (0.11) 2 16.65% Penalty for Withdrawal Mean St. dev. 0.97* (0.09) 0.64* (0.09) 3 24.05% Revenue Mean St. dev. 1.71* (0.08) 0.33* (0.09) 4 50.23% Level Scale Values Coefficient Insurance Mean St. dev. 0.03 (0.05) 0.09 (0.06) 6.56% Risk Pool Effects coded 0.03 7.00% No Penalty Mean St. dev. 0.2* (0.05) 0.01 (0.05) 8.29% Penalty Effects coded 0.2 5.54% $5 acre per year Effects coded 0.9 6 2.60% $10 acre per year Mean St. dev. 0.76* (0.09) 0.03 (0.07) 3.16% $20 acre per year Mean St. dev. 1.37* (0.12) 0 .07 (0.1) 26.78% $30 acre per year Mean St. dev. 0.35* (0.1) 0.04 (0.08) 9.61% 5 year contract Effects coded 0.18 5. 64% 10 year contract Mean St. dev. 0.2* (0.08) 0.02 (0.09) 5.52% a One (*), two (**), and (***) asterisk represent 0.01, 0.05 0.10 level of statistical significance, respectively. b Number in parentheses are standard errors. c Chi square statist ic associated with a test of the hypothesis that all model parameters are zero.
94 Table 4 2. Continued Level Scale Values Coefficient Share of preference 40 year contract Mean St. dev. 0.39* (0.09) 0.12 (0.09) 2.74% 100 year contract Mean St. dev. 0.77* (0.08) 0.04 (0.05) 8.76% Number of Respondents 93 Number of Choices 17856 Log Likelihood 2482.26 Chi Square Statistic c 228.06 a One (*), two (**), and (***) asterisk represent 0.01, 0.05 0.10 level of statistical significance, re spectively. b Number in parentheses are standard errors. c Chi square statistic associated with a test of the hypothesis that all model parameters are zero. e Effects coding: negative sum of the above level scale values corresponding to this attribute T he importance of the levels of contract length show a similar order of Table 4 3 shows another model specification for BWS_RPL excluding impact significant at 1%, but in contrast with the results of Tabl e 4 2, the standard deviation of this level was also significant at 10%. The mean was .28, and the standard deviation per
95 most important of this attribute. The most impor evaluating the importance of each value that results from MNL and RPL models is that the BWS estimations using multinomial logit and RPL remedy this problem by computing parameters into the logistical equation. (Column four of Table 4 3) The shares of preference were (which add up to one given the nature of the logistical equation) are presented Figure 4 1, and they show that on average, 13% of preferences were not estimated for BWS_clogit, because it was a conditional estimation of BWS. Table 4 3. Results from best worst scaling: r andom parameters logit model e stimations excluding impact variables Attribute Impacts Coefficient Share of Preference Insurance Reference Case 0 No Penalty Mean St. dev. 0.08** (0.04) 0.06 (0.1) 10.67% Penalty Effects coded 0. 08 9.05% $5 acre per year Effects coded 0.31 7.21% a One (*), two (**), and (***) asterisk represent 0.01, 0.05 0.10 level of statistical significance, respectively. b Number in parentheses are standard errors. c Chi square statistic associated wit h a test of the hypothesis that all model parameters are zero.
96 Table 4 3. Continued Attribute Impacts Coefficient Share of Preference Insurance Reference Case 0 No Penalty Mean St. dev. 0.08** (0.04) 0.06 (0.1) 10.67% Penalty Effects coded 0 .08 9.05% $5 acre per year Effects coded 0.31 7.21% $10 acre per year Mean St. dev. 0.10 (0.06) 0.00 (0.09) 8.88% $20 acre per year Mean St. dev. 0.26* (0.06) 0.00 (0.08) 12.81% $30 acre per year Mean St. dev. 0.15** (0.06) 0.00 (0.13) 11 .38% 5 year contract Effects coded 0.22 7.87% 10 year contract Mean St. dev. 0.09 (0.07) 0.01 (0.08) 8.98% 40 year contract Mean St. dev. 0.03 (0.07) 0.22** (0.12) 10.13% 10 0 year contract Mean St. dev. 0.28* (0.07) 0.21*** (0.12) 13.0 3% a One (*), two (**), and (***) asterisk represent 0.01, 0.05 0.10 level of statistical significance, respectively. b Number in parentheses are standard errors. c Chi square statistic associated with a test of the hypothesis that all model parameters are zero. The shares of preference were also estimated for BWS_RPL and presented in Table 4 2. Figure 4 2 uses estimates from the RPL model using impact weights. An advantage of BWS is that it allows for the separation of impact weights and scale values ( Flynn et al., 2007).
97 Figure 4 1. Relative desirability of carbon offset institutional elements: estimated with BWS_RPL excluding impact variables. Figure 4 2. Relative desirability of carbon offset institutional e lements: estimated with BWS_RPL
98 Fig ure 4 3. Relative desirability of carbon offset i nstitutional attributes : estimated with BWS_RPL In Figure 4 3 illustrate the share of prefer ence of Table 4 4 presents the results from a BWS Frequency Count model (BWS_Freq). Given that BWC elicits direct res elements of a carbon offset program via the use of BWS, these methods allows us to explore the data from this study without pairing the most and least responses. The procedure adjusts for the availability of individual attribute levels. The analysis of these frequency counts typically evaluates estimates by subgroups and categories (e.g., Loureiro and Arcos, 2012), but in this chapter I restrict the scope of this study to the aggregate population of respondents. Each respondent answered 16 questions and 97 participants completed the entire
99 attribute levels. Table 4 4. Results from b e st worst s ca ling: frequency counts Attribute level N a Most a Least a Most Least a Mean b Insurance 1504 63 258 195 0.02 (0.06) c Risk Pool 1504 81 269 188 0.02 (0.06) No Penalty 1504 157 217 60 0 .00 (0.06) Penalty 1504 132 251 119 0.01 (0.06) $5 acre/year 1504 156 46 110 0.02 (0.12) $10 acre/year 1504 179 40 139 0.02 (0.12) $20 acre/year 1504 228 22 206 0.03 (0.11) $30 acre/year 1504 220 24 196 0.03 (0.11) 5 years 1504 46 94 48 0.01 (0.12) 10 years 1504 46 89 43 0.01 (0.12) 40 years 1504 84 96 12 0 .00 (0.12) 100 years 1504 112 98 14 0.00 (0.12) a The number of times an attribute level was chosen across all choice sets and respondents. b Total most least counts divided by the availability of each principle (calculated as the number of times it appeared across the design, see Figure B 3 Option 1 in Appendix B for availability design). c Standard deviations from the adjusted mean. e Effects coding: negative sum of the above level scale values corresponding to this attribute Figure 4 4 i The analysis of this study focuses on the adjusted mean estimates from column 6 on Table 4 4, which are graphically displayed in Figure 4 5. Figure 4 4 shows the mean frequency scores before adjustment. From Figure 4 5 I see that the relative estimate of per per
100 Figure 4 4. Best levels of carbon program institution al components Figure 4 5. Best of carbon program institutional components
101 Figure 4 6. Best of carbon prog ram institutional components, divided by the number of times they were available Table 4 5 presents the results of DCE using random parameters model. As previously discussed, the model was specified to assume that all non price attibutes were independently normally distributed, and the revenue was not allowed to vary. The specification was similar to the logistical analysis of DCE in Chapter 3, but the model is respecified to allow for the coefficients to vary in th population rathan being fixed within the conditional logit (see Train, 1998). Fixing the price coefficient enables the estimation of WTA measurements that are ensured to be normally distributed, and it forces respondents to have a positve revenue coefficient, which could not happen under a normal quantitatively coded for all estimates, and two different specification are presented in Table 4 5. Model 1 in this table uses effects coding for all attribute levels (expect revenue), and Model 2 quantitatively codes Revenue and Contract Length (Time).
102 Table 4 5. Results from discrete c hoice experimentation model: r andom parameters logit model e stimations Attribute Specification Model 1 Revenue quantitative Model 2 Revenue & Tim e quantitative Revenue Quantitative Fixed 0.19* a (0.02) 0.18* (0.02) Insurance Mean St. dev. 0.01 (0.13) 0.54* (0.11) 0.51* (0.12) 0.66* (0.15) Risk Pool Effects coded 0.01 0.51 No Penalty Mean St. dev. 0.81* (0.17) 0.83* (0.24) 0.81* (0.28) 0.64* (0.3) Penalty Effects coded 0.81 0.81 Time Quantitative Mean St. dev. 0.12* (0.01) 0.08* (0.01) 5 year contract Effects coded 2.69 10 year contract Mean St. dev. 1.81* (0.25) 1.02* (0.23) 40 year contract Mean St. dev. 0.99* (0.3) 1.07* (0.13) 100 year contract Mean St. dev. 3.51* (0.51) 0.64* (0.26) ASC Mean St. dev. 22.47* (3.23) 17.21* (3.18) 10.84* (1.06) 11.88* (1.23) Number of Respondents 85 85 Number of Choices 5440 5440 Log Likelihood 305.79 328.40 Chi Square Statistic c 1144.54 1104.52 a One (*), two (**), and (***) asterisk represent 0.01, 0.05 0.10 level of statistical significance, respectively. b Number in parentheses are standard errors. c Chi square statist ic associated with a test of the hypothesis that all model parameters are zero.
103 The majority of estimates from Model 1 were significant at a 1% level, except for and th standard deviations from Model 2 were significant at a 1% level, execpt for the standard means o absence of a penalty for withdrawal would increase participation in carbon offset markets. The lower two levels of contract length from Model 1 have an association with participati on, while the lower two are negative at the mean values. The estimate of from an estimated mean of 0.12. The ASC paramter estimate was negative and significant to a 1% level. This may indicate the presense of the protest answers to this management o f their forest land, then ASC would have a clearer behavioral interpretation. Table 4 6. Mean willingness to a ccept ($/choice) e stimates for discrete choice experimentation: random parameters logit (Model 1) Attribute Mean WTA Median WTA [95% Confidence I nterval] Insurance 0.06 0.06 [ 1.41 1.29] b Risk Pool 0.06 0.06 No Penalty 4.27 4.24 [2.57 6.18] Penalty 4.27 4.24 5 year contract 14.24 14.04 10 year contract 9.53 9.46 [6.61 12.9] 40 year contract 5.18 5.13 [ 8.67 1.91] 100 year contract 18.59 18.36 [ 25.67 12.9] b Number in parentheses are confidence intervals, estimated using 10,000 repetitions of the Krinsky Robb Method
104 Table 4 7. Mean willingness to a ccept ($/choice) e stimates for discrete cho ice experimentation: random parameters logit (Model 2) Attribute Mean WTA Median WTA [95% Confidence Interval] Insurance 2.9 2.83 [ 5.34 0.91] Risk Pool 2.9 2.83 No Penalty 4.58 4.51 [3.69 5.86] Penalty 4.58 4.51 Time Quantitative 0.68 0.67 [ 2.15 0.71] b Number in parentheses are confidence intervals, estimated using 10,000 repetitions of the Krinsky Robb Method The confidence intervals from Table 4 6 and Table 4 7 were estimated using 10,000 repetitions of the Krinsky R in Table 4 6 indicates that the inclusion of this attribute is associated with a $18.59 acre ase in mean compensation needed for participation. The exclusion of penalty for withdrawal would decrease participation costs by an ($4.58). Table 4 7 estimates that t he overall effect of a one unit increase in contract length is an associated mean increase of $0.68 acre per year in enrollment costs. Table 4 8. Differences in mean willingness to a ccept ($/choice) for discrete choice experimentation: random parameters l ogit (Model 1) e stimates Attribute Difference in WTA Absolute value Order of impact WTA to go from Insurance to Risk Pool $0.12 $0.12 1 WTA to go from No Penalty Penalty $8.54 $8.54 3 WTA to go from a 5 to 10 year contract $4.71 $4.71 2 WTA to go fr om a 10 to 40 year contract $14.71 $14.71 5 WTA to go from a 40 to 100 year contract $13.41 $13.41 4 Table 4 8 details differences in mean WTA across attribute levels for Model 1 of DCE_RPL. These results indicate that it would take an average of $13.41 acre per year to have a landowner accept a change from a 40 year contract to a 100 year contract,
105 the biggest absolute difference between levels came from highest levels of contract length and the lowest (of the significant estimates) came from also from this attribute. Figure 4 7. Willingness to quantitatively coded Figures 4 8 compares WTA estimates of Binary and DCE_clogit from Chapter 3, with DCE_RPL and finds relative agreement from all models. For the most part, Binary appears to be over estimating WTA measurements in 3/8 attribute leve ls, while estimates of DCE_RPL seem to be in between the estimates of the other two models. The quantitatively specified DCE_RPL was compared in Figure 4 8, and finds similar results as Figure 4 7, but seems to have more agreement with DCE_clogit.
106 Figu re 4 8. Willingness to Comparisons of Importance Measures Figure 4 9 compares the estimates of BWS using Random Parameter Logit, Conditional Logit Model, and Adj usted Mean Frequency Counts (the upper horizontal axis corresponds to the adjusted Mean of Frequency Counts). This figure shows more agreement between estimations of BWS_clogit and BWS_RPL, and slightly less with BWS_freq. The latter model does not estimat e impact weights and therefore seems absent in the lower part of this Figure. There is strong disagreement among models overall did not carry a lot of significance. There is strong agreement with dominant importance of the highest two levels of revenue, and some disagreement between BWS_Freq and the other two models. There is general agreement of the relative
107 importance of the levels within the attribute contract len gth, which generally state that _clogit_demographics, which are compared in Figure 4 10. Figure 4 9. Best worst scaling model comparison: estimated using random parameter logit, conditional logit model, adjusted mean frequency counts (the upper horizontal axis corresponds to the adjus ted mean of frequency counts, and the lower to the paired BWS estimations of clogit and RPL)
108 Figure 4 10. Best worst scaling attribute impact estimate comparison of random parameter logit and conditional logit models Table 4 9. Order of differences b et ween best w orst scaling level scale attributes for BWS_clogit, BWS_RPL, and BWS_Freq (1 having the lowest distance) Level Scale Attribute Order of difference between levels (difference) BWS_RPL BWS_clogit BWS_Freq. Difference between Insurance to R isk Pool 6 5 1 Difference between No Penalty Penalty 1 6 6 Difference between $5 to $10 acre per year 7 1 5 Difference between $10 to $20 acre per year 3 7 8 Difference between $20 to $30 acre per year 2 3 3 Difference between 5 to 10 year contract 5 4 2 Difference between 10 to 40 year contract 8 8 7 Difference between 40 to 100 year contract 4 2 4 Table 4 9 presents the differences between BWS level scale values and orders them according to their magnitude (1 having the least magnitude). Result s show there is almost general agreement between the three models that the biggest difference 10 to 40 year contract important attribute with respect to the estimations of impact values. Figur e 4 11
109 graphically displays this ordinal comparison, and shows more agreement between BWS_RPL and BWS_clogit. Figure 4 11. Best worst scaling order of difference between attribute levels of random parameter logit and conditional logit models Table 4 10. Ordering of relative impact of attribute/scale level ranges across best worst scaling models estimated using random parameter logit, conditional logit, adjusted mean frequency counts (1 having the least impact) Level Scale Attribute Order of relative im pact (range) BWS_RPL BWS_clogit BWS_Freq Risk Tool 1 (0.07) 1 (0.1) 1 (0.00) Penalty 2 (0.4) 3 (0.63) 3 (0.00) Time 3 (1.16) 2 (0.57) 2 (0.01) Revenue 4 (2.33) 4 (0.69) 4 (0.02) Table 4 10 presents the o rdering of relative impact of attribute/s cale level ranges across methods BWS estimations (1 having the least impact) This table illustrates a general agreement between methods regarding the ranking of the attributes with the biggest and shortest ranges, but shows disagreement on the rankings in between. Figure 4 12 displays these ordinal rankings and shows relative agreement between the three models, but more between BWS_freq and BWS_RPL.
110 Figure 4 12. Best w orst s caling order of relative impact of attribute/scale level ranges across methods T able 4 11 level ranges across methods. The values from the last third columns correspond to th e ordering of WTA ranges of DCE models and Binary Model 1, respectively; the middle three column s reflects the hierarchy of level scale value range in BWS These values range from 1 to 4, where 4 match the attribute with biggest range. As explained by Louviere and Islam (2008), attribute level ranges of DCE and Binary parameter estimates are c onfounded with weight and utility scale, and therefore potentially misleading in terms of assessing the importance of the attribute. By transforming these estimates to WTA, the values take on a common metric scale ($/choice), which allows for the estimates of WTA ranges to provide a more accurate measure importance. BWS has the advantage of providing estimates on a common underlying scale. By design,
111 Table 4 11. Comparison between ran k and ordering of relative impact of attribute/scale level ranges a cross all m ethods used in Chapter 3 and Chapter 4 (1 having the least impact) BWS Rank Order of relative impact Attribute Using WTA at tribute range BWS (RPL) BWS (clogit) BWS_clogit (demographics) BWS (RPL) BWS (clogit) BWS (Freq) DCE (RPL) DCE Model 2 (clogit) Binary Model 2 (REL) Risk Tool 1 (0) 2 (0) 3 (0.78) 1 1 1 1 (0.12) 1 (2.52) 1 (0.04) Penalty 3 (0.97) 3 (0.44) 2 (0.38) 2 3 3 2 (8.54) 2 (12.92) 2 (20.28) Time 2 (0.61) 1 ( 0.01) 1 (0) 3 2 2 Revenue 4 (1.71) 4 (1.89) 4 (3.62) 4 4 4 3 (32.83) 3 (56.19) 3 (60.00) Table 4 12. Comparison of order of differences between levels a cross all m ethods used in Chapter 3 and Chapter 4 (1 having the least impact) Order of differences between levels (1 having the lowest magnitude) Level Scale Attribute BWS (RPL) BWS (clogit) demographics BWS (clogit) BWS (Freq) DCE (RPL) DCE Model 2 (clo git) Binary Model 2 (REL) Difference between Insurance to Risk Pool 6 8 5 1 7 8 1 Difference between No Penalty Penalty 1 1 6 6 2 1 4 Difference between $5 to $10 acre per year 7 5 1 5 1 6 7 Difference between $10 to $20 acre per year 3 6 7 8 6 2 2 Difference between $20 to $30 acre per year 2 3 3 3 3 3 5 Difference between 5 to 10 year contract 5 4 4 2 4 4 3 Difference between 10 to 40 year contract 8 2 8 7 5 7 8 Difference between 40 to 100 year contract 4 7 2 4 8 5 6
112 Table 4 11 shows gene ral agreement between all models indicating that BWS_clogit_demographics and BWC_clogit, w and DCE_clogit. The exclusion of covariates from the model BWS_clogit produces a slightly different order of impact than BWS_clo git_demographics, but if the impact f BWS_clogit are in more agreement with DCE_clogit. The ranges estimated using WTA attribute measurments seem to be similar for all DCE, but the magnitudes of DCE_RPL apper slightly smaller than DCE_clogit and DCE_REL. Overall, BWS Rank appears to producin g similar measurements of order of relative impact of DCE using WTA and BWS using level scale values. Figure 4 13. Order of attribute impact
113 Figure 4 13 shows the order of impact from all models. BWS_clogit using demographics seems to be the only model Table 4 across all methods used in C hapters 3 and 4. The differences we re ranked from 1 to eight, based on the largest magnitude differences between levels, where 1 indicates the lowest difference. Figure 4 14 graphically displays these ordinal comparisons. This figure shows a general disagreement of between all models, which concurs with the BWS comparison of these ordering indicators shown in Figure 4 11. Figure 4 14. Order of difference between levels Limitations This comparison chapter did not include other BWS estimations, such as weighted least squares, and other marg inal methods proposed in Flynn et al. (2007). Further analysis of covariates would have provided more information regarding the
114 influence of their interaction with level scale values and impact variables in the BWS. The analysis of sub populations, which a re common in the use of Frequency Counts, was not factored into these comparisons, which may also limit the scale of this analysis. Discussion and Summary To better select the most appropriate discre te choice experiment, Louviere e t al. (2000) advise rese archers to consider the following design objectives: identification, precision, cognitive complexity, and market realism. For some applications, such as referendum type questions, BWC may provide the necessary market realism to estimate welfare measures an d BWS analysis importance. The identification and precision of BWS estimates from BWC have been found to be sensitive to model specification and estimation, but to generally agree with the paired estimations of conditional logit and random parameters logit There was also general agreement with the ordering of attribute impacts, especially with regards to the highest and lowest ordered attributes across most DCE, BWS, and Binary models, but not within the middle range areas. Most BWS specifications (includi ng the use specifications using demographic covariates) did agree fully with this ordering. This study did not determine the reason for these differences, but this is the subject of further research exploring the effects of relaxing some assumptions of con ditional logit using random parameters logit in BWS estimations, as well as potential problems of price endogeneity and non linearity in BWS. The welfare estimates of Binary from BWC were found to be in general agreement with all DCE estimations in regards to sign and significance, but they seem to be overestimating some (about 37%) of t he WTA measurements of attribute levels of carbon offset programs institutional components. Train (2003) explains the fact that
115 binary logit models that fail to account for other available (or expected) choices have fewer parameters in the denominator of the probabilistic equation, which is estimated with the same numerator that assumes only two options (yes/no), may affect the probability of choosing a carbon program. The o verall results of this study suggest that analysts in Forestry and Applied Economics exploring areas of market realism that resemble binary choices (i.e. referendum type studies), will significantly benefit from using best worst choice, which produces best worst scaling measurements, along with consistent welfare estimates from a binary discrete choice experimentation tool.
116 CHAPTER 5 E STIMATING THE SUPPLY OF FOREST CARBON OFF SETS : STATED PREFERENCE MEASUREMENTS OF CARBON SEQUESTRATION IN FLORIDA An improvem ent of forest management practices (IFM) in Florida has been estimated to yield an additional $116.8 million 1 to producers of pine plantations who engage in forest carbon markets (Mulkey et al., 2008) The use of stated preference methodologies to estimate willingness to accept (WTA) compensation for producing forest carbon offsets has been successfully applied to predict participation in these programs, and to examine the supply elasticity for various prices of carbon (Markowski Lindsay, 2011; Kline et al. 2000; Fletcher et al., 2009). In this chapter I apply a Linear offset plot data from the Forest I nventory Analysis (FIA) of the USDA Forest Service. 2 From a policy perspective, it is useful to present the findings of stated preference (SP) research in a manner that allows for policy makers to assess the influence of attributes on aggregate participat ion or adoption rates. Recent studies of hypothetical carbon offset programs in the US have used this format to present their results of multiple scenarios with fewer or greater institutional requirements (see Markowski Lindsay et al., 2011). Kline et al. (2000) estimated the potential supply of NIPF land enrolling in 10 year forest conservation reserve programs in Western Oregon and Western Washington, by using a random utility framework (RUM) to estimate WTA, compute the probability of enrollment, and mul tiplying this probability by the estimated 1 These estimates were done using $20 per metric ton CO2 equivalent and do not reflect costs of creation and maintenance of mitigation projects. 2 http://www.fia.fs.fed.us/
117 area of NIPF forestland in these regions. A recent study focusing on land conversion, surveyed Florida FSP members to elicit responses to bids offering compensation for renting their woodlands for corn production. This study used similar methods to measure WTA and estimate a hypothetical supply curve of NIPF forested land corn production in Florida (Pancholy et al., 2011). This approach may suffer from over generalizations of survey results from subpopulations, but if appropriately implemented, this method has shown to offer an interesting policy interpretation of stated preference research. Several studies have used revealed preference (RP) data to estimate marginal costs to supply carbon sequestration in the US, b ased on landowner behavior (see Stavins, 1999). Lubowski et al. (2006) use RP micro data of landowner behavior to model six major private land uses in the US, by treating commodity prices as endogenous, and predicting carbon storage changes with carbon opt imization models. Their simulation results from their study, which introduces a subsidy/tax rate ranging from $0 to $350 per acre to estimate a baseline and policy scenario for multiple land uses, finds lower marginal costs of carbon sequestration when tim ber harvesting is prohibited on lands enrolled in this type of subsidy/tax program. Their findings show that land use preferences of NIPF landowners would be responsive to incentives (e.g. subsidy and more), and that institutional components, such as taxes for carbon leakages, and restrictions are relevant considerations in RP simulations. Stainback and Alavalapati (2002) also focused on similar carbon sequestration subsidy/taxation policy in the southern US, using Land Expectation Values (LEV) from timber benefits. Their
118 estimations suggest that a carbon sequestration benefits and emissions cost policy would benefit private forest landowners. The use of simulation tools is often required in carbon sequestration certification programs to evaluate current car bon stocks and future sequestration rates. The Climate Action Reserve (Chapter 2) for example, under the forest protocol of IFM, requires a qualitative chara cterization of likely vegetation conditions and activities that would have occurred without the pro jec t (including laws, statutes, regulations or other legal mandates), along with 20 sample plots to perform a compute r simulation for 100 years. This chapter extends the results from previous chapters by augmenting the analysis of carbon market preference s with measured forest plot data of current carbon sequestration rates from Florida Forests (USDA Forest Services Forest Inventory Analysis). The probability of participation in multiple carbon offset program scenarios is used to estimate the carbon seques tration supply of respondents who are assumed to enroll in either an IFM program or reducing emissions from deforestation and degradation (REDD). Conceptual Framework Following the framework from Louviere et al. (2002), I implement a linear logit model to estimate the probability of choosing to participate in multiple carbon offset certification programs. Using the Binary data from BWC, I estimate the participation probabilities as follows: Let the probability of choosing to enroll equal to where V 1 q and V 2q are the linear characteristic associated with alternative 1 (enroll) and 2 (not enroll). Then,
119 (5 1) where X k is the value of the explanatory variables and the par ameter estimates. The left hand side is known as the logit of the probability of choice, and it represents the logarithm of the odds that individual q will choose to (enroll) 1. Given that I have observations of repeated choices for each attribute, this ap proach will yield consistent parameters (Louviere et al., 2002). This study estimates the probabilities of enrollment for 82 survey respondents who correctly reported their zip code and/or address of their largest forest plot in Florida using BWC. A range of $5 to $130 acre per year carbon prices was used to trace the (Chapter 2) to deal with uncertainty, and supply shifts were estimated using the presence and absence of a penalty for withdrawal (Chapter 3). Three programs were considered to match contract durations of currently available carbon offset certification programs for Florida landowners (Chapter 2): Program 1 used a 5 year contract (e.g. OTC), Program 2 used a 40 year contract (ACR and VCS), and Program 3 used 100 year contract (e.g. CAR). Estimating WTA Compensation for Producing Carbon Offsets This study continues the work from Chapter 3, where two survey elicitation methods were implemented to elicit responses o f BWC and DCE (Figure 3 1 from Chapter 3). Survey participants were randomly assigned to either survey type, differing only in terms of the conjoint questions (Chapter 3). Table 5 1 describes the results from a logistical estimation of the binary BWC (Bina ry_logit from now on) survey responses
120 using the attributes described in Table 3 2 from Chapter 3. Table 3 4 from Chapter 3 presents a comparison of survey respondents from BWC, DCE, and the most recent (2006) National Woodland Ownership Survey 3 (the most comprehensive NIPF landowners survey in Florida). As seen in Table 3 4, BWC and DCE respondents were properly randomized into each survey, and their demographic responses were very similar. The results from Table 5 1 included observations who responded to a question regarding the location of their biggest plot of land, and who provided identifiable information (e.g. zipcode, address, etc). The BWC data was reduced by 15 observations, and the map of these plots is shown in Figure 5 1. Figure 5 1. Map of b iggest forestland plot owned by BWC survey participants. 3 http://www.fi a.fs.fed.us/nwos/
121 The results from Table 5 evenue Quantitative) (Chapter 3 details effects coding). The results were very similar to Chapter 3 in terms of sign and significance. Namely, the level, while the re was negative. The marginal WTA estimates were also similar to their counterparts from the previous c costs of enrollment in a carbon offset program by $10.07 acre per year, while an increase of one year of contract duration would increase marginal WTA by $0.69 acre per year. Table 5 1. Results from b inary c hoice m odel: logit m odel e stimations for survey respondents who reported a correct with address and/or zip code Attribute Coefficient WTA Insurance 0.03 (0.06) $1.04 Risk Pool 0.03 e $1.04 No Penalty 0.29* (0.06) $10.07 Pen alty 0.29 e $10.07 Revenue Quantitative 0.03* (0.01) Time Quantitative 0.02 (0) $0.69 Constant 0.13 (0.13) Number of Respondents 82 Number of Choices 1312 Log Likelihood 831.10 Chi Square Statistic c 156.17 Pseudo R 2 0.09 a One (*), t wo (**), and (***) asterisk represent 0.01, 0.05 0.10 level of statistical significance, respectively. b Number in parentheses are standard errors. c Chi square statistic associated with a test of the hypothesis that a ll model parameters are zero. e Effe cts coding: negative sum of the above level scale values corresponding to this attribu te.
122 Figure 5 2 presents a comparative graph of WTA estimates from all discrete choice experimental models of this study, using the coding described for the measurements of Table 5 1. This figure shows that Binary_logit estimates of this sample are in agreement with all other models, especially with Binary_REL, but they re very similar in all models. Figure 5 2. Willingness to accept ($/attribute level) comparison of various models from
123 Estimating the Probability of Enrollment Figure 5 3 shows the estimated supply response by a landowner, with a 95% confidence interval, using Binary_logit for a program with a 5 year contract dur ation, the per year. The blue diamond points in this figure are the estimated probabilities of enrollment and the red and green lines show the upper and lower bounds of the 95% confidence interval. The figure shows that 95% of this sample would be predicted to participate within these bounds, which lie very close to the estimated participation pr obability line. Hence, there seems fairly elastic for the revenue range of $5 to $70 acre per year, and somewhat inelastic for higher revenues. Figure 5 3. Supply response with 95% confidence interval. Supply response by landowner using Binary_logit for a program with 5 years contract duration,
124 Figure 5 5 illustrates the supply response for Scenario 1 (5 ye ar contract, risk 5 3 seem to be consistent in this graph, namely, lower estimates of participation probability seem to be more elastic for revenue that is less than $70 acre per year. Figure 5 4. Scenario 1 supply response. Supply response by landowner using Binary_lo git for a program with 5 Figure 5 5 shows Scenario 2 (40 year contract with risk pool), which seems to be more elastic than Scenario 1, and the supply shift commitments are more decrease sensitive to the inclusion of a penalty for withdrawal.
125 Figure 5 5. Scenario 2 supply respons e. Supply response by landowner using Binary_logit for a program with 40 Figure 5 6 shows the supply response for Scenario 3 (100 year contract with risk po ol). This scenario is less elastic than scenarios 1 and 2. The supply shift of inclusion of a penalty for withdrawal is significantly greater than scenario with lower contractual commitment. The lower revenue values of $5 to $50 acre per year are less elas tic for participation by almost 13%. This figures illustrates the sensitivity of en rollment to changes in contract restrictions (penalty), which appear to be far greater for programs with higher contract durations.
126 Figure 5 6. Scenario 3 supply response. Supply response by landowner using Binary_logit for a program with 100 year contr Figures 5 7 and 5 8 compares four scenarios with different contract durations (5,30,40, and 100 years). These graphs examine the shifts arising contract duration. Fig ure 5 figure indicates a 100 year contract elicits a significant shift to the left of other supply response curves. The 100 year contract supply curve seems to be more elastic than the other time commitments of 5 to 40 year contracts. At the revenue price of $30 acres per year, a commitment of 100 years is estimated to enroll about 28% of participants, while 40 years would enroll 65%, 30 years 68%, and 5 years 78%. There seems to be a 50% difference in estimated participation rates of 5 to 100 years of contract commitment.
127 Figure 5 7. Multi cenario supply comparison. Supply response by landowner using Binary_logit for programs with several levels of contract duration, the use of Figure 5 8. Multi scenario supply comparison. Supply response by landowner using Binary_logit for programs with several levels of contract duration, the use of
128 Figure 5 8, compares the four different time commitments for supply response similar elasticities for contract duration of 5 to 100 years. The shifts are progressively going to the right with less contract commitments. At the revenue price of $30 acres per year, a commitment of 100 years is estimated to enroll about 42% of participants, while 40 years would enroll 51%, 30 years 55%, and 5 years 68%. There was a differen ce of almost 26% in estimated participation rates of 5 to 100 years of contract commitment. Figures 5 7 and 5 8 seem to indicate that at high prices, program features do not appear to be influencing participation probabilities, but at lower prices they bec ome very significant in determining the participation. Figures 5 8 and 5 9 show programs with higher time commitments tend to shift the supply response curves or probability of enrollment to the left, which decrease shifts to the left of the supply response curves, arising from higher commitment contracts by approximately 24% at a $30 acre per year revenue from producing carbon offsets. Combining BWC and DCE Survey Dat a with FIA Sample Plots In this section I described the augmentation of the survey data from BWC and DCE with field plot data from the USDA Forest Services Forestry Inventory Analysis program. The FIA is a nationwide continuous forest census, designed to e valuate current forest management practices. The surveys described in Chapter 3 included demographic and institutional questions to identify landowner preferences, and a question regarding the percentage of their Florida forestland they would be willing to enroll in a specific carbon offset program (Q17 from now on). The sample population
129 used to estimate the WTA and supply response curves in this chapter, come from landowners who reported either the zip code or address of their biggest plot of forest. In t his section I combine both DCE and BWC data to estimate current carbon sequestration rates of all participants, using matched data from FIA. In Table 5 1 presents the summary statistics of the combined responses for both BWC and DCE (Combined from here on ). Survey participants were randomly distributed to BWC and DCE and Table 3 4 from Chapter 3 shows that they are very similar in characteristics. Table 5 2 shows the summary statistics of demographic and institutional questions for the Combined dataset. T able 5 2. Summary statistic of variables from carbon program survey: mean and Variable Definition Mean (std dev) Acresf Number of forest acres own ed in F lorida 837 80 ( 3885. 6 3 ) Q17 Percent of FL forestland willing to enroll in a program with: either Insurance or Risk Pool, No penalty for early withdrawal, $30 acre per year, 5 year contract 0.79 (0.34) Q17_ a cres Interaction of variables acresf & Q17 603.54 (2691.26 ) Survey type If data collected in BWC =1, DCE=0 0.52 (0.5) NGO government carbon credit 1 = Very Unimportant 2 = Somewhat Unimportant 3 = Neutral 4 = Somewhat Important 5 = Very Important 3.13 (1.37) Verifier government program vs a government verifier 1 = Very Unimportant 2 = Somewhat Unimportant 3 = Neutral 4 = Somewhat Important 5 = Very Important 3.12 (1.4)
130 Table 5 2. Continued Variable Definition Mean Cost shar e 1 = if ever used a state or federal sponsored cost share program 2 = otherwise 0.39 (0.53) Type of CS 1= if ever used a State cost share program 2= if ever used a Federal cost share program 3= if ever used both State and Federal cost share program 4= none (Follow 2.14 (1) Home 1 = if primary residence within one mile from any of their forestland 2 = if primary residence is not within one mile from any of their forest land 3 = Not applicable 1.59 (0.59) Income 1 = less than $25,000 2 = $25,000 to $49,999 3 = $50,000 to $99,999 4 = $100,000 to $199,999 5 = $200,000 or more 3.99 (1.7) Sex 1 = Male; 2 = Female 1.25 (0.61) Age 1 = under 25 years 2 = 25 to 34 years 3 = 35 to 44 years 4 = 45 to 54 years 5 = 55 to 64 y ears 6 = 65 to 74 years 7 = 75 years or over 5.09 (1.21) Education 1 = Less than 12 th grad 2 = High school graduate or GED 3 = Some college 4 = Associate or technical degree 6 = Graduate degree 4.86 (1.3) Using the zip code of thei r biggest reported plot of land, FIA data was used to pair these observations with estimates of two observations of compiled for the 2009 FIA report and the 2010 FIA data (Tables 5 3 and 5 4), with measurements periods of
131 roughly 7 years apart. These Combi ned data resulted in 136 observations (Figure 5 9). There were multiple FIA plot observations per zip code for some the survey data, and the systematic rule for these situations was to take the average per zip code of the values provided for a given plot. FIA data included above ground and belowground carbon stocks for two measur e ment periods The above and belowground sequestration ( units tons/ha) per plot was esti mated by taking the annualized difference between these two observation periods for above and below ground carbon stock, and d ividing the m by the amount of years between observations. (Tables 5 3 and 5 4) Figure 5 9 shows the map of the biggest reported plot of forestland from the Combined dataset. As seen in this figure, this survey sample appear s to be evenly distributed in the Northeast, Northwest, and North Central areas of Florida. Figure 5 9. Map of the biggest reported plot of forestland from Combined data
132 The third columns of Table 5 2 shows that 65% of Combined matched (FIACombined from hereafter) observations had no observable management treatments learing, slash burning, chopping, disking, bedding ), or other practices clearly intended to prepare a site for either natu ral or artificial regeneration, and 4% had used some type of site preparation (e.g. c learing, slash burning, chopping, disking, bedding, or other practices ) clearly intended to prepare a site for either natu ral or artificial regeneration. Table 5 3. Forest Inventory Analysis Code DCE Model 2: Ordered by absolute WTA range Frequency 0 No observable treatment. 88 (65%) 10 Cutting The removal of one or more trees from a stand. 42 (31%) 20 Site preparation Cle aring, slash burning, chopping, disking, bedding, or other practices clearly intended to prepare a site for either natural or artificial regeneration. 6 (4%) 30 Artificial regeneration Following a disturbance or treatment (usually cutting), a new stand where at least 50 percent of the live trees present resulted from planting or direct seeding. 0 40 Natural regeneration Following a disturbance or treatment (usually cutting), a new stand where at least 50 percent of the live trees present (of any siz e) were established through the growth of existing trees and/or natural seeding or sprouting. 0 50 Other silvicultural treatment The use of fertilizers, herbicides, girdling, pruning, or other activities (not covered by codes 10 40) designed to improve the commercial value of the residual stand, or chaining, which is a practice used on western woodlands to encourage wildlife forage. 0 Table 5 4, shows that 68 of observations were slash p ine forest types, 21 loblolly, 38 mixed forest, 6 longleaf, and 2 southern scrub oak. The average difference between measurements was 6.95 years, the mean stand age was about 39 years, and the average site productivity was close to 85 cubic feet/acre/ye variable in this table indicates that the vast m ajority of observations came from areas that were not r eserved or withdrawn by law(s) prohibiting the ma nagement of the land
133 for production variable informs us that 93% of these plots were from of national forests system, state, local, etc.). Table 5 4. Summary statistics of variables from Forest Inventory Analysis. Means and Variable Definition Mean Tree 1 = Slash Pine ( n= 68) 2 = Longleaf ( n= 6) 3 = Loblolly ( n= 21) 4 = Mixed ( n= 38) 5 = Southern Scrub Oak ( n= 2) Year Diff Difference between plot measurments (years) 6.95 (0.6) Stand Age Stand Age (years) 38.72 (16.95) C_AG_Seq Annual above ground carbon sequestration. Estimated by taking the annualized difference betwe en above ground carbon stock measurements of FIA data (2009 2010). The observations were annualized by dividing the corresponding differences in carbon stock, by the variable 2.61 (12.88) C_BG_Seq Annual below ground carbon sequestra tion. Estimated by taking the annualized difference between above ground carbon stock measurements of FIA data (2009 2010). The observations were annualized by dividing the corresponding differences in carbon stock, by the variable 0. 39 (2.82) SITECLCD Site productivity class co de. A classification of forest land in terms of inherent capacity to grow crops of industrial wood. 1 = 225+ cubic feet/acre/year 2 = 165 224 cubic feet/acre/year 3 = 120 164 cubic feet/acre/year 4 = 85 119 c ubic feet/acre/year 5 = 50 84 cubic feet/acre/year 6 = 20 49 cubic feet/acre/year 7 = 0 19 cubic feet/acre/year 4.69 (0.58) RESERVCD Reserved land is land that is withdrawn by law(s) prohibiting the management of the land for the production of wood pr oducts. 0 = if not reserved 1 = if reserved 0.01 (0.04)
134 Table 5 4. Continued Variable Definition Mean OWNCD Owner class code: 1 = National Forest System ( n= 3) 2 = Bureau of Land Management ( n= 1) 3 = Department of Defense/Energy ( n= 1) 4 = Other fede ral ( n= 1) 5 = State ( n= 3) 6 = Undifferentiated private ( n= 120) 5.76 (0.88 ) Treatment 10 = if cutting since last measurement (5 years if new plot) 20 = Site preparation (clearing, slash burning, etc) 3.97 (5.75) Carbon Sequestration Estimates Table 5 4 indicates that average above ground carbon sequestration for these plots is 2.61 tons/ha/year, with standard deviation of 12.88 tons/ha, and belowground average carbon sequestration is 0.30 tons/ha, with a standard deviation of 2.82 tons/ha. These numbers indicate a very large dispersion in this data. Figure 5 10. Example of Question 17 of survey. Each of the respondents in the BWC and DCE surveys were asked to respond to the question in Figure 5 10. Their responses to Question 17 (Q17) of the survey ar e graphed in Figure 5 11. This horizontal axis of this figure displays the percent of their
135 Florida forestland willing to enroll in the carbon program described in Figure 5 10. Close to 60% of respondents were willing to enroll 100% of their land in Q17, w hich is Scenario 1 at Revenue $30 acre per The mean of Q17 was 79% of land willing to enroll in Scenario 1, with a standard deviation of 34%. (Figure 5 11) Figure 5 11. Answers to Question 17 of the BWC and DCE sur veys: the horizontal axis displays the percent of their Florida forestland willing to enroll in a carbon program described in Figure 5 6. Figure 5 12 displays the interaction between reported forest acreage with Q.17, which indicates the amount acres of F lorida forestland willing to enroll in Q.17. This figure shows that the majority of respondents would enroll less than 2,000 acres.
136 Figure 5 12. Interaction between forest acreage reported and the percent of forestland willing to enroll in Scenario 1 Figure 5 13 displays the supply response for Scenario 1 (5 years contract interval, using the same method from the previous section. If I assume that Q.17 implies per cent participation in this program, then the mean of Q17 would fall within the 95% CI of Scenario 1 supply response curve (Figure 5 13). The purple line indicates predicted probability of enrollment and the blue and green lines are the 95% confidence inter val. The red dot in this figure is the mean of Q17, with a range of one standard deviation (red line). The red dot in Figure 5 13 representing Q17 indicates that 79% (mean of Q17) of Florida forestland reported in this survey (FIACombined) would have been e nrolled at this program for $30 acre per year. If I were to assume that prediction enrollment curve
137 implied a supply of acres enrolled in this particular program (Scenario 1), then, at $30 acre per year, the estimation would have been very close to the mea n acreage reported in Q17. Continuing with this assumption, and looking at the first standard deviation of Q17 displayed on this graph (red line), at least 34.1% of Q17 (half of the standard deviation) falls within the 95% confidence interval (CI) of the p redicted probability of enrollment (within the red and blue lines). The rest is arguably either inside of the 95% CI of this curve, or to the right of the upper bound, which means that if I were to assume a one to one relationship between predicted enrollm ent and acreage enrollment, not only would this be fairly accurate assumption for Q17 in Scenario 1 at $30 acres per year, but a slightly conservative one. Figuer 5 13. S upply response with 95% confidence interval. Supply response by landowner using Bin ary_logit for a program with 5 years contract duration,
138 Given these results, this analysis will assume that the supply response curves imply a percent of acres enrolled in the particular program f or the respondents in the FIACombined data set. Additionality Figure 5 14 shows the total aboveground and belowground carbon sequestration (Ctot) estimates of the sum of C_AG_Seq and C_BG_Seq estimates from Table 5 4, for respondents in the FIACombined dat aset. The horizontal axis is the survey ID (ordered by Ctot) and the units of the vertical axis are in metric tons per year. This figure shows that almost 50 plots have a negative rate of carbon sequestration. Figure 5 14. Total carbon sequestration est imates for respondents in FIACombined. The horizontal axis is the survey ID and the units of the vertical axis are in metric tons per year. The analysis of the previous section indicates that most of the plots in this study land is not withdrawn by l aw(s) prohibiting the
139 management of the land for the production of wood products. Thus, I model additionality for the participants with positive carbon sequestration rates with an Improved Forest Management (IFM) carbon certificat ion protocol (which in this case assumes that landowners manage their land for commercial wood products) and the rest using a reduce deforestation and degradation protocol. (Chapter 2 defines and analyzes IFM and REDD) IFM Additionality Estimation As not ed in the introduction of this chapter, the CAR certification program, requires the use of simulation models in their protocols of IFM (Chapter 2). To estimate IFM additionality, I employ the results from Mulkey et al. (2008), which simulate changes in man agement intensities in commercial pine plantations in Florida. Their study simulated slash pine and loblolly forest, which comprise represent 66% (another forest types i n this data (Table 5 4). Mulkey et al. (2008) assume that management activities and intensities in Florida uthors also assume that these practices are implemented in private forestlands. Average volume harvested per unit area is estimated by aggregating the combination of site productivity and management intensity. Their simulation models changes from low to me dium management intensity, and from medium to high management intensity. They also assume that fertilizer is only applied for medium to high changes in IMP. The high management intensity scenario is assumed to use fertilizer every 4 years on sites with a r otation age of 25 years, whereas the medium scenario uses fertilizer at years 2 and 16. To estimate aboveground carbon,
140 21% of stem biomass. Carbon in roots was estimate a s twice the annual carbon production in roots. Table 5 5 shows carbon pool simulation estimates in commercial pine plantations potentially yield 0.39 tonnes (MMt) carbon per acre per year. Namely, going from medium to high intensity would increase sequestration by approximately 35% per acre per year, and 29.4% for a low to medium switch. Table 5 5. Carbon pool in commercial pine plantations in Florida. Management intensity % of land under each intensity # of acres (millions) Biomass acre 1 (MMt) Carbon acr e 1 (MMt) Rotation age (years) Carbon acre 1 year 1 (MMt) Low 0.37 5 63.5 25.42 30 0.85 Medium 0.58 5 68.5 27.42 25 1.10 High 0.05 5 92.8 37.14 25 1.49 Note: 1.4 cubic meters equals 1 MMt; in converting biomass to carbon, the authors assumed that moist ure content would be 20% an d 50% of dry biomass is carbon. [ Table ad a pted fro m Mulkey S., J. Alavalapati, A. Hodges, A.C. Wilkie and S. Grunwald (2008), Opportunities for greenhouse gas reduction by agriculture and forestry in F lorida (Page 17, Table 4) University of Florida, School of Natural Resources and Environment Department of Environmental Defense Washington D.C. ] To estimate additionality for IFM, I assume a baseline scenario of current carbon sequestration rates based on Figure 5 14, and I make the assumption that, these lands would be producing commerci al timber, and that a change in IFM for low productivity lands would yield a 29% increase in carbon sequestration (addtionality) and from medium to high a 35% increase in carbon sequestration. As seem in Table 5 4, the variable SITECLCD provides an estimat e of s ite productivity The FIACombined data was classified this data in three groups: 1) low productivity if the land produces 0 to 49 cubic feet/acre/year, 2) medium if 50 to 164 cubic feet/acre/year, and 3) high
141 productivity if greater than 165 cubic fe et/acre/year. In all 96.32% of observations were of medium productivity, and 3.68% were low productivity, and none were highly productive. consists of multiplying acres of forest by current carbon sequestration rates, 2) if classified as medium productivity, assume that a 35% increase in sequestration occurred by switching to from a medium to high management intensity, and 29% for switching from low to medium, and 3) subtract the new rate of carbon sequestration by the baseline. REDD Additionality Estimation REDD estimations for areas currently having a negative sequestration rate are REDD (Chapter 2). A sensitivity analysis is performed for different rates of REDD. Figures 5 15 and 5 16 show the maps of forestlands assumed to be under IFM and REDD, respectively. Figure 5 15 shows a map of forestland assumed to be under IFM. This map indicates that the majority of forestland assumed to be enrolled under IFM, appears to be in North Florida. Figure 5 16 shows the forestland areas assumed to be under REDD do not seem to be clustering in any particular region of Florida. Figures 5 15 and 5 16 do not vis ually indicate a particular clustering around major Florida cities, but Jacksonville appears to have more IFM present than REDD.
142 Figure 5 15. Map of forestland assumed to be under IFM Figure 5 16. Map of forestland assumed to be under REDD
143 Results Figures 5 18 to 5 20 show IFM estimates of carbon additionaly with varying levels of REDD. The blue lines indiate the increase in management intensity implemented for the purpose of IFM and REDD, and the maroon lines are the base line or current sequestra tion rates. The observations of approximately the first 50 individuals in these graphs show a negative sequestration for the current sequestration rate, which is assumed to come from deforestation. The blue dots corresponding to these obserervations become positive by assuming they stop deforestation by participation in REDD. Figure 5 16, shows REED participation at 10%, Figure 5 17 at 20% and Figure 5 18 at 50%. The rest of blue curve comes from the IFM estimates described in the previous section. Figu re 5 17. Carbon sequestration and additionality for IFM and REDD at 10% IFM & REDD at 10%
144 Figure 5 18. Carbon sequestration and additionality for IFM and REDD at 20% Figure 5 19. Carbon sequestration and additionality for IFM and REDD at 50% IFM & REDD at 10% IFM & REDD at 10%
145 By subtrac ting the bas eline in Figure 5 17 with the estimated change s in management intensity and REDD at 10%, I estimate an aggregate carbon additionality of 19 ,776 tons per year. This additionality was multiplied by the supply response scenarios from the previous sections to estimate the supply curves in metric tons (MMt) of carbon per year shown in Figures 5 21 to 5 23. These figures produce the same elasticity framework discussed in the previous sections. Figure 5 20. Supply of carbon additionality for Scenario 1 usi ng IFM and REDD at 10% MMt of carbon per year
146 Figure 5 21. Supply of carbon additionality for Scenario 1 using IFM and REDD at 10% Figure 5 22. Supply of carbon additionality for Scenario 1 using IFM and REDD at 10% MMt of carbon per year MMt of carbon per year
147 Discussion and Summary This chapter provided an appr oximation of the supply of carbon sequestration for survey participants using direct forest plot measurements from the Forest Inventory Analysis, and the predicted participation probabilities of hypothetical carbon offset programs. The simulation estimates from Mulkey et al. (2008) provided a way to estimate additionality from a change in improved forest management intensity, but the simulation of land expectation values can be used to model carbon policy scenarios that may be more suitable for the stated p references elicited in the BWC and DCE surveys (e.g. Stainback and Alavalapati, 2002). The supply response curves from 5 20 to 5 23 show that respondents to this survey can potentially provide up to 19 .78 MMt per year of carbon, and they may serve as an in teresting reference for policy makers examining effects of various institutional factors of carbon sequestration programs.
148 CHAPTER 6 CONCLUSION S This dissertation characterized forest carbon offset markets in Florida using a hypothetical carbon program survey, administered to one the most comprehensive lists of non industrial private landowners. In December 2011, 920 Florida Forest Stewardship Program affiliates were surveyed to eliciting response to different carbon offset programs that varied according to composition of institutional components. The survey was administered electronically and 310 responded to the requests, of which 189 completed the entire survey. Additionally, Chapter 2 of this dissertation provided an analysis of current available carb on certification options for Florida forestland owners, and reviewed their institutional compositions to qualitatively identify potential barriers to participate. Results from Chapter 2 indicated a lack of access to regional cap and trade markets in North America, but described specific options to engage in voluntary carbon markets. The dissertation in Chapter 3 quantitatively analyzed the results from Chapter 2 by examining the influence of four particular attributes of carbon market program references: Co ntract Duration, Revenue, Penalty for Early Withdrawal from a program, and the use of different Risk Tools to reduce the risks associated with producing offsets. The results from this chapter indicate that landowners in this study prefer programs with comp ensations of $20 to $30 acres per year, contract durations of 5 to 20 years, no penalty for withdrawal. Next, Chapter 4 implemented an empirical comparison of a relatively new stated preference tool that elicits responses using best worst scaling and binar y discrete choice experimentation. The results from this chapter indicated a general agreement among estimations of best worst scaling, and identified potential measurements of disagreement. Finally, Chapter 5 provided an estimation of current
149 carbon seque stration rates for survey respondents, and estimated the supply of carbon additionality for various scenarios that approximate current existing carbon certification programs, using stated preference and available forest plot data from the Forest Inventory Analysis.
150 APPENDIX A T HE SURVEYS Introduction and Filter Questions Survey
151 Instructions for Discrete Choice Experimentation
152 Instructions for Best Worst Choice
153 D emographic Questions
155 APPENDIX B NON LINEAR CONSIDERATIONS Figure B 1. Discr ete choice experimentation estimates of Mode 1 vs. attribute levels Figure B 2. Binary estimates of Mode 1 vs. attribute levels
156 Figure B 3. Orthogon al main effects design. [ Figure a d a pted from Street, D.J., Burgess, L., and J.J. Louviere (2005), Qui ck and easy choice sets: constructing optimal and nearly optimal stated choice experiments (Page 465, Table 9) Int. J. Res. Marketing, 22 (2005), pp. 459 470. ]
157 LIST OF REFERENCES Acosta, Y. (2009), County Energy Efficiency and Conservation Efforts. Flori da Association of Counties Retrieved from: http://www.fl counties.com/Docs/Legislative%20Division/Energy/County Energy Efficiency and Conservation Efforts 2 11 09.pdf Adams, D.C., (2009). Agriculture and Greenhouse Gas Cap and Trade. Policy Issues, PI3, 1 5 Adamowicz W., Boxall R, Louviere, Willi ams M. (1994), Combining Reveal and Stated Prefences Methods for Valuing Environmental Amenities, Journal of Environmental Economics and Management 26: 271 92. Adamowicz W., Boxall R, Williams M., Louviere (1998), Stated Preference Approaches for Measuring Passive Use Values: Choice Experiments and Contingent Valuation, American Journal of Agricultural Economics 80:64 75 American Carbon Registry (2010), American Carbon Registry Forest Carbon Project, version 2.1. Winrock International, Little Rock, Arkansas. Bell, D. C., Roberts, K. R. English C. B. Park M. W. (1994), A logit analysis of and Applied Economics 26: 463 72. Berry, S. (19 94), Estimating discrete choice mo dels of product differentiation RAND Journal of Economics 25, 242 262. Broder, J.M. and Krauss, C. (2010, January 26). Advocates of Climate Bill Scale Down Their Goals. The New York Times. Retrieved from http://www.nytimes.com/2010/01/27/science/earth/27climate.html Butler, Brett J., Earl C. Leatherberry (2004), America's Family Forest Owners. Journal of Forestry Volume 102, Number 7, October/November, pp. 4 14(11). Butterfield, R. Ha nsen, E., Fletcher, R., Nikinmaa, H. (2005), Forest certification and small forest enterprises: key trends and impacts, benefits and barriers. Washing ton, DC: Forest Trends. Cent er for Climate Strategies (2009), Final Florida Greenhouse Gas Inventory and Reference Case Projections 1990 2025. Charnley, S., Diaz, D., Gosnell, H. (2010), Mitigating Climate Change through Small Scale Forestry in the USA : Opportunities and Challenges. Small scale Forestry 9(4):445 462.
158 Climate Action Reserve Forest Project Protocol (Version 3.2), retrieved from http://www.climateactionreserve.org/wp content/uploads/2009/03/Forest_Project_Protocol_Version_3.2.pdf Climate Action Reserve, Projects. Retrieved March 2, 2011, from https ://thereserve1.apx.com/myModule/rpt/myrpt.asp?r=111 Coase, Ronald (1960), The Problem of Social Cost, Journal of Law and Economics, 3:1 44. Coast, J., Salisbury, C., de Berker, D., Noble, A., Horrocks, S., Peters, T. J., et al. (2006), Preferences for asp ects of a dermatology consultation. British Journal of Dermatology, 155, 387 392. Diaz, D., Ham ilton, K., Johnson E. (2011), State of the Forest Carbon Markets 2011: From Canopy to Currency Ecosystem Marketplace. Dickinson, Brenton J. (2010), Massachusetts landowner participation in forest management programs for carbon sequestration: an ordered logit analysis of of Resource Economics. Dillman, D.A., Smyth, J.D., Christian, L.M. (2009), Internet, Mail, and Mixed Mode Surveys: The Tailored Design Method. Hoboken, NJ: John Wiley & Sons, Inc. Elrod T Louviere J J Davey K S (1992), An empirical comparison of ratings based and choice based conjoint models. Journal of Marketing Research XXIX (3):368 377. Fin n, A., and Louviere J.J. (1992 ) Determining the Appropriate Response to Evidence Journal of Public Policy & Marketing 11:12 25. Fischer A.P and Charnley, S (2010), Social and Cultural Influences on Management for Carbon Sequestration on US Family Forest lands: A Literature Synthesis, International Journal of Forestry Research, vol. 2010, Article ID 960912, 14 pages, 20 10. doi:10.1155/2010/960912 Fletcher, L.S., D. Kittridge, Jr., Stevens T. ( 2009 ) carbon credits: a pilot study. Northern Journal of Applied Forestry 26 (1), 35 37. Flynn, T.N., Louviere J.J. P eters, T.J. Coa st J. (2007), Best Worst Scaling: What It Can Do for Health Care Research and How to Do It. Journal of Health Economics 26 : 171 89. Flynn T .N. Louviere J .J Peters T .J. Coast J. ( 2008 ) Estimating preferences for a dermatology consultation using Best Worst Scalin g: comparison of various methods of ana lysis, BMC Med Res Methodology, 8:76 88.
159 Foster, V. and Mourato, S. (2002), Testing for consistency in contingent ranking experiments, Journal of Environmental Economics and Management Vol. 44 No. 2, pp. 309 28. Fri edman, CS and Glinn JG (2010), Florida Is Keeping Pace: House Bill 7179. The Florida Bar Journal. Volume 84, No. 8 Gronewold, N. (2011, January 3). Chicago Climate Exchange Closes Nation's First Cap And Trade System but Keeps Eye to the Future. The New Yor k Times. Retrieved from http://www.nytimes.com/cwire/2011/01/03/03climatewire chicago climate exchange closes but keeps ey 78598.html Hansen, J. Mki. Sato, Ruedy R. (2012), Perception of clim ate change Proceedings of the National Academy of Sciences 109, 14726 14727, E2415 E2423, doi:10.1073/pnas.1205276109. Hanemann, W. M. (1984), Welfare evaluation in contingent valuation experiments with discrete responses. American Journal of Agricultur al Economics 66: 332 41. Intergovernmental Panel on Climate Change (IPCC) Solomon S, et al., ed. (2007) Climate Change 2007: The Physical Science Basis (Cambridge press, New York, NY). Jaffe, A.B., Newell R.G., Stavins, R.N. (2005), A tale of two market f ailures: T echnology and environmental policy, Ecological Economics, 54 pp. 164 174 Johnson, LR et al. ( 200 5), Lifecycle impacts of forest resource activities in the Pacific North West and Southeast United States, Wood and Fiber Scien ce, 37 CORRIM Special Issue: 30. Kline, J.D., Alig, R.J ., Johnson, R.L. (2000) Fostering the production of non timber services among forest owners with heterogeneous objectives. Forest Science, 46, 302 311. Krinsky, I. and Robb A.L. (1986), On Approximating the Statistical P roperties of Elasticities, Review of Economics and Statistics 64: 715 19. Lancsar, E., J. Louviere Flynn T. (2007), Several methods to investigate relative attribute impact in stated preference experiments, Social Science and Medicine 64(8):1738 1753. Li, Yanshu (2010), Environmental Credit marketing Survey Report, Texas Forest Service Loureiro M.L., and Dominguez Arcos F. (2012), Applying Best Worst Scaling in a stated preference analysis of forest management programs, Journal of Forest Economics, ht tp://dx.doi.org/10.1016/j.jfe.2012.06.006
160 Louviere, J., Hen sher D. Swait J. (2000), Stated Choice Methods Analysis and Application. Cambridge: Cambridge University Press. Louviere, J. J., & Islam, T. (2008) A comparison of importance weights and willi ngness to pay measures derived from choice based conjoint, constant sum scales and best worst scaling, Journal of Business Research 61 (9), pp. 903 911. Lusk, J.L., and Briggeman B.C. (2009) Food Values, American Journal of Agricultural Economics 91(1): 184 196 Lusk, JL, Roosen J., Fox J.A. (2003), Demand for Beef Cattle Administered Growth Hormones or Fed Genetically Modified Corn: A Comparison of Consumers in France, Germany, The United Kingdom, and the United States, Amer. J. Agr. Econ. 85(1): 16 29 Lusk, Jason L. and Parker N. ( 2009 ) Consumer Preferences for Fat in Ground Beef. Journal of Agricultural and Appl ied Economics, 41,1, April, pp. 75 90. Markowski Lindsay M, Stevens T, Kittredge DB, Butler BJ, Catanzaro P, Dickinson BJ (2011) Barriers to Massachusetts forest landowner participation in carbon markets. Ecol Econ 71:180 190 M arley, A.A.J., T.N. Flynn, Louviere J.J. ( 2008) Probabilistic Models of set dependent and at tribute level best worst choice, Journa l of Mathematical Psychology 52, pp. 281 296. Zarembka, ed., Frontiers in Econometrics, Academic Press, New York, pp. 105 142. Mc Fadden D. (2001), Disaggregate behavioral travel demand's rum side: a 30 year retrospective. In: Hensher DA, editor. Travel behavioral research: the leading edge. Amsterdam: Pergamon. Mulkey S., Alavalapati J. Hodges, A. Wilkie A.C. Grunwald S. (2008), Opportunities for greenhouse gas reduction by agriculture and forestry in Florida. University of Florida, School of Natural Resources and Environment Department of Environmental Defense Washington D.C. Nicholls, OECD Environment Working Papers No. 1, OECD Publishing. DOI: 10.1787/011766488208. Retrieved from http://www.oecd ilibrary.org/environment/ranking port cities with high exposure and vulnerability to climate extremes_011766488208 Ohler T, Le A, Louviere J, Swait J., (2000), Attribute range effects in binary response tasks. Marketing Letters ; 11(3): 249 60.
161 Pancholy a N., Thomas, M.H., Solis, D., Stratis N. (2011) The impact of biofuels on the propensity of land use conversion among non industrial private forest landowners in Florida. Forest Policy and Economics 13 pp. 570 574. Peters Stanley M, Hamilton K (2012) State of the Voluntary Carbon Markets 2012: Developing Dimension. Ecosystem Marketplace and Bloomberg New Energy Finance, Washington, DC Potoglou D, Burge P, Flynn T, Netten A, Malley J, J. Forder et al. (2011), Best worst scaling vs. discrete choice experiments: an empirical comparison usin g social care data, Social Science Medicine, 72 (2011), pp. 1717 1727 Raymond, L., and Shively, G. (2008), Market Based Approaches to CO 2 Emissions Reductions. Choices 23(1), 38 40. Rodriguez G, Elo I. (2003), Intra class correlation in random effects mo dels for binary data, Stata Journal, 3: 32 46. Shaikh, S., L. Sun, van Kooten G. C. (2007), Are agricultural values a reliable guide in eate forest carbon sinks? Canadian Journal of Agricultural Economics 55,1:97 114. St ainback, G A and Alavalapati J R R ( 2002 ), Econo mic analysis of slash pine forest carbon sequestration in the southern U.S. Journal of Forest Economics 8: 105 117. Stavins R N (1999), The cost of carbon sequestration: A reveled preference approach. The American Economic Review 89: 994 1009. Street, D.J., Burgess, L., Louviere J.J. (2005), Quick and easy choice sets: constructing optimal and nearly optimal stated choice exper iments, Int. J. Res. Marketing, 22 (2005), pp. 459 470. Train, K.E. 2003. Discrete Choice Methods with Simulation. Cambridge, UK: Cambridge University Press. Voluntary Carbon Standard Guidance for Agriculture, Forestry and Other Land Use Projects (VCS 20 07.1, 2008). Van Kooten G.C., Shaikh S.L., Suchanek P. (2002), Mitigating climate change by planting trees: the transactions costs trap. Land Economics 78, 559 572
162 BIOGRAPHICAL SKETCH Jos R. Soto was born along the US/Mexico border in the town of Noga les, Sonora (Ambos Nogales). T hroughout his life, he become a scientist or observer of the grand or small schemes of the economic realities and relations of the border His personal growth was accompanied by a growing aware ness of many disparities of the e nvironmental, social, cultural, and economic interactions along the border. He therefore intuitively came to understand the politics of the border fence that preclude long lines of produce trucks and trains full of chemicals, as well as people waiting to c ross official checkpoints. In addition the drug culture and undocumented migrants, all attempting to get t for him it was just a fence. At 15, his family migrated to the US side of this wall After that, he would sometimes climb up the Nogales hills in order to watch the traffic jams, listen to the sirens and look at the awesome disparity of two countries. He could see the colonias ful l of makeshift shacks where poor Mexican workers live around the industrial parks ( maquilas ) As a result of this vantage point, he came to understand the topology of dividing line s of many types and reasons. This contemplation invited him to think about that twenty foot tall, corroded, metal wall dividing, while at the same time, orga nically making him a global citizen His interests and aspirations in economics are therefore not i nnate or superficially acquired. T hey stem from a need to know why or how it is that our natural resources, our workforce our pollution, and our incentiv es are managed the way they are. But more importantly, a need to intellectually know how they can be improved. His career goals are to seek an academic job that enables him to r esearch both sides of the fence that opens into a global market. T he impact and p otential of trade and
163 environmental policies along the Mexico/US border, the effects and contributions of undocumented migrant labor in the agric ultural and industrial sectors; and to utilize this position to mentor underrepresented Hispanic students, and encourage them to seek graduate degrees, and subsequent participation in academia.