<%BANNER%>

Potential Cost Benefit from Early Detection of Critical Congenital Heart Disease

Permanent Link: http://ufdc.ufl.edu/UFE0045476/00001

Material Information

Title: Potential Cost Benefit from Early Detection of Critical Congenital Heart Disease
Physical Description: 1 online resource (106 p.)
Language: english
Creator: Archer, Jeremy Michael
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2013

Subjects

Subjects / Keywords: cardiac -- cardiology -- congenital -- cost -- heart -- malformation -- newborn -- oximetry -- pediatric -- screening -- selection
Health Outcomes and Policy -- Dissertations, Academic -- UF
Genre: Medical Sciences thesis, M.S.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Introduction. CHD is the most common type of birth defect, and the largest cause of birth-defect related mortality. NBS using pulse oximetry is effective at detecting much CCHD, and detection of CCHD before initial hospital discharge has been associated with better short-term survival and other outcomes. Methods. Using Medicaid and CHIP claims data from infants born in 2008-2009 in Texas, with follow-up data through 2011, we conducted a quasi-experimental study using a multilevel generalized linear modeling approach to test the relationship of timing of diagnosis with long-term cost of care, controlling for multiple explanatory variables. Both instrumental variable and treatment effects modeling were used to evaluate for group selection bias. We also evaluated the overall disease burden of both groups using the CRG score, and examined differences in race/ethnicity and SES between groups. Results. Early detection of CCHD predicted a long-term health care cost reduction of 16%, or $18,000, when compared to late detection. Non-white race, extracardiac anomalies, aortic arch obstruction, and single ventricle physiology also predicted increased cost. More of those with late detection of CCHD were in the most complex CRG category. Conclusion. Early detection of CCHD predicts reduced health care cost, which could offset the increased cost associated with pulse-oximetry based NBS programs. Early detection may also be associated with increased overall disease burden and health care utilization. The association of non-white race with increased cost is unexplained.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Jeremy Michael Archer.
Thesis: Thesis (M.S.)--University of Florida, 2013.
Local: Adviser: Shenkman, Elizabeth Ann.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2015-05-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2013
System ID: UFE0045476:00001

Permanent Link: http://ufdc.ufl.edu/UFE0045476/00001

Material Information

Title: Potential Cost Benefit from Early Detection of Critical Congenital Heart Disease
Physical Description: 1 online resource (106 p.)
Language: english
Creator: Archer, Jeremy Michael
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2013

Subjects

Subjects / Keywords: cardiac -- cardiology -- congenital -- cost -- heart -- malformation -- newborn -- oximetry -- pediatric -- screening -- selection
Health Outcomes and Policy -- Dissertations, Academic -- UF
Genre: Medical Sciences thesis, M.S.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: Introduction. CHD is the most common type of birth defect, and the largest cause of birth-defect related mortality. NBS using pulse oximetry is effective at detecting much CCHD, and detection of CCHD before initial hospital discharge has been associated with better short-term survival and other outcomes. Methods. Using Medicaid and CHIP claims data from infants born in 2008-2009 in Texas, with follow-up data through 2011, we conducted a quasi-experimental study using a multilevel generalized linear modeling approach to test the relationship of timing of diagnosis with long-term cost of care, controlling for multiple explanatory variables. Both instrumental variable and treatment effects modeling were used to evaluate for group selection bias. We also evaluated the overall disease burden of both groups using the CRG score, and examined differences in race/ethnicity and SES between groups. Results. Early detection of CCHD predicted a long-term health care cost reduction of 16%, or $18,000, when compared to late detection. Non-white race, extracardiac anomalies, aortic arch obstruction, and single ventricle physiology also predicted increased cost. More of those with late detection of CCHD were in the most complex CRG category. Conclusion. Early detection of CCHD predicts reduced health care cost, which could offset the increased cost associated with pulse-oximetry based NBS programs. Early detection may also be associated with increased overall disease burden and health care utilization. The association of non-white race with increased cost is unexplained.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Jeremy Michael Archer.
Thesis: Thesis (M.S.)--University of Florida, 2013.
Local: Adviser: Shenkman, Elizabeth Ann.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2015-05-31

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2013
System ID: UFE0045476:00001


This item has the following downloads:


Full Text

PAGE 1

1 POTENTIAL COST BENEFIT FROM EARLY DETECTION OF CRITICAL CONGENITAL HEART DISEASE By JEREMY MICHAEL ARCHER A THESIS PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE UNIVERSITY OF FLORIDA 2013

PAGE 2

2 2013 Jeremy Michael Archer

PAGE 3

3 To the children and families whose lives have been chan ged by congenital heart disease

PAGE 4

4 ACKNOWLEDGMENTS I would like to thank all those who made this work possible: Betsy Shenkman, for her encouragement and guidance throughout providing access to a wonderful faculty and staff and financial support; Barry Byrne for sparking my interest in newborn screening; Bruce Vogel for his patient teaching in health economics and econometric techniques; I Chan Huang for his mentorship in patient centered outcomes Keith Muller for his assistance with the sample size calculation; Sarah Lynn e Landsman for reviewing an early draft of my proposal and suggesting methods to address selection bias; Melanie Sberna Hinojosa and John Reiss for opening my eyes to the health disparities in my clinical field; Mildred Maldonado Molina Alex Wagenaar, and Kelli Komro for teaching fundamental research methodology ; Jay Fricker and Arwa Saidi, for their support of my program of study; Katie Eddleton, for her assistance with preparin g the Institutional Review Board protocol; Chris Shaffer, Chun He, Lei Zhang, Ashley Sanders, and Deepa Ranka for p rogramming and data preparation; Jackie Hall, for geocoding ; Doug Livingston, for an excellent discussion of multilevel modeling and intra cl ass correlation; Renata Shih, Mike Mattingly, and Srini Badugu for their support and encouragement; and the faculty and staff of both the Department of Health Outcome and Policy and the Division of Pediatric Cardiology for their teaching and other contribu tions Finally, I would like to acknowledge and thank my wife, Jess Archer, for her support and patience during this work.

PAGE 5

5 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ .. 4 LIST OF TABLES ................................ ................................ ................................ ............ 8 LIST OF FIGURES ................................ ................................ ................................ .......... 9 LIST OF ABBREVIATIONS ................................ ................................ ........................... 10 ABSTRACT ................................ ................................ ................................ ................... 11 CHAPTER 1 INTRODUCTION ................................ ................................ ................................ .... 13 Critical Congenital Heart Disease ................................ ................................ ........... 14 Early and Late detection ................................ ................................ ......................... 14 Newborn Screening for CCHD ................................ ................................ ................ 14 Racial/Ethnic Disparities in the Diagnosis of CHD ................................ .................. 15 2 BACKGROUND ................................ ................................ ................................ ...... 16 Scientific and P olicy Context of Pulse Oximetry Screening for CCHD .................... 16 Early Studies ................................ ................................ ................................ .... 16 Recent Studies ................................ ................................ ................................ 18 Technical Considerations ................................ ................................ ................. 20 Satu ration cutoff ................................ ................................ ......................... 20 Confirmation of abnormal saturation ................................ .......................... 21 Testing time ................................ ................................ ............................... 21 Pre and post ductal saturation ................................ ................................ .. 21 Oximetry equipment ................................ ................................ ................... 22 Prenatal Diagnosis ................................ ................................ ........................... 22 Policy And Implementation ................................ ................................ ............... 23 Previous Economic Analyses ................................ ................................ ........... 23 Su mmary ................................ ................................ ................................ .......... 24 Racial/Ethnic Disparities in CHD Care ................................ ................................ .... 25 Diagnosis ................................ ................................ ................................ .......... 25 Overall Outcomes ................................ ................................ ............................. 26 Operative Timing and Quality ................................ ................................ ........... 27 Neurodevelopmental Outcomes ................................ ................................ ....... 29 Worldwide Access to Care ................................ ................................ ............... 29 Summary ................................ ................................ ................................ .......... 30 3 METHODS ................................ ................................ ................................ .............. 32

PAGE 6

6 Theoretical Framew ork ................................ ................................ ........................... 32 Specific Aims ................................ ................................ ................................ .......... 32 Primary Hypothesis ................................ ................................ .......................... 33 Secondary Aims ................................ ................................ ............................... 33 Racial/ethnic and SES disparity ................................ ................................ 33 Overall disease burden ................................ ................................ .............. 34 Survival ................................ ................................ ................................ ...... 34 Research Design ................................ ................................ ................................ .... 34 Data and Variable Specifications ................................ ................................ ............ 34 Data Sources ................................ ................................ ................................ .... 34 Research Subjects ................................ ................................ ........................... 35 Birth cohort ................................ ................................ ................................ 35 Follow up data ................................ ................................ ........................... 35 Data Workflow and Security ................................ ................................ ............. 35 Variable Definitions ................................ ................................ .......................... 36 Determining early versus late detection ................................ ..................... 36 Geocoding ................................ ................................ ................................ .. 36 Covariates ................................ ................................ ................................ .. 37 Sample Size Determ ination ................................ ................................ .................... 37 Preliminary Analysis ................................ ................................ ................................ 38 Missingness ................................ ................................ ................................ ...... 38 Normality Testing ................................ ................................ ............................. 38 Descriptive Statistics ................................ ................................ ........................ 39 Intra class Correlation ................................ ................................ ...................... 39 Park Test for Heteroskedasticity ................................ ................................ ....... 39 Selection Bias and Endogeneity ................................ ................................ ............. 40 Instrumental Variable Analysis ................................ ................................ ......... 41 Heckman Selection Model ................................ ................................ ................ 43 Modeli ng the Primary Hypothesis ................................ ................................ ........... 44 Testing Secondary Aims ................................ ................................ ......................... 45 Racial/ethnic and SES Disparity ................................ ................................ ....... 45 Overall Disease Burden ................................ ................................ ................... 46 Survival ................................ ................................ ................................ ............. 46 4 RESULTS ................................ ................................ ................................ ............... 53 Sample Size Determination ................................ ................................ .................... 53 Pr eliminary Analysis ................................ ................................ ................................ 53 Missingness ................................ ................................ ................................ ...... 53 Normality Testing ................................ ................................ ............................. 54 Descriptive Statistics ................................ ................................ ........................ 55 Intra class Correlation ................................ ................................ ...................... 55 Samp le Size Recalculation ................................ ................................ ............... 55 Heteroskedasticity ................................ ................................ ............................ 56 Selection Bias ................................ ................................ ................................ ......... 56 Instrumental Variable Analysis ................................ ................................ ......... 56 Heckman Selection Model ................................ ................................ ................ 57

PAGE 7

7 Measured Variables That Influenced Group Selection ................................ ..... 57 Modeling the Primary Hypothesis ................................ ................................ ........... 57 Testing Secondary Aims ................................ ................................ ......................... 59 Racial /Ethnic and SES Disparity ................................ ................................ ...... 59 Outcome variables ................................ ................................ ..................... 59 Selection variable ................................ ................................ ....................... 60 Covariates ................................ ................................ ................................ .. 60 Overall Disease Burden ................................ ................................ ................... 61 Surviv al ................................ ................................ ................................ ............. 62 5 DISCUSSION ................................ ................................ ................................ ......... 72 Primary Hypothesis ................................ ................................ ................................ 72 Secondary Aims ................................ ................................ ................................ ...... 73 Racial/Ethnic and SES Disparities ................................ ................................ .... 73 Overall Disease Burden ................................ ................................ ................... 74 Limitations and Threats to Validity ................................ ................................ .......... 74 Threats to Statistical Conclusion Validity ................................ .......................... 74 Threats to Internal Validity ................................ ................................ ................ 75 Threats to Construct Validity ................................ ................................ ............ 76 Threats to External Validity (Generalizability) ................................ ................... 77 Additional study limitations ................................ ................................ ............... 77 Future Directions ................................ ................................ ................................ .... 78 Conclusion ................................ ................................ ................................ .............. 78 A PPENDIX A ADDITIONAL MODEL RESULTS ................................ ................................ ........... 79 B CODE USED IN ANALYSES ................................ ................................ .................. 81 SAS Code ................................ ................................ ................................ ............... 81 STATA Code ................................ ................................ ................................ ........... 95 REFERENCES ................................ ................................ ................................ .............. 96 BIOGRAPHICAL SKETCH ................................ ................................ .......................... 106

PAGE 8

8 LIST OF TABLES Table page 2 1 Vulnerable populations in CHD ................................ ................................ ........... 31 3 1 Outcome variables ................................ ................................ .............................. 47 3 2 Pr imary independent variable ................................ ................................ ............. 47 3 3 Covariates ................................ ................................ ................................ .......... 48 3 4 Internal variables ................................ ................................ ................................ 49 3 5 Algorithm for assigning early versus late detection ................................ ............. 49 3 6 Diagnosis list A: congenital heart disease ................................ .......................... 50 3 7 Diagnosis list B: critical illness ................................ ................................ ............ 50 3 8 Prematurity coding ................................ ................................ .............................. 52 3 9 Low birth weight coding ................................ ................................ ...................... 52 4 1 Missing data ................................ ................................ ................................ ....... 63 4 2 Descriptive statistics for continuous data ................................ ............................ 63 4 3 Descriptive characteristics of categorical data ................................ .................... 64 4 4 Final outcome model results ................................ ................................ ............... 69

PAGE 9

9 LIST OF FIGURES Figure page 3 1 Data workflow ................................ ................................ ................................ ..... 47 3 2 Hospitals in Texas. ................................ ................................ ............................. 51 4 1 Distribution of raw cost ................................ ................................ ....................... 65 4 2 Distribution of log transformed cost ................................ ................................ .... 65 4 3 Comparison of log transformed cost ................................ ................................ ... 66 4 4 Comparison of CRG ................................ ................................ ........................... 66 4 5 Comparison of interhospital distance ................................ ................................ .. 67 4 6 Park test for heteroskedasticity ................................ ................................ ........... 68 4 7 Log transformed cost by racial/ethnic category ................................ .................. 70 4 8 Median income by racial/ethnic category ................................ ............................ 71 4 9 Percent of children in poverty in census tract by race/ethnicity .......................... 71

PAGE 10

10 LIST OF ABBREVIATIONS CCHD Critical congenital heart disease CHD Congenital heart disease CHIP ealth Insurance Program CRG Clinical Risk Group DHHS Department of Health and Human Services GLM Generalized linear model ICC Intra Class Correlation Coefficient ICD 9 International Classification of Disease, 9 th edition ICHP Institute for Child Health Policy IRB Institutional Review Board IV Instrumental variable NBS Newborn screening OLS Ordinary least squares PHI Protected h ealth i nforma tion PI Principal i nvestigator SES Socioeconomic Status UF Univer sity of Florida US United States

PAGE 11

11 Abstract of Thesis Presented to the Graduate School of the University of Florida in Partial Fulfillment of the Requirements for the Degre e of Master of Science POTENTIAL COST BENEFIT FROM EARLY DETECTION OF CRITICAL CONGENITAL HEART DISEASE By Jeremy Michael Archer May 2013 Chair: Elizabeth Shenkman Major: Medical Sciences Health Outcomes and Policy Introduction. CHD is the most common type of birth defect, and the largest cause of birth defect related mortality. NBS using pulse oximetry is effective at detecting much CCHD, and detection of CCHD before initial hospital discharge has been associated with better sho rt term survival and other outcomes. Methods. Using Medicaid and CHIP claims data from infants born in 2008 2009 in Texas, with follow up data through 2011, w e conducted a quasi experimental study using a multilevel generalized linear model ing approach to test the relationship of timing of diagnosis with long term cost of care, controlling for multiple explanatory variables. Both instrumental variable and treatment effects modeling were used to evaluate for group selection bias. We also evaluated the over all disease burden of both groups using the CRG score, and examined differences in race/ethnicity and SES between groups. Results Early detection of CC HD predicted a long term health care cost reduction of 16%, or $18, 0 00, when compared to late detection. Non white race, extracardiac anomalies, aortic arch obstruction, and single ventricle physiology also predicted increased cost. More of those with late detection of CCHD were in the most

PAGE 12

12 complex CRG category. Conclusion Early detection of CCHD predicts reduced health care cost, which could offset the increased cost associated with pulse oximetry based NBS programs. Early detection may also be associated with increased overall disease burden and health care utilization. The association of non white rac e with increased cost is unexplained.

PAGE 13

13 CHAPTER 1 INTRODUCTION CHD is the most common type of birth defect, affecting between 0.8 and 1% of intervention in the first month of life. 1 4 CHD is also the single b iggest cause of infant mortality attributable to birth defects. 5 6 Moreover, neonates who are discharged into the community prior to diagnosis o f their CCHD and then return to medical care have worse short term outcomes including preoperative condition and operative mortality 7 8 To investigate a potential means of avoiding delayed diagnosis, several large studies have shown pulse oximetry screening in the newborn nursery can d iagnose most C CHD with reasonable accuracy. 9 11 Indeed, universal screening programs have been implemented in sever al European countries and at least eight U.S. states with pending legislation in 22 and pilot programs in seven others 12 13 and both the US DHHS and the American Academy of Pediatrics have recommended universal NBS with pulse oximetry 14 While there have been cost analyses of limited screening programs 10 11, 15 18 and cost effectiveness models based upon anticipated program implementation 19 21 there is little data on the actual difference in health care cost from early diagnosis of CCHD whether by a NBS program or through other means We used the administrative claims databases of the Texas Medicaid and CHIP programs to compare paid amounts in the first four years of life between infants whose C CHD was detected before or during their birth hospitalization and those with delayed diagn osis of CCHD Such a direct cost comparison has not been reported before. Although this study does not address a universal NBS program or pulse oximetry per se our hope is to provide insight into the direct effect of timely diagnosis on actual costs i n a Medicaid/CHIP

PAGE 14

14 population in a large and diverse state. We believe that this study will inform clinical and policy decisions as individual hospitals and states enact NBS programs for CCHD. Secondary analyses included the overall disease burden a s well as racial/ethnic and SES disparities in diagnosis and cost. Critical Congenital Heart Disease CCHD is defined as CHD requiring surgery or transcatheter intervention within the first 28 days of life, although some authors extend this window to one year and concept. 1, 21 The U.S. Department of Health and Human Services defines seven specific defects as CCHD hypoplastic left heart syndrome, pulmonary atresia (with intact septum), tetralogy of Fallot, total anomalous pulmonary v enous return, 22 focusing screening efforts on cyanotic disease 23 despite clear evidence that left sided obstructive disease, which is often non c yanotic and therefore often missed on physical exam ination, may have worse outcomes if diagnosis is delayed. 8, 24 27 Early and Late detectio n We define d early detection as diagnosis of CCHD prior to discharge from the birth hospitalization. This can occur antenatally or postnatally. We define late detection as diagnosis of CCHD after discharge from the birth hospitalization, or initial diagn osis if birth was not in a hospital setting. Newborn Screening for CCHD Those with late detection of CCHD typically have worse survival and physiologic outcomes than those with early detection, and this is particularly true in cases of left sided obstructive disease, where the presentation following closure of the ductus

PAGE 15

15 arteriosus can include acidosis, severe cardiogenic shock, and death 7 8, 24 27 Cyanosis is a common, although not universal, clinical feature of CCHD and when mild may not be detectable on physical examination. Pulse oximetry offers a simp le and relatively inexpensive adjunct to clinical examination as a potential screening tool for C CHD, and multiple studies have demonstrated that pulse oximetry combined with clinical examination has moderate sensitivity and high specificity with relativel y few false positives for detecting CCHD 9 1 2, 18, 28 38 Racial/ E thnic Disparities in the Diagnosis of CHD Important outcomes in CHD include diagnosis rates, survival, surgical timing and mortality, transplantation results, and neurocognitive development. While there has been significant investigation into the disparities found in adult cardiovascular care and outcomes, far less work has been done to investigate the determinants of and disparities in care and outcomes for children with CHD. 39 Study of this topic is complicated by the fact that both overall and lesion specific prevalence rates vary by both gender and race/ethnicity, so analyses must adjust for these poorly understood but well re cognized differences. 40 41 As a secondary aim of this paper, we will investigate the role of racial and ethnic disparities, in particular, in the timing of diagnosis and cost of care for children with CCHD

PAGE 16

16 CHAPTER 2 BACKGROUND In this chapter we provide a scientific and policy context for the move towards universal NBS for CCHD in the United States, as well as detailing some its technical aspects We then r eview the current status of health disparities research as it relates to the diagnosis and long term outcomes of CHD. Scientific and Policy Context of Pulse Oximetry Screening for CCHD Early Studies In the early 2000s, studies of pulse oximetry focused on feasibility and test performance characteristics. Hoke and colleagues performed pre and post ductal saturation measurements on 2876 infants, demonstrating a sensitivity of 85% for left sided obstructive lesions and a sensitivity of 79% for other forms of disease, using a relatively low saturation cutoff of <92% and a wide pre /post ductal difference cutoff of 28 They also showed that in left sided obstructive disease, pre and post ductal saturations differed wh ile in cyanotic, CHD, the two measurements did not differ. screening post (75%) cases of isolated congenital hear t disease. 29 Importantly, this study also demonstrated the utility of saturation screening in detecting non cardiac disease as well. Kopell and colleagues used the same cutoff but screened at 24 hours of age or old er, achieving a false positive rate of <0.001%. 30 The low false positive rate was presumably because waiting until after 24 hours of life to screen allowed for a more complete transition to extra uterine physiology, although it should be noted that their population sample also had a lower prevalence of CHD than has been published

PAGE 17

17 elsewhere. Multiple other single center studies in the U.S., Italy, Switzerland, Thailand, and Saudi Arabia confirmed the feasibility of implementing pulse oximetry as an adjunct to clinical examination, as well as its moderate to high sensitivity (67 100%), and high specificity (99 100%) when using post 31 34 Meberg and colleagues reported the first multicenter study of pulse oximetry screening, in Norway, using a post ductal saturation measurement on admission to the nursery with a cutoff of <95%. 9 This study incorporated a repeat measurement to confirm abnormal result s, demonstrating a sensitivity o f 77.1%, a specificity of 94%, and a false pos itive rate of 0.6%. As in prior studies, 29 30 more than half of the abnormal screens detected non cardiac disease, although this is likely reflected testing during the physiologic transition period from fetal to neonatal life. When this study was re analyzed with non screening regions of Norway as a control group, 12% of CCHD was missed in the screened group versus 23% in the control group; of the missed cases 82% were left sided obstructive disease. 36 There were no deaths from unrecognized CCHD, and cost was not stu died. A multicenter Swedish study combined pulse oximetry, with a post ductal cutoff /post ductal difference cutoff of >3% and clinical examination to report a sensitivity of 8 2.8% and a specificity of 97.9% with a false positive rate of 2.1%. 11 This study also compared results with a non randomized control cohort, finding fewer deaths complications in the screen ed region although the outcomes were too rare to draw conclusions Unfortunately, both of these quasi experimental approaches with non randomized control groups failed to meaningfully address group selection bias

PAGE 18

18 in their desi gn and analysis and so their inferences conclusions regarding outcome in the screened versus unscreened groups are relatively weak. Recent Studies At least one other multicenter study substantiated early results regarding the test performance characteristi cs of pulse oximetry, 17, 42 although two studies of moderate size (15233 and 7672 patients) did show that the addition of routine pulse oximetry failed to yield any additional diagnoses of CCHD 18, 43 Systematic reviews of early studies with pooled data analysis corroborated the moderate though variable sensitivity and high specificity found in early individual studies, and suggested further research 44 45 as did government sponsored evidence reviews in the U.S. (Tennessee) 46 and the U.K. 19 In the Middle Tennessee region, one such further study was performed after mandatory pulse oximetry screening was proposed in the state legislature. 12, 46 Because screening was voluntary, seven of thirty hospitals, representing over half the births during the study period, did not participate. Of the 15,564 infants screened, 111 had an abnormal screen but only three were referred for further evaluation, a signific ant failure of the screening program. This pilot study utilized a cutoff of <94% with a single post ductal measurement after 24 hours of age, with no repeat measurements, and points to the need for confirmatory repeat testing as well as a well developed r eferral network. Tennessee has gone on to pass a law mandating universal pulse oximetry screening. 13 In a large multicenter study in Poland with 52,993 infants, Turska Kmiec and colleagues demonstrated that a post ductal measurement with a repeat confirmation of all positive results had good sensitivity and sp ecificity, as well as nearly universal

PAGE 19

19 acceptability to parents. 38 Nearly 20% of CCHD was detected solely on the basis of the pulse oximetry screening alone in this stud y. Bradshaw and colleagues demonstrated the feasibility of implementing a screening program with pre and post ductal measurement in a community hospital, reporting minimal additional cost and no additional staffing needs, a mean screening time of 3.5 minu tes, and barriers to screening identified in only 2.4% of patients. 47 The great majority of barriers reported related to sc reening equipment, staff workload, and crying or active newborns making the screening difficult. In Great Britain, the landmark PulseOx study was a prospective evaluation of 20,055 newborns, in which all received both antenatal screening ultrasound and new born pulse oximetry screening (pre and post >2% difference cutoffs, with a repeat confirmation of abnormal screens). 10 All infants were followed to 12 months of age to confirm diagnosis and outcome. Their findings incl uded an overall sensitivity of 75% and specificity of 99% 48 This study was particularly valuable because it was done in the context of a rigorously measured antenatal detection rate of 50%, reflecting the added value of newborn pulse oximetry even with relatively high rates of prenatal diagnosis consistent with other areas. 49 52 A related parent survey demonstrated that screening did not increase parental anxiety, althou gh this effect was different among different ethnic groups. 53 Finally, Thangaratinam and colleagues performed a meta analysis based upon all published studies including the PulseOx study, find ing an overall sensitivity of 76.5% and specificity of 99.9%, with false positive rates of 0.05% when screening was conducted after 24 hours of life. They concluded based upon the evidence to date that

PAGE 20

20 etry screening for CCHD Based upon the extant evidence and an advisory committee recommendation, 16 the U S DHHS approved the implementation of pulse oximetry as part of the NBS panel routinely performed for all newborn infants. 54 Technical Considerations As the studies of pulse oximetry h ave highlighted, there are at least five important technical considerations involved in testing: the saturation cutoff, confirmatory measurement for abnormal tests, testing time, the measurement of a pre and post ductal saturation difference, and the oxi metry equipment itself. Saturation cutoff the single saturation or post ductal measurement, and from >3% to > 7% for the pre /post ductal difference. In 2005, de Wahl Granelli and colleagues measured saturation at >12 hours of life in 66 newborns with CCHD and 200 normal newborns, to compare test performance characteristics using various cutoffs. They found that using a cutoff of and post specificity of 96.0%, an positive predictive value of 88.9%. 55 Unfortunately this study d id not report receiver operator curves, limiting analysis to cutoffs previously published in the literature. Still, most recent studies have used a post a pre /post ductal difference cutoff of >2 3%. 9 11, 38, 47 Testing at higher altitudes may require the development of different saturation cutoffs. 15, 56

PAGE 21

21 Confirmation of abnormal saturation Repeat measurements were introduced in the Swedish study 9 and have been incorporated into most large studies since 10 11 Confirmatory measurements are felt to increase te st accuracy and reduce the potential for false posi tive results. In fact, the US DHHS advisory committee report recommends two separate confirmations of saturations between 90 and 95%, stressing that in infants with saturation of <90% evaluation should no t be delayed for repeat measurement. 16 Testing time Although one study found that at least six minutes of testing was necessary for maximum reliability 18 others have demonstrated the feasibility and low cost of testing lasting 1 5 minutes 16 and the average testing time in the PulseOx study was 3.5 minutes. 10 Pre and post ductal saturation A difference between pre and post ductal saturation (typically measured on the right hand and right foot) may reflect differential perfusion due to left sided obstructive disease, which may not necessarily be cyanotic. 10 11, 28, 31 32 The effort to detect pre and post ductal differences and is clearly worthwhile, although focusing on relative pulse amplitude (reflecting peripheral perfusion) may be a better approach. 55, 57 Considering that even the earliest studies in this area focused on the severe consequences and relatively high rate of non detection of left sided obstructive disease, 8, 11, 24 28, 58 the DHHS decision to define CCHD purely on the basis of cyanotic disease is puzzling. 22 23

PAGE 22

22 Ruegger and colleagues 59 studied the utility of the left hand as a preductal site, and found it useful in normal infants, but there were no infants with CHD included. The ear and nose are also candidates for pre ductal measurements. Oximetry equipment Pulse oximetry technology has evolved significantly over the last decade. New generation oximeters are more accurate and have less measurement error than older generation or conventional oximeters. 11, 57 Studies of screening programs should ensure uniformity of oximetry equipment, and implementation of screening programs should, ideally, use the new gener ation equipment. Prenatal Diagnosis A major factor influencing the accuracy, cost effectiveness, and importance of NBS for CCHD is prenatal diagnosis, usually in the form of obstetric screening ultrasound with refe rral for fetal echocardiography when avail able. In the United States, prenatal diagnosis rates for critical or significant CHD range from 31.7 to 50%, 49 51 although higher rates are possible. 52 In the PulseOx study, in which every participant received a screening obstetric ultrasound leading to fetal echocardiography if warranted, 50% (12 of 24) of cases of CCHD were prenatally diagnosed. 1 0 Peiris and colle agues found that both increasing socioeconomic status and private health insurance were associated with higher likelihood of having received prenatal diagnosis. 51 All of these results point to the fact that improved NBS still has a major role to play, given that fetal ultrasound is neither highly accurate nor universally ava ilable for the prenatal diagnosis of critical congenital heart disease.

PAGE 23

23 Policy And Implementation Both the American Heart Association and the American Academy of Pediatrics hav e endorsed the US DHHS recommended the addition of CCHD to the recommended uniform screening panel. 14, 54, 60 Screening, even in the absence of a state mandate, is being performed in many hospitals; in Wisconsin, for example, 28% of hospitals representing 35% of births in 2010 were performing routine pulse oximetry screening. As of this writing, at least ni ne U.S. States have passed legislation to mandate statewide screening programs, although the Centers for Disease Control report that the majority of these are still in the planning stages and face significant financial obstacles. 13, 61 A survey in 2007 demonstrated that just over half of responding U.S. pediatric cardiologists supported mandated pulse oximetry screening, but a more recent survey in light of new evidence and recommendations has not been repeated 62 Surveys have evaluated the degree of implementation of routi ne pulse oximetry screening in other countries as well. A Swiss survey conducted in 2008, 3 years after recommendation of universal screening in that country, demonstrated that 76% of maternity units representing 85% of newborns performed routing pulse ox imetry screening. 63 In the U nited Kingdom in 2010, 7% of neonatal units reported the practice and in 2012, 18% did. 64 66 Previous Economic Analyses Typical cost estimates for pulse oximetry screening range from $0 $10 per test 10 11, 15 17 However, Reich and colleagues, in their study of non tertiary care hospitals in Florida, found a per test cost of $1,410 $2,128 per six minute screen by a Licensed Practicing Nurse, the minimum time and level of training associated with the best test

PAGE 24

24 performance. 18 Although their cost model may have been over inclusive, their very different estimate points to the possibility that actual per test costs may vary significantly. Using a decision analytic model to predict cost effectiveness in Great Britain, Knowles, Griebsch, and colleagues found an incremental cost effectiveness ratio for pulse oximetry screening program of $7,500 (in 2009 currency assuming an exchange rate of US D$0.65 per British pound) per additional timely diagnosis of CCHD 19 20 When Ewer and colleagues later adapted this decision analytic model using the test performance characteristics from the PulseOx study, this incremental cost effectiveness ratio was between $ 10 ,000 and $42,300 per additional timely diagnosis depending upon model assumptions and the incl usion of prenatal diagnosis 21 This model use s typical per test cost figures, 18 but could certainly overestimate cost effectiveness if per test cost is underestimated. A rigorous economic analysis in the Unites States has not been reported although this would certainly advisable as programs are being implemented on a state by state basis. Several autho rs point out that CCHD is more prevalent than many other diseases recommended for routine NBS in the U.S. 30, 46 The cost is substantially less than that of newborn hearing screening, the prototype for NBS using a non blood spot methodology. 16 Further cost analysis will be necessary to determine the differe nce in long term or lifetime health care cost of those with CCHD detected by screening versus those whose disease is not diagnosed in a timely fashion. Summary Multiple studies have demonstrated the effectiveness of routine pulse oximetry as an adjunct to clinical examination to screen for CCHD Given the U S DHHS

PAGE 25

25 recommendation for implementation and a number of calls for universal screening, and the fact that at least eight states have already mandated it 13, 67 69 pediatricians, cardiolog ists and others need to focus careful attention on proper implementation. This includes using an appropriate saturation cutoff, incorporation of pre ductal saturation measurement, adequate testing time, standardized and modern equipment, and the availabi lity of both echocardiography and timely referral. Although the current recommendation focuses on cyanotic disease, we should not abandon the goal of developing more effective ways to detect left sided obstructive disease, whether prenatally or by NBS Racial/Ethni c Dispari ties in CHD Care Important d omains in the care of patients with CHD include diagnosis, overall outcome, the timing and quality of operative int ervention, and longer term neurodevelopmental outcomes. All of these factors have the poten tial to affect both cost and quali ty of life in children with CHD, and all are vulnerable to health disparity. In not due to access 70 although we remain cognizant of broader definitions incorporating access to care. 71 Diagnosis Diagnosis of CHD is a prerequisite to the surgical and medical therapies that have dramatically decreased mortality over the last several decades, and thus important to outcomes. 72 Diagnosis can occur at many points of care: before birth (prenatal or fetal diagnosis), during the birth hospitalization, after an infant is discharged home but during childhood, during adulthood, and after death on autopsy. In the U.S., the rate of

PAGE 26

26 screening in both the prenatal and newborn periods is increasing, but by no means universal. 73 Additionally, timing of referral to the pediatric cardiologist or t he pediatric cardiology center directly affects diagnosis. 40 Fetal or prenatal diagnosis is acknowledged to be widely disparate in terms of access, technique, and rate of diagnosis of CHD. 51, 73 Peiris and colleague studied the interaction of race, socioeconomic quartile, neighborhood poverty level, and prenatal diagnosis in a hospital with an extremely high prenatal diagnosis rate. They found that although the rate of prenatal diagnosis was correlated with all economic predictor variables, only private health insurance w as a strong predictor (odds ratio 3.7) of prenatal diagnosis when all variables were modeled together 51 Because universal NBS for CCHD is in the early stages of implementation and not all CHD is critical 61 referral by pediatricians and family physicians for diagnosis by a pediatric cardiologist i s the nex t important time point in diagnosis. In 1993, Fixler and colleagues reported population based data from Dallas County, Texas, showing that race/ethnicity and socioeconomic status were unrelated to average age at postnatal diagnosis for nine of the most co mmon lesions. 40 Perlstein and colleagues, in 1997, reported that non urban location was associated with later referral to a pediatric cardiologist in the neonatal period, while after the neonatal period non urban location, and managed care insurance were associated with later referral. 74 This study did not include race/ethnicity in its analysis. So far, the evidence poin ts to a rural/urban disparity, but not necessarily one related to race or socioeconomic status. Overall Outcomes A large review of U.S. death certificates from 1979 1997 reveals that although population based mortality rates from CHD have declined from 2.5 to 1.5 per 100,000, a

PAGE 27

27 39% decrease, the approximately 20% mortality gap between blacks and whites has not changed to any significant degree 72 Nembhard and colle agues, in two separate large studies with population based denominators, demonstrated higher rates of childhood and early childhood mortality for nonwhites relative to whites from CHD across several lesion types 75 76 In some subgroups and in the overall analysis of the 2008 paper, Hispanic males and females had lower mortality than whites 75 However, the picture of racial disparity in CHD is not all straightforward. Recently, the Centers for Disease Control reported on 2,256 neonatal deaths attributable to CHD, out of 11 mil lion live births from 2003 2 006. 5 A higher proportion of neonatal deaths in children of white mothers (5.4%) were due to CHD versus those in children of black mothers (2.3%) In preterm infants, the neonatal death r ate due to CHD was lower for children of black versus white mothers (4.5 versus 6.8 per 10,000), while in term births, neonatal mortality rates were higher for children of black versus white mothers (1.5 versus 1.0 per 10,000). These data illustrate the c omplexity of analyzing just three factors: race, prematurity, and neonatal death. Operative Timing and Quality Age at operation has been used as an indicator of both quality and access to care in several studies of health disparities in CHD. 77 79 Additionally, hospital based surgical mortality is an important quality indicator for pediatric cardiovascular surgery centers. Erikson and colleagues analyzed California discharge data in 1992 1994 for 5071 patients undergoing congenital heart surgery, finding that those with managed care insurance had a lower chance of having an operation at a hospital with lower

PAGE 28

28 surgical mortality. 80 Interestingly, after stratification by race/et hnicity, there was no difference in access to low mortality hospitals between Medicaid and traditional private insurance, although the difference between managed care traditional private insurance persisted. Chang and colleagues also examined statewide dat a in California in 1995 1996, finding that commercial (private) versus public insurance was associated with earlier repair of four specific defects, while urban location was associated with later repair. 77 Race had little to no impact on age of repair, except that Asian ch ildren underwent later repair. These results again illustrate the complexity of untangling the factors influencing diagnosis and referral. One could theorize that urban children tend to have predominantly public insurance or some other latent factor related to less access or lower socioeconomic status, and are thus repaired later; one could also theorize that cardiologists simply referred these children for surgery later because they were close by and easier to watch medically. Similarly, there may be cultural or economic factors later referral or later family consent to surgery. In 2002, Milazzo and co lleagues reported a relatively small, single center study of timing of second and third stage palliative surgeries for single ventricle physiology performed from 1997 2000. 79 They found that African American children underwent both surgeries significantly later than their white counterparts, by a difference of six months for the cavopulmonary anastomosis (Glenn) and two years for the Fontan completion surgery. However, when a similar study was repeat using data from 29

PAGE 29

29 hospitals and a much larger sample size, no difference in median age was found, although Hispanic patients did have a longer length of ICU stay after one of t he procedures. 78 Neurodevelopmental Outcomes The study of neurodevelomental outcomes is at the forefront of pediatric cardiology, especially in the current era, in wh ich most patients with CHD are expected to survive. 81 Unfortunately, many of the landmark trials in this area have focused almost exclusively on physiologic or surgical factors, with less attention to healt h disparities or the influence of socioeconomic or demographic factors. 82 83 This has been the case even when th e focus is on social determinants of behavioral issues. 84 Atallah and colleagues did address socioeconomic status in a landmark study examining outcomes of two palliative surgery techniques for hypoplastic left heart syndrome, used in two hi storically different time periods 85 They found that socioeconomic status was a predictor of lower mental development scores with one technique, used in the earlier time period, although not with the other, used in the later time period. Male gender, on the other hand, was associated with lower psychomotor development scores in the more recent surgical time period. Worldwide A ccess to Care Finally, although this portion of the current study focuses primarily on the disparity landscape in the United States, worldwide disparities in access to pediatric cardiovascular care are significant. There is a gross mismatch in pediatric c ardiac surgery centers relative to the birth rate of infants with CHD, and a call has been made for worldwide access to improve such access in conjunction with the United Nations 2 000 Millennium Development Goals. 86

PAGE 30

30 Summary Current evidence reveals multiple disparities i n the care and outcomes of patients with CHD, disparities that cut across race, gender, socioeconomic status, and geographic location. Groups identified as vulnerable by the current evidence are summarized in Table 2 1 While several disparities have been shown in large studies and replicated multiple times, particularly the racial disparities in overall and postoperative survival, others require further and focused study. To this end, we include a descriptive analysis of racial/ethnic disparity in our st udy.

PAGE 31

31 Table 2 1. Vulnerable populations in CHD Process or Outcome Vulnerable Group(s) Prenatal Diagnosis Publicly Insured Postnatal Diagnosis Rural Access to Care Developing World Overall Mortality Non white Age at Operation Publicly Insured Urban Asian and African American Operation at Low Mortality Center Managed Care Insured Neurocognitive Outcomes Low SES

PAGE 32

32 CHAPTER 3 METHODS The study protocol was approved by the UF Health Science Center IRB (protocol IRB201200064) on November 30, 2012. A waiver of informed consent was granted. Neither the author nor any study participant declared any potential conflict of interest related to the study The SAS System version 9.3 87 was used for all statistical analysis except whe re noted otherwise. Theoretical Framework The theoretical framework for our study was based upon the current literature showing that late detection of CCHD can lead to increased physiologic derangement of multiple organ syst ems 8, 24 27 which could lead to differential outcomes. We included additional explanatory variables addressing race/ethnicity, SES, specific lesion physiology, additional birth defects, and birth related factors, each of which could potentially influence group selection, outcomes, or both. Specific Aims Based upon the review of existing literature, we chose to focus on the following questions for this study: 1. D o children with early detection of CCHD have a different health care cost in the first four years of li f e when compared with those with late detection of CCHD ? 2. What other measurable variables affect health care cost in these groups? 3. How do differences in race/ethnicity and SES affect the timing of detection and the health care cost in the first four years of life? 4. Compared with infants with late detection of CCHD, do infan ts with early detection have a different overall disease burden affecting health care utilization ?

PAGE 33

33 5. Compared with infants with late detection of CCHD, do infants with early detection have i ncreased survival in the first five years of life? Our goal was to address gaps in the literature using a methodology that rigorously evaluated and controlled for selection bias, heteroskedasticity, and clustering effects. Primary Hypothesis In children with CCHD enrolled in Medicaid and CHIP and born in Texas in 2008 and 2009, we hypothesized that those whose CCHD was detected prior to initial hospital discharge will have a lower health care cost in the first four years of life when compared with those w hose CCHD was detected after initial discharge. (Hypothesis 1) We did not distinguish between costs due directly to CCHD and costs due to other medical condition and co st. As a complementary analysis, we hypothesized that the re would be relationship between cost and multiple variables known to affect outcome in CHD 8, 88 90 or neonat al survival : 91 93 gender, multiple gestation, delivery via Cesarean section, prematuri ty, low birth weight, the presence of an extracardiac birth defect race/ethnicity, and SES as reflected in the median annual income and percent of children living below poverty by census tract or zip code midpoint. (Hypothesis 2) Secondary Aims Racial/eth nic and SES disparity We hypothesized that there would be a significant relationship between race/ethnicity and cost, possibly via a role in group selection. (Hypothesis 3A) Furthermore, we hypothesized that there would be a significant relationship between SES and cost of care, possibly via a role in group se lection. (Hypothesis 3B)

PAGE 34

34 Overall disease burden We hypothesized that the overall disease burden af fecting health care utilization, as measured in the pediatric CRG score 94 95 would be lower in those with early detection of CCHD versus those with late detection. (Hypothesis 4) Survival W e hypothesized that the overall survival would be higher in those with early detection of CCHD versus those with late detection. (Hypothesis 5). This expected result would be consistent with prior published quasi experimental work 11, 36 Resear ch Design This is a quasi experimental, retrospective d esign: NR X O NR O randomized group selection, in this case early versus late measure, paid cl aims from birth through December 31, 2011. To perform th e analysis, we used a multilevel GLM, preceded by prospective determination of sample size, testing for heteroskedasticity, and evaluation for selection bias. To perform secondary analyses, we adapted the same general design as well as using descriptive and simple comparative statistics for alternative outcomes Data and Variable Specifications Data Sources The state of Texas provide s health claims, encounter, a nd enrollment data at the person level for its Medicaid and CHIP programs on a quarterly basis. This includes

PAGE 35

35 F ee F or S er vice, and Primary Care Case Management, and Health Maintenance Organization (STAR) structures The Medicaid and CHIP enrollment and c laims data are housed by UF for quality assurance activities carried out by the ICHP pursuant to its quality assurance contract with Texas (Evalu a ting Healthcare Quality in Texas Medicaid and CHIP, UF Project Number 0068240, Elizabeth Shenkman, PI). Elizab eth Shenkman, Ph D is the Professor and Chair of the Department of Health Outcomes and Policy and the Director for the ICHP. Research Subjects Birth cohort The research subjects consisted of a ll babies born in 2008 2009 in TX who met our diagnostic criteria for CCHD and were enrolled in Medicaid of any type or CHIP at the time of birth. Follow up data The data analyzed consi st ed of all Medicaid and CHIP enrollment and claims data for all children in the birth cohort from birth through December 31, 2011. Data Workflow and Security The original database s described above, reside on a secure server in the ICHP, under contracts approved by the state of Texas and approved by the UF IRB. These databases are maintained und er strict security and data is not permitted to leave the server. For our study, the only study team members with access to the original data were the programming and geocoding team members. The programmers extracted all operational variables to an operat ional dataset at the individual study participant level. The geocoder used address and zip cod e information to extract income and poverty variables from American Community Survey 96 data and hospital distance variables from

PAGE 36

36 the American Hospital Association data, 97 and these were re merged with t he operational dataset. At this point, all PHI was removed from the dataset to form the de identified analysis dataset. The de identified analysis dataset was accessible only to the primary author (JA) who performed all data analysis and did not have acc ess to the original database s or any PHI containing data for this study. The operational workflow is summarized in Figure 3 1. Variable Definitions Seventy one v ariables were oper ationalized into outcome variables, the primary independent variable, covari ates, and internal variables as described in Tables 3 1, 3 2, 3 3, and 3 4, respectively Determining early versus late detection The algorithm used to determine early or late detection is given in Tables 3 5, 3 6, and 3 7. Infants born at home or out of t he hospital setting and diagnosed with CCHD w ere automatically considered to have late detection. Geocoding Three variables were obtained using geocoding with ArcMap 10.1 using premium street data 98 Addresses were mapped to census tract, and the median income and the American Community Survey dataset from 2010. 96 Because calculating income and poverty variables required census tract level information, subjects with no address, an address outside of Texas, a Post Office Box, or a zip code only were not assigned values for these variables. The interhospital distance was calculated by subtracting the driving time in

PAGE 37

37 hospital hospital that reported providing pediatric cardiac surgery services in the American Hospital Association survey for fiscal year 2007 97 (see Figure 3 2) For interhospital distance, subjects with no address or an address outside of Texas were not assigned a value but subjects with a zip code only or a Post Office Box were assigned a value calculated using the midpoint of t heir zip code instead of a street address. Covariates Covariates were chosen based upon established risk models for neonata l survival and disease severity, 91 93 as well as upon studies of factors influencing CCHD outcomes 8, 88 90 Aortic arch obstructi on and in particular isolated coarctation are sometimes difficult to diagnose after birth and are more common in th ose with delayed diagnosis. S ingle ventricle physiology is typically associated with a minimum of two to three palliative surgeries and so impacts both survival and cost. The birth hospital was used not strictly as a covariate, but as a clustering variable, as discussed below The interhospital distance, similarly, was used as an instrumental variable to test and control for selection bias due to unmeasured variables. Finally, the total number of months enrolled during the study period was used as a covariate after log transformation. The operational specifications of all covariates are given in Table 3 3, with diagnosis lists for prematurity and low birth weight given in Tables 3 8 and 3 9. Sample Size Determination A preliminary dataset consisting of infants born and enrolled in Texas Medicaid (all types) in 2011, was prepared for to estimate the sample size needed for the full analysis. The online program GLIMMPSE 99 based in part upon work by Creidler, Glueck and Muller, 100 can calculate power and sample size for multilevel data in a

PAGE 38

38 generalized linear model. We calculated the descriptive statistics, ICC, and the harmonic mean of the cluster size of the preliminary data, and used the GLIMMPSE program with the default Hotelling Lawley Trace method to prospectively calculate an initial sample size. We then rep eated the calculation assuming half of the difference in outcome variable and twice the variance as seen in the preliminary data to compute our final desired sample size. After the final dataset was prepared we retrospectively calculated another sample si ze based upon the actual characteristics of the data, to detect a difference in outcome half as large as was actually observed. This was used to confirm that our actual sample size was adequate. Preliminary Analysis Missingness The percentage of all varia bles missing was determined. For variables with more than 5% missingness, a chi squared test was used to determine if the degree of missingness was significantly different between the early and late selection groups. Missing value s were not filled in or imputed, and observations with missing values were dropped from the final analysis. Normality Testing To determine if cost would satisfy the assumption of normality it was analyzed using the SAS univariate procedure to determine if the variable fit a stan dard normal distribution using visual analysis of a histogram as well as the kurtosis and skewness.

PAGE 39

39 Descriptive Statistics For continuous variables, the SAS ttest procedur e was used to compare means between the ear ly and late detection groups. Similarly, the SAS freq procedure was used to determine the overall proportion of subjects with each value of categorical variables, and then to compare proportions between detection groups using the chi square test. Intra class Correlation ICC for clustered data is given by (3 1) where is the between group variance and is the within group variance. For this study, the SAS mixed procedure was used to determine the within and between group variances and to calculate the ICC. This was performed once while ignoring subjects with a missing birth hospital, and repeated with all subjects having a missing birth hospital treated as if they were in a single cluster. Park Test for Heteroskedasticity Manning and Mullah y 101 describe the Park test for heteroskedasticity in log transformed models of health care expenditures. This test regresses log transformed predicted value against a log trans formed residual uses the estimated coefficient to choose the distribution function used in the final model. The simplified form of the procedure described by Manning and Mullahy 101 is given by (3 2) (3 3)

PAGE 40

40 where ( 3 2) is either a GLM of a raw cost with a log link or an OLS model of log transformed cost, and ( 3 3) is a linear regression with the predicted value if a GLM was used initially or to correct for retransformation bias if an OLS model was used initially The value of determines the probability distribution for the final GLM: if there is no appreciable heteroskedasticity and a Gaussian distribution is appropriate; if Poisson is used; if a gamma distribution is used; if an inverse Gaussian distribution is used. In this study, the Park test was performed three times The first was with an OLS regression using the SAS glm procedure with a the appropriate correction for retransformation bias ; the second was with a GLM using the SAS genmod procedure and a Poisson distribution, and the third was with a GLM again using the SAS genmod procedure and a gamma distribution. In all three cases the SAS reg procedure was used to model equation ( 3 3) and determine For the Park test, multilevel modeling was not used. Selection Bias and Endogeneity Selection bias refers generally to the nonrandom sampling of the population. One specific case of this is group selec tion bias, in which nonrandom samples of the population are selected into groups for analysis. Although these are often referred to as investigating a treatment. When we r efer to selection bias, we mean group selection bias, although we do address the general issue of nonrandom sampling in Chapter 5. Because group selection was neither randomized nor under experimental control, there

PAGE 41

41 was felt to be a high potential for g roup selection bias in our study, due to both measured and unmeasured variables. Endogeneity refers to the correlation between the selection variable and the error term, which can incorporate unmeasured variables measurement error or dynamic features o f the equation system Selection bias and endogeneity are related, but not identical concepts, and endogeneity of the selection variable can be one re ason for group selection bias. In the absence of randomization, multiple approaches exist to detecting an d c ontrolling for selection bias and endogeneity Many health outcomes researchers have used propensity scoring, 102 105 in which the probability of selection into one group is determined based upon measured variables. The propensity score is then used to match the groups, as a control variable in the final model, or both. The disadvantage o f this approach is that it fails to account for unmeasured variables, and may not control for endogeneity due to unmeasured variables. Instrumental Variable Analysis An alternative approach to endogeneity and selection bias, which accounts for both measure d and unmeasured variables, is the use of instrumental variables. The instrumental variable, or instrument, should be strongly related to the selection variable and account for unmeasured variables in that relationship, and it should be otherwise independ ent of the outcome variable. 106 Using the differential distance from the nearest hospital to a hospital offering specialty care as an instrument has been successful in previous neonatal 106 107 and i n the cardiovascular 108 109 studies. For that reason, we chose differential distance to a pediatri c cardiac surgery center as an instrument.

PAGE 42

42 A standard instrumental variable approach involves predicting the probability of selection into a reference group based on the instrument and other variables. The predicted probability is typically used in place of the actual group selection variable in the final model, although it has been used for group matching as well 106 A standard instrumental variable model is represented by (3 4) (3 5) where is the outcome variable, is the group selection variable represents the predicted probability of being selected into a reference group, is the instrumental variable and represents all additional exogenous covariates some of which may influence group selection. In ( 3 5), note that the outcome variable i s calculated based on the predicted probability, not the actual group selection var ia ble Wooldridge 110 describes a three stage instrumental variable approach given by (3 6) (3 7) (3 8) in which the predicted probability of selection into a reference group calculated in ( 3 6) is used as an instrument with the same covariates as in ( 3 7), producing a second predicted probability of selection into a reference group that is then used in the final outcome model ( 3 8). In both ( 3 6) and ( 3 7), the selection variable is the left hand side of the equation; in ( 3 8) it is the outcome variable. The primary advantages of this three stage model are increased accuracy and efficiency if the model for is mis specified.

PAGE 43

43 Our instrumental variable approach involved computing predi ctor variables using both the two and the three stage methodology. The SAS logistic procedure was used to model the predictor variables. The strength and significance of the parameter estimate for the instrument in predicting the selection variable was e xamined to assess the strength of the instrument itself. Heckman Selection Model In addition to evaluating the suitability of the chosen instrument during the modeling process we used the Heckman selection model 111 as a treatment effects model to test the endogeneity of the selection variable and the potential for its influence on group selection This involves treating both the selection an outcome variable as endogenous variables, simultaneously modeling a prob it regression of the selection variable and a linear regression of the outcome variable 112 as given b y (3 9) (3 10 ) Note that represents a set of exogenous variables modeled with the group selection variable, while represents a separate set of exogenous variables modeled with the outcome variable. A nonzero relationship between the disturbance terms and is evidence of group selection bias Adding exogenous variables (regressors) in the outcome equation that are not present in the prediction equation can make this approach more robust. 112 The Heckman selection model simultaneously tests and controls for selection bias.

PAGE 44

44 In our study we used the STATA 113 treatreg command with a maximum likelihood model to perform a Heckman selection model In this model, the statistic represents the correlation of disturbance terms. Modeling the Primary Hypothesis After assessing missingness, data normality, heteroskedasticity, and endogeneity and selection bias, we modeled our data using a multilevel generalized linear model with a log link. We ran multiple candidate models using all covariates and narrowed this d own to a parsimonious model. Based upon the results of the preliminary analyses (presented in Chapter 4), we chose a gamma distribution function and did not incl ude the instrument al variable in the final model. Because preliminary modeling did not find a significant relationship between months enrolled and cost, and because we did not use a Poisson distribution function in the final model, we included the log transformed version of total months enrolled as a covariate rather than an offset. Detailed desc riptions of candidate models, the final model, and alternative models are presented in Chapter 4 and Appendix A. Although others have used mixed models to account for random effects, 114 we were unable to achieve convergence with a generalized linear mixed model using the SAS glimmix procedure, presumably due to the complexity of the nonuniform clustering. A GLM or generalized estimating equation without random effects terms, using the SAS genmod procedure with an exchangeable covariance matrix, was found to be more robust and was us ed for all final data modeling. The final model allowed us to test Hypothesis 1, while the progression from the maximal to parsimonious model allowed us to address hypothesis 2.

PAGE 45

45 Estimates of change in cost by individual independent variables can be calculated using the parameter estimate for variable and the overall sample mean cost: ( 3 11) Because a GLM is used there is no need to adjust for retransformation bias when re exponentiating the parameter estimate Testing Secondary Aims Racial/ethnic and SES Disparity The continuous variables including cost, interhospital distance, income, poverty, and enrolled month were compared across race/ethnicity groups using analysis of variance with the SAS anova procedure. The categorical variables including gender, prematurit y, low birth weight, multiple gestation, Cesarean section, aortic arch obstruction, aortic coarctation, single ventricle physiology, and clinical resource group were compared across race/ethnicity groups using chi squared with the SAS freq procedure. To further assess for hidden or subtle disparities, these analyses were repeated comparing only two categories: white and non white race/ethnicity. This was accomplished using the SAS ttest procedure for continuous variables and the SAS freq procedure for categorical variables. These analyses addressed Hypothesis 3A. Finally, to assess the relationship of income and poverty to the outcome and the selection variable s both of these variables were included in several candidate models of the outcome variable Additionally, their means were compared in the early versus late detection group using the SAS ttest procedure. Because of missingness, however, they were not included in the final model. These analyses addressed hypothesis 3B.

PAGE 46

46 Overall Disease Burde n The CRG scoring system was developed to predict health care resource utilization 94 and then adapted for pediatric use. 95 It is readily calculable using claims data and may broadly reflect overall disease burden as it relates to healthcare utilization. We must be clear that it has not, to our knowledge, been used in this fashion before, and this is a preliminary investigation. Our original intent had been to use maximum likelihood estimation will be used to compare the CRG score between detection groups, with a similar group of covariates as used in the final primary outcomes model. However, because of a high degree of missingness and concerns over the standardization of timing of the CRG given the different ages of the patients at the conclusion of the study, we opted inst ead to describe the differences in CR G between the early and late detection groups, analyzing it both as a continuous variable using the SAS ttest procedure and as a polychotomous, ordinal variable using chi square with the SAS freq procedure. This addressed Hypothesis 4 but did not test it rigorously. Survival We originally planned to conduct a proportional hazards regression using the SAS phreg procedure. However, this analysis not performed due to incomplete and presumably inaccurate survival data. Hypothesis 5 was therefore not tested.

PAGE 47

47 Figure 3 1. Data workflow Table 3 1. Outcome variables Variable Description Type Operational Specification Cost Total paid amount in the first five years of life. Continuous Exp_Med: total Medicaid expenditures from birth through end of CY2011 ($0.00) Comorbidity Index Clinical Risk Group 95 Ordinal CRG SUM (code as integer): from CY2011 or most recent available year 1: 1Healthy 2: Significant Acute 3: SHCN Minor 4: SHCN Moderate 5: SHCN Major Missing : Unassigned Survival Age at death (789.x), or survival to age 5 Continuous Survival: survival days at death (the death date is as the first date with 789.x), or survival to December 31 2011 (integer) F_Death: flag of death (integer) 0=N 1=Y Table 3 2. Primary independent variable Variable Description Type Operational Specification Detection Group Early or late detection of CCHD. Dichotomous F_grp: Early/Late (see Tables 3 5, 3 6, 3 7)

PAGE 48

48 Table 3 3. Covariates Variable Description Type Operational Specification Gender Gender. Dichotomous Sex: (M/F) Race Race. Categorical Race_TXT: White, Non Hispanic Black, Non Hispanic Hispanic Asian, Pacific Islander American Indian or Alaskan Unknown / Other Income Median annual income in census tract. Continuous ($0.00) from geocoding Poverty Index Percent of children below poverty in census tract. Continuous (xx%) from geocoding Prematurity Gestational age < 37 weeks at birth. ( See Appendix B) Categorical F_PReM_0 F_PreM_9 (see Table 3 8 ) (integer) 0=No ; 1=Yes Cesarean Section Cesarean section as del ivery route ( 763.4 ) Dichotomous F_CS: (integer) 0=N ; 1=Y Low Birth Weight Low birth weight. ( See Appendix C) Categorical F_LBW0 F_LBW9 ( See Table 3 9 ) (integer) 0=N ; 1=Y Extracardiac Anomaly Additional non cardiac congenital conditions present at birth. (Any 740.x 744.x, 747.4 9, 748.x 759.x ) Dichotomous F_EA: (integer) 0=N ; 1=Y Multiple Gestation Member of multiple gestations. ( V31 V37) Dichotomous F_MG (integer) 0=N ; 1=Y CHD diagnosis or diagnoses ( Table 3 6 ) F_CHD_1 F_CHD_18 (integer) 0=N ; 1=Y Birth Hospital Birth Hospital Categorical BLNG_Prov_Name Months Enrolled Number of months enrolled from birth through Dec 2011 Integer Enroll_Mo Interhospital Distance Distance in miles (driving miles) from closest hospital to closest pediatric heart surgery center) Continuous from geocoding

PAGE 49

49 Table 3 4. Internal variables Variable Description Type Operational Specification Age_last_day Age at last day of analysis (days) Continuous Age (days) on Dec. 31, 2011 Birthdate Birth date Date bthdate: date at birth YYYY MM DD Death Date Death date Date d_date: date at death YYYY MM DD Critical Illness diagnoses Dichotomous F_CRIT_1 F_CRIT_14 (see Table 3 7 ) (integer) 0=N ; 1=Y Address Patient Address Of Residence at Birth String Internal variable to be used for geocoding removed after geocoding. Table 3 5. Algorithm for assigning early versus late detection Group 1. Infants with Late Diagnosis of Critical Congenital Heart Disease Conceptual Definition: Neonates presenting with symptoms of critical illness, who have congenital heart disease that was not diagnosed during the birth hospitalization. Operational Case Definition: At least one encounte r with diagnosis from Diagnosis List A (congenital heart disease) during days 1 28 of life. AND Birth hospitalization does NOT have a diagnosis from Diagnosis List A (congenital heart disease). AND At least one diagnosis from Diagnosis List B (critical illness) at any time during first 28 days of life. Group 2. Infants with Early Diagnosis of Critical Congenital Heart Disease Conceptual Definition: Neonates with critical congenital heart disease diagnosed during the birth hospitalization. Operational Case Definition: At least one diagnosis from Diagnosis List A (congenital heart disease) made during the birth hospitalization.

PAGE 50

50 Table 3 6 Diagnosis l ist A : congenital heart disease ICD 9 code Diagnosis 745.1x transposition or double outlet right ventricle 745.2 Tetralogy of Fallot 745.3 single ventricle 745.6 atrioventricular canal defect 746.0x pulmonary stenosis or atresia 746.1 tricuspid atresia or stenosis 746.2 746.3 aortic stenosis 746.5 mitral stenosis 746.7 hypoplastic left heart syndrome 746.81 subaortic stenosis 746.83 subpulmonic stenosis 746.84 747.1x coarctation, hypoplasia, interruption of aortic arch 747.22 aortic atresia 747.3 pulmonary atresia or hypoplasia Table 3 7 Diagnosis list B : critical illness ICD 9 code Diagnosis 276.2 metabolic and lactic acidosis 427.5 cardiac or cardiorespiratory arrest 775.8 acidosis in newborn 785.51 cardiogenic shock 785.59 circulatory shock 785.5 shock unspecified 785.9 other symptoms involving cardiovascular symptom 786.03 apnea 786.06 tachypnea 786.09 respiratory other 798.x death 799.0x asphyxia and hypoxemia 799.1 respiratory arrest 799.82 ALTE

PAGE 51

51 Figure 3 2. Hospitals in Texas. Hospitals with pediatric cardiac surgery services in red; all other hospitals in blue. Data from American Hospital Association survey, 2007.

PAGE 52

52 Table 3 8. Prematurity coding ICD 9 c odes Category Assigned 765.29 Not premature, >37 weeks GA (default) 765.1, 765.20 Premature, unspecified: 765.21 Premature, < 24 weeks GA 765.22 Premature, 24 weeks GA 765.23 Premature, 25 26 weeks GA 765.24 Premature, 27 28 weeks GA 765.25 Premature, 29 30 weeks GA 765.26 Premature, 31 32 weeks GA 765.27 Premature, 33 34 weeks GA 765.28 Premature, 35 36 weeks GA Table 3 9. Low birth weight coding ICD 9 c odes Category Assigned 765.09 Normal birthweight >=2499 gm (default) V21.30, 764.10, 765.10 LBW, unspecified 764.01, V21.31 LBW, < 500gm 764.02, 765.02 LBW, 500 749 gm 764.03, 765.03, V21.32, 765.00 LBW, 750 999 gm 764.04, 765.14 LBW, 1000 1249 gm 764.05, 765.14, V21.33, 765.10 LBW, 1250 1499 gm 765.06, 765.16 LBW, 1500 1749 gm 765.07, 765.17, V21.34 LBW, 1750 1999 gm 765.08, 765.19 LBW, 2000 2499 gm

PAGE 53

53 CHAPTER 4 RESULTS Sample Size Determination Preliminary data consist ed of subjects with CCHD (n=1711) born in 2011 with cost follow up data through December 1, 2011 Based on analysis of the preliminary data, we calculated clustering by birth hospital with 12 members per group (in this case obtained by taking the harmonic mean of the cluster sizes in the preliminary dataset), an ICC of 0.16, an early detection group 3 times the size of the late detection group, a mean log t ransformed cost difference detection threshold of 0.6 between early and late groups and a standard deviation of 1.78. Based upon th ese assumptions, a power of 0.9 vari ance, would require a sample size of 940 subjects. Further uncertainty was introduced by the challenge of unbalanced cluster sizes, so our goal for final sample size was to have at least twice the calculated size, or 1880 subjects. Preliminary Analysis Th e final dataset contained data from 3267 subjects with CCHD 2602 ( 79.6 %) of whom had early detection of CCHD and 665 ( 20.4 %) of whom had late detection The overall rate of CCHD based upon 806,8 41 live births in Texas in 2008 and 2009 115 116 was 4 per 1000. This is consistent with previously published estimates. 1 Miss ingness Missingness is summarized in Table 4 1. The primary outcome variable was missing in only one case, but it was zero in 69 additional cases that were therefore unsuitable for subsequent logarithmic

PAGE 54

54 transformations These observations were assumed to be erroneous as a cost of $0 for a child with CCHD is implausible, and they were dropped from all outcome analyses The CRG was missing in 414 (12.7%) cases 326 ( 12.5 %) in the early detection group and 88 ( 13.2 %) in the late detection group. The propor tion missing was not significantly different between groups (p=0.6, chi square). The interhospital distance was missing in 104 cases (3.2%) 91 (3.5%) in the early detection group and 13 (2.3%) in the late detection group. These data were missing in subje cts for whom addresses or zip co des in Texas were not available from the original claims data. The proportion missing was not significantly different between groups (p=0.1, chi square). These cases were dropped from all instrumental variable and treatment effect models but analyzed in the final models. Income and poverty data were missing in 443 (13.6%) subjects, 352 (13.5%) in the early detection group and 91 (13.7%) in the late detection group. These data were missing in those who not have street address es that mapped appropriately to the census tracts used for geocoding. The proportion missing was not significantly different between groups (p =0.9, chi square ). Gender was missing for one subject. All other variables were extracted and coded with a defaul t value and thus had no missing values. Normality Testing The distribution of raw cost is shown in Fi gure 4 1. Based on unacceptably high skewness (5.1) and kurtosis (39.5) it was felt to be unsuitable for models that assume a normal distribution of the o utcome variable. The distribution of the natural logarithm of the raw cost (log transformed cost) is shown in Figure 4 2. It had very low skewness (

PAGE 55

55 0.2) and kurtosis ( 0.01 ). Outcome modeling used the log transformed co st or the raw cost with a log lin k function, depending upon the specific model. Descriptive Statistics The baseline characteristics of all variables and differences between early and late detection groups are given in Table 4 2 for continuous variables and in Table 4 3 for categorical variables. CRG was analyzed as a continuous and a categorical variable f or this purpose, but is omitted from Table 4 3 for simplicity. C ost, log transformed cost (see Figure 4 3) CRG (see Figure 4 4), interhospital distance (see Figure 4 5 ) gender, low birth weight as a binary variable, the presence of multiple gestation, t he presence of extracardiac anomaly, and aortic arch obstruction all showed a significant difference between detection groups Median income by census tract, percent of children in poverty by census tract, enrolled months, race, prematurity as a binary va riable, and single ventricle physiology did not show a significant difference between detection groups. The rate s of prematurity and low birth weight were 51.5% and 67.8% in the overall study sample. The baseline rate of prematurity in the US is one in n ine (11.1%) 117 Intra class Correlation There were 177 (5.4%) of subjects with no birth hospital available. When ignori ng these observations, the ICC was 0. 0 55 When counting these observations as clustered around a single subject, the ICC was 0. 0 48. Sample Size Recalculation When compared with the preliminary dataset, the final dataset had a much lower intra class correl ation coefficient (0. 0 48 0. 0 55), a different ratio of early to late detection

PAGE 56

56 (4:1), a different average cluster size (harmonic mean 2.3), and a smaller difference between log transformed cost between the early and late detection groups (0.42), an addition al post hoc sample size calculation was performed based upon the actual harmonic mean cluster size of 2, an ICC of 0.57, power of 0.9 to detect a difference of one half subjects. Heteroskedasticity The Park test based on a first stage OLS outcomes model yielded a of 1.9. The regression is shown in Figure 4 6 The Park tests using a first stage GLM with Poisson and gamma distributions yielded values of 1.6 and 1.8 respectively. The poverty and income variables were left out of the first stage models due to missingness; all other covariates were included. Based on the consistent result of the Park tests in characterizing the heteroskedasticity present in the predicted outcome variable we chose to use a gamma distribution for our final GLM approach. Selection Bias Instrumental Variable Analysis Using the first stage of instrumenta l variable the model given in Equation ( 3 6 ) was run using four sets of covariates: all covariates, all covariates except the highly missing income and poverty variables, a minimal model consisting of non white race, prematurity, low birth weight, extrac ardiac malformation, single ventricle physiology, and arch obstruction, and a model with no covariates. In each case, the instrumental variable was found to be only weakly although significantly, related to the group

PAGE 57

57 selection variable. Parameter est imates for ranged from 0.004 and 0.00408. Using baseline probability of early detection of 0.8, a change in the interhospital distance by one standard deviation, or 45 minutes, the change in probability of being in the early detection group would be given by (4 1 ) or a 2.8% absolute change in probability Despite the weak instrument, we prepared the second stage of the three stage instrumental variable analysis as planned. After modeling the second stage given in Equation ( 3 7), the predictor variable which incorporated the effect of regressors included in Equation ( 3 6) was now strongly related to the selection variable. Heckman Selection Model The Heckman selection model yielded a statistic of 0.16 (95% C I 1.0 0.40). This was not significantly different than 0 (p=0.32), and thus we found no evidence of group selection bias via endogeneity of the selection variable. Measured Variables That Influenced Group Selection In both the instrumental variable modeling and the treatment effects model, variables that consistently and significantly predicted a higher probability of early detection were female gender, multiple gestation, delivery via Cesarean section, and univentricular physiology. Variables that significantly predicted a lower probability of early detection were the presence of an extracardiac lesion and aortic arch obstruction. Modeling the Primary Hypothesis Using a GLM with a log link, a gamma distribution an exchangeable covariance matrix an d clustering by birth hospital, we first modeled the outcome equation using all covariates, including the months enrolled. This model had an high level of missing data,

PAGE 58

58 with 501 (15.3%) of observations dropped, primarily due to the SES variables. Based o n the results of this model as well as the preliminary OLS model used in the Park test we selected significant and near significant covariates to include in the parsimonious model This model represented our best effort at incorporating all preliminary a nalyses and testing into a single model with minimal missingness Only 70 observations (2.1%) were dropped from this final analysis, all due to missing or zero cost data. We then repeated the parsimonious model without accounting for clustering effects, and also compared the results to the OLS model using the same parameters that was previously run as part of the Park test. Because of a weak instrument and no evidence of selection bias based on the Heckman model, we did not include the predicted selectio n probability from the instrumental variable approach in the final modeling process. In th model, early detection was associated with a decreased cost (p<0.00 9 ). The parameter estimate for this effect was 0.1 8 which when exponentiated yields 0 .8 4 or a 1 6 % reduction in base cost. This would give a cost savings of approximately $ 18, 0 00 compared to late detection, using the observed mean cost as a baseline S ignificant predictors of increased cost were non white race ($34,000 cost increase) th e presence of an extracardiac lesion ($223,000) single ventricle physiology ($135,000) and aortic arch obstruction ($56,000) Other s ignificant predictors of decreased cost were prematuri ty ($37,000 cost reduction) and low birth weight ($44,000 cost red uction) although concerns about the robustness of the prematurity and low birth weight are discussed below. The log transformed months enrolled did not significantly

PAGE 59

59 predict cost variation. The results of models are presented together in Table 4 4 For the sake of transparency, Appendix A also reports the general results of additional exploratory models that were run with instrumental variables, alternative distributions, and/or an offset variable. Each of these models was not used to inform the final analysis of the primary hypothesis for reasons given above. Appendix B gives the SAS and STATA code for all data manipulation and analyses performed during this study. Testing Secondary Aims Racial/ E thnic a nd SES Disparity We conducted analysis of racial and ethnic disparities across race/ethnicity and SES for the outcome variables, the selection variable, and all covariates. Outcome variable s Raw and log transformed cost both differed across racial /ethnic c ategories (p<0.0001, ANOVA). The log transformed cost by racial /ethnic category is shown in Figure 4 7 Although t he raw cost was higher in nonwhite vs. white subjects ($117,800 versus $93,500, p=0.04, t test), the log transformed cost was not significan tly different using this simplified grouping. In the outcomes model used for the primary hypothesis, non white race/ethnicity was positively associated with increased cost using the simplified binary classification but tent predictor when all categories were analyzed. The SES variables, m edian income and the percen t of children living in poverty, were not significant predictors of cost.

PAGE 60

60 The CRG score as a continuous variable was not significantly different across racia l categories or when comparing white and non white subjects. There was however, a difference across racial categories when consider ing CRG as a categorical variable. I Unknown CRG categ subjects were in this category. Selection variable The proportion of subjects in each racial/ethnic category did not differ significantly between the early and late detection groups, n or did the proportion of white vs. non white subjects differ between the groups (see Table 4 3) Similarly, r ace /ethnicity whether considered by category or as a binary white/non white variable, was not significantly associated with the selection variabl e in the instrumental variable or treatment effects model. Income and poverty did not significantly predict the sele ction variable in the maximally specified instrumental variable model. Covariates There was no significant difference in gender across racia l/ethnic categories. Prematurity was different across racial categories (p<0.0001) and appeared lower in 55.2%). Low birth weight was similarly different across racial/ethnic categories (p=0 .0005) and appeared 63.4%) than in all others (69.8 78.3%). Multiple gestation was also significantly different across racial/ethnic categories, occurring His (16.5%) than in all others (7.7 10.2%). Note that all pairwise comparisons are post hoc

PAGE 61

61 and were not analyzed for statistical significance. There was no significant difference in Cesarean section rates across racial/ethnic groups. The pr oportion of those with aortic arch obstruction and single ventricle physiology differed among groups but with no discernible pattern. When using the binary white/non white binary race/ethnicity variable, no categorical covariates were significant; however, nonwhite versus white subjects had a lower mean median income for census tract ($37,900 versus $49 ,000, p < 0.0001, see Figure 4 8 ) and a higher percent of children living in poverty for census tract (26.1% versus 1 5.6%, p < 0.0001, see Figure 4 9 ). Final ly, the interhospital distance was significantly different across racial/ethnic categories (p<0.0001) by ANOVA and significantly shorter in nonwhite versus white subjects (36.4 versus 46.3 minutes, p<0.0001). SES variables were not tested directly against additional covariates b eyond what was described above. Overall Disease Burden The mean CRG was slightly higher in the late versus early detection group (3.4 versus 3.2, p < 0.002 by t test). When considering the CRG as a categorical variable, the difference in distributions between detection groups was significantly different.(p=0.001 by chi square). Of those with an available CRG score, 31.7% of the late detection group ve Health Care Needs mi ssingness of the CRG as well as uncertainty regarding its timing, regression analyses were not performed for this variable.

PAGE 62

62 Survival Only 10 (0.3%) subjects were coded as non survivors using the algorithm given in Table 3 1. This method, using the 798 .x g roup of ICD 9 code s identified far fewer deaths than would be expected based on well established survival rates. 5 Our assumption is that death was inconsistently rec orded as a diagnosis code, and w e were thus unable conduct this analysis using the available claims data.

PAGE 63

63 Table 4 1. Missing data Total Missing Variable N o. No. % Cost 3266 1 0.0 Log Transformed Cost 3197 70 2.1 Interhospital Distance 3163 104 3.2 Income 2824 443 13.6 Poverty 2824 443 13.6 Enrolled Months 3267 0 0.0 CRG 2853 414 12.7 Gender 3266 1 0.0 Race/Ethnicity 3267 0 0.0 Gestational Age 3267 0 0.0 Birth Weight 3267 0 0.0 Multiple Gestation 3267 0 0.0 Cesarean 3267 0 0.0 Extracardiac Lesion 3267 0 0.0 Arch Obstruction 3267 0 0.0 Coarctation 3267 0 0.0 Single Ventricle 3267 0 0.0 Table 4 2. Descriptive statistics for continuous data Overall Early Detect. Late Detect. Variable Mean SD Mean Mean p b Cost a 114.7 233.8 108.3 138.4 0.003 Log Transformed Cost 10.4 1.7 10.3 10.8 < .0001 Interhospital Dist. (min.) 37.7 45.4 35.9 44.9 < .0001 Income a 40.3 18.3 40.1 41.2 0.22 Poverty (%) 24.6 14.1 24.7 24.0 0.27 Enrolled Months (mo.) 27.2 12.3 27.1 27.4 0.55 CRG 3.3 1.5 3.2 3.4 0.002 a t housands of U.S. Dollars b by t test with equal or unequal variances as indicated by Folded F test

PAGE 64

64 Table 4 3. Descriptive characteristics of categorical data Total Early Detect. Late Detect. Characteristic N (%) N (% of group) N (% of group) p a Male 1758 (53.8) 1364 (52.4) 394 (59.3) 0.0014 Female 1508 (46.2) 1238 (47.6) 270 (40.7) White /Non Hispanic 450 (13.8) 353 (13.6) 97 (14.6) 0.69 Black /Non Hispanic 431 (13.2) 299 (11.5) 83 (12.5) Hispanic 1972 (60.4) 1581 (60.8) 391 (58.8) Asian/Pacific Isl. 23 (0.7) 17 (0.7) 6 (0.9) Amer. Indian/Alaskan 6 (0.2) 4 (0.2) 2 (0.3) Unknown/Other 385 (11.8) 299 (11.5) 86 (12.9) White 450 (13.8) 353 (13.6) 97 (14.6) 0.50 Non White 2817 (86.2) 2249 (86.4) 568 (85.4) 37 weeks) 1584 (48.5) 1249 (48.0) 335 (50.4) 0.27 Preterm (<37 weeks) 1683 (51.5) 1353 (52.0) 330 (49.6) Birth Weight 2.5 kg 1052 (32.2) 793 (30.5) 259 (39.0) <0.0 0 01 Birth Weight < 2.5 kg 2215 (67.8) 1809 (69.5) 406 (61.1) Singleton 2956 (90.5) 2330 (89.5) 626 (94.1) 0.0003 Multiple gestation 311 (9.5) 272 (10.5) 39 (5.9) Vaginal Delivery 2820 (86.3) 2224 (85.5) 596 (89.6) 0.005 Cesarean Delivery 447 (13.7) 378 (14.5) 69 (10.4) No Extracardiac Anomaly 1590 (48.7) 268 (40.3) 1322 (50.8) <0.0001 Extracardiac Anomaly 1677 (51.3) 397 (59.7) 1280 (49.2) Normal Arch 2530 (77.4) 2037 (78.3) 493 (74.4) 0.02 Arch Obstruction 737 (22.6) 565 (21.7) 172 (25.9) Two Ventricle 2274 (84.9) 2203 (84.7) 571 (85.9) 0.44 Single Ventricle 493 (15.1) 399 (15.3) 94 (14.1) a greater than chi square

PAGE 65

65 Figure 4 1 Distribution of raw cost Figure 4 2. Distribution of log transformed cost

PAGE 66

66 Figure 4 3. Comparison of log transformed cost (0=late detection, 1=early detection) Figure 4 4. Comparison of CRG (0=late detection, 1=early detection)

PAGE 67

67 Figure 4 5. Comparison of interhospital distance (0=late detection, 1=early detection)

PAGE 68

68 Figure 4 6 Park test for heteroskedasticity (using OLS as first stage model )

PAGE 69

69 Table 4 4. Final outcome model results Comparison Models Variable Parameter Model a Maximal b Non Multilevel c OLS d Intercept 0 10.9 10.8 10.9 9.6 Early Detection 1 0.17 0.13 0.16 0.22 p 0.009 0.08 0.002 0.0002 Female 1 0.003 0.09 Premature 1 0.39 0.39 0.48 0.63 Low Birth Weight 1 0.49 0.47 0.47 0.48 Multiple Gestation 1 0.04 0.04 Cesarean Section 1 0.10 0.10 Extracardiac Lesion 1 1.08 1.05 1.09 1.06 Single Ventricle 1 0.78 0.78 0.78 1.02 Aortic Arch Obstruction 1 0.40 0.41 0.52 0.47 Median Income e 1 0.0 % Children in Poverty e 1 0.0004 Black, Non Hispanic f 1 0.24 0.004 Hispanic f 1 0.16 0.12 Amer. Indian / Alaskan Native f 1 0.47 1.07 Asian/Pacific Islander f 1 0.40 0.25 Other /Unknown f 1 0.52 0.20 All Non White f 1 0.26 0.13 Log of Months Enrolled 1 0.04 0.04 0.01 0.28 a Parsimonious GLM with significant covariates, multilevel model b Maximal GLM with all covariates, multilevel model c Parsimonious GLM with significant covariates, no multilevel model d OLS with all covariates except SES, no multilevel model e by Census Tract f with "White" race/ethnicity as referent group p < 0.05 indicates not modeled

PAGE 70

70 Figure 4 7 Log transformed cost by racial/ethnic category

PAGE 71

71 Figure 4 8 Median income by racial/ethnic category Figure 4 9 Percent of children in poverty in census tract by race/ethnicity

PAGE 72

72 CHAPTER 5 DISCUSSION Primary Hypothesis Our results support our primary hypothesis, demonstrating that early detection of CCHD predicts an average reduction in long term health care cost of approximately 1 6 %. In our study this tran slated to approximately $18,300 in unadjusted U.S. dollars. This could partially or completely offset the incremental cost effectiveness ratio of an NBS program. We reached this conclusion using a modeling strategy t hat quanti tatively evaluated (1) missingness, (2) distribution of the outcome variable, (3) clustering and intra class correlation, (4) heteroskedasticity, and (5) endogeneity and group selection bias. In addition, our primary result reached significance in two com parison models using different modeling methodologies, and reached near significance in a model that differed only by the incorporation of additional covariates. We also found that the presence of an extracardiac lesion, single ventricle physiology, aortic arch obstruction and non white race were strong predictors o f increased cost across models, with particularly strong effects from extracardiac lesions and single ventricle physiology. (Hypothesis 2) Prematurity and low birth weight predicted decreased co st. Given the significantly higher mortality of premature infants with CCHD, in particular those weighing less than 1500 grams at birth, 89, 118 there may be have been earlier mortality and thus decreased long term cost in this group. However, t here was also an unexpectedly high prevalence of prematurity (51.5%) and low b irth weight (67.8%) in our study sample. There is a higher frequency of CCHD in very low birth weight

PAGE 73

73 infants 8 9 and a clinical practice of preterm delivery or induction persists despite strong evidence against it 118 so some degree of overrepresentation would be expected in this population. However, the rates of both prematuri ty and low birth weight in our study w ere approximately five times baseline population rates 117 T here may be a flaw in recording the diagnoses of prematurity and l ow birth weight in claims data and/or a flaw in our algorithm for converting it to a binary variable. For this reason we urge caution when interpreting these particular relationships. Secondary Aims R acial/Ethnic and SES Disparities Non white race was a predictor of increased cost in the final outcomes model. (Hypothesis 3A). The reason for this is unclear. However, while many covariates including prematurity, multiple gestation, Cesarean section, aor tic arch obstruction, and single ventricle physiology were different across racial/ethnic categories, these differences were not significant across a binary white/non white comparison. This would suggest that a binary comparison does not capture the textur e of racial variation and that the minimal model was perhaps too reductive. This is further reinforced by the difficulty in identifying a clear qualitative pattern across the full range or racial/ethnic categories for multiple variables, and the Un known racial/ethnic classification would be preferred if available. Race/ethnicity did not appear to predict early versus late detection of CCHD. SES variables did not appear to predict either th e selection or primary outcome, cost, although the degree of missingness limited their utility significantly and they were not included in the final modeling strategy. (Hypothesis 3B)

PAGE 74

74 Finally, our results confirmed the presence of baseline economic dispari ties across racial/ethnic groups, demonstrating lower median incomes and a higher percent of children living in poverty near the homes of Black and Hispanic subjects in particular. The interhospital distance varied by race/ethnicity, suggesting that subje cts in some categories tend to live closer to pediatric cardiac surgery centers. Overall Disease Burden The significantly higher mean CRG score in the late detection group, the higher percentage of late versus early detection subjects in the most severe CR G category, and the lower percentage of late versus early detection subjects in the least severe CRG category all suggest that late detection of CCHD is associated with an increased long term disease burden affecting resource utilization. Because the pair wise percentage comparisons were post hoc and quantitative, we regard this as evidence to support but not prove Hypothesis 4. Limitations and Threats to Validity Multiple threats to validity affect any quasi experimental study We follow Shadish, Cook, an and how they affected or were addressed in this study 119 Threat s to Statistical Conclusion Validity We used a prospective power calculation and verified it with a post hoc calculation to avoid low statistical power To avoid violating specific statistical assumptions we used methods to test and account for distribut ional assumptions, heteroskedasticity, and clustering. To avoid inflating the Type I error rate we accounted for clustered data and avoided multiple comparisons except where clearly stated. Unreliability of measures was an issue with geocoded data, survival,

PAGE 75

75 prematurity and low birth weight diagnoses, and a small proportion of the cost variables. Heterogeneity of units was also suggested by differences in baseline data among groups, and was addressed by controlling for covariates in the final outcomes model Threats related to treatment implementation unreliability and experimental setting variability did not apply to this quasi experimental and retrospective study. Threats to Internal Validity Selection was the bigge st source of potential bias in this study. Sample selection was an issue both in limiting the study to one state and limiting it to Medicaid and CHIP clients. This may have skewed the data towards groups who differ in characteristics influencing the outc ome. Race/ethnicity and SES are two means by which this could occur, both of which are analyzed in some detail in the text. Future studies can avoid this issue by including geographically and demographically diverse states as well as using all payer clai ms databases. Group selection has already been addressed in detail. Ambiguous temporal precedence was not an issue as all explanatory variables were present at birth and all outcome variables developed after birth. History may have been an issue in this particular birth cohort from 2008 2009 although preliminary results from infants born in 2011 are very similar. More generally, continuous advances and changing technology in the diagnosis and treatment of CHD make it difficult, as in any field, to compar e estimates of cost, survival, or comorbidity across time. Maturation is not a major issue as our subjects are followed from birth on. Similarly, regression to the mean does not apply as our subjects are selected based upon a relatively fixed diagnosis. Attrition due to disenrollment in Medicaid/CHIP is a potential threat to validity here, but inclusion of a term for total months enrolled in the final outcome model

PAGE 76

76 controlled for this issue. Testing is not an issue although in this study although would be for future prospective studies involving long term neurocognitive outcomes. Instrumentation in the form of changing trends in recording ICD 9 codes, may affect results although this is unlikely in such a short period of time. In the future, however, t he use of ICD 10 codes may make comparison with ICD 9 based data very challenging. Finally, specific threats related to groups undergoing treatments, including treatment diffusion, compensatory equalization, compensatory rivalry, and resentful demoraliza tion do not apply to this study. Threats to Construct Validity There are several potential and important threats to construct validity in this study. Inadequate explication of constructs may affect our definition and operationalization of variables; indeed, this appeared to be the case with the prematurity and low birth weight variables. We guarded against this by carefully defining variable and data specifications as well as carefully e xamining a preliminary data set, making specificat ion adjustments, and re examin ing the final dataset. Mono operation bias is unlikely as children were treated at multiple hospitals. Although we identified several outcome measures in part to avoid mono m ethod bias we completed the full, planned analysis only on cost, so this remains a threat to validity. Threats not applicable to this study methodology, which did not have treatment, study participants per se, or any staff inter actions with subjects, incl uded hypothesis guessing, evaluation apprehension, experimenter expectancies, novelty or disruption effects, interaction of testing or treatment effects confounding constructs with levels of constructs, interaction of testing and treatment, and interactio n of different treatments.

PAGE 77

77 Threats to External Validity (Generalizability) The primary threats to the external validity of this study is the interaction of selection with the causal relationship. Selecting based upon enrollment in Medicaid and CHIP makes it imperative to repeat this type of study with an all payer claims database, as well as with subjects whose racial/ethnic and economic diversity reflects the makeup of the population to which we wish to generalize. In a more limited sense, interaction of the set ting Texas with the causal relationship should be addressed in similar fashion in the future by studying other states or regions Future studies could also include non US locations with qualitatively different health care systems, to avoid context dependent mediation The interaction of the causal relationship with treatment variations could become an issue in diagnosis of CCHD. Finally, i nteraction of the causa l re lationship with units is less of an issue with this study as individual people are the sampling unit of interest Additional study limitations W e did not adjust cost values for inflation by year. Although the time span of this study was short enough that it may not have affected, future and more comprehensive studies should adopt this methodology, particularly if researchers hope to inform or affect policy. The study also would have been significantly strengthened by adopting a more robust method to eval uate survival, such as a link with a publicly available death registry. Similarly, obtaining hospital records for a sample of patients to verify the diagnoses and other information would have helped assure the quality of the claims database variable extra ction process

PAGE 78

78 Future Directions Given the legislative landscape in the United States, the time is ripe for prospective experimental and/or quasi experimental studies on a larger scale to determine the potential cost benefit realized by increasing the earl y detection of congenital heart disease, as well as the potential benefit to long term chronic disease burden, which in turn affects resource utilization. Long term neurocognitive outcomes are also important areas of study and could potentially impact hea lth care cost and resource utilization in this important population. We envision a multi state study of a newborn screening program, combining all payer claims data over time with a population sub sample undergoing periodic neurocognitive screening and rep orting on health related quality of life using an instrument designed to meet the needs of children and young adults with CCHD. 120 122 This analysis would be compared to actual economic data on the cost of implementation of the screening program, including an incremental cost effectiveness ratio based on measured, rather than forecast, costs. Conclusion Diagnosis of CCHD before initia l hospital discharge, when compared with later diagnosis, predicted a reduction in long term health care costs in this sample of subjects with CCHD enrolled in Texas Medicaid and CHIP and born in 2008 2009. Non white race, extracardiac lesions, aortic arc h obstruction, and single ventricle physiology all predicted significantly increased cost. In addition, the long term disease burden, which potentially affects resource utilization, appears to be less in the group with early than with late d iagnosis. Prospective studies with more diverse populations, payer mixes, and geographic regions are needed to confirm these findings.

PAGE 79

79 APPENDIX A ADDITIONAL MODEL RESULTS

PAGE 80

80 Multi Instru Co Off Cov. l evel ment variates set D istr. Str uct 1 p Significant Covariates yes 2 stage maximal no gamma exch. 0.44 0.65 premature, lowbirthwt, extracardiac, univentricular, archobstruction, black or Am. Indian race yes 3 stage maximal no gamma exch. 0.15 0.85 premature, lowbirthwt, extracardiac, univentricular, archobstruction, black, Am. Indian, unknown race no 2 stage minimal no gamma exch. 0.28 0.48 premature, lowbirthwt, extracardiac, univentricular, archobstruction, nonwhite no 3 stage minimal no gamma exch. 0.44 0.27 premature, lowbirthwt, extracardiac, univentricular, archobstruction, nonwhite yes 2 stage minimal no gamma exch. 0.92 0.1 premature, lowbirthwt, extracardiac, univentricular, archobstruction, nonwhite yes 3 stage minimal no gamma exch. 0.52 0.065 premature, lowbirthwt, extracardiac, univentricular, archobstruction, nonwhite yes no ne minimal yes gamma exch. 0.08 0.51 premature, extracardiac, univentricular, archobstruction yes 2 stage minimal yes gamma exch. 1.83 0.02 premature, extracardiac, univentricular yes 3 stage minimal yes gamma exch. 0.72 0.009 premature, extracardiac, univentricular yes no ne minimal no gamma unstruct. 0.34 a 0.22 lowbirthwt, extracardiac yes 2 stage minimal no gamma unstruct. 0.16 a 0.86 lowbirthwt, extracardiac, univentricular yes 3 stage minimal no gamma unstruct. 0.16 a 0.86 lowbirthwt, extracardiac, univentricular yes no ne minimal yes gamma unstruct. -b -N/A yes 2 stage minimal yes gamma unstruct. 0.14 a 0.3 extracardiac, univentricular, archobstruction yes 3 stage minimal yes gamma unstruct. 0.68 a 0.58 lowbirthwt, extracardiac, univentricular yes no ne minimal no Poisson exch. 0.08 0.23 premature, lowbirthwt, extracardiac, univentricular, archobstruction, nonwhite yes 2 stage minimal no Poisson exch. 0.46 0.4 premature, lowbirthwt, extracardiac, univentricular, archobstruction, nonwhite yes 3 stage minimal no Poisson exch. 0.49 0.31 premature, lowbirthwt, extracardiac, univentricular, archobstruction, nonwhite yes no ne minimal yes Poisson exch. 0.08 0.21 premature, lowbirthwt, extracardiac, univentricular, archobstruction yes 2 stage minimal yes Poisson exch. 0.83 0.22 premature, extracardiac, univentricular, archobstruction yes 3 stage minimal yes Poisson exch. 0.63 0.17 premature, extracardiac, univentricular, archobstruction All models are GLM models. using the SAS genmod procedure a Met convergence criteria and provided parameter es timates but terminated in error; b Terminated in error before providing parameter estimates.

PAGE 81

81 APPENDIX B CODE USED IN ANALYSES SAS Code /**********************************************************/ /***** POTENTIAL COST BENEFIT FROM EARLY DETECTION OF *****/ /***** OF CRITICAL CONGENITAL HEART DISEASE *****/ /**********************************************************/ /***************************/ /*****MACRO DEFINITIONS*****/ /***************************/ data _null_; %LET categ_covars_all =race; %LET categ_stmt_all = class &categ_covars_all; %LET covars_all = female premature lowbirthwt multgest cesarean extracardiac univentricular archobstruction income poverty &categ_covars_all; %LET categ_covars_noincome =&categ_covars_all; %LET categ_stmt_noincome = class &cat eg_covars_noincome; %LET covars_noincome = female premature lowbirthwt multgest cesarean extracardiac univentricular archobstruction &categ_covars_noincome; %LET categ_covars_minimal =; %LET categ_stmt_minimal =; %LET covars_minimal = nonwhite premature lowbirthwt extracardiac univentricular archobstruction &categ_covars_minimal; %LET class_st_covar_prelim=&categ_stmt_noincome; %LET covar_list_prelim=&covars_noincome; %LET offset_var = logenrollmo; %LET of fset_opt = offset=&offset_var; %put _all_; run; /**************************/ /*****DATA PREPARATION*****/ /**************************/ libname archer 'p: \ data'; %macro CLEAROUTPUTWINDOW; ods html close; /* close previous */ ods html; /* open new */

PAGE 82

82 %mend CLEAROUTPUTWINDOW; %CLEAROUTPUTWINDOW; options mprint; /* Code the independent variables for analysis */ data transformeddata; set Archer.CHD_Final_v_2_23_13; /* lesion type */ univentricular=0; archobstruc tion=0; if f_CHD_4=1 then univentricular=1; /*"single ventricle"*/ if f_CHD_7=1 then univentricular=1; /*tricuspid atresia or stenosis*/ if f_CHD_9=1 then archobstruction=1; /*aortic stenosis*/ if f_CHD_10=1 then archobstruction=1; /*mitral stenosis*/ if f_CHD_11=1 then univentricular=1; /*HLHS*/ if f_CHD_11=1 then archobstruction=1; /*HLHS*/ if f_CHD_12=1 then archobstruction=1; /*subaortic stenosis*/ if f_CHD_14=1 then archobstruction=1; /*Shone's complex*/ if f_CH D_15=1 then archobstruction=1; /*coarc or IAA*/ if f_CHD_16=1 then univentricular=1; /*aortic atresia*/ if f_CHD_16=1 then archobstruction=1; /*aortic atresia*/ /* binary variables with new names for coding clarity */ if f_grp='Late' then ear lydetect=0; else if f_grp='Early' then earlydetect=1; else earlydetect=.; if sex='M' then female=0; else if sex='F' then female=1; else female=.; /*additional transformations*/ if exp_med <= 0 then logcost=.; else logcost=log(exp_med); /*l og transform cost */ enrollmo=months; /* prepare offset variable for modeling log cost*/ enrollmocorr=months; logenrollmo = log(enrollmo); logenrollmocorr = log(enrollmocorr); difftime = min_cardsrg min_hosptl; /* prepare the instrument al variable */ if f_Prem_0=1 then premature=0; else premature=1; if premature=0 then premcat="37 or more wks"; else if /*f_Prem_1=1 or*/ f_Prem_9=1 then premcat="35 36 wks"; else if f_Prem_1=1 then premcat="Premature Unspecified"; else if f_Prem_2=1 or f_Prem_3=1 then premcat="24 or less wks"; else if f_Prem_4=1 then premcat="25 26 wks"; else if f_Prem_5=1 then premcat="27 28 wks"; else if f_Prem_6=1 then premcat="29 30 wks"; else if f_Prem_7=1 then premcat="31 32 wks "; else if f_Prem_7=1 then premcat="33 34 wks"; else premcat=""; if f_LBW_0=1 then lowbirthwt=0; else lowbirthwt=1;

PAGE 83

83 if lowbirthwt=0 then birthwtcat = "E. NORMAL (>=2500g)"; else if f_LBW_7=1 or f_LBW_8=1 or f_LBW_9=1 /*or f_LBW_1=1*/ then b irthwtcat="D. LBW (1500 2499g)"; else if f_LBW_1=1 then birthwtcat="LBW Unspecified"; else if f_LBW_5=1 or f_LBW_6=1 then birthwtcat="C. VLBW (1000 1.499g)"; else if f_LBW_3=1 or f_LBW_4=1 then birthwtcat="B. ELBW (500 999g)"; else if f_LBW_2=1 then birthwtcat = "A. <500g"; else birthwtcat=""; format birthhospwithother $char50.; if BLNG_Prov_Name='' then birthhospwithother='OTHER'; else birthhospwithother=BLNG_Prov_Name; if race_txt='White' then nonwhite=0; else nonwhite=1; format race $char25.; if race_txt = '' then race='Unknown / Other'; else race=race_txt; monthsoldonlastday = (age_last_day / 365) 12; rename BLNG_Prov_Name=birthhosp f_MG=multgest f_EA=extracardiac f_CS=cesarean exp_med=rawcost crgsum_txt=crg crgsum=clinicalresourcegroup; run; /* Lesion classification*/ data transformeddata; set transformeddata; lesiontype=1; if univentric ular=0 and archobstruction=1 then lesiontype=2; else if univentricular=1 and archobstruction=0 then lesiontype=3; else if univentricular=1 and archobstruction=1 then lesiontype=4; if f_CHD_15=1 then coarctation=1; else coarctation=0; run; /* Keep variables of interest and label them for clarity of output */ data cleandata; set transformeddata; keep rawcost logcost earlydetect female cesarean extracardiac lowbirthwt premature multgest race nonwhite birthhosp birthhospwithother income poverty difftime enrollmo enrollmocorr univentricular archobstruction lesiontype logenrollmo logenrollmocorr difftime clinicalresourcegroup premcat birthwtcat coarctation monthsoldonlastday; label rawcost='Raw Cost' logco st='Natural Log of Raw Cost' earlydetect='Early Detection' female='Female' cesarean='Cesarean Section' extracardiac='Extracardiac Anomaly' lowbirthwt='Low Birth Weight' premature='Premature' multgest='Multiple Gestation' nonwhite='Non white Race/Ethnicity'

PAGE 84

84 birthhosp='Birth Hospital' birthhospwithother='Birth Hospital (missing values grouped together)' income='Median Income for Census Tract' poverty='Percent of Children in Poverty for Census Tract' difft ime='Interhospital Driving Time (min)' enrollmo='Months Enrolled (uncorrected)' enrollmocorr='Months Enrolled (with administrative correction)' logenrollmo='Ln(Months Enrolled uncorrected)' logenrollmocorr='Ln(Months Enrolled with administrativ e correction)' race='Race/Ethnicity' univentricular='Single Ventricle' archobstruction='Arch Obstruction' coarctation='Aortic Arch Coarctation, Hyoplasia, or Interruption' lesiontype='Lesion Type' clinicalresourcegroup='Clinical Res ource Group' premcat='Prematurity Category' birthwtcat='Birth Weight Category' monthsoldonlastday='Age (mo) at End of Study'; run; /* Export to STAT for selection bias procedure */ proc export data=cleandata oufile="P: \ data \ CHD_data"; run; /***************************************************************************/ /*****BASIC DATA DIAGNOSTICS, NORMALITY TESTING, EXPLORATORY STATISTICS*****/ /***************************************************************************/ /* Test normality for raw cost and ln(cost) */ /* note: ignores months enrolled*/ title 'Expenditure Normality Testing Overall'; ods graphics on; ods select BasicMeasures ExtremeObs Quantiles Histogram Moments TestsForNormality ProbPlot; proc univariate data =cleandata normaltest; var rawcost logcost; histogram rawcost / kernel(color=red) name='Overall Raw Cost Distribution'; inset mean std / format=6.4; probplot rawcost / normal (mu=est sigma=est) square; histogram logcost / kernel(color=red ) name='Overall Log Cost Distribution'; inset mean std / format=6.4; probplot logcost / normal (mu=est sigma=est) square; run; title 'Expenditure Normality Testing by Group'; proc univariate data=cleandata normaltest; var rawcost logcost; c lass earlydetect; histogram rawcost / kernel(color=red) name='Raw Cost Distribution by Group';

PAGE 85

85 inset mean std / format=6.4; probplot rawcost / normal (mu=est sigma=est) square; histogram logcost / kernel(color=red) name='Log Cost Distribution by Group'; inset mean std / format=6.4; probplot logcost / normal (mu=est sigma=est) square; run; /* Diagnostic data for numerical covariates */ title 'Covariate Baseline Diagno stics'; proc univariate data=cleandata normaltest; var difftime income poverty enrollmo clinicalresourcegroup; histogram difftime / kernel(color=red); inset mean std / format=6.4; probplot difftime / normal (mu=est sigma=est) square; hist ogram income / kernel(color=red); inset mean std / format=6.4; probplot income / normal (mu=est sigma=est) square; histogram poverty / kernel(color=red); inset mean std / format=6.4; probplot poverty / normal (mu=est sigma=est) square ; histogram enrollmo / kernel(color=red); inset mean std / format=6.4; probplot enrollmo / normal (mu=est sigma=est) square; histogram clinicalresourcegroup / kernel(color=red); inset mean std / format=6.4; run; proc ttest data=cle andata cochran ci=equal umpu; class earlydetect; var rawcost logcost difftime income poverty enrollmo clinicalresourcegroup; run; quit; proc freq data=cleandata order=formatted; tables female / plots(only)=freqplot(scale=percent); tables earlydetect*female / chisq cmh plots(only)=freqplot(scale=percent); tables race / plots(only)=freqplot(scale=percent); tables earlydetect*race / chisq cmh plots(only)=freqplot(scale=percent); tables nonwhite / plots(only)=freqplot(scale= percent); tables earlydetect*nonwhite / plots(only)=freqplot(scale=percent); tables premature / plots(only)=freqplot(scale=percent); tables earlydetect*premature / chisq cmh plots(only)=freqplot(scale=percent); tables premcat / plots(only)=freq plot(scale=percent); tables earlydetect*premcat / chisq cmh plots(only)=freqplot(scale=percent); tables lowbirthwt / plots(only)=freqplot(scale=percent); tables earlydetect*lowbirthwt / chisq cmh plots(only)=freqplot(scale=percent); tables birt hwtcat / plots(only)=freqplot(scale=percent); tables earlydetect*birthwtcat / chisq cmh plots(only)=freqplot(scale=percent); tables multgest / plots(only)=freqplot(scale=percent); tables earlydetect*multgest / chisq cmh plots(only)=freqplot(scale= percent);

PAGE 86

86 tables cesarean / plots(only)=freqplot(scale=percent); tables earlydetect*cesarean / chisq cmh plots(only)=freqplot(scale=percent); tables archobstruction / plots(only)=freqplot(scale=percent); tables earlydetect*archobstruction / chi sq cmh plots(only)=freqplot(scale=percent); tables coarctation / plots(only)=freqplot(scale=percent); tables earlydetect*coarctation / chisq cmh plots(only)=freqplot(scale=percent); tables univentricular / plots(only)=freqplot(scale=percent); t ables earlydetect*univentricular / chisq cmh plots(only)=freqplot(scale=percent); tables clinicalresourcegroup / plots(only)=freqplot(scale=percent); tables earlydetect*clinicalresourcegroup / chisq cmh plots(only)=freqplot(scale=percent); tables extracardiac / plots(only)=freqplot(scale=percent); tables earlydetect*extracardiac / chisq cmh plots(only)=freqplot(scale=percent); run; quit; title 'Variable Differences by Race/Ethnicity'; proc anova data=cleandata; class race; model rawcost logcost difftime income poverty enrollmo clinicalresourcegroup = race; run; quit; proc freq data=cleandata order=formatted; tables race*female / chisq cmh plots(only)=freqplot(scale=percent); tables race*premature / chisq cmh plots(only)=fre qplot(scale=percent); tables race*premcat / chisq cmh plots(only)=freqplot(scale=percent); tables race*lowbirthwt / chisq cmh plots(only)=freqplot(scale=percent); tables race*birthwtcat / chisq cmh plots(only)=freqplot(scale=percent); tables ra ce*multgest / chisq cmh plots(only)=freqplot(scale=percent); tables race*cesarean / chisq cmh plots(only)=freqplot(scale=percent); tables race*archobstruction / chisq cmh plots(only)=freqplot(scale=percent); tables race*coarctation / chisq cmh plo ts(only)=freqplot(scale=percent); tables race*univentricular / chisq cmh plots(only)=freqplot(scale=percent); tables race*clinicalresourcegroup / chisq cmh plots(only)=freqplot(scale=percent); run; quit; title 'Variable differences by White/Non white Race/Ethnicity'; proc ttest data=cleandata cochran ci=equal umpu; class nonwhite; var rawcost logcost difftime income poverty enrollmo clinicalresourcegroup; run; quit; proc freq data=cleandata order=formatted; tables nonwhite*fema le / chisq cmh plots(only)=freqplot(scale=percent);

PAGE 87

87 tables nonwhite*nonwhite / plots(only)=freqplot(scale=percent); tables nonwhite*premature / chisq cmh plots(only)=freqplot(scale=percent); tables nonwhite*premcat / chisq cmh plots(only)=freqplot (scale=percent); tables nonwhite*lowbirthwt / chisq cmh plots(only)=freqplot(scale=percent); tables nonwhite*birthwtcat / chisq cmh plots(only)=freqplot(scale=percent); tables nonwhite*multgest / chisq cmh plots(only)=freqplot(scale=percent); t ables nonwhite*cesarean / chisq cmh plots(only)=freqplot(scale=percent); tables nonwhite*archobstruction / chisq cmh plots(only)=freqplot(scale=percent); tables nonwhite*coarctation / chisq cmh plots(only)=freqplot(scale=percent); tables nonwhite* univentricular / chisq cmh plots(only)=freqplot(scale=percent); tables nonwhite*clinicalresourcegroup / chisq cmh plots(only)=freqplot(scale=percent); run; quit; title 'Frequency listing of birth hospitals'; proc freq data=cleandata order=freq; tabl es birthhosp birthhospwithother; run; quit; title 'Relationship Between Months Enrolled and Age at End of Study'; proc reg data=cleandata noprint; model enrollmo = monthsoldonlastday; plot enrollmo*monthsoldonlastday; run; quit; /***************************************************/ /***** TEST INTRACLASS CORRELATION COEFFICIENT *****/ /***************************************************/ /* Intraclass correlation coefficient code without grouping missing values */ title 'ICC (Mi ssing Values Not Grouped, Method 1)'; ods output CovParms = covp; proc mixed data = cleandata; class birthhosp; model rawcost = ; random intercept /subject= birthhosp; run; data icc; set covp end=last; retain bvar; if subject~="" then bvar = estimate; if last then icc = bvar/(bvar+estimate); run; proc print data = icc; run;

PAGE 88

88 title 'ICC (Missing Values Not Grouped, Method 2)'; proc mixed ratio data = cleandata; class birthhosp; model rawcost = ; random intercept /subject= birthhosp; run; /* Intraclass correlation coefficient code WITH grouping missing values */ title 'ICC (Missing Values Grouped, Method 1)'; ods output CovParms = covp; proc mixed data = cleandata; class birthhospwithother; model rawcost = ; random intercept /subject= birthhospwithother; run; data icc; set covp end=last; retain bvar; if subject~="" then bvar = estimate; if last then icc = bvar/(bvar+estimate); run; proc print data = icc; run; title 'ICC (Missi ng Values Grouped, Method 2)'; proc mixed ratio data = cleandata; class birthhospwithother; model rawcost = ; random intercept /subject= birthhospwithother; run; quit; /*********************************/ /*****TEST HETEROSKEDASTICITY*****/ /*********************************/ /* The next sections prepare for Park test for hetereoskedasticity using two methods */ /*OLS model using proc glm to generate residuals and predicted values on log transformed cost */ /* note: need offset equivalent here */ title 'OLS Model to Generate Residuals'; proc glm data=cleandata; &class_st_covar_prelim; model logcost = earlydetect &covar_list_prelim &offset_var / solution; output out=residualdatabyols p=logcosthat r=logcostresid stdr=stderrofresid; run; quit;

PAGE 89

89 /*Generalized linear models using proc genmod to generate predicted values of cost using log link */ /* GLM with Poisson distribution */ /* note: the predicted value is re exponentiated */ /* note: need to specify covariance matrix */ title 'GLM (Poisson distribution) to Generate Residuals'; proc genmod data=cleandata; &class_st_covar_prelim; model rawcost = earlydetect &covar_list_prelim / link=log dist=poisson &offset_opt; output out=residualdatabyglmpoisson p=costhatpoisson; run; quit; /* note: the predicted value is re exponentiated */ /* note: need to specify covariance matrix */ /* GLM with gamma distribution */ title 'GLM (gamma distribution) to Generate Re siduals'; proc genmod data=cleandata; &class_st_covar_prelim; model rawcost = earlydetect &covar_list_prelim / link=log dist=gamma &offset_opt; output out=residualdatabyglmgamma p=costhatgamma; run; quit; /* Create Park test variables (OLS method) */ data residualdatabyols; set residualdatabyols; yhati = exp(logcosthat + (0.5 (stderrofresid ** 2))); parkpred = log(yhati); parkresid = log((rawcost yhati) ** 2); label parkresid='Ln(Yi Yhati)^2' parkpred='Ln(Yhati)'; run; /* Create Park test variables for GLM with Poisson distribution */ data residualdatabyglmpoisson; set residualdatabyglmpoisson; parkpred = log(costhatpoisson); parkresid = log((rawcost costhatpoisson) ** 2); label parkresid='Ln(Yi Yhati)^2' parkpred='Ln(Yhati)'; run; /* Create Park test variables for GLM with gamma distribution */ data residualdatabyglmgamma; set residualdatabyglmgamma; parkpred = log(costhatgamma);

PAGE 90

90 parkresid = log((rawcost costhatgamma) ** 2); label parkresid='Ln(Yi Yhati)^2' parkpred='Ln(Yhati)'; run; /* Park test by OLS */ title 'Park Test for Heteroskedasticity (OLS model)'; proc reg data=residualdatabyols noprint; model parkresid = parkpred; plot parkresid*parkpred; run; quit; /* Park test by GLM with Poisson distriubtion */ title 'Park Test for Heteroskedasticity (GLM with Poisson distribution)'; proc reg data=residualdatabyglmpoisson noprint; model parkresid = parkpred; plot parkr esid*parkpred; run; quit; /* Park test by GLM with gamma distriubtion */ title 'Park Test for Heteroskedasticity (GLM with Gamma distribution)'; proc reg data=residualdatabyglmgamma noprint; model parkresid = parkpred; plot parkresid*parkpred; run; quit; /***************************************/ /*****PREPARE INSTRUMENTAL VARIABLE*****/ /***************************************/ /* Instrumental variable model Logistic Regression with following parameters: inputdataset predictorvar: left hand side of equation categ_class_stmt: name of macro for class statement for categorical variables; should be null if none covariate_list: covariate list including categoricals; outputdataset; outputvariabl e: name of variable predicted probability */ %macro IVMODEL(inputdataset,predictorvar,categ_class_stmt,covariate_list,outputdatas et,outputvariable); proc logistic data=&inputdataset; &categ_class_stmt; model earlydetect (ref=first) = &predictorvar &c ovariate_list / link=logit; output out=&outputdataset p=&outputvariable; run; quit;

PAGE 91

91 %mend IVMODEL; title 'First stage Instrumental Variable preparation Maximum Model'; %IVMODEL(cleandata,difftime,&categ_stmt_all,&covars_all,iv1_max,predictorvar phi); title 'Second Stage Instrumental Variable preparation Maximum Model'; %IVMODEL(iv1_max,predictorvar_phi,&categ_stmt_all,&covars_all,iv2_max,predict orvar_w); title 'First stage Instrumental Variable preparation Maximum Model without Income/Pover ty'; %IVMODEL(cleandata,difftime,&categ_stmt_noincome,&covars_noincome,iv1_noinc,p redictorvar_phi); title 'Second Stage Instrumental Variable preparation Maximum Model without Income/Poverty'; %IVMODEL(iv1_noinc,predictorvar_phi,&categ_stmt_noincome,& covars_noincome,iv2 _noinc,predictorvar_w); title 'First stage Instrumental Variable preparation Minimum Model'; %IVMODEL(cleandata,difftime,&categ_stmt_minimal,&covars_minimal,iv1_min,predi ctorvar_phi); title 'Second Stage Instrumental Variable preparat ion Minimum Model'; %IVMODEL(iv1_min,predictorvar_phi,&categ_stmt_minimal,&covars_minimal,iv2_min ,predictorvar_w); title 'First stage Instrumental Variable preparation No Covariates'; %IVMODEL(cleandata,difftime,,,iv1_min,predictorvar_phi); title 'Sec ond Stage Instrumental Variable preparation No Covariates'; %IVMODEL(iv1_min,predictorvar_phi,,,iv2_min,predictorvar_w); /************************/ /*****FINAL MODELING*****/ /************************/ /* Generalized Linear Model (no random effects) /GEE with Parameters: inputdata multilevel_flag: 1 if accounting for cluster, 0 if not clustervar: name of clustering var, insert dummy if not selectionvar: the selection variable or IV predictted probability categ_class_stmt: name of macro for class statement for categorical variables; should be null if none covariate_list: covariate list including categoricals; offset_flag: 1 if there is an offset variable; 2 if it goes in the covariate list; 0 if none; offs et_name: in form offset=offsetvariable; distributiontype: typically gamma or poisson covarstruct: IND, CS (exchangeable), UN, AR(1); */ %macro GENLINEARMODEL(inputdata,multilevel_flag,clustervar,selectionvar,categ_class_ stmt,covariate_list,offset_flag,offset_name,distributiontype,covarstruct); proc genmod data=&inputdata; &categ_class_stmt; %if &multilevel_flag=1 %then class &clustervar; ;

PAGE 92

92 mo del rawcost = &selectionvar &covariate_list %if &offset_flag=2 %then &offset_name; / link=log dist=&distributiontype %if &offset_flag=1 %then offset=&offset_name; type3 maxiter=100 corrb covb; %if &multilevel_flag=1 %then repeated subject = &clustervar / corr=&covarstruct; ; run; quit; %mend GENLINEARMODEL; title 'Generalized Linear Model with Multilevel Modeling Maximal Model No IV, offset as covariate, gamma distribution, exchangeable covariance matrix'; %GENLINEARMODEL(cleandata,1,bir thhospwithother,earlydetect,&categ_stmt_all,&c ovars_all,2,&offset_var,gamma,CS); title 'Generalized Linear Model with Multilevel Modeling Maximal Model Two Stage IV, offset as covariate, gamma distribution, exchangeable covariance matrix'; %GENLINEARMO DEL(iv1_noinc,1,birthhospwithother,predictorvar_phi,&categ_stmt_a ll,&covars_all,2,&offset_var,gamma,CS); title 'Generalized Linear Model with Multilevel Modeling Maximal Model Three Stage IV, offset as covariate, gamma distribution, exchangeable covari ance matrix'; %GENLINEARMODEL(iv2_noinc,1,birthhospwithother,predictorvar_w,&categ_stmt_all ,&covars_all,2,&offset_var,gamma,CS); title 'Generalized Linear Model WITHOUT Multilevel Modeling Minimal Model No IV, offset as covariate, gamma distribution, exchangeable covariance matrix'; %GENLINEARMODEL(cleandata,0,birthhospwithother,earlydetect,&categ_stmt_minima l,&covars_minimal,2,&offset_var,gamma,CS); title 'Generalized Linear Model WITHOUT Multilevel Modeling Minimal Model Two Stage IV, offset as c ovariate, gamma distribution, exchangeable covariance matrix'; %GENLINEARMODEL(iv1_noinc,0,birthhospwithother,predictorvar_phi,&categ_stmt_m inimal,&covars_minimal,2,&offset_var,gamma,CS); title 'Generalized Linear Model WITHOUT Multilevel Modeling Minima l Model Three Stage IV, offset as covariate, gamma distribution, exchangeable covariance matrix'; %GENLINEARMODEL(iv2_noinc,0,birthhospwithother,predictorvar_w,&categ_stmt_min imal,&covars_minimal,2,&offset_var,gamma,CS); title 'Generalized Linear Model with Multilevel Modeling Minimal Model No IV, offset as covariate, gamma distribution, exchangeable covariance matrix'; %GENLINEARMODEL(cleandata,1,birthhospwithother,earlydetect,&categ_stmt_minima l,&covars_minimal,2,&offset_var,gamma,CS); title 'Gene ralized Linear Model with Multilevel Modeling Minimal Model Two Stage IV, offset as covariate, gamma distribution, exchangeable covariance matrix'; %GENLINEARMODEL(iv1_noinc,1,birthhospwithother,predictorvar_phi,&categ_stmt_m inimal,&covars_minimal,2,& offset_var,gamma,CS); title 'Generalized Linear Model with Multilevel Modeling Minimal Model Three Stage IV, offset as covariate, gamma distribution, exchangeable covariance matrix';

PAGE 93

93 %GENLINEARMODEL(iv2_noinc,1,birthhospwithother,predictorvar_w,&categ_ stmt_min imal,&covars_minimal,2,&offset_var,gamma,CS); title 'Generalized Linear Model with Multilevel Modeling Minimal Model No IV, offset as offset, gamma distribution, exchangeable covariance matrix'; %GENLINEARMODEL(cleandata,1,birthhospwithother, earlydetect,&categ_stmt_minima l,&covars_minimal,1,&offset_var,gamma,CS); title 'Generalized Linear Model with Multilevel Modeling Minimal Model Two Stage IV, offset as offset, gamma distribution, exchangeable covariance matrix'; %GENLINEARMODEL(iv1_no inc,1,birthhospwithother,predictorvar_phi,&categ_stmt_m inimal,&covars_minimal,1,&offset_var,gamma,CS); title 'Generalized Linear Model with Multilevel Modeling Minimal Model Three Stage IV, offset as offset, gamma distribution, exchangeable covariance matrix'; %GENLINEARMODEL(iv2_noinc,1,birthhospwithother,predictorvar_w,&categ_stmt_min imal,&covars_minimal,1,&offset_var,gamma,CS); title 'Generalized Linear Model with Multilevel Modeling Minimal Model No IV, offset as covariate, gamma distribution, unstructured covariance matrix'; %GENLINEARMODEL(cleandata,1,birthhospwithother,earlydetect,&categ_stmt_minima l,&covars_minimal,2,&offset_var,gamma,UN); title 'Generalized Linear Model with Multilevel Modeling Minimal Model Two Stage IV, offset as cov ariate, gamma distribution, unstructured covariance matrix'; %GENLINEARMODEL(iv1_noinc,1,birthhospwithother,predictorvar_phi,&categ_stmt_m inimal,&covars_minimal,2,&offset_var,gamma,UN); title 'Generalized Linear Model with Multilevel Modeling Minimal Mo del Three Stage IV, offset as covariate, gamma distribution, unstructured covariance matrix'; %GENLINEARMODEL(iv2_noinc,1,birthhospwithother,predictorvar_w,&categ_stmt_min imal,&covars_minimal,2,&offset_var,gamma,UN); title 'Generalized Linear Model with Multilevel Modeling Minimal Model No IV, offset as offset, gamma distribution, unstructured covariance matrix'; %GENLINEARMODEL(cleandata,1,birthhospwithother,earlydetect,&categ_stmt_minima l,&covars_minimal,1,&offset_var,gamma,UN); title 'Generalized Linear Model with Multilevel Modeling Minimal Model Two Stage IV, offset as offset, gamma distribution, unstructured covariance matrix'; %GENLINEARMODEL(iv1_noinc,1,birthhospwithother,predictorvar_phi,&categ_stmt_m inimal,&covars_minimal,1,&offset_var ,gamma,UN); title 'Generalized Linear Model with Multilevel Modeling Minimal Model Three Stage IV, offset as offset, gamma distribution, unstructured covariance matrix'; %GENLINEARMODEL(iv2_noinc,1,birthhospwithother,predictorvar_w,&categ_stmt_min imal, &covars_minimal,1,&offset_var,gamma,UN); title 'Generalized Linear Model with Multilevel Modeling Minimal Model No IV, offset as covariate, poisson distribution, exchangeable covariance matrix'; %GENLINEARMODEL(cleandata,1,birthhospwithother,earlydetect,&categ_stmt_minima l,&covars_minimal,2,&offset_var,poisson,CS); title 'Generalized Linear Model with Multilevel Modeling Minimal Model Two Stage IV, offset as covariate, poisson distribution, e xchangeable covariance matrix';

PAGE 94

94 %GENLINEARMODEL(iv1_noinc,1,birthhospwithother,predictorvar_phi,&categ_stmt_m inimal,&covars_minimal,2,&offset_var,poisson,CS); title 'Generalized Linear Model with Multilevel Modeling Minimal Model Three Stage IV, offset as covariate, poisson distribution, exchangeable covariance matrix'; %GENLINEARMODEL(iv2_noinc,1,birthhospwithother,predictorvar_w,&categ_stmt_min imal,&covars_minimal,2,&offset_var,poisson,CS); title 'Generalized Linear Model with Multilevel Modeling M inimal Model No IV, offset as offset, poisson distribution, exchangeable covariance matrix'; %GENLINEARMODEL(cleandata,1,birthhospwithother,earlydetect,&categ_stmt_minima l,&covars_minimal,1,&offset_var,poisson,CS); title 'Generalized Linear Model with M ultilevel Modeling Minimal Model Two Stage IV, offset as offset, poisson distribution, exchangeable covariance matrix'; %GENLINEARMODEL(iv1_noinc,1,birthhospwithother,predictorvar_phi,&categ_stmt_m inimal,&covars_minimal,1,&offset_var,poisson,CS); titl e 'Generalized Linear Model with Multilevel Modeling Minimal Model Three Stage IV, offset as offset, poisson distribution, exchangeable covariance matrix'; %GENLINEARMODEL(iv2_noinc,1,birthhospwithother,predictorvar_w,&categ_stmt_min imal,&covars_minima l,1,&offset_var,poisson,CS); /*************/ /*****END*****/ /*************/

PAGE 95

95 STATA Code use "P: \ Data \ CHD_data.dta", clear treatreg logcost nonwhite premature lowbirthwt extracardiac univentricular archobstruction logenrollmo, treat(earlydetect = diff time female multgest cesarean)

PAGE 96

96 REFERENCES 1. Flanagan MF YS, Weindling SN. Cardiac Disease. In: Avery GB F, MacDonald MG, ed. Neonatology: Pathophysiology and Management of the Newborn Philadelphia, PA: Lippincott William and Wilkins; 1999:577 646. 2. Ferencz C, Loffredo CA, Rubin JD. Epidemiology of Congenital Heart Disease: The Baltimore Washington Study 1981 1989 Mount Kisco, NY: Futura Publishing; 1993. 3. Lorenzo DB, Goldmuntz E, Lin AE. Epidemiology and prevention of congenital heart defects. In: Allen HD, Driscoll DJ, Shaddy RE, Feltes TF, eds. Moss and Adams' Heart Disease in Infants, Children, and Adolescents Including the Fetus and Young Adult. Vol 1. 7 ed. Philadelphia, PA: Lippincott Williams & Wilkins; 2008:524 545. 4. Hoffman JI. Congenital heart disease: incidence and inheritance. Pediatr Clin North Am. Feb 1990;37(1):25 43. 5. CDC. Racial differences by gestational age in neonatal deaths attributa ble to congenital heart defects --United States, 2003 2006. MMWR Morb Mortal Wkly Rep. Sep 24 2010;59(37):1208 1211. 6. Schlingmann TR, Thiagarajan RR, Gauvreau K, et al. Cardiac Medical Conditions Have Become the Leading Cause of Death in Children with Heart Disease. Congenit Heart Dis. 2012;7(6):551 558;. 7. Brown KL, Ridout DA, Hoskote A, Verhulst L, Ricci M, Bull C. Delayed diagnosis of congenital heart disease worsens preoperative condition and outcome of surgery in neonates. Heart. Sep 2006;92(9):12 98 1302. 8. Schultz AH, Localio AR, Clark BJ, Ravishankar C, Videon N, Kimmel SE. Epidemiologic features of the presentation of critical congenital heart disease: implications for screening. Pediatrics. Apr 2008;121(4):751 757. 9. Meberg A, Brugmann Pieper S, Due R, Jr., et al. First day of life pulse oximetry screening to detect congenital heart defects. J Pediatr. Jun 2008;152(6):761 765. 10. Ewer AK, Middleton LJ, Furmston AT, et al. Pulse oximetry screening for congenital heart defects in newborn infant s (PulseOx): a test accuracy study. Lancet. Aug 27 2011;378(9793):785 794. 11. de Wahl Granelli A, Wennergren M, Sandberg K, et al. Impact of pulse oximetry screening on the detection of duct dependent congenital heart disease: a Swedish prospective screen ing study in 39,821 newborns. BMJ. 2009;338:a3037. 12. Walsh W. Evaluation of pulse oximetry screening in Middle Tennessee: cases for consideration before universal screening. J Perinatol. Feb 2011;31(2):125 129.

PAGE 97

97 13. Newborn Coalition. CCHD Screening Map. 2012; http://www.cchdscreeningmap.com/ Accessed March 1, 2013. 14. Mahle WT, Martin GR, Beekman RH, 3rd, Morrow WR. Endorsement of Health and Human Services recommendation for pulse oximetry screening for c ritical congenital heart disease. Pediatrics. Jan 2012;129(1):190 192. 15. Mahle WT, Newburger JW, Matherne GP, et al. Role of pulse oximetry in examining newborns for congenital heart disease: a scientific statement from the AHA and AAP. Pediatrics. Aug 2 009;124(2):823 836. 16. Kemper AR, Mahle WT, Martin GR, et al. Strategies for implementing screening for critical congenital heart disease. Pediatrics. Nov 2011;128(5):e1259 1267. 17. Riede FT, Worner C, Dahnert I, Mockel A, Kostelka M, Schneider P. Effectiveness of neonatal pulse oximetry screening for detection of critical congenital heart disease in daily clinical routine -results from a prospective multicenter study. Eur J Pediatr. Aug 2010;169(8):975 981. 18. Reich JD, Connolly B, Bradley G, et al. The reliability of a single pulse oximetry reading as a screening test for congenital heart disease in otherwise asymptomatic newborn infants. Pediatr Cardiol. Sep 2008;29(5):885 889. 19. Knowles R, Griebsch I, Dezateux C, Brown J, Bull C, Wren C. Newborn screening for congenital heart defects: a systematic review and cost effectiveness analysis. Health Technol Assess. Nov 2005;9(44):1 152, iii iv. 20. Griebsch I, Knowles RL, Brown J, Bull C, Wren C, Dezateux CA. Comparing the clinical and economic effects of clinical examination, pulse oximetry, and echocardiography in newborn screening for congenital heart defects: a probabilistic cost effectiveness model and value of information analysis Int J Technol Assess Health Care. Spring 2007;23(2):192 204. 21. Ewer AK, Furmston AT, Middleton LJ, et al. Pulse oximetry as a screening test for congenital heart defects in newborn infants: a test accuracy study with evaluation of acceptability and cos t effectiveness. Health Technol Assess. 2012;16(2):v xiii, 1 184. 22. Pulse Oximetry Screening for Critical Congenital Heart Defects. 2012; http://www.cdc.gov/ncbddd/pediatricgenetics/p ulse.html Accessed November 12, 2012. 23. Knapp AA, Metterville DR, Kemper AR, Prosser L, Perrin JM. Evidence review: critical congenital cyanotic heart disease, final draft September 3 2010. 24. Chang RK, Gurvitz M, Rodriguez S. Missed diagnosis of criti cal congenital heart disease. Arch Pediatr Adolesc Med. Oct 2008;162(10):969 974. 25. Abu Harb M, Hey E, Wren C. Death in infancy from unrecognised congenital heart disease. Arch Dis Child. Jul 1994;71(1):3 7.

PAGE 98

98 26. Abu Harb M, Wyllie J, Hey E, Richmond S, W ren C. Presentation of obstructive left heart malformations in infancy. Arch Dis Child Fetal Neonatal Ed. Nov 1994;71(3):F179 183. 27. Kuehl KS, Loffredo CA, Ferencz C. Failure to diagnose congenital heart disease in infancy. Pediatrics. Apr 1999;103(4 Pt 1):743 747. 28. Hoke TR, Donohue PK, Bawa PK, et al. Oxygen saturation as a screening test for critical congenital heart disease: a preliminary study. Pediatr Cardiol. Jul Aug 2002;23(4):403 409. 29. Richmond S, Reay G, Abu Harb M. Routine pulse oximetry in the asymptomatic newborn. Arch Dis Child Fetal Neonatal Ed. Sep 2002;87(2):F83 88. 30. Koppel RI, Druschel CM, Carter T, et al. Effectiveness of pulse oximetry screening for congenital heart disease in asymptomatic newborns. Pediatrics. Mar 2003;111(3):451 455. 31. Reich JD, Miller S, Brogdon B, et al. The use of pulse oximetry to detect congenital heart disease. J Pediatr. Mar 2003;142(3):268 272. 32. Bakr AF, Habib HS. Combining pulse oximetry and clinical examination in screening for congenital heart disease. Pediatr Cardiol. Nov Dec 2005;26(6):832 835. 33. Rosati E, Chitano G, Dipaola L, De Felice C, Latini G. Indications and limitations for a neonat al pulse oximetry screening of critical congenital heart disease. J Perinat Med. 2005;33(5):455 457. 34. Arlettaz R, Bauschatz AS, Monkhoff M, Essers B, Bauersfeld U. The contribution of pulse oximetry to the early detection of congenital heart disease in newborns. Eur J Pediatr. Feb 2006;165(2):94 98. 35. Ruangritnamchai C, Bunjapamai W, Pongpanich B. Pulse oximetry screening for clinically unrecognized critical congenital heart disease in the newborns. Images Paediatr Cardiol. Jan 2007;9(1):10 15. 36. Meb erg A, Andreassen A, Brunvand L, et al. Pulse oximetry screening as a complementary strategy to detect critical congenital heart defects. Acta Paediatr. Apr 2009;98(4):682 686. 37. Bradshaw EA, Cuzzi S, Kiernan SC, Nagel N, Becker JA, Martin GR. Feasibilit y of implementing pulse oximetry screening for congenital heart disease in a community hospital. J Perinatol. Sep 2012;32(9):710 715. 38. Turska Kmiec A, Borszewska Kornacka MK, Blaz W, Kawalec W, Zuk M. Early screening for critical congenital heart defect s in asymptomatic newborns in Mazovia province: experience of the POLKARD pulse oximetry programme 2006 2008 in Poland. Kardiol Pol. 2012;70(4):370 376.

PAGE 99

99 39. IOM. Introduction and Literature Review. In: Smedley BD, Stith AY, Nelson AR, eds. Unequal Treatmen t: Confronting Racial and Ethnic Disparities in Health Care : National Academies Press; 2003:29 79. 40. Fixler DE, Pastor P, Sigman E, Eifler CW. Ethnicity and socioeconomic status: impact on the diagnosis of congenital heart disease. J Am Coll Cardiol. Jun 1993;21(7):1722 1726. 41. Williams RV, Ravishankar C, Zak V, et al. Birth weight and prematurity in infants with single ventricle physiology: pediatric heart network infant single ventricle trial screened population. Congenit Heart Dis. Mar Apr 2010;5(2): 96 103. 42. Riede FT, Dahnert I, Schneider P, Mockel A. Pulse oximetry screening at 4 hours of age to detect critical congenital heart defects. Pediatrics. Mar 2009;123(3):e542; author reply e542 543. 43. Sendelbach DM, Jackson GL, Lai SS, Fixler DE, Stehe l EK, Engle WD. Pulse oximetry screening at 4 hours of age to detect critical congenital heart defects. Pediatrics. Oct 2008;122(4):e815 820. 44. Thangaratinam S, Daniels J, Ewer AK, Zamora J, Khan KS. Accuracy of pulse oximetry in screening for congenital heart disease in asymptomatic newborns: a systematic review. Arch Dis Child Fetal Neonatal Ed. May 2007;92(3):F176 180. 45. Valmari P. Should pulse oximetry be used to screen for congenital heart disease? Arch Dis Child Fetal Neonatal Ed. May 2007;92(3):F 219 224. 46. Liske MR, Greeley CS, Law DJ, et al. Report of the Tennessee Task Force on Screening Newborn Infants for Critical Congenital Heart Disease. Pediatrics. Oct 2006;118(4):e1250 1256. 47. Bradshaw EA, Cuzzi S, Kiernan SC, Nagel N, Becker JA, Marti n GR. Feasibility of implementing pulse oximetry screening for congenital heart disease in a community hospital. J Perinatol. Jan 26 2012. 48. Thangaratinam S, Brown K, Zamora J, Khan KS, Ewer AK. Pulse oximetry screening for critical congenital heart defe cts in asymptomatic newborn babies: a systematic review and meta analysis. Lancet. Jun 30 2012;379(9835):2459 2464. 49. Friedberg MK, Silverman NH, Moon Grady AJ, et al. Prenatal detection of congenital heart disease. J Pediatr. Jul 2009;155(1):26 31, 31 e21. 50. Pinto NM, Keenan HT, Minich LL, Puchalski MD, Heywood M, Botto LD. Barriers to prenatal detection of congenital heart disease: a population based study. Ultrasound Obstet Gynecol. Oct 2012;40(4):418 425. 51. Peiris V, Singh TP, Tworetzky W, Chong EC, Gauvreau K, Brown DW. Association of socioeconomic position and medical insurance with fetal diagnosis of

PAGE 100

100 critical congenital heart disease. Circ Cardiovasc Qual Outcomes. Jul 2009;2(4):354 360. 52. Levy DJ, Pretorius DH, Rothman A, et al. Improved Pre natal Detection of Congenital Heart Disease in an Integrated Health Care System. Pediatr Cardiol. Nov 2 2012. 53. Powell R, Pattison HM, Bhoyar A, et al. Pulse oximetry screening for congenital heart defects in newborn infants: an evaluation of acceptabili ty to mothers. Arch Dis Child Fetal Neonatal Ed. May 18 2012. 54. Sebelius KT. 2011; http://www.hrsa.gov/ad visorycommittees/mchbadvisory/heritabledisorders/recommendat ions/correspondence/cyanoticheartsecre09212011.pdf 55. de Wahl Granelli A, Mellander M, Sunnegardh J, Sandberg K, Ostman Smith I. Screening for duct dependant congenital heart disease with pulse oximetry: a critical evaluation of strategies to maximize sensitivity. Acta Paediatr. Nov 2005;94(11):1590 1596. 56. Bakr AF, Habib HS. Normal values of pulse oximetry in newborns at high altitude. J Trop Pediatr. Jun 2005;51(3):170 173. 57. de Wahl Granelli AW, Ostman Smith I. Noninvasive peripheral perfusion index as a possible tool for screening for critical left heart obstruction. Acta Paediatr. Oct 2007;96(10):1455 1459. 58. Riede FT, Schneider P. Most wanted, least found: coarctation. Co ncerning the article by J.I.E. Hoffman: It is time for routine neonatal screening by pulse oximetry [Neonatology 2011;99:1 9]. Neonatology. 2012;101(1):13; author reply 13. 59. Ruegger C, Bucher HU, Mieth RA. Pulse oximetry in the newborn: is the left hand pre or post ductal? BMC Pediatr. 2010;10:35. 60. Pulse oximetry screening in newborns: a policy position from the American Heart Association : American Heart Association; June 2012. 61. CDC. Newborn screening for critical congenital heart disease: potenti al roles of birth defects surveillance programs United States, 2010 2011. MMWR Morb Mortal Wkly Rep. Oct 26 2012;61:849 853. 62. Chang RK, Rodriguez S, Klitzner TS. Screening newborns for congenital heart disease with pulse oximetry: survey of pediatric cardiologists. Pediatr Cardiol. Jan 2009;30(1):20 25. 63. Kuelling B, Arlettaz Mieth R, Bauersfeld U, Balmer C. Pulse oximetry screening for congenital heart defects in Switzerland: most but not all maternity units screen their neonates. Swiss Med Wkly. No v 28 2009;139(47 48):699 704.

PAGE 101

101 64. Shastri AT, Clarke P, Roy R. Pulse oximetry screening for detection of critical congenital heart disease in newborns: a survey of current practices in the United Kingdom. Acta Paediatr. May 2011;100(5):636 637. 65. Kang SL, Tobin S, Kelsall W. Neonatal pulse oximetry screening: a national survey. Arch Dis Child Fetal Neonatal Ed. Jul 2011;96(4):F312. 66. Singh A, Ewer AK. Pulse oximetry screening for critical congenital heart defects: a UK national survey. Lancet. Fe b 16 2013;381(9866):535. 67. Hoffman JI. It is time for routine neonatal screening by pulse oximetry. Neonatology. 2011;99(1):1 9. 68. A new milestone in the history of congenital heart disease. Lancet. Jun 30 2012;379(9835):2401. 69. Swenson AK, Brown D, Stevermer JJ. PURLs: Pulse oximetry for newborns: should it be routine? J Fam Pract. May 2012;61(5):283 286. 70. IOM. Unequal Treatment: Confronting Racial and Ethnic Disparities in Health Care. In: Smedley BD, Stith AY, Nelson AR, eds. Washington, D.C.: T he National Academies Press; 2003: http://www.iom.edu/Reports/2002/Unequal Treatment Confronting Racial and Ethnic Disparities i n Health Care.aspx Accessed March 12, 2013. 71. DHHS. Healthy People 2010, 2nd ed. With understanding and Improving Health and Objectives for Improving Helath, 2 vols. Washington, D.C.: U.S. Department of Health and Human Services; 2000: http://www.healthypeople.gov/2010/ Accessed March 12, 2013. 72. Boneva RS, Botto LD, Moore CA, Yang Q, Correa A, Erickson JD. Mortality associated with congenital heart defects in the United States: trends and racial dis parities, 1979 1997. Circulation. May 15 2001;103(19):2376 2381. 73. Sharland G. Fetal cardiac screening: why bother? Arch Dis Child Fetal Neonatal Ed. Jan 2010;95(1):F64 68. 74. Perlstein MA, Goldberg SJ, Meaney FJ, Davis MF, Zwerdling Kluger C. Factors i nfluencing age at referral of children with congenital heart disease. Arch Pediatr Adolesc Med. Sep 1997;151(9):892 897. 75. Nembhard WN, Pathak EB, Schocken DD. Racial/ethnic disparities in mortality related to congenital heart defects among children and adults in the United States. Ethn Dis. Autumn 2008;18(4):442 449. 76. Nembhard WN, Salemi JL, Ethen MK, Fixler DE, Dimaggio A, Canfield MA. Racial/Ethnic disparities in risk of early childhood mortality among children with congenital heart defects. Pediatr ics. May 2011;127(5):e1128 1138.

PAGE 102

102 77. Chang RK, Chen AY, Klitzner TS. Factors associated with age at operation for children with congenital heart disease. Pediatrics. May 2000;105(5):1073 1081. 78. Ingaramo OA, Khemani RG, Markovitz BP, Epstein D. Effect of race on the timing of the Glenn and Fontan procedures for single ventricle congenital heart disease. Pediatr Crit Care Med. Mar 2012;13(2):174 177. 79. Milazzo AS, Jr., Sanders SP, Armstrong BE, Li JS. Racial and geographic disparities in timing of bidirectional Glenn and Fontan stages of single ventricle palliation. J Natl Med Assoc. Oct 2002;94(10):873 878. 80. Erickson LC, Wise PH, Cook EF, Beiser A, N ewburger JW. The impact of managed care insurance on use of lower mortality hospitals by children undergoing cardiac surgery in California. Pediatrics. Jun 2000;105(6):1271 1278. 81. Marino BS, Lipkin PH, Newburger JW, et al. Neurodevelopmental outcomes in children with congenital heart disease: evaluation and management: a scientific statement from the American Heart Association. Circulation. Aug 28 2012;126(9):1143 1172. 82. Bellinger DC, Wypij D, duPlessis AJ, et al. Neurodevelopmental status at eight ye ars in children with dextro transposition of the great arteries: the Boston Circulatory Arrest Trial. J Thorac Cardiovasc Surg. Nov 2003;126(5):1385 1396. 83. Shillingford AJ, Glanzman MM, Ittenbach RF, Clancy RR, Gaynor JW, Wernovsky G. Inattention, hyper activity, and school performance in a population of school age children with complex congenital heart disease. Pediatrics. Apr 2008;121(4):e759 767. 84. Visconti KJ, Saudino KJ, Rappaport LA, Newburger JW, Bellinger DC. Influence of parental stress and soc ial support on the behavioral adjustment of children with transposition of the great arteries. J Dev Behav Pediatr. Oct 2002;23(5):314 321. 85. Atallah J, Dinu IA, Joffe AR, et al. Two year survival and mental and psychomotor outcomes after the Norwood pro cedure: an analysis of the modified Blalock Taussig shunt and right ventricle to pulmonary artery shunt surgical eras. Circulation. Sep 30 2008;118(14):1410 1418. 86. Leblanc JG. Creating a global climate for pediatric cardiac care. World J Pediatr. May 2009;5(2):89 92. 87. The SAS System, version 9.3 [computer program]. Cary, NC: SAS Institute, Inc.; 2011. 88. Benavidez OJ, Gauvreau K, Jenkins KJ. Racial and ethnic disparities in mortality following congenital heart surgery. Pediatr Cardiol. May Jun 2006 ;27(3):321 328.

PAGE 103

103 89. Archer JM, Yeager SB, Kenny MJ, Soll RF, Horbar JD. Distribution of and mortality from serious congenital heart disease in very low birth weight infants. Pediatrics. Feb 2011;127(2):293 299. 90. Jenkins KJ, Gauvreau K, Newburger JW, Spr ay TL, Moller JH, Iezzoni LI. Consensus based method for risk adjustment for surgery for congenital heart disease. J Thorac Cardiovasc Surg. Jan 2002;123(1):110 118. 91. Zupancic JA, Richardson DK, Horbar JD, Carpenter JH, Lee SK, Escobar GJ. Revalidation of the Score for Neonatal Acute Physiology in the Vermont Oxford Network. Pediatrics. Jan 2007;119(1):e156 163. 92. Richardson DK, Corcoran JD, Escobar GJ, Lee SK. SNAP II and SNAPPE II: Simplified newborn illness severity and mortality risk scores. J Pedi atr. Jan 2001;138(1):92 100. 93. The CRIB (clinical risk index for babies) score: a tool for assessing initial neonatal risk and comparing performance of neonatal intensive care units. The International Neonatal Network. Lancet. Jul 24 1993;342(8865):193 1 98. 94. Hughes JS, Averill RF, Eisenhandler J, et al. Clinical Risk Groups (CRGs): a classification system for risk adjusted capitation based payment and health care management. Med Care. Jan 2004;42(1):81 90. 95. Neff JM, Sharp VL, Muldoon J, Graham J, Po palisky J, Gay JC. Identifying and classifying children with chronic conditions using administrative data with the clinical risk group classification system. Ambul Pediatr. Jan Feb 2002;2(1):71 79. 96. United States Census Bureau American Community Survey. United States Department of Commerce; 2010. http://www.census.gov/acs/www/ 97. Annual Survey custom dataset [computer program]. Washington, DC: American Hospital Association; 2006 2007. 98. ArcGIS Desktop: Release 10 [computer program]. Redlands, CA: Environmental Systems Research Institute; 2011. 99. Kreidler SK, Muller KE, Glueck DH. GLIMMPSE. 2012; htt p://glimmpse.samplesizeshop.org/ Accessed 24 December, 2013. 100. Glueck DH, Muller KE. Adjusting power for a baseline covariate in linear models. Stat Med. 2003;22(16):2535 2551. 101. Manning WG, Mullahy J. Estimating log models: to transform or not to t ransform? J Health Econ. Jul 2001;20(4):461 494.

PAGE 104

104 102. D'Agostino RB, Jr. Propensity score methods for bias reduction in the comparison of a treatment to a non randomized control group. Stat Med. Oct 15 1998;17(19):2265 2281. 103. Linden A, Adams JL. Using propensity score based weighting in the evaluation of health management programme effectiveness. J Eval Clin Pract. Feb 2010;16(1):175 179. 104. Seeger JD, Kurth T, Walker AM. Use of propensity score technique to account for exposure related covariates: an example and lesson. Med Care. Oct 2007;45(10 Supl 2):S143 148. 105. Ounpraseuth S, Gauss CH, Bronstein J, Lowery C, Nugent R, Hall R. Evaluating the effect of hospital and insurance type on the risk of 1 year mortality of very low birth weight infants: co ntrolling for selection bias. Med Care. Apr 2012;50(4):353 360. 106. Lorch SA, Baiocchi M, Ahlberg CE, Small DS. The differential impact of delivery hospital on the outcomes of premature infants. Pediatrics. Aug 2012;130(2):270 278. 107. Wehby GL, Ullrich F, Xie Y. Very low birth weight hospital volume and mortality: an instrumental variables approach. Med Care. Aug 2012;50(8):714 721. 108. McClellan M, McNeil BJ, Newhouse JP. Does more intensive treatment of acute myocardial infarction in the elderly reduc e mortality? Analysis using instrumental variables. JAMA. Sep 21 1994;272(11):859 866. 109. Beck CA, Penrod J, Gyorkos TW, Shapiro S, Pilote L. Does aggressive care following acute myocardial infarction reduce mortality? Analysis with instrumental variable s to compare effectiveness in Canadian and United States patient populations. Health Serv Res. Dec 2003;38(6 Pt 1):1423 1440. 110. Wooldridge J. Econometric Analysis of Cross Section and Panel Data 2 ed. Cambridge, MA: MIT PRess; 2010. 111. Heckman JJ. Sample Selection Bias as a Specification Error. Econometrica. 1979;47(1):153 161. 112. Ettner SL. Methods for Addressing Selection Bias in Observational Studies. NRSA Trainees Research Conference Slide Presentation. 2004; http://www.ahrq.gov/fund/training/ettnertxt.htm Accessed February 28, 2013. 113. Stata Statistical Software: Release 10 [computer program]. College Station, TX: StataCorp, LP; 2007. 114. Liu L, Strawderman RL, Cow en ME, Shih YC. A flexible two part random effects model for correlated medical costs. J Health Econ. Jan 2010;29(1):110 123.

PAGE 105

105 115. TXDSHS. Births by Public Health Region, County, and City of Residence, Texas, 2008. 2011; http://www.dshs.state.tx.us/chs/vstat/latest/t09t.shtm Accessed March 12, 2013. 116. TXDSHS. Births by Public Health Region, County, and City of Residence, Texas, 2009. 2012; http://www.dshs.state.tx.us/chs/vstat/vs09/t09t.shtm Accessed March 12, 2013. 117. CDC. CDC Features: National Prematurity Awareness Month. 2013. Accessed March 5, 2013. 118. Costello JM, Polito A, Brown DW, Et.al. Birth before 39 weeks' gestation is associated with worse outcomes in neonates with heart disease. Pediatrics. 2010;126(2):375 461. 119. Shadish WR, Cook TD, Campbell DT. Experimental and Quasi experimental Designs for Generalized Causal Inference : Houghton Mifflin; 2003. 120. Marino BS, Shera D, Wernovsky G, et al. The development of the pediatric cardiac quality of life inventory: a quality of life measure for children and adolescents with heart disease. Qual Life Res. May 2008;17(4):613 626. 121. Macran S, Birks Y, Parsons J, et al. The development of a new measure of quality of life for children with congenital cardiac disease. Cardiol Young. Apr 2006;16(2):165 172. 122. Kendall L, Lewin RJ, Parsons JM, Veldtman GR, Quirk J, Hardman GE. Factors associated with self perceived state of health in adolescents with congenital cardiac disease attending paediatric cardiologic clinics. Cardiol Young. Jul 2001;11(4):431 438.

PAGE 106

106 BIOGRAPHICAL SKETCH The author was born in Delaware and grew up in Gainesville, FL. He received his Bachelor of Arts, with highest honors, from Princeton University in 1998, with an additional Certificate in Applications of Computing. H e earned his Doctor of Medicine, with honors, from the University of Florida College of Medicinein 2004. He completed residency training in Pediatrics at the University of Vermont in 2009, and served there as Chief Resident from 2009 2010. He is currently completing a fellowship in Pediatri c Cardiology and a Master of Science degree with a concentration in Health Outcomes and Policy at the University of Florida He lives in Gainesville, Florida with his wi fe and three wonderful children. His future plans involve Pediatric Cardiology practi ce in Montana with continued involvement in child health policy and research