University of Southern California Dissertations and Theses

The development of the prognostic propensity score: an introduction to a method to identify optimal treatment according to individual tailoring variables when heterogeneous treatment effects are ...

(USC Thesis Other)

The development of the prognostic propensity score: an introduction to a method to identify optimal treatment according to individual tailoring variables when heterogeneous treatment effects are ...

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

THE DEVELOPMENT OF THE PROGNOSTIC PROPENSITY SCORE:
AN INTRODUCTION TO A METHOD TO IDENTIFY OPTIMAL TREATMENT
ACCORDING TO INDIVIDUAL TAILORING VARIABLES WHEN
HETEROGENEOUS TREATMENT EFFECTS ARE PRESENT

by

Dana Renée Stafkey-Mailey

A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirement for the Degree
DOCTOR OF PHILOSOPHY
(PHARMACEUTICAL ECONOMICS & POLICY)

December 2008

Copyright 2008 Dana Renée Stafkey-Mailey
ii
DEDICATION

This dissertation is dedicated to my wonderful family, especially…

my husband, Mark, whose love and support provided me the strength and perseverance I
needed to achieve my goals.

my parents, Ruby and Bruce Stafkey, who have always had confidence in me and offered
me encouragement and support in all my endeavors.

my sister, Kara, and brother, Jared, whose continuous harassment of a life long student
has only proven to be inspirational.

iii
ACKNOWLEDGEMENTS

I wish to thank my committee members who were more than generous with their
expertise and precious time. A special thanks to Dr. Jeff McCombs, my committee
chairman, for his countless hours of reflecting, reading, and encouraging. Thank you Dr.
Jeonghoon Ahn, and Dr. Geert Ridder for providing the econometric support and
feedback necessary to write a paper of this nature. And thank you Dr. Kathleen Johnson,
and Dr Glenn Stimmel for providing invaluable clinical feedback during the initial review
of this dissertation proposal.
iv
TABLE OF CONTENTS

DEDICATION .................................................................................................................... ii

ACKNOWLEDGEMENTS ............................................................................................... iii

LIST OF TABLES ............................................................................................................ vii

LIST OF FIGURES ............................................................................................................ x

ABSTRACT ..................................................................................................................... xix

CHAPTER 1: INTRODUCTION ....................................................................................... 1
1.1 Ambiguous Evidence ............................................................................................... 2
1.2 Inappropriate Care ................................................................................................... 4
1.3 Financial Impact on Society..................................................................................... 6
1.4 Drug Market Impact ................................................................................................. 7
1.5 Dissertation Objective .............................................................................................. 9

CHAPTER 2: BACKGROUND ...................................................................................... 11
2.1 Homogeneous Treatment Effects ........................................................................... 11
2.2 Heterogeneous Treatment Effects (HTE) .............................................................. 12
2.2.1 Quantitative HTE ............................................................................................ 12
2.2.2 Qualitative HTE .............................................................................................. 13
2.2.3 Focus of this Dissertation................................................................................ 14
2.3 Sources of Data for Determining HTE .................................................................. 15
2.3.1 Randomized Controlled Trials (RCTs) ........................................................... 15
2.3.2 Observational Studies ..................................................................................... 16
2.3.3 Source of Data for this Dissertation ................................................................ 18
2.4 Methods.................................................................................................................. 19
2.4.1 Interaction ....................................................................................................... 19
2.4.2 Stratification .................................................................................................... 21
2.4.3 Most Appropriate Method to Use for this Dissertation .................................. 22
2.5 Techniques to Stratify Patients and to Control for Treatment Selection Bias ....... 23
2.5.1 Instrumental Variable (IV) .............................................................................. 23
2.5.2 Stratifying by Potential Confounders.............................................................. 25
2.5.3 Propensity Score (PS) ..................................................................................... 26
2.5.4 Multivariate Confounder Score (MCS) .......................................................... 27
2.5.5 Most appropriate Technique for Our Analysis ............................................... 29

CHAPTER 3: METHOD DEFINITION ......................................................................... 31
3.1 Challenges in Evaluating Treatment Effects .......................................................... 31
3.2 Definition of Prognostic Propensity Score (PPS) .................................................. 33
3.3 Controlling for Confounding ................................................................................. 34
3.4 Calculation of Prognostic Propensity Score (PPS) ................................................ 36
v
3.4.1 Overlap (or Support) Condition ...................................................................... 38
3.4.2 Balancing Condition ....................................................................................... 39
3.5 Identifying Heterogeneous Treatment Effects ....................................................... 40
3.6 Identifying Tailoring Variables.............................................................................. 43
3.7 Next Steps .............................................................................................................. 44

CHAPTER 4: APPLICATION OF METHOD ................................................................. 46
4.1 Application of the PPS Method in Schizophrenia ................................................. 46
4.2 Treatments Used in Schizophrenia ........................................................................ 48
4.3 Unit of Analysis ..................................................................................................... 51
4.4 Definition of Analysis Period ................................................................................ 57
4.5 Inclusion & Exclusion Criteria .............................................................................. 57
4.6 Outcome Measure .................................................................................................. 58
4.7 Covariates .............................................................................................................. 61
4.7.1 Type of schizophrenia ..................................................................................... 61
4.7.2 Demographic Factors ...................................................................................... 62
4.7.3 Comorbidities .................................................................................................. 63
4.7.4 Concomitant Therapy ...................................................................................... 65
4.7.5 Prior Antipsychotic Drug Use ......................................................................... 66
4.7.6 Episode type .................................................................................................... 66
4.7.7 Prior Health Services Utilization .................................................................... 67
4.8 Statistical Methods ................................................................................................. 67
4.8.1 Descriptive Statistics ....................................................................................... 67
4.8.2 Prognostic Propensity Score (PPS) ................................................................. 68
4.8.2.1 Calculation ............................................................................................... 68
4.8.2.2 Heterogeneous Treatment Effects ............................................................ 69
4.8.2.3 Tailoring Variables .................................................................................. 70
4.9 Next Steps .............................................................................................................. 72

CHAPTER 5: RESULTS ................................................................................................. 73
5.1 Data ........................................................................................................................ 73
5.2 Descriptive Statistics .............................................................................................. 74
5.3 Prognostic Propensity Score .................................................................................. 75
5.3.1 Calculation ...................................................................................................... 75
5.3.2 Overlap (or Support) Condition ...................................................................... 77
5.3.3 Balancing Condition ....................................................................................... 82
5.4 Heterogeneous Treatment Effects .......................................................................... 84
5.4.1 Crude Treatment Effect ................................................................................... 84
5.4.2 Adjusted Treatment Effect .............................................................................. 87
5.5 Tailoring Variables ................................................................................................ 96
5.6 Sensitivity Analyses ............................................................................................... 96
5.6.1 Separate Model with Propensity Score Calculated Prior to Stratifying .......... 97
5.6.2 Separate Model with Propensity Score Calculated After Stratifying ............. 99

vi
CHAPTER 6: CONCLUSION & DISCUSSION .......................................................... 101
6.1 Summary .............................................................................................................. 101
6.2 Prognostic Propensity Score ................................................................................ 102
6.3 Identify if Heterogeneous Treatment Effects are Present .................................... 104
6.3.1 Crude Treatment Effect ................................................................................. 104
6.3.2 Adjusted Treatment Effect ............................................................................ 106
6.4 Identify if Heterogeneous Treatment Effects are Quantitative or Qualitative ..... 107
6.5 Identify Tailoring Variables ................................................................................. 108
6.6 Implications of the Study Results ........................................................................ 109
6.7 Contributions to the Field .................................................................................... 110
6.8 Next Steps ............................................................................................................ 110

REFERENCES ............................................................................................................... 111

APPENDIX ..................................................................................................................... 119

vii
LIST OF TABLES

Table 4.1: Typical Antipsychotic Doses .......................................................................... 49

Table 4.2: Atypical Antipsychotic Year to Market & Doses ........................................... 50

Table 4.3: Covariates Representing Type of Schizophrenia ............................................. 61

Table 4.4: Demographic Covariates ................................................................................. 63

Table 4.5: Comorbidity Covariates ................................................................................... 63

Table 4.6: Concomitant Therapy Covariates .................................................................... 65

Table 4.7: Prior Antipsychotic Drug Use Covariates ....................................................... 66

Table 4.8: Prior Health Services Utilization Covariates ................................................... 67

Table 5.1: PPS Model Utilizing Olanzapine as the CT ................................................... 76

Table 5.2: Common Region of Support Considering each Treatment as the CT ............ 81

Table 5.3: Cutpoints for Quintiles when Olanzapine is Utilized as the CT ..................... 82

Table 5.4: Descriptive Statistics for Olanzapine Quintile 1 ............................................ 83

Table 5.5: Adjusted Treatment Effect within the Olanzapine Quintiles .......................... 90

Table 5.6: Adjusted Treatment Effect within the CT Quintiles ....................................... 92

Table A.1: Baseline Descriptive Statistics ..................................................................... 119

Table A.2: The PPS Model Considering Risperidone as the CT ................................... 122

Table A.3: The PPS Model Considering Quetiapine as the CT ..................................... 123

Table A.4: The PPS Model Considering Low Potency TAPs as the CT ....................... 124

Table A.5: The PPS Model Considering Medium Potency TAPs as the CT ................. 126

Table A.6: The PPS Model Considering High Potency TAPs as the CT ...................... 127

Table A.7: Cutpoints for Risperidone Quintiles ............................................................ 134
viii

Table A.8: Cutpoints for Quetiapine Quintiles .............................................................. 134

Table A.9: Cutpoints for Low Potency TAPs Quintiles ................................................ 134

Table A.10: Cutpoints for Medium Potency TAPs Quintiles ........................................ 134

Table A.11: Cutpoints for High Potency TAPs Quintiles.............................................. 134

Table A.12a: Descriptive Statistics for Olanzapine Quintile 1 ...................................... 135

Table A.12b: Descriptive Statistics for Olanzapine Quintile 2 ...................................... 137

Table A.12c: Descriptive Statistics for Olanzapine Quintile 3 ...................................... 139

Table A.12d: Descriptive Statistics for Olanzapine Quintile 4 ...................................... 141

Table A.12e: Descriptive Statistics for Olanzapine Quintile 5 ...................................... 143

Table A.13a: Descriptive Statistics for Risperidone Quintile 1 ..................................... 145

Table A.13b: Descriptive Statistics for Risperidone Quintile 2 .................................... 147

Table A.13c: Descriptive Statistics for Risperidone Quintile 3 ..................................... 149

Table A.13d: Descriptive Statistics for Risperidone Quintile 4 .................................... 151

Table A.13e: Descriptive Statistics for Risperidone Quintile 5 ..................................... 153

Table A.14a: Descriptive Statistics for Quetiapine Quintile 1 ...................................... 155

Table A.14b: Descriptive Statistics for Quetiapine Quintile 2 ...................................... 157

Table A.14c: Descriptive Statistics for Quetiapine Quintile 3 ...................................... 159

Table A.14d: Descriptive Statistics for Quetiapine Quintile 4 ...................................... 161

Table A.14e: Descriptive Statistics for Quetiapine Quintile 5 ...................................... 163

Table A.15a: Descriptive Statistics for Low Potency TAPs Quintile 1 ......................... 165

Table A.15b: Descriptive Statistics for Low Potency TAPs Quintile 2 ........................ 167

Table A.15c: Descriptive Statistics for Low Potency TAPs Quintile 3 ......................... 169

ix
Table A.15d: Descriptive Statistics for Low Potency TAPs Quintile 4 ........................ 171

Table A.15e: Descriptive Statistics for Low Potency TAPs Quintile 5 ......................... 173

Table A.16a: Descriptive Statistics for Medium Potency TAPs Quintile 1 .................. 175

Table A.16b: Descriptive Statistics for Medium Potency TAPs Quintile 2 .................. 177

Table A.16c: Descriptive Statistics for Medium Potency TAPs Quintile 3 .................. 179

Table A.16d: Descriptive Statistics for Medium Potency TAPs Quintile 4 .................. 181

Table A.16e: Descriptive Statistics for Medium Potency TAPs Quintile 5 .................. 183

Table A.17a: Descriptive Statistics for High Potency TAPs Quintile 1 ........................ 185

Table A.17b: Descriptive Statistics for High Potency TAPs Quintile 2 ........................ 187

Table A.17c: Descriptive Statistics for High Potency TAPs Quintile 3 ........................ 189

Table A.17d: Descriptive Statistics for High Potency TAPs Quintile 4 ........................ 191

Table A.17e: Descriptive Statistics for High Potency TAPs Quintile 5 ........................ 193

x
LIST OF FIGURES

Figure 2. 1: Homogeneous Treatment Effects ................................................................. 11
Figure 2. 2: Quantitative Heterogeneous Treatment Effects ........................................... 13
Figure 2. 3: Qualitative Heterogeneous Treatment Effects ............................................. 14
Figure 2. 4: Treatment Covariate Interaction ................................................................... 21
Figure 4. 1: Receptor Binding Profile of AAPs ............................................................... 51
Figure 4. 2: Continuous Therapy ..................................................................................... 52
Figure 4. 3: Example of a First Observed Episode .......................................................... 53
Figure 4. 4: Example of a Restart Episode ...................................................................... 54
Figure 4. 5: Example of a Swtich Episode ....................................................................... 54
Figure 4. 6: Example of an Augmentation Episode ......................................................... 56
Figure 4. 7: Example of a Late Switch Episode............................................................... 57
Figure 4. 8: Example Demonstration of TTAD Identification ........................................ 59
Figure 5.1: Inclusion & Exclusion Criteria ...................................................................... 73
Figure 5.2: Number of Episodes by Treatment ................................................................ 74
Figure 5.3: Box Plot of Olanzapine PPS – Success vs No Success ................................. 78
Figure 5.4: Histogram of Olanzapine PPS – Success vs No Success .............................. 79
Figure 5.5: Histogram of Olanzapine PPS by Treatment ................................................ 80
Figure 5.6: Refined Breakdown of Episodes by Treatment Type When Utilizing
Olanzapine as the CT ..................................................................................... 81
Figure 5.7: Crude Treatment Effect When Olanzapine is Utilized as the CT ................. 85
Figure 5.8: Adjusted Treatment Effect Utilizing Olanzapine as the CT
Olanzapine vs Risperidone ............................................................................ 89

xi
Figure 5.9: Adjusted Treatment Effect Utilizing Olanzapine as the CT
Olanzapine vs Medium Potency TAPs .......................................................... 91

Figure 5.10: Adjusted Treatment Effect Utilizing Risperidone as the CT
Risperidone vs Olanzapine .......................................................................... 94

Figure 5.11: Adjusted Treatment Effect Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Risperidone ............................................................ 95

Figure 5.12: Adjusted Treatment Effect with PS (Prior)
Utilizing Olanzapine as the CT Olanzapine vs Risperidone ........................ 98

Figure 5.13: Adjusted Treatment Effect with PS (Post)
Utilizing Olanzapine as the CT Olanzapine vs Risperidone ...................... 100

Figure A.1: Histogram of Risperidone PPS by Treatment ............................................ 129

Figure A.2: Refined Breakdown of Episodes by Treatment Type When Utilizing
Risperidone as the CT ................................................................................. 129

Figure A.3: Histogram of Quetiapine PPS by Treatment .............................................. 130

Figure A.4: Refined Breakdown of Episodes by Treatment Type When Utilizing
Quetiapine as the CT ................................................................................... 130

Figure A.5: Histogram of Low Potency TAPs PPS by Treatment ................................ 131

Figure A.6: Refined Breakdown of Episodes by Treatment Type
When Utilizing Low Potency TAPs as the CT ........................................... 131

Figure A.7: Histogram of Medium Potency TAPs PPS by Treatment .......................... 132

Figure A.8: Refined Breakdown of Episodes by Treatment Type When Utilizing
Medium Potency TAPs as the CT ............................................................... 132

Figure A.9: Histogram of High Potency TAPs PPS by Treatment ................................ 133

Figure A.10: Refined Breakdown of Episodes by Treatment Type
When Utilizing High Potency TAPs as the CT ........................................ 133

Figure A.11: Crude Treatment Effect by Risperidone Quintile ..................................... 195

Figure A.12: Crude Treatment Effect by Quetiapine Quintile ...................................... 195

Figure A.13: Crude Treatment Effect by Low Potency TAPs Quintile ......................... 196
xii
Figure A.14: Crude Treatment Effect by Medium Potency TAPs Quintile .................. 196

Figure A.15: Crude Treatment Effect by High Potency TAPs Quintile ........................ 197

Figure A.16a: Adjusted Treatment Effect Utilizing Olanzapine as the CT
Olanzapine vs Quetiapine ........................................................................ 197

Figure A.16b: Adjusted Treatment Effect Utilizing Olanzapine as the CT
Olanzapine vs Low Potency TAPs .......................................................... 198

Figure A.16c: Adjusted Treatment Effect Utilizing Olanzapine as the CT
Olanzapine vs High Potency TAPs ......................................................... 198

Figure A.17a: Adjusted Treatment Effect Utilizing Risperidone as the CT
Risperidone vs Quetiapine ...................................................................... 199

Figure A.17b: Adjusted Treatment Effect Utilizing Risperidone as the CT
Risperidone vs Low Potency TAPs ......................................................... 199

Figure A.17c: Adjusted Treatment Effect Utilizing Risperidone as the CT
Risperidone vs Medium Potency TAPs .................................................. 200

Figure A.17d: Adjusted Treatment Effect Utilizing Risperidone as the CT
Risperidone vs High Potency TAPs ........................................................ 200

Figure A.18a: Adjusted Treatment Effect Utilizing Quetiapine as the CT
Quetiapine vs Olanzapine ........................................................................ 201

Figure A.18b: Adjusted Treatment Effect Utilizing Quetiapine as the CT
Quetiapine vs Risperidone ....................................................................... 201

Figure A.18c: Adjusted Treatment Effect Utilizing Quetiapine as the CT
Quetiapine vs Low Potency TAPs ........................................................... 202

Figure A.18d: Adjusted Treatment Effect Utilizing Quetiapine as the CT
Quetiapine vs Medium Potency TAPs .................................................... 202

Figure A.18e: Adjusted Treatment Effect Utilizing Quetiapine as the CT
Quetiapine vs High Potency TAPs ......................................................... 203

Figure A.19a: Adjusted Treatment Effect Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Olanzapine ......................................................... 203

Figure A.19b: Adjusted Treatment Effect Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Quetiapine .......................................................... 204
xiii

Figure A.19c: Adjusted Treatment Effect Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Medium Potency TAPs ...................................... 204

Figure A.19d: Adjusted Treatment Effect Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs High Potency TAPs ............................................ 205

Figure A.20a: Adjusted Treatment Effect Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Olanzapine ................................................... 205

Figure A.20b: Adjusted Treatment Effect Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Risperidone .................................................. 206

Figure A.20c: Adjusted Treatment Effect Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Quetiapine ..................................................... 206

Figure A.20d: Adjusted Treatment Effect Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Low Potency TAPs ...................................... 207

Figure A.20e: Adjusted Treatment Effect Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs High Potency TAPs ...................................... 207

Figure A.21a: Adjusted Treatment Effect Utilizing High Potency TAPs as the CT
High Potency TAPs vs Olanzapine ......................................................... 208

Figure A.21b: Adjusted Treatment Effect Utilizing High Potency TAPs as the CT
High Potency TAPs vs Risperidone ......................................................... 208

Figure A.21c: Adjusted Treatment Effect Utilizing High Potency TAPs as the CT
High Potency TAPs vs Quetiapine .......................................................... 209

Figure A.21d: Adjusted Treatment Effect Utilizing High Potency TAPs as the CT
High Potency TAPs vs Low Potency TAPs ............................................. 209

Figure A.21e: Adjusted Treatment Effect Utilizing High Potency TAPs as the CT
High Potency TAPs vs Medium Potency TAPs ..................................... 210

Figure A.22a: Adjusted Treatment Effect with PS (Prior)
Utilizing Olanzapine as the CT Olanzapine vs Risperidone ................... 210

Figure A.22b: Adjusted Treatment Effect with PS (Prior)
Utilizing Olanzapine as the CT Olanzapine vs Quetiapine ..................... 211

Figure A.22c: Adjusted Treatment Effect with PS (Prior)
Utilizing Olanzapine as the CT Olanzapine vs Low Potency TAPs ........ 211
xiv
Figure A.22d: Adjusted Treatment Effect with PS (Prior)
Utilizing Olanzapine as the CT Olanzapine vs Medium Potency TAPs . 212

Figure A.22e: Adjusted Treatment Effect with PS (Prior)
Utilizing Olanzapine as the CT Olanzapine vs High Potency TAPs ....... 212

Figure A.23a: Adjusted Treatment Effect with PS (Prior)
Utilizing Risperidone as the CT Risperidone vs Olanzapine ................... 213

Figure A.23b: Adjusted Treatment Effect with PS (Prior)
Utilizing Risperidone as the CT Risperidone vs Quetiapine ................... 213

Figure A.23c: Adjusted Treatment Effect with PS (Prior)
Utilizing Risperidone as the CT Risperidone vs Low Potency TAPs ..... 214

Figure A.23d: Adjusted Treatment Effect with PS (Prior)
Utilizing Risperidone as the CT Risperidone vs Medium Potency TAP . 214

Figure A.23e: Adjusted Treatment Effect with PS (Prior)
Utilizing Risperidone as the CT Risperidone vs High Potency TAPs ..... 215

Figure A.24a: Adjusted Treatment Effect with PS (Prior)
Utilizing Quetiapine as the CT Quetiapine vs Olanzapine ...................... 215

Figure A.24b: Adjusted Treatment Effect with PS (Prior)
Utilizing Quetiapine as the CT Quetiapine vs Risperidone ..................... 216

Figure A.24c: Adjusted Treatment Effect with PS (Prior)
Utilizing Quetiapine as the CT Quetiapine vs Low Potency TAPs ......... 216

Figure A.24d: Adjusted Treatment Effect with PS (Prior)
Utilizing Quetiapine as the CT Quetiapine vs Medium Potency TAPs .. 217

Figure A.24e: Adjusted Treatment Effect with PS (Prior)
Utilizing Quetiapine as the CT Quetiapine vs High Potency TAPs ....... 217

Figure A.25a: Adjusted Treatment Effect with PS (Prior)
Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Olanzapine ......................................................... 218

Figure A.25b: Adjusted Treatment Effect with PS (Prior)
Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Risperidone ........................................................ 218

xv
Figure A.25c: Adjusted Treatment Effect with PS (Prior)
Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Quetiapine .......................................................... 219

Figure A.25d: Adjusted Treatment Effect with PS (Prior)
Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Medium Potency TAPs ...................................... 219

Figure A.25e: Adjusted Treatment Effect with PS (Prior)
Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs High Potency TAPs ............................................ 220

Figure A.26a: Adjusted Treatment Effect with PS (Prior)
Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Olanzapine ................................................... 220

Figure A.26b: Adjusted Treatment Effect with PS (Prior)
Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Risperidone .................................................. 221

Figure A.26c: Adjusted Treatment Effect with PS (Prior)
Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Quetiapine .................................................... 221

Figure A.26d: Adjusted Treatment Effect with PS (Prior)
Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Low Potency TAPs ...................................... 222

Figure A.26e: Adjusted Treatment Effect with PS (Prior)
Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs High Potency TAPs .................................... 222

Figure A.27a: Adjusted Treatment Effect with PS (Prior)
Utilizing High Potency TAPs as the CT
High Potency TAPs vs Olanzapine .......................................................... 223

Figure A.27b: Adjusted Treatment Effect with PS (Prior)
Utilizing High Potency TAPs as the CT
High Potency TAPs vs Risperidone ......................................................... 223

Figure A.27c: Adjusted Treatment Effect with PS (Prior)
Utilizing High Potency TAPs as the CT
High Potency TAPs vs Quetiapine .......................................................... 224

xvi
Figure A.27d: Adjusted Treatment Effect with PS (Prior)
Utilizing High Potency TAPs as the CT
High Potency TAPs vs Low Potency TAPs ............................................. 224

Figure A.27e: Adjusted Treatment Effect with PS (Prior)
Utilizing High Potency TAPs as the CT
High Potency TAPs vs Medium Potency TAPs ...................................... 225

Figure A.28a: Adjusted Treatment Effect with PS (Post)
Utilizing Olanzapine as the CT
Olanzapine vs Risperidone ...................................................................... 225

Figure A.28b: Adjusted Treatment Effect with PS (Post)
Utilizing Olanzapine as the CT
Olanzapine vs Quetiapine ........................................................................ 226

Figure A.28c: Adjusted Treatment Effect with PS (Post)
Utilizing Olanzapine as the CT
Olanzapine vs Low Potency TAPs .......................................................... 226

Figure A.28d: Adjusted Treatment Effect with PS (Post)
Utilizing Olanzapine as the CT
Olanzapine vs Medium Potency TAPs .................................................... 227

Figure A.28e: Adjusted Treatment Effect with PS (Post)
Utilizing Olanzapine as the CT
Olanzapine vs High Potency TAPs .......................................................... 227

Figure A.29a: Adjusted Treatment Effect with PS (Post)
Utilizing Risperidone as the CT
Risperidone vs Olanzapine ...................................................................... 228

Figure A.29b: Adjusted Treatment Effect with PS (Post)
Utilizing Risperidone as the CT
Risperidone vs Quetiapine ....................................................................... 228

Figure A29c: Adjusted Treatment Effect with PS (Post)
Utilizing Risperidone as the CT
Risperidone vs Low Potency TAPs .......................................................... 229

Figure A.29d: Adjusted Treatment Effect with PS (Post)
Utilizing Risperidone as the CT
Risperidone vs Medium Potency TAPs ................................................... 229

xvii
Figure A.29e: Adjusted Treatment Effect with PS (Post)
Utilizing Risperidone as the CT
Risperidone vs High Potency TAPs ......................................................... 230

Figure A.30a: Adjusted Treatment Effect with PS (Post)
Utilizing Quetiapine as the CT
Quetiapine vs Olanzapine ........................................................................ 230

Figure A.30b: Adjusted Treatment Effect with PS (Post)
Utilizing Quetiapine as the CT
Quetiapine vs Risperidone ....................................................................... 231

Figure A.30c: Adjusted Treatment Effect with PS (Post)
Utilizing Quetiapine as the CT
Quetiapine vs Low Potency TAPs ........................................................... 231

Figure A.30d: Adjusted Treatment Effect with PS (Post)
Utilizing Quetiapine as the CT
Quetiapine vs Medium Potency TAPs .................................................... 232

Figure A.30e: Adjusted Treatment Effect with PS (Post)
Utilizing Quetiapine as the CT
Quetiapine vs High Potency TAPs .......................................................... 232

Figure A.31a: Adjusted Treatment Effect with PS (Post)
Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Olanzapine .......................................................... 233

Figure A.31b: Adjusted Treatment Effect with PS (Post)
Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Risperidone ......................................................... 233

Figure A.31c: Adjusted Treatment Effect with PS (Post)
Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Quetiapine ........................................................... 234

Figure A.31d: Adjusted Treatment Effect with PS (Post)
Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Medium Potency TAPs ....................................... 234

Figure A.31e: Adjusted Treatment Effect with PS (Post)
Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs High Potency TAPs ............................................. 235

xviii
Figure A.32a: Adjusted Treatment Effect with PS (Post)
Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Olanzapine .................................................... 235

Figure A.32b: Adjusted Treatment Effect with PS (Post)
Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Risperidone ................................................... 236

Figure A.32c: Adjusted Treatment Effect with PS (Post)
Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Quetiapine ..................................................... 236

Figure A.32d: Adjusted Treatment Effect with PS (Post)
Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Low Potency TAPs ....................................... 237

Figure A.32e: Adjusted Treatment Effect with PS (Post)
Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs High Potency TAPs ...................................... 237

Figure A.33a: Adjusted Treatment Effect with PS (Post)
Utilizing High Potency TAPs as the CT
High Potency TAPs vs Olanzapine .......................................................... 238

Figure A.33b: Adjusted Treatment Effect with PS (Post)
Utilizing High Potency TAPs as the CT
High Potency TAPs vs Risperidone ......................................................... 238

Figure A.33c: Adjusted Treatment Effect with PS (Post)
Utilizing High Potency TAPs as the CT
High Potency TAPs vs Quetiapine .......................................................... 239

Figure A.33d: Adjusted Treatment Effect with PS (Post)
Utilizing High Potency TAPs as the CT
High Potency TAPs vs Low Potency TAPs ............................................. 239

Figure A.33e: Adjusted Treatment Effect with PS (Post)
Utilizing High Potency TAPs as the CT
High Potency TAPs vs Medium Potency TAPs ...................................... 240
xix
ABSTRACT

The foundation of this dissertation is built upon the belief that treatment effects are often
heterogeneous. Thus, different patients experience different outcomes on the same
medication. The existence of such heterogeneity gives indication that the current clinical
evidence may not be appropriate. This becomes increasingly evident when
heterogeneous treatment effects (HTE) prove to be qualitative. Thus basing clinical
decisions on averages could have implications on the patients’ well-being, the cost of
healthcare to society, and the availability of medications in the marketplace.

Hence, the Prognostic Propensity Score (PPS) method was developed with three goals in
mind; (1) To identify if HTE are present, (2) To identify if HTE are quantitative or
qualitative, and (3) To identify unique patient characteristics or tailoring variables when
qualitative HTE are present. Accomplishing these goals will provide physicians and
other decision makers with evidence that will allow them to treat patients more
efficiently.

Thus, we have created the prognostic propensity score (PPS) method. The PPS is defined
as the expected outcome (on control) given the individual’s covariates. To calculate the
PPS we regress the outcome of interest on the covariates for only those patients treated
with the control (Drug A). Using the coefficients from this model and the patient
characteristics, we then compute the PPS for each patient assuming that every patient is a
member of the control group. We identify if treatment effects vary across subgroups by
xx
partitioning patients by PPS into strata and calculating the treatment effect within each
stratum. We repeat this analysis using the alternative treatment (Drug B) as the control.
Identifying and comparing the stratum that receives the optimal benefit from each
treatment we determine which patient characteristics are uniquely associated with success
for the individual treatments.

To demonstrate the use of the PPS, we use a sample of California Medicaid beneficiaries
diagnosed with schizophrenia. Results of this study indicate that the PPS can adequately
identify HTE, sufficiently differentiate quantitative and qualitative HTE and has the
potential to identify tailoring variables. Ultimately, this method will allow physicians to
more accurately prescribe the most beneficial treatment for each and every patient.

1
CHAPTER 1: INTRODUCTION

The benefits of prescription drugs are indisputable. They have saved millions of lives
and have extended and improved the quality of millions more. In contrast, prescription
drugs are also the fourth leading cause of death in the United States after heart disease,
cancer, and stroke (Lauzarou J et al. 1998). Each year, approximately 3.1 billion
prescriptions are issued in the United States of which approximately 2.1 million result in
an adverse reaction (Starfield B, 2005). Of those approximately 2.1 million adverse
reactions, one million result in hospitalization and approximately 106,000 result in death
each year (Starfield B, 2005). Despite these risks, the use of prescription drugs in the
United States continues to rise. Between 1994 and 2005 the number of prescription drugs
purchased in the US increased by nearly 71% (Kaiser Family Foundation, 2007).
Consequently, the cost of prescription drugs to the nation is also on the rise. In 2005, the
nation spent an estimated $200.7 billion on prescription drugs, and costs are expected to
reach nearly $500 billion by 2016 (Kaiser Family Foundation, 2007). As a result, it is
becoming increasingly important to gain a thorough understanding of how to best
maximize the potential benefit of prescription drugs, while avoiding their adverse events
and managing their ever increasing cost.

Balancing the benefit, risk, and cost of prescription drugs is a difficult task because the
response to medications is not the same for all patients. A prescription drug may provide
great benefit in one patient, while having no benefit in another. Similarly, a medication
may cause side effects in one patient and no side effects in another (Popcock SJ, et al.
2
2002; Graham J, et al. 2007). Further complicating matters, evidence regarding the
effects of medications on the individual is very limited. Traditionally treatment benefits
have been expressed as an average benefit across a large population of patients (Hayward
RA, et al. 2005). This measure does not adequately reflect the benefits observed by any
one individual patient (Rothwell PM, 1995). Recognition of inter-individual differences
in drug response is an essential step towards optimizing therapy.

1.1 Ambiguous Evidence

Traditionally physicians reference the somewhat ambiguous results of drug research
studies when making prescribing decisions. This clinical evidence is often reported as an
average treatment effect across a large population of patients (Kravitz RL, et al 2004;
Hayward RA, et al. 2005). Reporting average effects is appropriate if all patients
experienced the same effect from a given treatment. However, this is often not the case.
More often, different patients experience different outcomes on the same medication. If
this is true, then averaging the treatment effects across a large population may obscure
the outcomes received by most patients. The average effect may be composed of a large
benefit for some, no benefit for others and harm for a few patients (Popcock SJ, 2002;
Kravitz RL, 2004; Hayward RA, 2005; Graham J, 2007).

Further complicating the ability of physicians to translate clinical evidence into practice
is the variety of measures that can be used to report treatment outcomes. Three
commonly used measures are relative risk reduction, absolute risk reduction, and number
3
needed to treat (Ghosh AK, et al. 2005). The physicians’ prescribing decision is often
influenced by which outcome measures are reported by a study and the context in which
they are reported.

Take for example Pravachol, a cholesterol-lowering medication. A clinical trial, studying
the use of Pravachol in otherwise healthy men with high cholesterol revealed a 31%
reduction in the relative risk of heart attacks among men receiving the drug (Shepherd J
et al. 1995). When examining the results in slightly more detail, the true indication is that
over the five-year study 5.3% of men taking Pravachol and 7.5% of men receiving usual
care had a heart attack. Thus, for any given patient the absolute risk of heart attack
actually dropped by only 2.2 percentage points. Depending upon how the study results
are reported, a relative risk reduction of 31% may be more persuasive for a physician to
prescribe Pravachol than an absolute risk reduction of 2.2 percentage points.

The results of the Pravachol study could also be reported as a number needed to treat
(NNT). That is, in order to prevent one heart attack we would need to treat 45 patients
with Pravachol for five years. The other 44 patients who are treated with Pravachol
would receive absolutely no benefit. Of those 44 patients, 2 would have a heart attack
despite taking Pravachol, and 42 would not have had a heart attack regardless of which
treatment they received. Although this information is informative, it does not provide
physicians with evidence to determine which 1 of those 45 patients Pravachol is most
likely to prevent from having a heart attack. In addition, this prevents physicians from
4
exploring other options for the 44 patients for which Pravachol would have no benefit.
Thus, treatment selection remains dependent upon the personal experience and judgment
of each physician.

1.2 Inappropriate Care

Research suggests that the interpretation of clinical evidence varies tremendously from
physician to physician (Ghosh AK, et al. 2005; Carnes M, et al. 2007). Consequently, the
same patient is likely to be prescribed different treatments depending upon which
physician they see. This is evident in the variation in treatment patterns across
geographic regions. The Michigan component of the Dartmouth Atlas (created by
Wennberg and colleagues in coordination with the Blue Cross Blue Shield of Michigan)
examined the utilization of nine classes of drugs across the state. Results of this study
indicate a two to four fold variation in the rate of prescription drug utilization across
geographically defined regions. For example, the utilization rates of stimulants for the
treatment of attention deficit hyperactivity disorder (ADHD) varied from 1.6% in low
utilization areas to 6.5% in high utilization areas (Cox ER, et al. 2003; Wennberg J, et al.
2000). This represents a substantial difference in prescribed treatment across what
should have been almost identical populations. Similarly, utilization rates of beta-
blockers varied from 42% to 71% across various regions (Wennberg J, et al. 2000).
Confirming these regional findings, similar results were found in numerous drug classes
in a national sample of patients using the Express Scripts Atlas (Motheral B, et al. 2002).
These findings support the notion that the prescription drugs a patient receives, and how
5
much of them they receive, is largely determined by where they live and where they seek
medical care – i.e. rather than by a generally accepted standard of care that is uniformly
adhered to by all physicians.

The actual variation in prescription drug utilization in itself may not be the problem.
However, unexplained differences in utilization may be an indication of inappropriate
prescribing (Schwartz JS, 1984; Wennberg J, et al. 2000). Large variations in utilization
rates may signify over-prescribing by some physicians and under-prescribing by other
physicians (Schwartz JS, 1984; Kravitz RL, et al. 2004). Over-prescribing, which leads
to over-utilization, is characterized as the use of prescription drugs in patients who
receive no benefit from their use. This prevents or delays the use of more effective
treatment. In extreme cases, over-utilization may actually harm the patient. Commonly
over-prescribed medications include tranquilizers and sleeping pills in patients who
technically could fall asleep naturally. In the extreme case of the elderly population, the
use of sleeping pills may provide no benefit and could lead to adverse outcomes such as
falls (Linjakumpu T, 2002).

In contrast to over-prescribing, much less documented and understood is the morbidity
and mortality associated with under-prescribing or under-utilization of effective
therapies. For example, numerous publications have identified patients with high blood
pressure who are not receiving anti-hypertensive medication. This lack of treatment
increases the risk of myocardial infarction and can result in unnecessary and lengthy
6
hospital stays (Egan BM, 2005). Clearly, under-utilization can deprive patients of
potentially beneficial treatments and subsequently lead to poor health outcomes.

1.3 Financial Impact on Society

Health-care costs in the US have been on the rise. Rising health care costs may be partly
attributed to inappropriate use of prescription drugs (Schwartz JS, 1984). Over-
utilization of prescription drugs results in the waste of societal resources that could have
been spent more effectively elsewhere (Schwartz JS, 1984). In extreme cases where
over-utilization actually harms the patient, not only will societal resources be wasted, but
additional resources are likely to be consumed in the treatment of adverse events
associated with their use (Schwartz JS, 1984). In addition, under-utilization may also be
associated with increased health care costs. Although the initial cost of prescription
drugs may be lower the cost to other healthcare sectors (e.g. physician visits, trips to the
emergency room, hospital care, etc.) may inadvertently be increased (Charatan F, 2002).

In response to the rising cost of health care many health insurance providers have
implemented cost containment strategies. Many of these strategies have been specifically
aimed at prescription drug utilization (Soumerai SB, 2004; Hoadley J 2005). These
strategies include formularies, prior authorization, quantity dispensing limits, and
enrollee cost-sharing. An implicit assumption of each of these methods is that patients
respond to medications equally. None of these initiatives are responsive to inter-
individual differences in treatment outcomes. Thus the preferred (or least expensive)
7
treatment is identical for all patients in a given plan. When treatment effects are
heterogeneous, it is possible that some individuals may have a poor outcome on the
preferred (or least expensive) medication in comparison to non-preferred (or expensive)
alternatives in the same class. These patients will either use the less effective preferred
treatment or be forced to pay a higher price for a more effective medication.

In order to efficiently control rising health care costs in the US we must focus on
initiatives that control costs in all sectors of the health-care system, while maintaining
quality of care. Current initiative to control the rising cost of health care may have
detrimental consequences to the patients and society when treatment effects vary from
patient to patient.

1.4 Drug Market Impact

Not only are their financial impacts to society, variation in treatment also impacts the
drugs available in the marketplace. The mentality of treating a population as opposed to
an individual has been a long standing practice in the pharmaceutical industry. Over the
years it has been seen as beneficial for drug companies to sell and market prescription
drugs to the masses rather than to a niche or select group of patients; this has been
referred to as the “blockbuster” drug mentality.

The first implication of this “blockbuster” drug mentality is the inadvertent over-
prescribing of many treatments. The reporting of average treatment effects falsely leads
8
physicians to believe that the medication is safe and effective in a large population of
patients, while it may truly only benefit a few. The downside of this practice is that there
is a greater potential for more patients to experience an adverse event. Over the past
decade the Food and Drug Administration (FDA) has been forced to recall numerous
drugs due to adverse events (Fung M, 2001). Some of the withdrawn drugs were
beneficial and potentially lifesaving for a number of patients. If the drug indication was
modified (i.e. to focus on the most opportune population) then over-prescribing by
physicians could be limited. This would prevent the unnecessary removal of many drugs
from the marketplace.

Another implication of the “blockbuster” mentality is the limited development and
distribution of drugs that are only effective in a small population. By focusing efforts on
the research and development of “blockbuster” drugs, many viable drugs may never
make it to market. For example, if the drug shows benefits for most patients in early
trials, companies are likely to continue with further tests. However, if only a small
number of patients improve, companies often cease further development. This policy
could have detrimental effects for the small number of patients who would receive great
benefit from this drug that never reaches the market. A Pulitzer prize-winning Wall Street
Journal article published in November 2004 best illustrates the negative effect of this
research and marketing strategy. The article tells the story of Elizabeth Grubesich, a
metastatic breast cancer patient who was a participant in a phase II clinical trial for the
cancer vaccine TriAb. Her treatment regimen consisted of monthly injections of TriAb
for two years. Ms. Grubesich, along with one-third of the trial participants, had
9
overwhelming success on the treatment and was able to return to a normal productive
existence. Unfortunately, for the other two-thirds of the trial participants the drug was
not effective. Due to the results of the trial in August of 2002 Titan Pharmaceuticals, the
makers of TriAb, decided to stop production of the medication. Their decision left Ms.
Grubesich and other successfully treated trial participants without access to the drug that
essentially had been keeping them alive. It also prevented an unknown number of future
patients from receiving this potentially life saving medication (Marcus AD, 2004). This
story is likely a common occurrence in many clinical trials in which the drug does not
make it to market.

1.5 Dissertation Objective

It is our contention that the geographic variation, inappropriate care, and the limited
success of cost containment strategies can be largely attributed to a lack of evidence.
Physicians need the means to identify which patients will be most likely to benefit from a
given treatment.

The purpose of this dissertation is to provide physicians with evidence regarding more
individualized outcomes, by linking patient characteristics to treatment success. This will
allow physicians to more accurately prescribe the most beneficial treatment for each and
every patient. As a result, inappropriate utilization of prescription medications and
10
subsequent healthcare cost will be reduced. In addition, the number of drugs brought to
market will increase and the number of drugs pulled from the market due to adverse
events will decrease.
11
CHAPTER 2: BACKGROUND
2.1 Homogeneous Treatment Effects

The majority of research available to decision makers today provides evidence on the
average effect of treatment across a large population. These measures assume that
treatment effects are homogeneous or constant across all patients. In figure 2.1 the
distance between
1
T and
0
T

represents the treatment effect across a covariate of
interest ) (X . Since this distance is constant across all values of X the treatment is
believed to affect each patient equally (Pagano M, 2003).

Figure 2. 1: Homogeneous Treatment Effects

X
Y
T
1

T
0

Y represents the outcome of interest. X represents a covariate of interest (defined by a
single covariate such as age or by multiple covariates combined into a single score
such as the PS or MCS scores). T
0
and T
1
represent the control and treatment,
respectively. Distance between T
1
and T
0
represent the treatment effect. (Pagano M,
2003)
12
However, in reality treatment effects are likely to vary from patient to patient. In the
econometrics literature, this phenomenon has been termed heterogeneous treatment
effects (HTE) (Heckman J, et al. 1997b; Longford NT, et al. 1999; Manski C, 2001;
Manski C, 2002; Lechner M, 2002a; Lechner M, 2002b; Manski C, 2004; Heckman J, et
al. 2005; Heckman J, et al. 2006; ).

2.2 Heterogeneous Treatment Effects (HTE)

HTE can be classified as either quantitative or qualitative.
2.2.1 Quantitative HTE

Quantitative HTE exist when there is a variation in the magnitude but not in the direction
of treatment effects among subsets of patients (Peto R, 1982; Bailey KR, 1994). These
subsets of patients can often be identified using certain patient baseline characteristics
denoted by the covariate X . The relationships (i.e. between a covariate and treatment)
that can be used to identify quantitative HTE are called quantitative or non-crossover
interactions (Azzalini A, et al. 1984; Gail M, et al. 1985; Ciminera JL, et al. 1993). A
quantitative interaction is demonstrated in Figure 2.2 (Pagano M, et al. 2003). In this
figure the outcome on treatment ) (T
1
is better than the outcome on control ) (T
0
for all
values of the covariate of interest ) (X . However, patients with a higher value for the
covariate X benefit more from receiving treatment ) (T
1
than those patients with a low
value. Thus, treatment
1
T is the optimal clinical choice for treating all patients.
13
Figure 2. 2: Quantitative Heterogeneous Treatment Effects

2.2.2 Qualitative HTE

In contrast, qualitative HTE exists when a treatment has both net beneficial and net
adverse effects (Peto R, 1982; Bailey KR, 1994). This means that treatment ) (T
1
will
have a positive effect on one subgroup of patients, and a negative effect on another
subgroup of patients. Similar to quantitative HTE these subsets of patients can also be
identified using certain patient baseline characteristics ) (X . The relationship between
these characteristics and treatment are termed qualitative or crossover interactions
(Azzalini A, et al. 1984; Gail M, et al. 1985; Ciminera JL, et al. 1993). Figure 2.3
demonstrates qualitative HTE (Pagano M, et al. 2003). In this figure patients with a low
value for covariate X would have better outcomes on the control ) (T
0
than they would
X
Y
T
1

T
0

Y represents the outcome of interest. X represents a covariate of interest (defined by a
single covariate such as age or by multiple covariates combined into a single score
such as the PS or MCS scores). T
0
and T
1
represent the control and treatment,
respectively. Distance between T
1
and T
0
represent the treatment effect. (Pagano M,
2003)
14
on the treatment ) (T
1
. In contrast, patients with a high value for covariate X would have
better outcomes on the treatment ) (T
1
than on the control ) (T
0
.

Figure 2. 3: Qualitative Heterogeneous Treatment Effects

2.2.3 Focus of this Dissertation

The focus of this dissertation is threefold. First, we wish to identify if HTE exist.
Second, we hope to determine if HTE are quantitative or qualitative. Finally, we hope to
identify patient characteristics associated with qualitative HTE. Qualitative HTE as
opposed to quantitative HTE will affect the clinical treatment selection decision. For
example, if qualitative HTE is present then the treatment selected by the physician will
vary based on the patient's characteristics. In contrast, if quantitative HTE is present a
clinician will still prescribe the same treatment to everyone. Therefore, little attention
X
Y
T
1

T
0

Y represents the outcome of interest. X represents a covariate of interest (defined by a
single covariate such as age or by multiple covariates combined into a single score
such as the PS or MCS scores). T
0
and T
1
represent the control and treatment,
respectively. Distance between T
1
and T
0
represent the treatment effect. (Pagano M,
2003)
15
needs to be paid to quantitative HTE and our focus should be on identifying patient
characteristics associated with qualitative HTE. These characteristics that we wish to
identify (i.e. which will indicate which patients would do best on a given treatment) have
been referred to as tailoring variables or prescriptive indices in the literature (Collins LM,
et al. 2004; Hollon SD, et al. 2005).

2.3 Sources of Data for Determining HTE

In order to accomplish our goals we must first select an appropriate source of data for our
analysis. The two sources of data available for determining HTE are randomized
controlled trials (RCTs) and observational studies. Each of these sources has their own
advantages and disadvantages, discussed in the following sections, which will make them
more or less attractive for HTE research.
2.3.1 Randomized Controlled Trials (RCTs)

The first source of data we will evaluate for our use is data derived from a randomized
clinical trial (RCT). Many experts call the RCT the gold standard of research and place it
at or near the top of the hierarchy of evidence. The hallmark of a RCT is the random
assignment of participants to treatment. Randomization will guarantee us that the
treatment groups are similar in their baseline characteristics, thus reducing the likelihood
of bias associated with treatment selection. In addition, a RCT is often double-blinded
(i.e. both the physician and the patient who participate in the study are unaware of which
treatment they will receive). A RCT will therefore allow us to produce more objective
16
results, since the expectations of the physician and the patient about the treatment will not
affect our outcome. Finally, the inclusion of a control group will allow us to determine
the causal effect of the treatment.

Despite its many favorable attributes, using data derived from a RCT has limitations that
make it less favorable for use in our analysis. Most notably is its poor external validity.
Patients and physicians who enroll in a RCT are often a homogeneous population seldom
representative of patients in the true clinical practice (Schmoor C, et al 1996; Longford
NT, et al. 1999). Strict inclusion and exclusion criteria explicitly restrict enrollment to
patients with limited disease, comorbidities, and concomitant medications. Additionally,
implicit selection criteria including volunteerism likely mean that RCT patients’ are more
compliant, more health-conscious, and more likely to have a favorable outcome on
treatment than non-participants. Finally, patients enrolled in a RCT often receive better
care than patients in the usual clinical practice setting. Due to the nature of RCTs,
patients enrolled in them often have more physician visits throughout the course of their
treatment and receive more intensive follow-up care. Consequently, results of analyses
using data from RCTs provide a measure of efficacy or evidence which can only be
achieved in the optimal environment, and do not relate to the typical clinical practice
(McKee M, et al. 1999; Rothwell PM, 2005, Steg PG, et al. 2007).
2.3.2 Observational Studies

In contrast, the limitations of using data from a RCT are the advantages of using data
from an observational study. First and foremost, data derived from an observational
17
study has good external validity. The patients and physicians represented in the data are
heterogeneous populations which better represent the patients and physicians in the true
clinical practice setting. Additionally, patients in observational studies have an unlimited
number of comorbidities and take an unlimited number of concomitant medications.
Thus data from observational studies are beneficial in studying the effect of drugs in
populations often excluded from RCTs. Finally, patients in observational studies receive
care typical of the care actually received by patients in the true clinical practice setting.
Results of analyses using data from observational studies provide a measure of
effectiveness or evidence relative to what actually occurs in real practice (Black N,
1996).

Unfortunately these benefits are not obtained without a cost. Observational data is not
collected for research purposes and therefore has some inherent problems. These
problems often arise because patients are not randomly assigned to treatment. Instead,
treatment is generally assigned to patients with the worst prognosis or in whom the
physician believes the treatment will have the greatest benefit. These patients are likely
to be different from those patients who are not treated or who are treated with an
alternative treatment. Thus, the selection of treatment is confounded with patient factors
which are also related to the outcome. When confounding effects are present in non-
randomized observational studies crude unadjusted estimates of the treatment effect will
be biased. These confounding factors can cause either an over estimation or an under
estimation of the treatment effect. This is commonly referred to as “treatment selection
bias (Heckman J, 1979a).”
18
2.3.3 Source of Data for this Dissertation

Given the advantages and disadvantages of data from both a RCT and observational
study, we can conclude that data derived from an observational study is optimal for our
analysis. The objective of our analysis is to provide physicians with information
necessary to choose the most appropriate treatment for their individual patients. Thus, it
is important for us to use data on patients similar to the patients seen in their clinical
practice. This can best be achieved by using data from observational studies. In addition,
we wish to determine if a treatment produces different outcomes in different patients.
Given that a RCT selects patients most likely to benefit from the treatment, these
differences would be difficult to determine using this type of data. Observational studies,
which contain data on larger and more diverse populations, will be more capable of
identifying these relationships between patient and treatment outcome. Finally, we wish
to identify patient characteristics, which may indicate on which treatment a patient may
do best. Data from observational studies which contain information on a large
heterogeneous population of patients is the best source of data for us to accomplish this.
Data from observational studies contain information on more diverse patient
characteristics such as comorbidities and concomitant medications, which may ultimately
be identified as tailoring variables.

Although observational studies are prone to bias (i.e. associated with non-randomization)
numerous econometric and statistical methods have been developed to control or adjust
for this bias (Rosenbaum PR, et al. 1983; Ashenfelter O, et al. 1985; Manski C, 1990;
19
Imbens G, et al. 1994; Rosenbaum PR, 1995; Angrist J, et al. 1996; VanderKlaauw 2002;
Abadie A, 2005; Athey S, 2006;). Implementing these methods into our analysis will
allow us to overcome the disadvantage of using this type of data.

2.4 Methods

Now that we have determined that observational studies are appropriate for our analysis
we must now establish the most suitable method to use for our study. In the following
sections we will discuss two commonly used methods to identify treatment covariate
interactions. For each method we will discuss the advantages and disadvantages of using
them to identify HTE and tailoring variables. Once we have identified the most
appropriate method for our analysis, we will describe how this method could be
combined with current techniques to control for any bias associated with the use of
nonrandomized data.
2.4.1 Interaction

In order to identify treatment covariate interactions the most commonly used technique is
to utilize the treatment effect model and include an interaction term between the
treatment and covariate (Pagano M, et al. 2003). This can be expressed in the model:

ε β β β β + + + + = TX X T Y
3 2 1 0
(eq. 2.1)

20
Where Y is the outcome of interest, T is a treatment dummy variable, X is a vector of
covariates, andTX is an interaction between the treatment and covariate. If this is a linear
model, the coefficient ) (
3
β on the treatment covariate interaction term is easy to
interpret. This coefficient gives the magnitude and the sign of the interaction effect, and
we can test the statistical significance of this coefficient using Student’s t-test. This
method can also be used in nonlinear models, however, we would have to implement the
formulas developed by Ai and Norton (2003) to comprehend the magnitude and statistical
significance of the treatment covariate interactions.

An interaction model is an easy method to identify treatment covariate interactions.
However, this method does have some limitations. First, the number of interaction terms
included in the model may be limited by the number of observations in the data.
Therefore, researchers are often required to use their a priori knowledge about treatment
covariate relationships in order to determine which interaction terms to include in the
model. Any interaction terms not included in the model cannot be identified as potential
treatment-covariate interactions.

Another important limitation of this method is that it can only identify if an interaction
exists. This method will not allow us to distinguish between quantitative and qualitative
interactions (Pagano M, et al. 2003). Take for example, the coefficient for an interaction
term between a treatment and a covariate of interest ) (X which is negative and
statistically significant. This coefficient tells us that as the value of the covariate
21
) (X increases the treatment effect decreases significantly. However, we will not be able
to distinguish which graph, seen in figure 2.4, best represents this relationship.

Figure 2. 4: Treatment Covariate Interaction

Since we cannot identify which interactions are qualitative using an interaction model we
will not be able to identify tailoring variables. Thus, for the purposes of our analysis this
will not be an appropriate method.
2.4.2 Stratification

An alternative technique used to identify treatment covariate interactions is the
stratification method (Pagano M, et al. 2003). This method requires us to stratify patients
into subgroups based upon their values of important covariates ) (X . Once the subgroups
X
Y
T
1

T
0

Y represents the outcome of interest. X represents a covariate of interest (defined by a
single covariate such as age or by multiple covariates combined into a single score
such as the PS or MCS scores). T
0
and T
1
represent the control and treatment,
respectively. Distance between T
1
and T
0
represent the treatment effect.
(a) Quantitative HTE
X
Y
T
1

T
0

(b) Qualitative HTE
22
have been identified we would then compute the average treatment effect within each
subgroup. This can be expressed in the model:

ε β β β + + + = X T Y
2 1 0
(eq. 2.2)

Where Y is the outcome of interest, T is a treatment dummy variable, and X is a vector
of covariates. Next, we would use a statistical test, such as the Gail-Simon test, to test if
the directions of the treatment effects vary significantly across the subgroups of patients
(Gail M, et al. 1985). We could then conclude if qualitative heterogeneous treatment
effects are present. In contrast, if the directions of the treatment effects do not vary
significantly across subgroups then we would conclude that qualitative HTE does not
exist.

Stratification methods are simple and easy to use and comprehend. Most importantly
they allow us to distinguish quantitative interactions from qualitative interactions. In
addition, distinguishing which variables cause a significant change in the direction of the
treatment effect across subgroups should allow us to identify tailoring variables.
2.4.3 Most Appropriate Method to Use for this Dissertation

For the purposes of our analysis, stratifying patients is the most appropriate method to
use. However, there are multiple methods that can be used to stratify patients which will
also allow us to control for treatment selection bias. Each of these methods creates
different subsets of patients. As a result, when treatment effects are heterogeneous, these
23
approaches yield estimates for dissimilar patient groups. Therefore, it is important for us
to understand how the subgroups were generated in order to comprehend how these
methods could or could not be used for the purpose of our analysis.

2.5 Techniques to Stratify Patients and to Control for Treatment Selection Bias
2.5.1 Instrumental Variable (IV)

The most commonly used technique implemented by econometricians to control for
treatment selection bias is the instrumental variable (IV) method (Imbens G, et al. 1994;
Angrist J, et al. 1996). An IV is defined as a variable that is highly correlated with the
treatment allocation but is unrelated to the outcome. The model assumes that the
selection of treatment
i
T is generated through a latent variable
*
i
T :

i i i
v Z T + + =
1 0
*
α α (eq. 2.3)

Where ) 0 ( 1
*
≥ =
i i
T T (eq. 2.4)

i i i
T Y ε β β + + =
1 0
(eq. 2.5)

24
Where
i
Y is the outcome of interest,
i
T is and indicator of treatment, and
i
Z is the IV. In
the analysis, to control for confounding, the actual treatment status that may be
confounded is substituted by the unconfounded IV. This can be seen by substituting
equation (2.3) into equation (2.5):

) ( ) (
1 1 1 0 1 0 i i i i
v Z Y ε β α β α β β + + + + = (eq. 2.6)

An unbiased estimate of the effect of treatment can then be obtained after adjusting for
the association between the IV,
i
Z and the treatment
i
T .

The IV method can be seen as stratifying patients based upon their likelihood of receiving
treatment ) 1 ( =
i
t . This method assumes that patients can be divided into 3 groups; in the
first group all physicians would treat these patients, in the 2
nd
group only some physicians
would treat these patients, and in the 3
rd
group no physicians would treat these patients.
If treatment effects are heterogeneous then the IV estimate provides measure of the
treatment effect for only those patients in group 2, who are often referred to as “marginal
patients”. The treatment status for these marginal patients will depend on the value of the
IV used in the analysis (Heckman J, 1997).

An inherent assumption implied by using the IV method is that physicians prescribe in a
uniform manner. That is each physician agrees upon which subgroup (i.e. 1
st
, 2
nd
or 3
rd
) a
patient would fall in. Given the evidence in Chapter 1 that a large variation in treatment
25
exists across geographic regions, this is highly unlikely. Therefore, stratifying patients
using an IV approach may not be an appropriate method for our analysis. In addition, it
is often not possible to identify the marginal subgroups of patients whose average
treatment effect is measured by the IV (Harris KM, 1998). Thus, in order to interpret the
IV estimate we would have to rely on our assumptions about the clinical practice to
identify which patients (i.e. based upon patient characteristics) are the marginal patients.
Given that our analysis relies heavily upon being able to identify patients, according to
characteristics, in a given subgroup this method may not be appropriate for our purposes.
2.5.2 Stratifying by Potential Confounders

Since the IV method is not suitable for our analysis, we must look for other methods that
can be used to stratify patients while also controlling for treatment selection bias. The
simplest method is to stratify patients based upon covariates. This is a manual
subgrouping of data into different strata according to common patient characteristics. As
opposed to the IV method, this gives us full knowledge of the characteristics of patients
within each subgroup. In addition, this method is simple to implement if we only
consider one covariate such as gender. However, as the number of covariates increases
the number of strata increases exponentially – i.e. since each stratum must be mutually
exclusive. Therefore, this method is difficult to execute when we want to consider a
large number of covariates (Rosenbaum PR, et al. 1983).
26
2.5.3 Propensity Score (PS)

To overcome the challenges of using multiple covariates to stratify patients, we could use
the propensity score (PS) method developed by Rosenbaum and Rubin (1983). Instead of
considering each covariate separately the PS combines all covariate into a single value,
defined as a conditional probability of being treated given the individuals covariates.

) | (
i i i
x X T E PS = = (eq. 2.7)

Where ) 1 , 0 ( =
i i
t T is an indicator of treatment, and
i
X is a vector of observable
covariates. Typically the PS function is estimated using a multivariate logistic regression
model for the entire study population. The estimated PS values range from zero to one;
with values close to zero indicating a low probability of receiving treatment and values
close to one indicating a high probability of receiving treatment. The distribution of
confounders for patients in the treatment and control groups with the same PS should be
similar. It can then be assumed that the mechanism for treatment assignment is now
independent of the outcome, and the difference across treatment groups at any value of a
PS is an unbiased estimate of the average treatment effect (Rosenbaum PR, et al. 1983;
Rosenbaum PR, et al. 1984).

Similar to the IV method, the propensity score (PS) method can be used to stratify
patients based upon their likelihood of receiving treatment. Stratifying is based on the
27
idea of dividing the patients by their propensity score into subgroups. Within each
subgroup patients receiving treatment ) 1 ( =
i
t and patients receiving control ) 0 ( =
i
t
would have on average the same propensity score. Consequently, within each stratum
patients receiving the treatment ) 1 ( =
i
t and patients receiving the control ) 0 ( =
i
t would
have similar baseline characteristics. Once we have stratified the patients, we can then
compute the difference between the average outcomes of the treatment and the control
groups within each stratum. This will allow us to obtain a less biased estimate of the
average treatment effect within each subgroup (Rosenbaum PR, et al. 1984).

A big benefit of the PS method over the IV method is that it allows us to identify the
characteristics of patients within each subgroup. However, similar to the IV method the
PS method depends upon the inherent assumption that all physicians would treat each
patient in the same way. As we have already established this is an assumption that is
likely untrue. Therefore, the PS method may not be the most fitting method for our
analysis.
2.5.4 Multivariate Confounder Score (MCS)

An alternative to the PS method is the multivariate confounder score (MCS) method
(Miettinen O, 1976). The MCS, also referred to in some literature as a disease risk score
(DRS), is similar to the PS in that it allows us to combine multiple covariates into a single
value. However, the MCS is defined as the expected outcome (on control) given the
individuals covariates. This value can be interpreted as the baseline risk of the outcome.
28
Unlike the PS and IV methods, the MCS is not dependent upon the prescribing practices
of physicians.

To calculate the MCS we would begin by regressing the outcome of interest on a
treatment dummy variable and all other covariates of interest. This is often calculated
using a logistic regression of the model:

i i i i
X T Y ε β β β + + + =
2 1 0
∀ i (eq. 2.8)

Where
i
Y is the outcome of interest for patient i (i.e. 1 =
i
Y means patient had
“success”),
i
T is a treatment dummy variable and
i
X is a vector of potential confounders.
Using the coefficients from this model and the characteristics of patient i , we can
compute the MCS for each patient i assuming that every patient is a member of the
control group ) 0 ( =
i
t . Thus:

i i i
X T MCS
2 1 0
ˆ ˆ ˆ
β β β + + = (eq. 2.9)

Assuming 0 =
i
t for every patient the
i
MCS is equivalent to:

i i
X MCS
2 0
ˆ ˆ
β β + = (eq. 2.10)

29
The estimated MCS values range from zero to one, with values close to zero indicate a
low probability of “success” on control and values close to one indicate a high probability
of “success” on control (Miettinen O, 1976).

Similar to the PS method we can stratify patients based upon their MCS. Within each
subgroup patients receiving treatment ) 1 ( =
i
t and patients receiving control ) 0 ( =
i
t will
on average have the same MCS and similar baseline characteristics. Once we have
stratified the patients, we can then compute an unbiased estimate of the average treatment
effect within each subgroup (Miettinen O, 1976).

There are several advantages of the MCS over the previously discussed methods. First,
similar to the PS this method allows us to combine multiple confounders into one
convenient value. Most importantly, as opposed to the PS and IV, the MCS does not
depend upon the prescribing behavior of physicians. Despite its advantages, the MCS
does have a major limitation. In previous studies, the MCS was found to over exaggerate
the significance of the treatment when compared to the propensity score. The
overestimate of the significance was more pronounced when the treatment and
confounders were highly correlated (Pike MC et al. 1979; Cook FE 1988). Given this
limitation the MCS has seen limited use in the current econometric literature.
2.5.5 Most appropriate Technique for Our Analysis

All of the techniques described above will allow us to stratify patients and reduce
possible treatment selection bias. However, they all have certain characteristics that
30
make them unfavorable for use in our analysis. Our approach will be to combine the
favorable attributes of the PS and MCS techniques in order to create a method suitable
for our purposes.

Specifically, this new method, which we will call the prognostic propensity score (PPS),
will be defined as the expected outcome (on a selected treatment) given the individual’s
covariates. Like the MCS and PS, the PPS will allow us to combine multiple
confounders into a single value. In addition, it will allow us to avoid making any
assumptions about the physicians prescribing behavior. In contrast to the MCS we will
compute the PPS using only those patients receiving the selected treatment. By using
only those patients in this treatment group to calculate the PPS, we believe we can
prevent the over exaggeration of significance associated with the current MCS method.

In Chapter 3, we will provide a more in-depth description of the PPS. This will be
followed, in Chapter 4, by its application to a convenient sample of patients diagnosed
with schizophrenia.
31
CHAPTER 3: METHOD DEFINITION

In the following sections we will provide an in-depth description of the Prognostic
Propensity Score (PPS). This method combines the best attributes of the MCS and
stratification methods to identify heterogeneous treatment effects (HTE) and tailoring
variables. This is accomplished by evaluating treatments over smaller, more
homogeneous sub-groups.

3.1 Challenges in Evaluating Treatment Effects

In order to understand the significance of the PPS, it is important to have an
understanding of the problems associated with evaluating treatment effects. We will
begin by explaining the simplest case of two treatments (0 and 1). Treatment received by
each patient ) ,..., 2 , 1 ( N i i = will be denoted by an indicator variable
i
T which will take the
value 0 if the unit receives the treatment 0, and 1 otherwise. That is, ) 1 , 0 ( = t T is a finite
set of mutually exclusive treatments. Each patient i can be described by a vector of
observable pre-treatment covariates
i
X . For each patient i there are two possible
responses;
i
Y
0
the response that would have resulted if treatment 0 was received, and the
response
i
Y
1
that would have resulted if the treatment 1 was received. Thus, the
treatment effect, or potential gain, of treatment 1 for patient i is ) (
0 1 i i i
Y Y − = Δ . Since
each patient can only receive one treatment at a given time, either
i
Y
0
or
i
Y
1
will be
observed for patient i , but never both. Thus
i
Δ cannot be directly observed for any
32
individual patient at any given point in time. This has been referred to as the
“fundamental problem of causal inference (Holland PW, 1986).”

Although initially seen as impossible, there are certain techniques and assumptions that
can be implemented to allow the “fundamental problem of causal inference” to be
overcome. The most commonly used technique is the counterfactual or potential
outcome model. In this model the treatment that patient i does not receive is called the
“counterfactual” treatment. The outcome on the counterfactual treatment can be obtained
if the treatments are "exchangeable." In order to be exchangeable we assume that
patients with similar characteristics would respond in a similar manner if given treatment
0, and likewise similar patients would respond in a similar manner if given the treatment
1. Thus, counterfactual or potential outcomes can be obtained for those patients taking
treatment 0 by referencing similar patients receiving treatment 1 (and visa-versa).
Through use of counterfactuals we can therefore calculate an average treatment effect for
a population or subgroup (Neyman J, 1923; Fisher RA, 1925; Rubin DB, 1973a; Rubin
DB, 1973b; Rubin DB, 1974; Rubin DB, 1977; Rubin DB, 1978).

In order to use the counterfactual framework we must assume that the response of patient
i to treatment t does not depend on the treatment given to patient j (where j i≠ ). This,
stable unit treatment value assumption (SUTVA), can be more clearly restated as: for
each unit i and treatment t there is a unique value
ti
Y . When accounting for this
assumption, the counterfactual framework successfully confirms our assertion that similar
33
patients will have a similar outcome on a given treatment, and will therefore serve as the
basis for our analyses in this dissertation (Rubin DB, 1978).

3.2 Definition of Prognostic Propensity Score (PPS)

The term prognostic comes from the word prognosis which is defined as the prediction of
an event before its possible occurrence (Merriam-Webster 2008). Typically prognostic
models are calculated as the predicted outcome given that the patient does not receive
treatment (control group). For our purposes we will refer to the control group as a
comparison treatment group (CT). The CT will be the treatment for which the researcher
wishes to identify heterogeneous treatment effects and possible tailoring variables. The
CT will be selected by the researcher at the onset of analysis and the prognosis will be
estimated with respect to the CT. Thus, the PPS will be defined as the conditional
probability of an outcome, on CT, given a vector of observable covariates.

) | (
0
x X Y E PPS = = (eq. 3.1)

Where X is a vector of observable covariates and
0
Y is a measure of treatment outcomes
on CT ) 0 ( = t .

34
3.3 Controlling for Confounding

Similar to the PS, the PPS will allow us to combine multiple covariates into a single score
(Rosenbaum PR, 1983; Rosenbaum PR, 1984). However, the PPS does so by controlling
for the association between the covariates and the potential outcome as opposed to
treatment assignment in the standard PS.
The basis for the PPS strategy is rooted in conditional probability theory. If the outcome
on CT ) 0 ( = t is conditionally independent given X , then the outcome will also be
conditionally independent given the PPS. For our analyses the PPS will be represented
by )] ( [ x e . It can therefore be shown that:
) ( |
0
x e y x⊥ (eq. 3.4)

Since ) ( | |
0 0
x e y x x y x ⊥ ⇒ ⊥ (eq. 3.5)
Show that: )] ( | 1 [ )] ( , | 1 [
0 0
x e y P x e x y P = = = or (eq. 3.6)
)] ( | [ )] ( , | [
0 0
x e y E x e x y E =
By Law of Iterated Expectations:

) (
)] ( | ) ( [
)] ( | ] | [ [
)] ( | ] ), ( | [ [
)] ( | [ )) ( | 1 Pr(
0
0
0 0
x e
x e x e E
x e x y E E
x e x x e y E E
x e y E x e y
=
=
=
=
= =
(eq. 3.7)
Hence: )) ( | 1 Pr( )) ( , | 1 Pr(
0 0
x e y x e x y = = = or (eq. 3.8)
)] ( | [ )] ( , | [
0 0
x e y E x e x y E =
35
So That: ) ( |
0
x e y x⊥ (eq. 3.9)

Therefore, the PPS removes any association between the observable baseline covariates
X and the potential outcome on CT. Thus, individuals with the same PPS will have the
same outcome on CT and baseline covariates.

If we assume strong ignorability of treatment assignment X T Y Y | ) , (
1 0
⊥ , such that the
weak ignorability of treatment assignment is also true, then X T Y |
0
⊥ (Rosenbaum PR,
1983; Imbens G, 2000). If we know that ) ( |
0
x e Y X ⊥ , then stratification by
) (X e allows treatment T to be unconfounded with the outcome
0
y . Such that:

) ( | |
0 0
X e T Y x T Y ⊥ ⇒ ⊥ (eq. 3.10)
Proven Accordingly:
If: )) ( ) ( , | Pr(
0
x e X e Y t T = = (eq. 3.11)

Then by law of iterated expectations:

) (
)] ( , | ) ( [
)] ( , | ] | [ [
)] ( , | ] , | [ [
)] ( , | ] ), ( , | [ [
)] ( , | [ )) ( , | 0 Pr(
0
0
0 0
0 0
0 0
x e
x e y x e E
x e y x T E E
x e y x y T E E
x e y x x e y T E E
x e y T E x e y T
=
=
=
=
=
= =
(eq. 3.12)
If the treatment assignment is strongly ignorable X T Y Y | ) , (
1 0
⊥ and
1 )) ( | 1 Pr( 0 < = < x e T , then the expected difference in observed responses to the two
36
treatments at ) (x e is equal to the average treatment effect at ) (x e (Rosenbaum PR, 1983;
Rosenbaum PR, 1985). Such that:

If: ) ( | ) , ( | ) , (
1 0 1 0
X e T Y Y x T Y Y ⊥ ⇒ ⊥ (eq. 3.13)

And: 1 )) ( | 1 Pr( 0 < = < x e T (eq. 3.14)

Then: )}] ( , 0 | { )} ( , 1 | { [ ] [
0 1
x e t Y E x e t Y E E Y Y E = − = = − (eq. 3.15)

Therefore, using the PPS we can control for HTE and possibly reduce the bias that may
occur in observational studies while calculating the average treatment effect.

3.4 Calculation of Prognostic Propensity Score (PPS)

The PPS will be estimated by regressing the outcome of interest on all other covariates of
interest, for only those patients treated with the CT. This can be calculated using a
logistic regression of the model:

i i i
X Y ε β β + + =
1 0 0
(eq. 3.2)

Where
i
Y
0
is the outcome of interest for patient i on the CT (i.e. 1
0
=
i
Y means patient
had “success” on the CT) and
i
X is a vector of potential covariates. Using the
37
coefficients from this model and the characteristics of patient i , we can then compute the
PPS for each patient i assuming that every patient is a member of the CT group ) 0 ( =
i
t .
Thus:

i i
X PPS
1 0
ˆ ˆ
β β + = (eq. 3.3)

The PPS is similar to the MCS in several ways. First, they both allow us to combine
multiple confounders into one convenient value. Second, neither method depends upon
the prescribing behavior of physicians.

However, there are some advantages of the PPS over the MCS method. These
advantages can be attributed to creating the PPS score using only those patients in the CT
group as opposed to both treatment groups. First, the PPS model allows us to avoid
making inferences about the effect of treatment. This is essential since stratifying
patients by the treatment effect could introduce bias (Rosenbaum PR, et al. 1984).
Second, by calculating the PPS using only the CT group, we do not force the effect of
covariates to be constant across treatments (i.e. even in the absence of interaction terms).
This is beneficial since forcing the effect of covariates to be constant across treatments
could cause an overestimation of the treatment effect, similar to the MCS (Pike MC, et al.
1979; Cook FE, 1988).
38
3.4.1 Overlap (or Support) Condition

In order to utilize the PPS method to measure treatment effects a few conditions must be
met. The first testable condition that must be satisfied is the overlap (or support)
condition. This condition requires that for every patient who has success on the CT
) 1 (
0
=
i
Y there is a similar patient with approximately the same PPS value (and prognostic
factors) who does not have success on the CT ) 0 (
0
=
i
Y and visa-versa. If a patient does
not have a similar patient in the opposing group then these patients must be dropped from
our analysis. Several methods to verify if the overlap condition holds have been
suggested in the literature; each of which can be applied to the PPS model to verify the
overlap condition. The simplest, and our proposed method, is to perform a visual
analysis of the density distribution of the PPS, via a box-plot or a histogram (Rosenbaum
PR, et al. 1983). This visual analysis will allow us to identify a region of common
support which represents a significant overlap between the PPS distribution of patients
who have success on the CT versus those who do not.

In addition, since we wish to compare the effect of the CT to alternative treatments we
must also ensure that there is sufficient overlap between those patients receiving the CT
and each of the alternative treatments. This will ensure that we are comparing similar
patients in each treatment group. Again this can be accomplished by visually analyzing
the box-plot or histogram of the PPS distributions across the CT and the alternative
39
treatments (Rosenbaum PR, et al. 1983). This allows us the opportunity to re-evaluate
our region of common support to ensure that significant overlap occurs across all
treatments.
3.4.2 Balancing Condition

The PPS model is assumed to generate stratum in which the distribution of important
prognostic factors are approximately the same for both the patients who have success on
the CT (Y
0i
=1) and those who do not have success on the CT (Y
0i
=0). This is a testable
assumption. Tests for significant differences between outcome groups within each
stratum can be conducted using t-test and chi-square test for continuous and categorical
variables, respectively. If important differences are found between patients who have and
do not have success on the CT within each stratum, then the PPS prediction model would
need to be reformulated. For example, if we found that the covariates differed
significantly between patients who have and do not have success on the CT, then the
square of that covariate could be included in the PPS model. The process of calculating
the PPS and checking that the balancing condition holds would continue until few if any
significant differences can be found between the outcomes groups within each stratum
(Rosenbaum PR, 1984). If, after multiple attempts, the balancing condition still does not
hold then we may have to conclude that the distribution of covariates does not overlap
sufficiently to allow for appropriate stratification (Rosenbaum PR, et al. 1984).
40
3.5 Identifying Heterogeneous Treatment Effects

The first purpose of this dissertation is to determine if a treatment produces effects that
vary across patients. To accomplish this we will use the PPS in combination with a
stratification method. The idea of stratification is to partition patients by PPS into strata
and to calculate the treatment effect within each stratum. This method is also known as
interval matching, blocking or sub-classification.

The technique used to determine strata is straightforward. To begin we would rank
patients based upon their PPS score from lowest to highest. Patients with a low PPS will
have a low probability of success on CT, and patients with a high PPS will have a high
probability of success on CT. Once we have ranked the patients, we will stratify patients
into quintiles. Quintiles are values that divide a sample of data into five groups
containing (as accurately as possible) equal numbers of observations. We choose to use
quintiles since they should suffice to remove over 90% of the bias associated with the
included covariates (Cochran WG, 1968). Quintile boundaries can be based on the
values of the PPS for the CT group, for the treated group, or for both groups combined.
For the purpose of this dissertation we will base our quintile boundaries on just the CT
group. This will ensure an adequate number of CT patients in each stratum.

The justification for using stratification is based on the fact that within each subgroup
patients receiving treatment and patients receiving CT should have on average the same
PPS and similar baseline characteristics. Therefore, the estimate of the ATE within each
41
stratum should reduce the heterogeneity and potential bias. The treatment effect is
calculated by taking the mean difference in the outcome between the treated and CT
patients. Such that:

)] 0 | ( ) 1 | [( )] , 0 | ( ) , 1 | [( ) 0 | ( ) 1 | (
0 0 1 1 0 1
= − = + = = − = = = = − = t Y t Y x X t Y x X t Y t Y t Y
i i i i i i
(eq. 3.16)

In which the first term on the right side of the equation
)] , 0 | ( ) , 1 | [(
1 1
x X t Y x X t Y
i i
= = − = = is the average treatment effect in the strata. The
second term )] 0 | ( ) 1 | [(
0 0
= − = t Y t Y
i i
represents the differences between the
counterfactual outcome on CT for patients receiving the treatment and the outcome on
CT for patients actually receiving the CT (i.e. within the strata). This typically represents
a potential source of bias when estimating the ATE. However, if the subgroups are
defined appropriately then the average potential effect on CT within each stratum should
be the same for both treatment groups. Therefore this term should equal zero and the
estimate of the ATE within each strata should be unbiased.

Within each stratum the average treatment effect is equivalent to the coefficient on the
treatment covariate T in the model:

X T Y
2 1 0
β β β + + = (eq. 3.17)

42
Where Y is the outcome of interest, T is a treatment dummy variable, and X is a vector
of observable covariates. If the treatment effects are heterogeneous then the magnitude
of the average treatment effect (i.e.
1
β ) will vary across strata. This can best be observed
by plotting the average treatment effect calculated within each quintile in a forest plot. If
HTE is present then we will not be able to draw a vertical line perpendicular to the x-axis
that crosses the estimate of the average treatment effect within each quintile. More
importantly, if qualitative heterogeneity is present then both the magnitude and direction
of the ATE will vary across the strata. Therefore, the ATE would be positive in some
subgroups and negative in other subgroups. Observing these results in a forest plot we
will be able to determine if the average treatment effect in at least one quintile is
significantly greater than 1, and in another quintile significantly less than 1. In order to
confirm qualitative heterogeneity, we will test if the direction of treatment effects varies
significantly across the strata using the Gail-Simon test for qualitative interaction. (Gail
M, et al. 1985) This is a likelihood ratio test that validates the null hypothesis that
treatment effects in all subgroups are in the same direction, either positive or negative. If
the null hypothesis is rejected, then we can conclude that qualitative HTE exists. If
qualitative HTE exists then the subgroup in which the CT would have the greatest benefit
would be the subgroup in which the treatment effect has the lowest negative value.

Once we have completed this analysis using treatment ) 0 ( = t as the CT, we will repeat
the methods discussed above using the alternative treatment ) 1 ( = t as the CT.
This is in preparation for the following steps where we will attempt to distinguish
tailoring variables from the prognostic factors which may be similar across treatments.
43
3.6 Identifying Tailoring Variables

The final objective of our analysis is to identify tailoring variables. These are patient
characteristics associated with qualitative HTE. These characteristics give indication as
to which patients will do best on a given treatment (Collins LM, 2004; Hollon SD, 2005).
For example, assume that a given risk factor is a tailoring variable; then patients with this
given risk factor will have better outcomes on treatment 0, and patients without this risk
factor will have better outcomes on treatment 1. Hence, identifying these tailoring
variables would allow physicians to provide their individual patients with the most
effective treatment.

If the Gail-Simon test concludes that qualitative HTE is not present then it will not be
necessary or possible to identify tailoring variables. If qualitative HTE does not exist than
the optimal clinical treatment would be the same for all subgroups. For example, if the
ATE is positive in ever subgroup than the treatment designated by treatment ) 1 ( = t in
our analysis will be optimal for all subgroups. Conversely, if the ATE is negative in
every subgroup than the treatment designated as the CT ) 0 ( = t , will be the optimal
clinical treatment for all subgroups.

If the Gail-Simon test concludes that qualitative HTE is present then we can proceed with
trying to identify tailoring variables. To identify these variables we would start with the
subgroups of patients, identified in the first step, in which the CT had the greatest benefit.
Since this analysis was repeated twice, using treatment 0 as the CT the first time and
44
treatment 1 as the CT the second time, we should have identified two subgroups (one for
each treatment) in which the patients received the greatest benefit. We will now compare
the characteristics of the patients across these two subgroups using the t-test and chi-
square test for continuous and categorical variables; respectively. The characteristics
which are not significantly different across the two treatment subgroups can be
considered prognostic factors. These variables would increase the probability of success
regardless of which treatment is received. More importantly, variables which are
significantly different across the subgroups can be considered tailoring variables. These
variables give true indication of a patient characteristic specific to treatment success. For
example, if the frequency of males is significantly higher for patients in the subgroup
receiving treatment 0 than in those receiving treatment 1, male gender will be consider a
tailoring variable. If this is true, male patients will have a greater benefit on treatment 0
and female patients will have a greater benefit on treatment 1.

3.7 Next Steps

In the following chapter we will apply the methods discussed above to a convenient
sample of MediCal patients diagnosed with Schizophrenia. We will begin by calculate a
PPS score that meets both the balancing and overlap conditions. Next we will stratify the
patients into quintiles based on their PPS. Then we will use the Gail-Simon test to
determine if qualitative HTE exist. If qualitative HTE are present, we will identify
tailoring variables that may predict a patient’s success on a given medication. In chapter
45
4, we will present the results of our study. Our hope is that the results of this study will
help physicians choose the most appropriate treatment for their individual patients
diagnosed with schizophrenia.
46
CHAPTER 4: APPLICATION OF METHOD

To demonstrate the use of the PPS, we will use a convenience sample of California
Medicaid (Medi-Cal) beneficiaries diagnosed with schizophrenia. The data will be
derived from a 100% sample of the California Medicaid (Medi-Cal) program fee-for-
service paid claims data from the period of 1994-2002. Beneficiaries with a diagnosis of
schizophrenia recorded on a paid claim (ICD-9 code 295xx), who filled at least one
prescription for an antipsychotic medication during this period, will be eligible for
inclusion in the study.

4.1 Application of the PPS Method in Schizophrenia

Schizophrenia has been identified as one of the ten most debilitating diseases affecting
human beings by the World Health Organization. It is a devastating brain disorder that
affects 1.1 percent of the population age 18 and older, or an estimated 2.2 million
American adults (Keith SJ, 1991; American Psychiatric Association 1997; Narrow WE,
2002).

The analysis of treatment effects using observational data is clearly applicable in
schizophrenia research. The characteristics of patients diagnosed with schizophrenia,
who are enrolled in clinical trials, are often different from the characteristics of patients
seeking treatment outside of the clinical trial setting. A recent publication compared
patients with schizophrenia who participated in a randomized clinical trial (RCT) of
47
ziprasidone, to patients with schizophrenia cared for by a group of volunteer and
randomly selected psychiatrists in the clinical setting. In this study it was shown that
38% of patients with schizophrenia seen in the clinical setting would not have been
eligible to participate in the RCT. Patients seen in the true clinical practice setting were
often older, more likely to be female, and more likely to be white compared to RCT
participants (Zarin DA, 2005). This discrepancy in realistic patient participation
illustrates why the results of RCTs may not provide physicians with relatable information
necessary to treat the patients within their clinical practice.

In addition, traditional methods of measuring treatment effect (i.e. as an average across a
large population of patients) may not be appropriate in a population of patients diagnosed
with schizophrenia. This is largely attributed to the fact that a diagnosis of schizophrenia
may represent a number of discrete syndromes, rather than a single disease. Evaluation
of the diagnostic criteria for schizophrenia supports this notion. The most widely used
criteria for diagnosing schizophrenia is the Diagnostic and Statistical Manual of Mental
Disorders, Fourth Edition (DSM-IV) (American Psychiatric Association, 1994). The
DSM-IV states that in order to obtain a diagnosis of schizophrenia a patient must show
evidence of 2 (or more) of 5 specified symptoms. This in itself could potentially lead to
more than 200 unique combinations of symptoms which are sufficient to receive a
diagnosis of schizophrenia. Therefore, it is very likely that patients diagnosed with
schizophrenia may be very different. In a population of patients diagnosed with
schizophrenia an average treatment effect may obscure the effect of treatment for patients
with a specific set of symptoms. Thus, it becomes difficult for physicians to interpret and
48
apply the results from average treatment effect models to their individual patients. The
objective of our analysis is to provide physicians with predictors of treatment success that
will allow them to select the most appropriate medication for each of their individual
patients.

More importantly, the use of the PPS method to identify the patients diagnosed with
schizophrenia will not only benefit patients, but will also benefit physicians and the
healthcare system as a whole. The cost of schizophrenia is estimated to range from $30
billion to $48 billion in direct medical costs, lost productivity, and Social Security
pensions annually in the United States (Norquist GS, et al. 1996). Identifying the
patients most likely to benefit from a given treatment will improve patient care and
ultimately reduce the high cost often associated with inappropriate care.

4.2 Treatments Used in Schizophrenia

The objective of our analysis is to: (1) Identify if HTEs exist in an analysis comparing
antipsychotics for the treatment of schizophrenia, (2) Identify if HTEs are quantitative or
qualitative, (3) Identify tailoring variables. Ultimately we hope to provide physicians
with predictors of treatment success that will allow them to select the most appropriate
medication for each of their individual patients. In order to identify the most appropriate
treatment for individual patients we must first understand the treatment options available
to them. For the treatment of schizophrenia, the available treatments can be divided into
49
two classes; typical antipsychotics (TAPs) and atypical antipsychotics (AAPs). Both of
these treatments are proven to be effective, but their mechanism of action and side effect
profiles vary.

TAPs, also referred to as first generation antipsychotics, were the first class of
prescription drugs marketed for the treatment of schizophrenia (Goff DC, 2001). These
drugs exert their effect primarily by blocking dopamine, subtype 2 (D
2
) receptors in the
mesolimbaocortical and nigrostriatal areas of the brain. Although the mechanism of
action for all TAPs is the same, their effects may be different. The effect of each
treatment may largely be due to its potency. Potency refers to the strength of the drug, or
how much of a drug must be taken in order for it to be effective. TAPs can be
categorized into three subgroups based on their potency; low, medium, or high. For a list
of TAPs and dosing information see Table 4.1 (Lehman AF, et al. 2004). Increased
potency is associated with an increased effect. Unfortunately, increased potency is also
directly associated with the onset of extra-pyramidal side effects (EPS) (Dipiro JT, 1999).
The EPS associated with TAPs are often the leading cause of discontinuation of their use
(American Psychiatric Association, 1997).

Table 4.1: Typical Antipsychotic Doses
Generic Name Brand Name Usual Dose
(mg/d)
Chlorpromazine
Equivalent Dose
(mg)
Potency
Chlorpromazine Thorazine 300-1000 100 Low
Fluphenazine Prolixin, Permitil 6-20 2 High
Haloperidol Haldol 6-20 2 High
Loxapine Loxitane 30-100 10 Medium
Molindone Moban 30-100 10 Medium
50
Table 4.1: Typical Antipsychotic Doses (Continued)
Generic Name Brand Name Usual Dose
(mg/d)
Chlorpromazine
Equivalent Dose
(mg)
Potency
Mesoridazine Serentil 150-400 50 Low
Perphenazine Trilafon 16-64 10 Medium
Pimozide* Orap 1-20 2-4 High
Prochlorperazine Compazine, Buccastem,
Stemetil, Phenotil
50-150 15 Medium
Thioridazine Mellaril 300-800 100 Low
Thiothixene Navane 15-50 4 High
Trifluoperazine Stelazine 15-50 5 High

* Usual doses were derived from the PORT study. (Lehman AF, et al. 2004).
* Pimozide usual dose was derived from the product label.

An attractive alternative to TAPs for the treatment of schizophrenia are AAPs. AAPs
reduce the risk of EPS and are proven to be equally effective in the treatment of
schizophrenia. For a list of AAPs and dosing information see Table 4.2. The exact
mechanism of action of these agents is unknown. It is known, however, that the receptor
binding profile of the AAPs varies substantially across agents (See figure 4.1). Thus, this
variability may lead to clinical differences in patients’ response to each AAP.

Table 4.2: Atypical Antipsychotic Year to Market & Doses
Generic Name Brand Name Year FDA
Approved
Usual Dose (mg/d)
Clozapine Clozaril 1989 150-600
Risperidone Risperdol 1993 2-8
Olanzapine Zyprexa 1996 10-30
Quetiapine Seroquel 1997 300-800
Ziprasidone Geodon 2001 120-200
Aripiprazone Abilify 2002 10-30
* Usual doses were derived from the PORT study. (Lehman AF, et al. 2004).

51
Figure 4. 1: Receptor Binding Profile of AAPs
D
1
D4.2
D
2
5-HT
2A
5-HT
2C
5-HT
1A
5-HT6
α α α α1
α α α α2
Musc
H1
Olanzapine Clozapine
Risperidone
Quetiapine
Ziprasidone Haloperidol
Adapted from Eli Lilly
Aripiprazole

As demonstrated above, there is not only significant variability between AAPs and TAPs,
but there is also variability between the treatments within each class. Thus, for a single
patient, there is little consistency in the effect of these treatments. In conjunction, the
medication that is optimal for one patient may not be the optimal treatment for another
patient. By identifying tailoring variables, through use of the PPS model, it is hoped that
the optimal clinical treatment for patients’ diagnosed with schizophrenia can be more
effectively identified.

4.3 Unit of Analysis

A key element in the proposed research is the definition of continuous drug therapy
which is used to differentiate between treatment episode types. Duration of therapy will
be defined as a continuous use of a medication. The end of continuous therapy will be
52
identified as a break of greater than 15 days between the end of the estimated days
supplied for a given prescription and the next prescription refill. The estimated days
supplied will be set equal to the sum of the reported days supplied (on the prescription
claim) and a running count of unused days supplied. The unused days supplied is the
supply that a patient may have on-hand due to the early refill of prescriptions. For
example, if the days supplied for a current prescription is 30 days, but a patient gets a
refill on day 25, then they will have accumulated 5 days of unused supply. The sum of
unused days supplied will be capped at 30 days.

Figure 4. 2: Continuous Therapy

The MediCal paid claims data will then be used to identify each episode of drug therapy
initiated by MediCal patients (McCombs et al., 1999; 2000a; 2000b; 2003). This
Rx-1
Rx-3
Rx-2
Rx-4
5 Unused
Days Supplied
15 Day Break (Actual)

10 Day Break
(When Accounting for 5
Unused Days Supplied)

20 Day Break

End of Continuous Therapy

Beginning of
Continuous
Therapy

53
approach was taken to facilitate the documentation of each patient’s antipsychotic drug
use history and clinical status at the time of treatment. Use of episodes is critical in
deriving unbiased estimates of differences across alternative treatments. Five different
types of episodes of care have been defined based on the available data. Patients may
have multiple episodes on one medication and/or across multiple medications.

First Observed Episode: The first observed episode of drug therapy begins with the
patient’s earliest claim for an antipsychotic medication appearing in the data set and ends
just prior to the first 15 day break from antipsychotic treatment.

Figure 4. 3: Example of a First Observed Episode

Gap in Therapy >15 Day
If a patient’s first medication claim was for Olanzapine and they subsequently had a break
in therapy of greater than 15 days, then this initial episode is considered the First
Observed.
Olanzapine
First Observed Episode
54
Restart Episode: An episode will be defined as a restart episode if the patient restarted
on the same antipsychotic after a break in therapy of greater than 15 days from the end of
the previous episode (i.e. as long as the patient is not already on another active
antipsychotic drug therapy at the time the restart treatment is initiated).

Figure 4. 4: Example of a Restart Episode

Switch Episode: An episode will be defined as a switch episode if the patient adds a
second antipsychotic to an active antipsychotic regimen (after a break in therapy of less
than 15 days from the end of the previous treatment episode) and discontinues the
preceding medication within 60 days.
Olanzapine
Gap in Therapy >15 Day
If a patient was originally on Olanzapine then had a break in therapy of greater than 15
days, after which they started Olanzapine this episode would be considered a Restart
Episode.
Olanzapine
55
Figure 4. 5: Example of Switch Episode

Augmentation Episode: An episode will be defined as an augmentation episode if the
patient adds a second antipsychotic to an active antipsychotic regimen and does NOT
discontinue the preceding medication within 60 days.

Risperidone
Gap in Therapy <15 Days
If a patient was originally on Olanzapine then had a break in therapy of less than 15 days after
which they started Risperidone then this episode would be considered a Switch Episode.
Olanzapine
A.
Risperidone
Overlap in Therapy <60 Days
Olanzapine
B.
If a patient was originally on Olanzapine and then started Risperidone before discontinuing
the Olanzapine then the episode would be considered a Switch Episode; so long as the
therapies did not overlap for greater than 60 days.
56
Figure 4. 6: Example of an Augmentation Episode

Late Switch Episode: A patient treatment episode will be defined as a late switch
episode if the patient starts on a different antipsychotic from the previous treatment
episode after a break in therapy of greater than 15 days (i.e. as long as the patient is not
already on another active antipsychotic drug therapy at the time the new treatment is
initiated).

Risperidone
Overlap in Therapy >60 Days
Olanzapine
If a patient was originally on Olanzapine and then started Risperidone before
discontinuing the Olanzapine then the episode would be considered an Augmentation
Episode; so long as the therapies overlapped for greater than 60 days.
57
Figure 4. 7: Example of a Late Switch Episode

4.4 Definition of Analysis Period

The Medi-Cal paid claims data for each patient episode of antipsychotic therapy was
divided into three time periods. The month in which an episode of therapy was initiated
was designated as the “treatment month” for the episode. Two additional periods are
then defined around the treatment month: a 6-month pre-treatment period and a one-year
(12 month) post-treatment period.

4.5 Inclusion & Exclusion Criteria

Episodes which were excluded from the analysis:
1. If the episode did not have a 6-month pre-treatment period and a one-year post-
treatment period.
2. If the episode was designated as the first observed episode. Since these episodes
are the first episodes in our data we are unaware of the previous treatment
Risperidone
Gap in Therapy >15 Day
Olanzapine
58
attempts or medical care received. Many patients with schizophrenia become
eligible for Medicaid after having been diagnosed and treated for this disease.
Therefore, it is likely that if prior treatment history were available these episodes
could be more accurately categorized as restart, switch, augmentation, or late
switch episodes.
3. Episodes for patients less than 18 years of age.
4. Episodes for patients greater than 100 years of age. The age is likely reported
inappropriately for these patients.
5. If the episode involved combination therapy. This is by definition an episode in
which two antipsychotic medications where initiated at the start of the episode
simultaneously. These episodes were eliminated because they violate the
assumption that treatments are mutually exclusive in the PPS method.
6. If the episode of treatment was initiated with a depot agent. Multiple dose vials
make it difficult to accurately measure the days supply for these medications.
7. If the episode of treatment was initiated on clozapine, ziprasidone or
aripiprazole. Clozapine has already been identified to be effective in refractory
patients only. Ziprasidone and aripiprazole are new to the market and only a
small number of patients have utilized them.

4.6 Outcome Measure

Our primary outcome variable for this analysis will be time to all cause discontinuation
(TTAD). TTAD will be calculated using the first, or index medication, for the episode.
59
TTAD will begin at onset of the index treatment and conclude prior to the first break in
index treatment in excess of 15 days. The 15-day gap is consistent with findings by
Weiden, et al. that the risk of hospitalization increased significantly in patients with
schizophrenia after breaks in therapy as short as 10 days (Weiden PJ, 1995).

Figure 4. 8: Example Demonstration of TTAD Identification

Although this outcome measure is not a typical endpoint used in the majority of studies, it
is a common end point used in evaluations of schizophrenia (Gilmer TP, 2004; Beasley
CM, et al. 2007). In CATIE trial, which was funded by a grant from the National
Institute of Mental Health (NIMH), TTAD was selected by a panel of external
consultants and trial center investigators as the primary endpoint. The measure was
selected for its ability to collectively assess multiple aspects of effectiveness, including
Olanzapine
Risperidone

Risperidone
5 Day Overlap
<15 Day Break

Start of
Episode 2

TTAD for Episode 2

TTAD for Episode 1

Start of
Episode 1

20 Day Break

End of
Both Episodes

60
both patient and physician judgments about efficacy and tolerability. Proponents of the
measure also cited its relevance to “real world” clinical practice, as an advantage (Stroup
TS, et al. 2003). Thus, for our analysis we feel TTAD is an appropriate measure of
treatment effect in this population.

For the purposes of our analysis we will define treatment success as TTAD of greater
than or equal to 360 days (1 year). An associated dummy variable, ‘success’, will be
created and set equal to ‘one’ for TTAD success, and ‘zero’ otherwise. The rational for
defining success in this manner is supported by the Schizophrenia Patient Outcomes
Research Team (PORT). This team was developed in 1992, by the Agency for Health
Care Policy and Research and the National Institute of Mental Health. The objective of
the team was to develop and disseminate recommendations for the treatment of
schizophrenia based on existing scientific evidence. One of the recommendations of the
team was that persons who experience acute symptom relief with an antipsychotic
medication should continue to receive that medication for at least 1 year subsequent to
symptom stabilization. This was expected to reduce the risk of relapse. The rational for
this recommendation was based on more than 30 clinical trials. These trials confirmed
that maintenance therapy with an antipsychotic medication over a 1 year period was
beneficial. However, the value of maintenance therapy beyond the first year had not been
extensively studied (Lehman AF, et al. 2004).

61
4.7 Covariates

For the purposes of our analysis we will include covariates representing the type of
schizophrenia, patient demographic factors, comorbidities, concomitant drug use, prior
antipsychotic drug therapy, prior healthcare resource utilization and episode type. The
definitions of the covariates are as follows:
4.7.1 Type of schizophrenia

The clinical presentation of schizophrenia may vary significantly across patients.
Therefore the type of schizophrenia may provide us with one method of categorizing
patients based upon their clinical presentation. For the purposes of our analysis, the type
of schizophrenia will be defined using diagnosis codes (ICD-9 codes) in the treatment
month. If a diagnosis for a specific type of schizophrenia is not observed in the treatment
month then the diagnosis made in the pre-treatment period will be used. A list of the
types of schizophrenia used in our analysis is presented in Table 4.3.
Table 4.3: Covariates Representing Type of Schizophrenia
Definition ICD-9 Codes Variables
Simple Type 295.00-295.09 SIMPLE
Disorganized Type 295.10-295.19 DISORG
Catatonic Type 295.20-295.29 CATA
Paranoid Type 295.30-295.39 PARAS
Acute Schizophrenic Episode 295.40-295.49 ACUTE
Latent Schizophrenia 295.50-295.59 LATENT
Residual Schizophrenia 295.60-295.69 RESID
Schizo-affective type 295.70-295.79 SAFF
Other specified types of
schizophrenia
295.80-295.99 OTHNOS

62
Patients without a diagnosis for a specific type of schizophrenia or who have a diagnosis
of more than one type of schizophrenia will be classified as having an “unspecified” type
of schizophrenia.
4.7.2 Demographic Factors
Demographic factors may provide us with valuable information on the effect of treatment
in individuals. Age and gender may be good predictors of how long a patient has been
living with the disease. Three-quarters of persons with schizophrenia develop the disease
between 16 and 25 years of age; onset is uncommon after age 30 and rare after age 40
(Loranger AW, 1984). However, onset of schizophrenia is more common in males
between the ages of 16-25 and in females between the ages of 25 to 30 (Loranger AW,
1984).
Ethnic/racial groups may provide us information about the metabolism of the
medications. Research has indicated that ethnic groups may differ in their frequencies of
certain gene changes that affect drug metabolism (Johnson JA, 1997). Therefore, it is
possible that a patient’s age, gender and/or ethnic/racial groups could cause them to have
different outcome on different antipsychotic medications. Definition of the dummy
variables used to describe the demographic factors of each patient can be found in Table
4.4.
63
Table 4.4: Demographic Covariates
Definition Variables
18 ≤ Age in years ≤ 25 AGE_CAT18
25 < Age in years ≤ 35 AGE_CAT25
35 < Age in years ≤ 45 AGE_CAT35
45 < Age in years ≤ 55 AGE_CAT45
55 < Age in years ≤ 65 AGE_CAT55
65 ≤ Age in years AGE_CAT65
Male MALE
White, Non-Hispanic WHITE
African American BLACK
Hispanic HISP
Asian ASIAN
Any other Race/Ethnic Group OTHER_RACE

4.7.3 Comorbidities

Comorbidities are highly likely to affect a patient’s response to a given medication.
Therefore, we will use the diagnosis codes (ICD-9 codes) to create an exhaustive list of
comorbidities that may affect the patient’s response to treatment. If a patient has an ICD-
9 code indicating a diagnosis of a given comorbidity in either the pre-treatment or
treatment month they will be designated as having that comorbidity. A list of
comorbidities and their definition are presented in Table 4.5.
Table 4.5: Comorbidity Covariates
Definition ICD-9 Codes Variables
Accident E800-949;
E960-999
ACC
Paranoid, Anxiety, Phobia, Hysteria, Obsessive
Compulsive Disorder, or Other Neurotic Disorder
297.00-297.99; 300.00-
300.09; 300.20-300.29;
300.10-300.19;
300.30-300.39; 300.40-
300.49
ANX
Arrythmia Diagnosis 426.00-427.99 ARRTDX
Bipolar Affective Disorder 296.4-296.89 BIP
Blood System Disease 280.00-289.99 BLOOD
64
Table 4.5: Comorbidity Covariates (Continued)
Definition ICD-9 Codes Variables
Acute Myocardial Infarct, Angina, Congestive Heart
Failure
410.00-410.99; 413.00-
413.99; 428.00-428.99
CARDIO
Psychosis with Origin Specific to Childhood or
Hyperkinetic Syndrome of Childhood
299.00-299.99;
314.00-314.99
CHILD
Congenital Anomalies 740.00-759.99 CONGE
Dementia 290.00-290.99;
294.10-294.19;
319.00-319.99;
292.82; 090.40; 094.10-
094.19;
DEMEN
Major Depressive Disorder or Neurotic Depression 296.20-296.30; 300.40-
300.49
DEP
Diseases of the Digestive System 520.00-579.99 DIGES
Diabetes Mellitus 250.00-259.99 DIADX
Endocrinology Disease (Except Diabetes Mellitus) 240.00-240.99; 251.00-
271.99;
273.00-279.99
ENDO
Diagnosis of Extrapyramidal Side Effects 333.xx; E853.xx;
E854.xx
EPSDX
Disease of the Genitourinary System 580.00-629.99 GENTI
Glaucoma 365.00-365.99; GLAUC
High Blood Pressure 401.00-405.99 HBP
Human Immunodeficiancy Virus 042.00 or V08 HIV
High Cholesterol 272.00-272.99 HLIPID
Infection 001.00-041.19; 043.00-
090.39;
090.41-094.09;
094.20-139.99
INFEC
Injury and Poisoning 800.00-999.99 INJUR
Manic Disorder 296.00-296.19 MANIC
Disease of the Musculoskeletal System and Connective
Tissue
710.00-739.99 MUSCL
Neoplasm 140.00-239.99 NEOPL
Disease of the Nervous System 320.00-364.99;
366.00-389.99
NERV
Circulatory Disorders 390-400.99;
406.00-409.99;
411.00; 412.99;
414.00; 425.99;
429.00-459.99
OCIRC
Alcohol Psychosis, Drug Psychosis, Transient Organic
Psychotic Conditions, or Other Nonorganic Psychoses
291.00-291.99;
292.00-292.81;
293.00-293.99; 294.00-
294.99;
294.20-294.99
OTHPSY
Conditions of the Perinatal Period 760.00-779.99 PERI
Personality Disorder 301.00-301.99 PERSO
Complications of Pregnancy, Childbirth, and the
Puerperium
630.00-679.99 PSTG
65
Table 4.5: Comorbidity Covariates (Continued)
Definition ICD-9 Codes Variables
Disease of the Respiratory System 460.00-519.99 RESPI
Substance Abuse, Alcohol Dependency, or Drug
Dependency
305.1-305.9;
303.00-303.99;
305.00-305.09;
304.00-304.99
SABUSE
Sex Deviations and Disorders 302.00-302.99 SEXD
Disease of the Skin and Subcutaneous Tissue 680.00-709.99 SKIN
Suicide Attempt Diagnosis E950-959;
NEC-959.9;
300.9; X60-X87; Y87.0
SUICDX

4.7.4 Concomitant Therapy

Drug-drug interactions are a common occurrence in pharmaceutical care. They are an
especially important issue for psychiatrists because

many commonly used psychotropic
agents have abundant P-450 and

P-glycoprotein effects that may affect the blood levels of
other

drugs. Similarly, many of the non-psychotropic drugs that

schizophrenic patients
often take can significantly increase or decrease the

blood levels of most psychotropic
agents (Ereshefsky L, 1997). For the purpose of our analysis we include all concomitant
medications available in our data. We will create concomitant drug dummy variables
using cost data from the treatment month. A list of these concomitant variables is
presented in Table 4.6.
Table 4.6: Concomitant Therapy Covariates
Definition Variables
Antidepressant Use AD
Hypnotics HYPNT
Anti-seizure Medications SEIZU
Anti-arrhymics ARRH
Anti-parkinsonism Medications APARK
Anxiety Medications ANXIE
66
Table 4.6: Concomitant Therapy Covariates (Continued)
Definition Variables
Mood Stabilizers MOOD
Cholesterol-Lowering Medications LIPID
Primary Anti-Extrapyramidal Side Effect Medications EPS1
Secondary Anti-Extrapyramidal Side Effect Medications EPS2
Medications Used to Treat Diabetes DIABRX
Medications Thought to Cause Diabetes DIABDX
Psychotherapy PTHER

4.7.5 Prior Antipsychotic Drug Use

It is our belief that the outcome on current treatment may be affected by the prior
antipsychotics used. Therefore, using data from the previous episode for each patient we
will create dummy variables for all prior antipsychotics. A list of these prior
antipsychotic variables is presented in Table 4.7.
Table 4.7: Prior Antipsychotic Drug Use Covariates

Definition Variables
Olanzapine OLA
Risperidone RIS
Quetiapine QUE
Ziprasidone ZIP
Other Atypical Antipsychotic OAP
Low Potency Typical Antipsychotic L_TAP
Medium Potency Typical Antipsychotic M_TAP
High Potency Typical Antipsychotic H_TAP

4.7.6 Episode type

Episode type may provide information on how best to use a given medication. Therefore,
in our analysis we will include dummy variables indicating episode type (i.e. restart,
67
switch, augmentation, late switch). The definition of episodes will follow the methods
discussed earlier in this chapter.
4.7.7 Prior Health Services Utilization

Prior health service utilization may be an indicator of treatment severity. Thus, we would
expect a patient with no prior hospitalization in the previous 6 months to be in better
health than a patient who was hospitalized. For our analysis, we will create dummy
variable for each health service sector. We will set each sector variable equal to ‘one’ if a
cost for such care was observed in the month the episode was initiated or in the 6 months
previous. Otherwise the variable will be set equal to zero. These health service variables
are presented in Table 4.8.
Table 4.8: Prior Health Services Utilization Covariates
Definition Variables
Long Term Care LTC
Acute Hospital Care AHOSP
Psychiatric Hospital Care PHOSP
Inpatient Rehabilitation REHAB
Community Mental Health Center Care CMHC
Psychologist Care PSYCH

4.8 Statistical Methods
4.8.1 Descriptive Statistics

One of the most important steps in any observational study is to understand the data you
are studying. Thus, we will begin the analysis of our data analyzing the effect of the
68
studies inclusion and exclusion criteria on the number of observations. We will present
the results of this analysis using a simple diagram displaying the number of episodes
dropped due to each exclusion criteria.

Our next step will be to gain a better understanding of the episodes remaining in our
study population. We will accomplish this by running some simple descriptive statistics
on all specified covariates discussed above. We will compare these statistics across
treatment groups using t-test or chi-square test for continuous and categorical measures,
respectively.

4.8.2 Prognostic Propensity Score (PPS)
4.8.2.1 Calculation

Once we have adequately identified the episodes included in our final analysis we will
begin our primary analysis utilizing the PPS. To start we will calculate the PPS score, as
defined in Chapter 3, for the CT utilizing success (TTAD > 360 days) as the outcome of
interest. We will use a stepwise discriminant analysis to identify significant prognostic
factors, from the list of covariates describe above, to include in our model. Once we have
defined our final model, we will use this model to calculate the PPS or predicted
probability of success on the CT for all other patients.

Next we will identify if the overlap (or support) condition is met. We will accomplish
this using both a box-plot and histogram to visually identify the region of common
69
support. This will ensure that we have sufficient overlap of PPS values for those patients
who have and do not have success on the CT. We will further refine this region of
common support by validating its accuracy across all treatments.

Finally we will stratify episodes into quintiles based on the CT PPS. Within each quintile
we will check that the balancing condition is met. If the balancing condition is not met
we will have to redefine the PPS model possibly including higher-order and interaction
terms. If the balancing condition is still not met, even after including higher-order and
interaction terms, we will explore increasing the number of strata. If after increasing the
number of strata the balancing condition is still not met, the next step would be to
redefine the treatment outcome success. For the purposes of this analysis we originally
define 1 = success if the TTAD exceeds 360 days, and zero otherwise. In the original
data the TTAD is a discrete variable measured in days. By transforming a discrete
variable into a binary measure we will have lost a significant amount of information.
Thus, if the balancing condition does not hold it may be necessary to reevaluate the use
of the success variable and utilize the TTAD variable in its place. Therefore, we would
calculate the PPS using a multinomial probit model utilizing the TTAD variable as our
outcome.
4.8.2.2 Heterogeneous Treatment Effects

To identify if heterogeneous treatment effects exist we will use the quintile method
discussed in chapter 3. Once quintiles have been defined, we will perform a crude
analysis of the average treatment effect. This analysis will provide us an initial indication
70
as to whether HTE exists, and more importantly an early indication as to whether
qualitative HTE exists. However, the crude analysis does not take into consideration the
differences in the distribution of covariates across treatment groups. Therefore it is
necessary to perform an adjusted analysis to compensate for these differences and
validate the crude indications. This will be accomplished by calculating the treatment
effect of the CT versus each alternative one-by-one, within each quintile. Such as:

X AltTx Y
3 1 1 0
) ( β β β + + = (eq. 4.1)
X AltTx Y
3 2 1 0
) ( β β β + + = (eq. 4.2)

If heterogeneous treatment effects exist, then we would expect that the coefficient of each
treatment to vary from quintile to quintile. If qualitative HTE exists than we would
expect the ATE in at least one of the quintiles to be the opposite sign (e.g. positive or
negative) of the other quintiles. We will test if the direction of the treatment effects are
significantly different across strata using the Gail-Simon test (Gail M, 1985).

Once we have completed this analysis using one treatment as the CT group, we will
repeat the above steps 5 times using each alternative treatment as the CT.
4.8.2.3 Tailoring Variables

If we can conclude that qualitative HTE exists using the Gail-Simon test then we can
proceed with trying to identify tailoring variables. If HTE exists then we would identify
71
the subgroup in which the treatment effect has the lowest negative value. This is the
subgroup that achieves the greatest benefit from the treatment that we designated as CT.
We repeat this analysis seven times; once for each treatment. We should have identified
no more than 7 subgroups (i.e. the most beneficial for each treatment). It is possible to
have less than 7 subgroups if the Gail-Simon test, in step 1, concluded that qualitative
HTE was not present during any one of the analyses. Take for example the analysis
using low potency typical antipsychotics as the CT. If within each stratum every other
treatment was superior (i.e. the ATE was positive and the Gail-Simon test concluded that
qualitative HTE was not present) then the low potency typical antipsychotics may be
dominated by all other treatment. Thus, we would not need to include low-potency
antipsychotics in our analysis to identify tailoring variables.

Once we have identified the pertinent subgroup for each non-dominated treatment we
will compare the characteristics of the patients across these subgroups. We will use the t-
tests and chi-square tests for continuous and categorical variables, respectively. The
characteristics which are not significantly different across the treatment subgroups can be
considered prognostic factors. These variables increase the probability of success
regardless of which treatment is received. More importantly, variables which are
significantly different across the subgroups will be considered tailoring variables. These
variables will ultimately provide physicians with the information necessary to prescribe
the optimal treatment to their patients diagnosed with schizophrenia.

72
4.9 Next Steps

The results of these models will be presented in Chapter 5. This will be followed by a
brief discussion of the results and the implications of the PPS method in Chapter 6.
73
CHAPTER 5: RESULTS
5.1 Data

The data for this analysis is derived from a 100% sample of the California Medicaid
(Medi-Cal) program fee-for-service paid claims data from the period of 1994-2002. The
initial data set contains 225,952 beneficiaries with a diagnosis of schizophrenia who
filled at least one prescription for an antipsychotic medication. From these patients
2,211,763 treatment episodes are created. Our final study sample is derived using the
exclusion and inclusion criteria outlined in Chapter 4. For details on how the inclusion
and exclusion criteria affect the final study sample size see Figure 5.1.
Figure 5.1: Inclusion & Exclusion Criteria

74
As noted in Figure 5.1, after accounting for the inclusion and exclusion criteria the final
study sample consists of 219,858 episodes. The break down of episodes by treatment
type can be seen in Figure 5.2.

Figure 5.2: Number of Episodes by Treatment
38,804
37,560
11,413
26,395
56,901
48,785 Olanzapine
Risperidone
Quetiapine
Low Potency TAPs
Medium Potency TAPs
High Potency TAPs

5.2 Descriptive Statistics

To further understand the final data used in our analysis we compute some simple
descriptive statistics (see Table A.1 in the Appendix). The results indicate that there are
significant differences between the characteristics of patients receiving each treatment at
baseline. The characteristics of patients receiving medium potency TAPs appear to be
the most unlike any other treatment. A greater proportion of patients receiving medium
potency TAPs have an unspecified type of schizophrenia, are over 65 years of age, are
Hispanic and are diagnosed with non-mental health comorbidities.

The descriptive statistics also reveal a disproportionate distribution of some pertinent
covariates in our analysis. For example, the majority of patients within each treatment
75
group are Caucasian and have an unspecified type of schizophrenia. In addition, specific
types of schizophrenia, certain comorbidities (i.e. HIV, manic disorders, conditions of
perinatal period, and sexual deviations and disorders), and prior inpatient rehabilitation
may be under-represented in our data set.

5.3 Prognostic Propensity Score
5.3.1 Calculation

As described in chapter 4, the approach for calculating the PPS is similar for all
treatments. Therefore to simplify the presentation of our results, we will focus on the
example in which olanzapine is utilized as the CT, and speak to it in greater detail. We’ll
then present any notable trends observed when considering the alternative treatments as
the CT, and include a full set of detailed results in the Appendix.

When utilizing olanzapine as the CT, our first step is to define a subset of the data
including only those 38,804 episodes initiated with olanzapine. Using this subset of the
data we calculate the olanzapine PPS. This is accomplished using a stepwise logistic
regression including all covariates with a significance value less than 0.30 (See Table
5.1).
76
Table 5.1: PPS Model Utilizing Olanzapine as the CT
Covariate Point Estimate 95% Wald Confidence Limits
DISORG 1.488 1.013 2.185
PARAS 1.330 1.219 1.450
LATENT 0.324 0.040 2.625
RESID 1.421 1.086 1.859
SAFF 1.152 1.048 1.265
OTHNOS 1.363 1.231 1.510
AGE_CAT25 1.224 1.067 1.405
AGE_CAT35 1.387 1.216 1.581
AGE_CAT45 1.476 1.293 1.687
AGE_CAT55 1.723 1.495 1.986
AGE_CAT65 1.547 1.336 1.792
MALE 1.111 1.054 1.172
BLACK 0.615 0.566 0.669
HISP 0.647 0.572 0.731
ASAIN 0.782 0.699 0.876
OTHER_RACE 0.926 0.870 0.986
ACC 0.719 0.587 0.880
ANX 0.829 0.756 0.910
BIP 0.810 0.743 0.884
CHILD 1.441 1.010 2.054
DEMEN 1.453 1.278 1.651
DEP 0.790 0.707 0.882
GENTI 0.952 0.885 1.025
HIV 0.300 0.105 0.854
HLIPID 1.082 1.001 1.170
INFEC 0.882 0.773 1.005
INJUR 0.929 0.864 0.999
MUSCL 0.853 0.800 0.910
NERV 1.071 1.010 1.134
OCIRC 0.913 0.842 0.989
PERSO 1.195 0.952 1.499
RESPI 0.956 0.898 1.018
SABUSE 0.581 0.514 0.657
SEXDE 0.561 0.233 1.351
SKIN 1.231 1.144 1.325
AD 1.092 0.978 1.220
HYPNT 0.749 0.663 0.846
MOOD 1.195 1.121 1.274
EPS1 1.338 1.256 1.424
EPS2 0.833 0.729 0.952
DIABRX 0.847 0.765 0.937
DIABDX 0.914 0.853 0.979
PTHER 1.099 1.010 1.196
RIS 0.926 0.853 1.005
QUE 0.688 0.594 0.797
OAP 1.260 1.086 1.463
77
Table 5.1: PPS Model Utilizing Olanzapine as the CT (Continued)
Covariate Point Estimate 95% Wald Confidence Limits
TAP 1.221 1.134 1.315
SWT 2.738 2.497 3.002
LATE_SWT 1.769 1.651 1.895
AUG 2.125 1.923 2.349
LTC 1.728 1.578 1.893
AHOSP 0.838 0.729 0.963
PHOSP 0.790 0.681 0.916
CMHC 0.949 0.893 1.008
PSYCH 1.432 1.223 1.678

Our final olanzapine PPS model calculated using this method contains 55 covariates. As
expected, numerous comorbidities are associated with a decreased probability of success
on olanzapine. These include, anxiety (ANX), bipolar (BIP), depression (DEP), and
substance abuse (SABUSE); to name just a few. In addition, patients who were
hospitalized in the prior 6-months (AHOSP & PHOSP) also had a lower probability of
success on olanzapine.

When considering the alternative treatments as the CT, the final PPS models ranged in
size from 29 to 64 covariates; for quetiapine and high potency TAPs, respectively.
Detailed information including point estimates and 95% confidence intervals for each
model can be seen in Tables A.2 to A.6 in the Appendix.
5.3.2 Overlap (or Support) Condition

In order to utilize the PPS method to measure treatment effects we begin by verifying that
the overlap (or support) condition is met. This requires that for every patient successfully
treated with the CT there is a similar patient unsuccessfully treated with the CT, and visa-
78
versa. If a patient in either the successful or unsuccessful group does not have a similar
patient in the opposing group then these patients are dropped from our analysis. To
accomplish this we use a visual method. The first approach we explore is to plot the
distribution using a box-plot design. An example of this approach, utilizing olanzapine
as the CT, can be seen in Figure 5.3. Using this method we could trim those patients
above the maximum and below the minimum of the opposing group to ensure significant
overlap. However, if we were to use this approach we fear that we may not adequately
identify irregular patterns in the distribution of the PPS. Therefore, we re-plot the same
data using a histogram. An example when utilizing olanzapine as the CT can be seen in
Figure 5.4. Using the histogram we identify the region of common support. This region
includes the PPS values for which there is an overlap in distributions between the PPS for
patients who have and do not have success on the CT. In the olanzapine example this
area is defined by a PPS greater than 0.025 and less than 0.675 (see visual delineation in
Figure 5.4).
Figure 5.3: Box Plot of Olanzapine PPS – Success vs No Success
0 1
0
0.2
0.4
0.6
0.8
E
s
t
i
m
a
t
e
d
P
r
o
b
a
b
i
l
i
t
y
success

79
Figure 5.4: Histogram of Olanzapine PPS – Success vs No Success
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
C
o
u
n
t
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
C
o
u
n
t
1
Estimated Probability

Our ultimate goal, however, is to compare the effect of the CT versus each alternative
treatment. Thus we must also verify that there is adequate overlap in the distribution of
the CT PPS across treatments. This requires that we identify patients with similar
characteristics by utilizing the CT model to extrapolate a CT PPS value for ALL patients
receiving an alternative treatment. Given that the PPS is representative of a combination
of covariates, we assume that patients with similar CT PPS scores are exchangeable. In
order to confirm that exchangeable patients exist, we compare the distribution of the CT
PPS values for the alternative treatments to the distribution of the original CT PPS.
Based on the determined overlap, we redefine our CT region of common support to
ensure adequate overlap across treatments. See Figure 5.5 for an example in which
olanzapine is utilized as the CT. In this example we compare the distribution of the
original olanzapine PPS to that of the olanzapine PPS values extrapolated to patients
Region of
Common Support
80
receiving each alternative treatment. So just to reiterate, an olanzapine PPS value is
generated for each patient receiving an alternative treatment (i.e. patients receiving
risperidone, quetiapine, etc.). Those patients are still identified/labeled within our
histogram using the alternative treatment they originally received. However it’s the
extrapolated olanzapine PPS values which are graphed and utilized for comparison to
identify the region of common support.

Figure 5.5: Histogram of Olanzapine PPS by Treatment
0 0. 05 0. 1 0. 15 0. 2 0. 25 0. 3 0. 35 0. 4 0. 45 0. 5 0. 55 0. 6 0. 65 0. 7 0. 75 0. 8 0. 85 0. 9 0. 95 1
0
2000
4000
6000
8000
10000
C
o
u
n
t
ol a_pps

0
5000
10000
15000
20000
C
o
u
n
t
0
5000
10000
15000
20000
C
o
u
n
t
L
o
w
_
T
A
P
0
5000
10000
15000
20000
C
o
u
n
t
M
e
d
_
T
A
P
0
5000
10000
15000
20000
C
o
u
n
t
0 0. 05 0. 1 0. 15 0. 2 0. 25 0. 3 0. 35 0. 4 0. 45 0. 5 0. 55 0. 6 0. 65 0. 7 0. 75 0. 8 0. 85 0. 9 0. 95 1
0
5000
10000
15000
20000
C
o
u
n
t
ol a_pps
* Accurate overlap across treatments may be difficult to see due to the scale of the diagrams.

Region of
Common Support
81
In this olanzapine example the redefined region of common support is designated by a
PPS greater than 0.1 and less than 0.55 (see visual delineation in Figure 5.5). Refining
our population based on this region of common support reduces our total number of
episodes. The impact to the overall number of episodes when utilizing olanzapine as the
CT is displayed in Figure 5.6.

Figure 5.6: Refined Breakdown of Episodes by Treatment Type When Utilizing
Olanzapine as the CT
33,680
32,608
8,383
24,981
49,090
46,697 Olanzapine
Risperidone
Quetiapine
Low Potency TAPs
Medium Potency TAPs
High Potency TAPs

We repeat this analysis considering each alternative treatment as the CT. The results of
these analyses can be found in detail in the Appendix in Figures A.1 to A.10. From these
figures we were able to determine the region of common support for each alternative
treatment as defined in Table 5.2 below.
Table 5.2: Common Region of Support Considering each Treatment as the CT
Comparison Treatment
(CT)
Lower Limit of the
Common Region of
Support
Higher Limit of the
Common Region of
Support
Olanzapine 0.1 0.55
Risperidone 0.1 0.5
Quetiapine 0.1 0.5
Low Potency TAPs 0.05 0.4
82
Table 5.2: Common Region of Support Considering each Treatment as the CT
(Continued)

Comparison Treatment
(CT)
Lower Limit of the
Common Region of
Support
Higher Limit of the
Common Region of
Support
Medium Potency TAPs 0 0.2
High Potency TAPs 0.1 0.3
* Note: Based upon the defined region of common support a detailed breakdown of episodes by
treatment type for each CT can be seen in the Appendix in Figures A.2, A.4, A.6, A.8 & A.10.

As evident in Table 5.2, the region of common support is similar when utilizing AAPs as
the CT. However the region of common support covers a much lower and smaller range
of PPS values when utilizing TAPs as the CT.
5.3.3 Balancing Condition

Once we identify our region of common support we stratify the refined population of
patients treated with the CT into 5 equal groups (or quintiles) based upon their PPS.
When olanzapine is utilized as the CT these quintile cutpoints are defined as displayed in
Table 5.3 below.
Table 5.3: Cutpoints for Quintiles when Olanzapine is Utilized as the CT
Olanzapine Quintile Lower Bound Upper Bound
1 0.10 0.14
2 0.14 0.17
3 0.17 0.23
4 0.23 0.33
5 0.33 0.55
* The designated cutpopints for each alternative CT treatment are presented in
Tables A.7 to A.11 in the Appendix

In order to verify the balancing condition we compare the characteristics between the
patients who had and did not have success on the CT within each quintile using a chi-
83
square test (see Table 5.4 for example utilizing olanzapine as the CT). If no significant
differences exist between these two groups of patients then the average treatment effect
calculated within each quintile is considered homogeneous.
Table 5.4: Descriptive Statistics for Olanzapine Quintile 1
Covariates Success
N=814
No Success
N=6002
Age
18-25 11.79 9.96
25-35 17.08 16.48
35-45 24.32 26.44
45-55 24.20 24.48
55-65 10.93 10.55
>65 11.67 12.10
Male 40.54 39.07
Race/Ethnicity
White, Non-Hispanic 38.33 34.99
African American 17.57 20.08
Hispanic 10.69 10.45
Asian 13.39 13.45
Other 20.02 21.04
* The full table can be seen in the Appendix Table A.12a

As can be seen in Table 5.4 the characteristics of patients who have and did not have
success on olanzapine are similar for age, gender and race. The full set of results for the
olanzapine quintiles can be seen in the Appendix in Tables A.12a thru A.12e. The
detailed results indicate that by stratifying patients within each quintile (i.e. utilizing the
olanzapine PPS) we are able to eliminate most of the significant differences between
those patients who had and did not have success on olanzapine. Thus, the balancing
condition is satisfied and the average treatment effect calculated within each quintile is
considered homogeneous.

84
The balancing condition is then verified considering the alternative treatments as the CT.
The results of these analyses can be seen in the Appendix in Tables A.13-A.17. The
balancing condition is satisfied in all cases. Therefore, our estimates of the average
treatment effect within each quintile, for all treatments, are considered homogeneous.

5.4 Heterogeneous Treatment Effects
5.4.1 Crude Treatment Effect

Once the PPS quintiles are appropriately defined we then calculate the treatment effects
between the CT and each alternative treatment within each quintile. We begin by using a
crude measure of the treatment effect. This is reported as a simple average time to all
cause discontinuation within each quintile for each treatment. We display the results in a
line graph. If the CT PPS follows our hypothesis the average time to all cause
discontinuation for the CT should increase as the quintiles increase from 1 to 5. We also
expect the crude treatment effects to be heterogeneous. Thus, the lines within the graph
displaying the crude treatment effects for each alternative treatment would not be parallel
to that of the CT. More specifically, we hypothesize that the heterogeneous treatment
effects will be qualitative. Thus, we would expect the lines displaying the crude
treatment effect for each alternative treatment to intersect the line displaying the crude
treatment effect for the CT. Thus, if our hypotheses hold true, the effect of the CT would
be greater than all other treatments in quintile 5 (i.e. where the quintiles are defined by
85
the CT PPS) and worse than all other treatments in quintile 1. The results when utilizing
olanzapine as the CT are displayed in Figure 5.7.

Figure 5.7: Crude Treatment Effect When Olanzapine is Utilized as the CT
0
50
100
150
200
250
300
350
400
450
500
1 2 3 4 5
Olanzapine Quintiles
Average Time to All Cause Discontinuation
Olanzapine
Risperidone
Quetiapine
Low TAP
Med TAP
High TAP

From Figure 5.7 it is evident that the average time to all cause discontinuation increases
for olanzapine from quintile 1 to 5, and thus displays the expected trend. Since the lines
of each alternative treatment are not parallel to that of olanzapine we are also able to
conclude that the treatment effects are heterogeneous. However, our hypothesis
regarding qualitative heterogeneity is not fully satisfied. As expected the average time to
all cause discontinuation is greatest for patients treated with olanzapine in quintile 5.
However, contrary to our hypothesis, patients treated with olanzapine in quintile 1 do not
always have the shortest duration of therapy.
86
The figure demonstrates that qualitative HTE is apparent between olanzapine and the
other AAPs (i.e. risperidone and quetiapine); evident from the intersection of the trend
lines within the graph. However, qualitative HTE does not appear to be present in the
comparison of olanzapine to the TAPs (low, medium or high potency). The average time
to all cause discontinuation is higher for olanzapine in comparison to the TAPs within
every quintile (i.e. the lines never intersect).

We repeat this analysis using each alternative treatment as the CT. In each case the CT
PPS follows the expected trend; the average time to all cause discontinuation for the CT
increases from quintiles 1 to 5. However, the rest of our results are not as expected (See
Appendix Figures A.11 to A.15). Heterogeneous treatment effects appear to be present
when risperidone and quetiapine are defined as the CT, respectively. However,
heterogeneity is less apparent when the TAPs are utilized as the CT (with the exception
of the comparison of low and high potency TAPs). This of course also limits our ability
to identify qualitative HTE in these comparisons. Qualitative HTE appears to only be
present in several of the comparisons; these include the comparisons of quetiapine as the
CT vs. low potency TAPs, quetiapine as the CT vs. high potency TAPs, and low potency
TAPs as the CT vs. high potency TAPs. The heterogeneity apparent in all other
comparison appears to be quantitative. The largest departure from our a prior hypothesis
is the fact that the time to all cause discontinuation for the treatment defined as the CT is
not greater than all other treatments in quintile 5 and less than all other treatments in
quintile 1. The trends in duration of therapy are actually similar across all graphs.
Regardless of the CT used to define the quintiles, patients treated with olanzapine always
87
have the longest duration of therapy in quintile 5, followed in order by risperidone,
quetiapine, low potency TAPs, high potency TAPs and medium potency TAPs. In
addition, medium potency TAPs appear to be dominated by all other treatments in every
quintile.

Although the results are not in full agreement with our a priori hypothesis, it does provide
us an early indication that heterogeneity may exist in the comparison of some treatments.
More importantly, there may be some instances where qualitative heterogeneity is
present. However, since this method fails to adjust for any other covariates its results are
only reliable if there are no differences in patients across treatments. Therefore it is
necessary to obtain a more reliable estimate by calculating an average treatment effect
that adjusts for such differences. With further analysis we’ll be able to determine if
heterogeneity and more specifically qualitative heterogeneity truly exists.
5.4.2 Adjusted Treatment Effect

To adjust for differences in patients treated with the alternative therapies we use a logistic
regression to calculate the treatment effect. This analysis will include comparing the CT
to each alternative treatment one-by-one (i.e. CT vs. Alternative 1, CT vs. Alternative 2,
CT vs. Alternative 3, etc). Within each quintile we will calculate the treatment effect
using a logistic regression that contains a dummy variable for the alternative treatment in
combination with all other covariates of interest. The effect of the alternative treatment
in comparison to the CT, within each quintile, will be expressed as a log odds ratio and
displayed in a forest plot. Our a priori hypotheses are identical to those we discussed in
88
the crude treatment effect section above (5.4.1). We expect that the log odds ratio will
decrease as the quintiles increase from 1 to 5. Thus, the effect of the CT will be greatest
in those patients in which the expected probability of success (PPS) is the highest. Next,
we hypothesize that the treatment effects will be heterogeneous. In the forest plots, if
treatment effects are homogeneous we would be able to draw a single line perpendicular
to the x-axis that crosses the treatment effect estimate for each quintile. Finally, we
hypothesize that the heterogeneous treatment effects will be qualitative. Thus, we expect
the log odds ratio to be greater than one in at least one quintile and less than one in at
least one quintile. This indicates that there is a distinct subpopulation of patients who
would benefit most from the alternative treatment, while a separate subpopulation of
patients would benefit most from the CT treatment. More specifically, we would expect
that the probability of success will be greatest for the alternative treatment in quintile 1
(log odds ratio >1) and greatest for the CT in quintile 5 (log odds ratio <1). In quintiles
2, 3 and 4 we expect that the probability of success will be similar for both treatments
(log odds ratio=1). We will confirm our findings using the Gail-Simon Test of
Heterogeneity. This is a log-likelihoods test which tests for differences in the direction of
treatment effects.

Now lets consider the case in which olanzapine is utilized as the CT. The analysis
includes 5 separate comparisons within each quintile (i.e. olanzapine vs. risperidone,
olanzapine vs. quetiapine, olanzapine vs. low potency TAPs, etc.). If we look
specifically at the comparison of olanzapine (CT) versus risperidone, we calculate the
89
treatment effect using a logistic regression containing a risperidone dummy variable in
combination with all other covariates of interest. The results of this analysis are
displayed in Figure 5.8.

Figure 5.8: Adjusted Treatment Effect Utilizing Olanzapine as the CT
Olanzapine vs Risperidone

As expected the log odds ratio decreases as the quintiles increase from 1 to 5. In
addition, we cannot draw a single line perpendicular to the x-axis that connects the
treatment effect estimates for each quintile; thus demonstrating the presence of HTE.
However, our a priori hypothesis that the HTE is qualitative is not satisfied. As we had
expected, the probability of success is indeed greatest for olanzapine in quintile 5, and in
quintiles 2, 3, and 4 there are no significant differences in the probability of success on
either olanzapine or risperidone. However, contrary to our a priori hypothesis, in quintile
1 the probability of success on risperidone is not significantly greater than that of
0 0.5 1 1.5 2
Ola Quintile 1
Ola Quintile 2
Ola Quintile 3
Ola Quintile 4
Ola Quintile 5
Log Odds Ratio
Olanzapine Better Risperidone Better
90
olanzapine. Since there is not a quintile in which risperidone is significantly better than
olanzapine, we suspect that quantitative and not qualitative heterogeneous treatment
effects are present. We further confirm this assumption using the Gail-Simon test of
heterogeneity where the results prove this to be true (p=0.0657).

We repeat this analysis comparing olanzapine to each alternative treatment (see Table
5.5).
Table 5.5: Adjusted Treatment Effect within the Olanzapine Quintiles
Olanzapine
Quintile
Risperidone
Log Odds
Ratio
Quetiapine
Log Odds
Ratio
Low TAPS
Log Odds
Ratio
Medium TAPs
Log Odds
Ratio
High TAPs
Log Odds
Ratio
1 1.139 0.918 0.789 0.188* 0.807
2 1.036 1.118 1.009 0.239* 0.750*
3 1.06 0.997 0.832* 0.198* 0.625*
4 0.909 0.9 0.456* 0.165* 0.395*
5 0.787* 0.86* 0.303* 0.135* 0.290*
Gail Simon
Test
0.0657 0.0972 0.9375 0.9375 0.9375
* p<0.05; Forest Plots of Information Appearing in this Table are provided in Figures A.22a-d in the
Appendix.

As we can see from Table 5.5 in all comparisons utilizing olanzapine as the CT the log
odds ratio follows the general trend, and decreases across quintiles 1 through 5.
However, in our comparison of olanzapine to quetiapine and medium potency TAPs the
decrease in log odds ratios is barely noticeable across quintiles and the treatment effect
appears to be homogeneous. Heterogeneous treatment effects do appear to be present in
the comparisons of olanzapine to risperidone, low potency TAPs and high potency TAPs.
However, when comparing these treatment effects across quintiles, we do not observe
qualitative HTE. The Gail-Simon Test further confirms these results.
91
During our comparisons utilizing olanzapine as the CT, we do identify one notable trend
worth further discussion. When comparing olanzapine to medium potency TAPs the
probability of success on olanzapine is significantly greater than the medium potency
TAPs in every quintile. This can be seen in Figure 5.9 below.

Figure 5.9: Adjusted Treatment Effect Utilizing Olanzapine as the CT
Olanzapine vs Medium Potency TAPs

Thus, we are able to determine that olanzapine is a superior treatment in comparison to
medium potency TAPs.

We replicate our analysis considering each treatment as the CT. The results can be seen
in Figures A.16 to A.21 and are displayed in summary in Table 5.6 below.

0 0.5 1 1.5 2
Ola Quintile 1
Ola Quintile 2
Ola Quintile 3
Ola Quintile 4
Ola Quintile 5
Log Odds Ratio
Olanzapine Better Risperidone Better
92
Table 5.6: Adjusted Treatment Effect within the CT Quintiles
Quintile Olanzapine
Log Odds
Ratio
Risperidone
Log Odds
Ratio
Quetiapine
Log Odds
Ratio
Low TAPS
Log Odds
Ratio
Medium
TAPs
Log Odds
Ratio
High TAPs
Log Odds
Ratio
Olanzapine 1 1.139 0.918 0.789 0.188* 0.807
Olanzapine 2 1.036 1.118 1.009 0.239* 0.750*
Olanzapine 3 1.06 0.997 0.832* 0.198* 0.625*
Olanzapine 4 0.909 0.9 0.456* 0.165* 0.395*
Olanzapine 5 0.787* 0.86* 0.303* 0.135* 0.290*
Gail Simon
Test
0.0657 0.0972 0.9375 0.9375 0.9375
HTE
Qualitative
HTE

Superior
Inferior
Risperidone 1 0.931 1.341* 0.492* 0.168* 0.551*
Risperidone 2 1.014 0.942 0.667* 0.179* 0.601*
Risperidone 3 1.087 1.221* 0.664* 0.168* 0.542*
Risperidone 4 1.041 0.858 0.526* 0.160* 0.465*
Risperidone 5 1.200* 0.918 0.402* 0.159* 0.358*
Gail Simon
Test
0.9375 0.1082 0.1082 0.9375 0.9375
HTE
Qualitative
HTE

Superior
Inferior
Quetiapine 1 0.904 0.878 0.811 0.243* 0.941
Quetiapine 2 0.999 0.938 0.992 0.190* 0.643*
Quetiapine 3 1.092 1.067 0.652* 0.243* 0.532*
Quetiapine 4 1.098 1.028 0.473* 0.176* 0.413*
Quetiapine 5 1.071 0.948 0.333* 0.147* 0.377*
Gail Simon
Test
0.8239 0.8239 0.6602 0.9375 0.9375
HTE
Qualitative
HTE

Superior
Inferior

93
Table 5.6: Adjusted Treatment Effect within the CT Quintiles (Continued)
Quintile Olanzapine
Log Odds
Ratio
Risperidone
Log Odds
Ratio
Quetiapine
Log Odds
Ratio
Low TAPS
Log Odds
Ratio
Medium
TAPs
Log Odds
Ratio
High TAPs
Log Odds
Ratio
Low TAPs 1 2.699* 2.736* 2.764* 0.455* 1.116
Low TAPs 2 2.645* 2.647* 2.046* 0.389* 1.077
Low TAPs 3 2.435* 2.366* 2.233* 0.359* 1.027
Low TAPs 4 1.998* 1.792* 1.936* 0.296* 0.901*
Low TAPs 5 1.530* 1.533* 1.452* 0.260* 0.660*
Gail Simon
Test
0.9375 0.9375 0.9375 0.9375 0.1027
HTE
Qualitative
HTE

Superior
Inferior
Medium
TAPs 1
25.911* 40.021* 34.082* 14.282* 12.941*
Medium
TAPs 2
8.568* 14.950* 18.749* 8.886* 7.668*
Medium
TAPs 3
7.380* 9.505* 6.731* 5.055* 4.161*
Medium
TAPs 4
4.988* 4.700* 3.963* 2868* 2.397*
Medium
TAPs 5
3.798* 3.532* 3.090* 1.776* 1.701*
Gail Simon
Test
0.9375 0.9375 0.9375 0.9375 0.9375
HTE
Qualitative
HTE

Superior
Inferior
High TAPs 1 2.475* 2.446* 1.950* 1.058* 0.375*
High TAPs 2 2.472* 2.473* 2.161* 1.068* 0.310*
High TAPs 3 2.618* 2.184* 2.083* 1.235* 0.359*
High TAPs 4 2.404* 2.123* 1.771* 1.134* 0.379*
High TAPs 5 2.373* 2.067* 1.948* 1.190* 0.467*
Gail Simon
Test
0.9375 0.9375 0.9375 0.9375 0.9375
HTE
Qualitative
HTE

Superior
Inferior

94
As anticipated, in the majority of our comparisons the trends are as expected and the log
odds ratio decreases from quintiles 1 to 5. However, in the comparison of risperidone as
the CT versus olanzapine the trend is in the opposite direction. This can be seen in figure
5.10.

Figure 5.10: Adjusted Treatment Effect Utilizing Risperidone as the CT
Risperidone vs Olanzapine

As can be seen from figure 5.10 the log odds ratio actually increases from quintiles 1 to
5. In addition, the log odds ratio is significantly greater than one in quintile 5. Thus, the
patients expected to have the highest probability of success on risperidone (quintile 5)
would actually benefit more from olanzapine. A similar trend is seen in our comparison
of quetiapine as the CT versus olanzapine (i.e. log odds ratio increases from quintiles 1 to
5). However, the log odds ratio is never found to be significantly different than one in
any quintile.
0 0.5 1 1.5 2
Ris Quintile 1
Ris Quintile 2
Ris Quintile 3
Ris Quintile 4
Ris Quintile 5
Log Odds Ratio
Risperidone Better
Olanzapine Better
95
Similar to our olanzapine example, there are several cases in which the log odds ratio is
less than one in all 5 quintiles and the CT is therefore found to be superior. In contrast,
we find several cases in which the log odds ratio is greater than one in every quintile and
therefore the CT is inferior. This can be seen in our utilization of low potency TAPs as
the CT versus risperidone (see Figure 5.11).

Figure 5.11: Adjusted Treatment Effect Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Risperidone

Now that we fully understand the major trends seen in our comparisons, we shift our
focus to confirming our a priori hypothesis that heterogeneous treatment effects are
present. To accomplish this we further examine the data displayed in our forest plots. In
over half of our analyses a single line could not be drawn (perpendicular to the x-axis)
that would intersect the estimated treatment effect for each quintile. Thus are hypothesis
0 1 2 3 4
Low Quintile 1
Low Quintile 2
Low Quintile 3
Low Quintile 4
Low Quintile 5
Log Odds Ratio
Low TAPs Better Risperidone Better
96
is confirmed, heterogeneous treatment effects are present in the majority of cases.
However, although we were successful in identifying HTE, we were unable to identify
any cases in which we could validate that qualitative HTE was present.

5.5 Tailoring Variables

Had we been able to identify qualitative heterogeneous treatment effects, our next step
would be to analyze the subpopulation in which the CT is superior, and attempt to
identify unique patient characteristics (i.e. tailoring variables). However, since our
results failed to identify qualitative heterogeneity there is no need for further analysis.
Identification of tailoring variables is not viable in this instance.

5.6 Sensitivity Analyses

Although our model adequately controls for heterogeneous treatment effects, there was
some question as to whether it adequately controls for treatment selection bias. Since we
are using a retrospective claims database for our analyses, we suspect that treatment
selection bias is present. In order to ensure that we are controlling for possible bias, we
will test the robustness of our study results by implementing the concurrent use of a
traditional propensity score in conjunction with our PPS. We will compute the traditional
PS in two different ways. First we compute the PS across the entire population prior to
stratifying. Next, we compute the PS within each quintile after stratification has already
97
occurred. We then compare the results of these two more traditional approaches to our
initial PPS results in order to confirm the robustness of our study.
5.6.1 Separate Model with Propensity Score Calculated Prior to Stratifying

In this analysis we use a propensity score calculated prior to stratification in order to
adjust for possible observable treatment selection bias. To begin we create a data set
containing only those episodes initiated with either the CT or the selected alternative
treatment (i.e. Alternative 1, or Alternative 2, or Alternative 3, etc). We then use a
logistic regression to calculate the probability of receiving the CT (i.e. the CT PS). Next,
we stratify the patients into quintiles based upon the cutpoints previously determined for
the CT and verify that the overlap and support condition of the PS are met. If the support
and balancing condition are satisfied we then compute the adjusted treatment effect using
a logistic regression within each quintile. Within the logistic regression we account for
the CT PS, a dummy variable for the alternative treatment, and all other covariates of
interest. The results utilizing olanzapine as the CT and risperidone as the selected
alternative treatment are displayed in Figure 5.12.

98
Figure 5.12: Adjusted Treatment Effect with PS (Prior) Utilizing Olanzapine as the
CT
Olanzapine vs Risperidone

As can be seen from this figure, patients in the 4
th
and 5
th
quintiles have a significantly
greater probability of achieving success on olanzapine, but, there is no significant
difference between olanzapine and risperidone in quintile 1 through 3. When comparing
the results from this figure to our initial results presented in Figure 5.8 the results are
remarkably similar. Thus, our original model comparing risperidone to olanzapine
appears to adequately control for observable treatment selection bias.

We repeat this procedure utilizing each treatment as the CT, comparing it to each
alternative treatment. The results of these analyses can be seen in the Appendix in
Figures A.22 to A.27. The results are consistent with our findings noted above when
0 0.5 1 1.5 2
Ola Quintile 1
Ola Quintile 2
Ola Quintile 3
Ola Quintile 4
Ola Quintile 5
Log Odds Ratio
Olanzapine Better Risperidone Better
99
comparing risperidone to olanzapine; the treatment effect models which include the PS
closely resemble our initial results.
5.6.2 Separate Model with Propensity Score Calculated After Stratifying

To further confirm the results obtained in our first sensitivity analysis we recalculated our
models using a PS calculated after stratification. We follow the steps of the original PPS
method up to the point where the CT and the alternative treatments have been stratified
into quintiles based upon the CT PPS. Once patients have been stratified we create a data
set within each quintile containing only those episodes initiated with either the CT or the
selected alternative treatment (i.e. Alternative 1, or Alternative 2, or Alternative 3, etc).
The process from this point forward mimics the process used to compute the treatment
effect in the first sensitivity analysis. We calculate the probability of receiving the CT
within each quintile and verify that the overlap and balancing condition for the PS are
satisfied. Once these conditions are met we estimated the treatment effect within each
quintile including the CT PS, a dummy variable for the alternative treatment, and all
other covariates of interest in our model. The results utilizing olanzapine as the CT and
risperidone as the alternative treatment are displayed in Figure 5.13.

100
Figure 5.13: Adjusted Treatment Effect with PS (Post) Utilizing Olanzapine as the
CT
Olanzapine vs Risperidone

The results of this analysis are similar to the results achieved using the PS calculated
prior to stratification (i.e. our first sensitivity analysis), and are thus similar to the results
generated utilizing our initial PPS model. When we repeat this analysis considering each
treatment as the CT, our results are once again in agreement with our initial PPS models
(see Figure A.28 to A.33 in the Appendix). Thus, our sensitivity analysis provides clear
indication that the initial results of this study adequately control for observable treatment
selection bias.
0 0.5 1 1.5 2
Ola Quintile 1
Ola Quintile 2
Ola Quintile 3
Ola Quintile 4
Ola Quintile 5
Log Odds Ratio
Olanzapine Better Risperidone Better
101
CHAPTER 6: CONCLUSION & DISCUSSION
6.1 Summary

The foundation of this dissertation is built upon the belief that treatment effects are often
heterogeneous. Thus, different patients experience different outcomes on the same
medication. If this phenomenon exists, then as we have explained, current clinical
evidence may not be appropriate. The average treatment effect may not adequately
reflect the benefit achieved by any one individual patient. Thus basing clinical decisions
on such data could have implications on the patients’ well-being, the cost of healthcare to
society, and the availability of medications in the marketplace. This is especially true
when heterogeneous treatment effects are qualitative.

Hence, the PPS method was developed with three goals in mind; (1) To identify if
heterogeneous treatment effects are present, (2) To identify if heterogeneous treatment
effects are quantitative or qualitative, and (3) To identify tailoring variables when
qualitative heterogeneous treatment effects are present. Accomplishing these three goals
will aid us in addressing our ultimate objective, which is to provide physicians and other
decision makers with evidence that will allow them to treat patients more efficiently.

102
6.2 Prognostic Propensity Score

Calculating the PPS is one of the most important steps in our analysis. Optimally the
range of PPS values should reflect both low and high probabilities of success on the CT.
This would be represented as a wide region of common support, with values ranging
from zero to one. A wide region of common support lends itself to a more heterogeneous
population throughout, but a more homogeneous population within each quintile.
Contrary to our expectations the region of common support in our analysis of
antipsychotics for the treatment of schizophrenia was rather narrow. As can be seen in
our results in Chapter 5 the region of common support contained values no greater than
0.55 for any treatment. This is equivalent to stating that the highest probability of
success experienced by any patient on the CT in our analysis was 55%. Obviously this
was much lower than we had anticipated. However, this small region of common support
may not be entirely unrealistic given our population of schizophrenia patients and our
chosen definition of success.

As is typical to studies in regards to schizophrenia, our definition of success was the time
to all cause discontinuation (TTAD) greater than one year. Although this definition of
success is based upon the schizophrenia prescribing guidelines, we find that it may not be
realistic for our population. We are analyzing a Medi-Cal population in which the
majority of cases are considered severe. In severe cases of schizophrenia it is less likely
that a patient would remain on treatment for the full year. Thus our ability to predict a
high probability of success on the CT may be hindered. Theoretically if we were able to
103
analyze a broader population, including privately insured patients with less severe forms
of schizophrenia, our region of common support may have been wider, and extended past
the 0.55 threshold of success noted in Chapter 5.

Aside from re-evaluating our definition of success it also necessary to evaluate how
effective our method was in calculating an accurate probability of success (i.e. the PPS).
Our region of common support could have been directly impacted by the covariates we
were able to include within our model. In the literature there are many different theories
in regards to the predictors of success in patients diagnosed with schizophrenia. Many of
these characteristics or covariates are somewhat subjective in nature and are therefore not
always captured in claims data. According to Lacro JP et al, 7 factors were consistently
associated with nonadherence of antipyschotics for the treatment of schizophrenia. These
included poor insight, negative attitude or subjective response toward medication,
previous nonadherence, substance abuse, shorter illness duration, inadequate discharge
planning or aftercare, and poorer therapeutic alliance. Thus, we would suspect that the
opposite of these factors could be predictors of success (i.e. good insight, positive
attitude, etc). When reviewing our Medi-Cal claims data, it is evident that only 1 of these
7 possible predictors are represented. Therefore it is possible that the lack of adequate
predictors within our data set contributed to the narrow region of common support
witnessed within our study.

104
6.3 Identify if Heterogeneous Treatment Effects are Present

As noted in Chapter 5 to identify HTE we used a two-step process. We first analyzed the
general trends utilizing a crude treatment effect method. We then refined our analysis
through the use of adjusted treatment effect models.
6.3.1 Crude Treatment Effect

When we analyzed our crude treatment effect results, we hoped to confirm several
expectations. The first of which was the existence of HTE, evident when the CT and
alternative treatment trend lines are not parallel. The second expectation was that the CT
would have the highest probability of success in quintile 5 and lowest probability of
success in quintile 1.

When we examined the trend lines across our crude analyses in Chapter 5, we were given
early indication that HTE was present. This trend was most prevalent when AAPs were
utilized as the CT. We did not expect this trend to be more obvious in AAPs, however
the observed trend may be justified. According to the results in Chapter 5 the region of
common support was wider when AAPs were utilized as the CT. Thus, it is likely that
the patients within each quintile represented a more homogenous population. This
served to emphasize the differences between quintiles, and thus resulted in the
observation of HTE.

105
When evaluating the trends in quintiles 5 and 1 our results varied according to the CT.
When olanzapine was utilized as the CT our results very closely matched our
expectations. Olanzapine patients in quintile 5 had the longest TTAD. When compared
to other AAPs, olanzapine had the shortest TTAD in quintile 1. However, regardless of
the CT utilized the TAPs were consistently inferior, and proved to maintain the lowest
TTAD across all quintiles. We did not foresee this outcome, however when re-
examining the treatment patterns associated with the use of TAPs in schizophrenia this
outcome seems plausible. TAPs are known to have significant extrapyramidal side
effects. For this reason, they are routinely discontinued by patients and therefore may not
have satisfied our specified criteria for success (TTAD>360).

Contrary to expectations olanzapine dominated all other treatments in quintile 5
regardless of the treatment utilized as the CT. Obviously, this was entirely unexpected.
Based on our findings noted above, we are able to justify why olanzapine would prove
superior to TAPs, however it is harder to explain why olanzapine would prove superior to
all other AAPs. We suspect that this is due to the fact that olanzapine was proven to have
the widest region of common support, in which the highest probability of success was
achieved. Therefore the results of our quintile 5 analysis may have been skewed towards
olanzapine since the alternative treatments never reached the same level of success, and
as a result were under represented in quintile 5. As we have previously noted, our ability
to adequately identify a wide region of common support, with a high probability of
success for all treatments, may have been hindered by our study sample.
106
6.3.2 Adjusted Treatment Effect

Although crude treatment effect methods are helpful in gaining insight into overall
trends, they do not account for the differences in covariates across treatments. Therefore
it’s necessary to utilize an adjusted treatment effect model to gain a more accurate
assessment of HTE.

We expected our adjusted treatment models to further clarify the presence of HTE. As
noted in Charpter 5, HTE is evident when we are unable to draw a vertical line within the
forest plot (perpendicular to the x-axis) that intersects with the estimates of the treatment
effect for each quintile. Furthermore we would expect the log odds ratio to decrease from
quintiles 1 to 5. Similar to our expectations of the crude method, this is evidence that the
CT is superior to alternate treatments in quintile 5 and inferior to alternative treatments in
quintile 1.

Comparable to our crude analysis, HTE was evident in the majority of models which
utilized AAPs as the CT. In addition, the adjusted model validated the existence of HTE
in models which utilized TAPs as the CT. These results are in-line with the benefit that
we would expect to see when implementing an adjusted model. As noted, the AAPs had
a wider region of common support, thus the adjusted model confirmed the indication of
HTE evident in the crude method. Within the TAPs models the region of common
support was much more narrow and therefore the population within each quintile more
heterogeneous. Thus the estimates of the crude model may have been biased. The
107
adjusted model properly compensated for the differences within the quintiles, and
validated that HTE was indeed present within the TAPs models.

When analyzing the forest plots, the majority of models which demonstrated HTE
showed a general trend where the log odds ratio decreased from quintiles 1 to 5. However
in a few cases the logs odds ratio actually increased from quintiles 1 to 5. Thus, this
demonstrated that in the majority of cases the effect of the CT was greatest in those
patients in which their expected probability of success (PPS) was determined to be the
highest. In the cases where we saw an opposing trend, we assume this to be a direct
impact of the covariates captured within our study. If the covariates/predictors of success
for the two drugs compared were too similar, it may not have been possible to
differentiate one from the other in regards to their probability of success. Thus these
opposing trend lines were witnessed.

6.4 Identify if Heterogeneous Treatment Effects are Quantitative or Qualitative

Although we’ve been able to identify HTE, and the majority of our trend lines are in-line
with our expectations, the ability to determine qualitative versus quantitative HTE is
dependant upon an achieved level of significance.

If we were to only evaluate HTE based upon our crude models, we would suspect that
qualitative HTE existed. This is evident by the intersection of the CT trend line with
those of alternative treatments. However, as we noted above, the crude method does not
108
account for the differences in the distribution of covariates across treatment groups.
Therefore, we must depend upon the results of our adjusted models to more accurately
determine the existence of qualitative HTE. If the adjusted model provides us with a
treatment effect in quintile 1 significantly greater than 1, and a treatment effect
significantly less than 1 in quintile 5, then we can assume that qualitative heterogeneity is
present.

When evaluating qualitative HTE for the population within our study, we were unable to
achieve an adequate level of significance to indicate the presence of qualitative HTE. In
a number of our models which displayed HTE, we were able to identify a treatment effect
greater than 1 in quintile 1 and less than 1 in quintile 5. However, the necessary level of
significance was simply not evident. Similar to some of the contrary results noted above,
we believe that this can be attributed to a region of common support that was simply too
narrow and did not represent patients with a high enough probability of success. Given a
population with a higher threshold of success, we would have been able to better
differentiate patients and thus provide ourselves a greater opportunity to validate the
existence of qualitative HTE.

6.5 Identify Tailoring Variables

The concept of tailoring variables was developed based upon the prerequisite that
qualitative heterogeneity exists within the measured treatment effects. Since, qualitative
HTE was not evident in our analysis we did not have the opportunity to explore the use of
109
the PPS to identify tailoring variables. However we feel confident that under different
circumstances the PPS method has the potential to be very effective in identifying these
key patient characteristics.

6.6 Implications of the Study Results

Based on the results of our study it is clear that there are several factors which must be
considered in order for the PPS model to be effective. We have touched on each of these
points throughout the discussion of our results. To reiterate these points, the most
important factor is the presence of a wide region of common support. This supports the
need for an adequate representation of both high and low probabilities of success for each
treatment being compared. This is our only means to ensure that a homogeneous
population exists within quintiles while a heterogeneous population is maintained across
the analysis. The second point to consider is the benefit of comparing treatments that
share a similar range of PPS values. This would serve to eliminate the bias within
quintile 5, which can occur when the maximum probability of success achieved by one
treatment is significantly greater than another. This would also seem the most
appropriate population for which to provide physicians with tailoring variables, since the
effects of each individual treatment within these populations would typically be
indistinguishable.
110
6.7 Contributions to the Field

As the use of prescription drugs continues to rise so to will the number of adverse events
associated with their use and thus the overall cost of healthcare. It is therefore becoming
only more imperative that healthcare providers choose the most efficient medication for
their patients. This is going to require researchers to move beyond reporting an average
treatment effect. There is going to be an increased need and desire for researchers to
provide evidence as to which patients will receive the greatest benefit for a given
treatment. The PPS method is one of the first methods created to address this complex
and difficult issue. We believe that the results are promising and that with further
research we will be able to identify tailoring variables using such a method.

6.8 Next Steps

The next logical step is to further test the use of the PPS method to identify qualitative
HTE and tailoring variables. We propose to use a more heterogeneous population of
patients and explore the use of this method in other disease states. In addition, we hope
to examine the financial impacts of efficient healthcare by further exploring the
implications of quantitative and qualitative HTE through cost effectiveness analyses.

111
REFERENCES

Abadie A (2005). Semiparametric Difference-in-Differenece Estimators. Review of
Economic Studies. 72(1):1-19.

Ai C, and Norton EC (2003). Interaction Terms in Logit and Probit Models.
Econometric Letters. 80(1):123-129.

American Psychiatric Association. (1994). Diagnostic and statistical manual of mental
disorders (Revised 4th ed.). Washington, DC.

American Psychiatric Association (1997). Practice guideline for the treatment of patients
with schizophrenia. American Journal of Psychiatry. 154 (suppl 4): 1–63.

Angrist J, Imbens G, and Rubin D (1996). Identification of Causal Effects Using
Instrumental Variables. Journal of the American Statistical Association. 91: 444-472.

Ashenfelter O, and Card D (1985). Estimating the Effect of Training Programs on
Earnings. Review of Economics and Statistics. 60:47-57.

Athey S and Imbens G (2006). Identification and Inference in Nonlinear Difference-in-
Difference Models. Econometrica. 74(2).

Azzalini A, and Cox DR (1984). Two New Tests Associated with Analysis of Variance.
Journal of the Royal Statistical Society, Series B. 46:335-343.

Bailey KR (1994). Generalizing the Results of Randomized Clinical Trials. Controlled
Clinical Trials. 15:15-23.

Beasley CM Jr, Stauffer VL, Liu-Seifert H, Taylor CC, Dunayevich E, and Davis JM
(2007). All-cause treatment discontinuation in schizophrenia during treatment with
olanzapine relative to other antipsychotics: an integrated analysis. Journal of
Psychopharmacology. 27(3):252-258.

Black N, (1996). Why We Need Observational Studies to Evaluate the Effectiveness of
Health Care. British Medical Journal. 312:1215-1218.

Carnes M, Howell T, Rosenberg M, Francis J, Hildebrand C, and Knuppel J (2007).
Physicians Vary in Approaches to the Clinical Management of Delirium. Journal of the
American Geriatrics Society. 52(2):234-239.

Charatan F (2002). US Report Blames Poor Quality Control for Soaring Healthcare
Costs. British Medical Journal. 324(7352):1478.

112
Ciminera JL, Heyse JF, Nguyen HH, and Tukey JW (1993). Tests for qualitative
treatment-by-centre interaction using a 'pushback' procedure. Statistics in Medicine. 12:
1033-1045.

Cochrane WG (1968). The Effectiveness of Adjustment by Subclassification in
Removing Bias in Observational Studies. Biometrics. 24:295-313.

Collins LM, Murphy SA, and Bierman KL (2004). A Conceptual Framework for
Adaptive Preventive Interventions. Prevention Science. 5(3):185-196.

Cook FE, Goldman L (1989). Performance of Tests of Significance Based on
Stratification by a Multivariate Confounder Score or by a Propensity Score. Journal of
Clinical Epidemiology. 42(4):317-24.

Cox ER, Motheral BR, Henderson RR, and Mager D (2003). Geographic Variation in the
Prevalence of Stimulant Medication Use Among Children 5 to 14 Years Old: Results
From a commercially Insured US Sample. Pediatrics. 11:237-243.

DiPiro JT, Talbert RL, Yee GC, Matzke GR, Wells BG, Posey LM,et al. (eds)
Pharmacotherapy. A Pathophysiologic Approach. Stamford: Appleton & Lange; 1999.

Egan BM, Basile J, Chilton RJ, and Cohen JD (2005). Cardioprotection: The Role of β-
Blocker Therapy. Journal of Clinical Hypertension. 7:409-416.

Ereshefsky L (1997). Pharmacokinetics and Drug Interactions: Update for New
Antipsychotics. Journal of Clinical Psychiatry. 57 (Suppl 11):12-25.

Fisher RA. (1925). The Design of Experiments, 1st ed, Oliver and Boyd, London.

Fung M, Thornton A, Mybeck K, Wu, JH, Hornbuckle K, and Muniz E (2001).
Evaluation of the Characteristics of Safety Withdrawal of Prescription Drugs from
Worldwide Pharmaceutical Markets – 1960 to 1999. Drug Information Journal. 35:293-
317.

Gail M, and Simon R, (1985). Testing for Qualitative Interactions Between Treatment
Effects and Patient Subsets. Biometrics 41:361-372.

Ghosh AK, Ghosh K (2005). Translating Evidence-Based Information into Effective
Risk Communication: Current Challenges and Opportunities. Journal of Laboratory and
Clinical Medicine. 145: 171-180.

Gilmer TP, Dolder CR, Lacro JP, Folsom DP, Lindamer L, et al (2004). Adherence to
Treatment With Antipsychotic Medication and Health Care Costs Among Medicaid
Beneficiaries With Schizophrenia." American Journal of Psychiatry. 161: 692-699.

113
Goff DC, Heckers S, and Freudenreich O (2001). Schizophrenia. Medical Clinics of
North America. 85(3):663-689.

Graham J, and Hu J (2007). The Risk-Benefit Balance in the United States: Who
Decides? Health Affairs. 26(3):625-635.

Harris KM, Remler DK (1998). Who is the Marginal Patient? Understanding
Instrumental Variables Estimates of Treatment Effects. Health Services Research.
33(5): 1337-1360.

Hayward RA, Kent DM, Vijan S, and Hofer TP (2005). Reporting Clinical Trail Results
to Inform Providers and Consumers. Health Affairs. 24(6):1571-1581.

Heckman J (1978). Dummy Endogenous Variables in a Simultaneous Equation System.
Econometrica .46: 931-960.

Heckman, J (1979). Sample Bias As A Specification Error. Econometrica, 47(1):153-
162.

Heckman J (1997a). Instrumental Variables: A Study of Implicit Behavioral
Assumptions Used in Making Program Evaluations. Journal of Human Resources.
32:441-462.

Heckman J, Smith J and Clements N (1997b). Making the Most Out of Programme
Evaluations and Social Experiments: Accounting for Heterogeneity in Programme
Impacts. Review of Economic Studies. 64: 487-535.

Heckman J, Urzua S, and Vytlacil E (2006). Understanding Instrumental Variables in
Models with Essential Heterogeneity. Review of Economics and Statistics. 88(3): 389-
432.

Heckman J and Vytlacil E (2005). Structural Equations, Treatment Effects, and
Econometric Policy Evaluation. Econometrica. 73(3):669-738.

Hoadley J (2005). Cost Containment Strategies For Prescription Drugs: Assessing The
Evidence In The Literature. Kaiser Family Foundation.
Http://www.kff.org/rxdrugs/upload/Cost-Containment-Strategies-for-Precription-Drugs-
Assessing-The-Evidence-in-the-Literature-Report.pdf. Accessed August 10, 2007.

Holland PW (1986). Statistics and Causal Inference. Journal of American Statistical
Association. 81(396):945-960.

Hollon SD, Jarrett RB, Nierenberg AA, Thase ME, Trivedi M, Rush AJ (2005).
Psychotherapy and medication in the treatment of adult and geriatric depression: which
monotherapy or combined treatment? J Clin Psychiatry. 66:455–468.
114

Horowitz RI, Singer BH, Makuch RW, and Viscoli CM (1996). Can Treatment That is
Helpful on Average Be Harmful to Some Patients? A Study of the Conflicting
Information Needs of Clinical Inquiry and Drug Regulation. Journal of Clinical
Epidemiology. 49(4):395-400.

Imbens G (2000). The Role of the Propensity Score in Estimating Dose-Response
Functions. Biometrika. 87(5):706-710.

Imbens G. and Angrist J (1994). Identification and Estimation of Local Average
Treatment Effects. Econometrica. 61(2): 467-476.

Johnson JA, (1997). Influence of Race or Ethnicity on Pharmacokinetics of Drugs.
Journal of Pharmaceutical Sciences. 86(12):1328-1333.

Kaiser Family Foundation (May 2007), Prescription Drug Trends, a factsheet,
http://www.kff.org/rxdrugs/upload/3057_06.pdf. Accessed August 8. 2007.

Keith SJ, Regier DA, and Rae DS. Schizophrenic Disorders, in: Psychiatric Disorders in
America: The Epidemiological Catchment Area Study. Edited by Robins LN, Regier
DA. New York, Toronto: The Free Press, 1991.

Kravitz RL, Naihu D, and Braslow J (2004). Evidence-Based Medicine, Heterogeneity
of Treatment Effects, and the Trouble with Averages. The Millbank Quarterly.
82(4):661-687.

Lacro JP, Dunn LB, Dolder CR, et al (2002). Prevalence of and Risk Factors for
Medication Nonadherence in Patients with Schizohrenia: A Comprehensive Review of
Recent Literature. Journal of Clinical Psychiatry. 63:892-909/

Lauzarou J, Pomeranz BH, Corey PN (1998). Incidence of Adverse Drug Reactions
Hospitalized Patients, A Meta-analysis of Prospective Studies. Journal of American
Medical Association. 279:1200-1205.

Lechner M (2002a). Program Heterogeneity and Propensity Score Matching: An
Application to the Evaluation of Active Labor Market Policies. Review Economics and
Statistics. 84(2):205-220.

Lechner M (2002b). Some Practical Issues in the Evaluation of Heterogenuous Labour
Market Programmes by Matching Methods. Journal of Royal Statistical Society, Series
A. 165:659-82.

Lehman AF, Kreyenbuhl J, Buchanan RW, Dickerson FB, Dixon LB, Goldberg R,
Green-Paden LD, Tenhula WN, Boerescu D, Tek C, Sandson N, and Steinwachs DM
(2004). Patterns of Usual Care for Schizophrenia: Initial Results From the Schizophrenia
115
Patient Outcomes Research Team (PORT) Client Survey. Schizophrenia Bulletin.
30(2):193-217.

Linjakumpu T, Hartikainen S, Klaukka T

, Koponen H, Kivelä SL

, Isoaho R (2002).
Psychotropics among the home-dwelling elderly - increasing trends. International
Journal of Geriatric Psychiatry. 17(9):874-883.

Longford NT, and Nelder JA (1999). Statistics Versus Statistical Science in the
Regulatory Process. Statistics in Medicine. 18(17-18): 2311-2320.

Loranger AW (1984). Sex Difference in Age at Onset of Schizophrenia. Archives of
General Psychiatry. 41(2):157-161.
Manski C (1990). Nonparametric Bounds on Treatment Effects. American Economic
Review Papers and Proceedings. 80:319-323.

Manski C (2001). Designing Programs for Heterogeneous Populations: The Value of
Covariate Information. American Economic Review Papers and Proceedings. 91:103-
106.

Manski C (2002). Treatment Choice Under Ambiguity Induced by Inferential Problems.
Journal of Statistical Planning and Inference. 105:67-82.

Manski C (2004). Statistical Treatment Rules for Heterogeneous Populations.
Econometrica. 72(4):1221-1246.

Marcus AD (2004). A Patient’s Quest to Save New Drug Hits Market Reality. The Wall
Street Journal. November 16.

McCombs JS, Nichol MB, Stimmel GL (1999). Use Patterns for Antipsychotic
Medications in Medicaid Patients with Schizophrenia. Journal of Clinical Psychiatry.
60(Suppl 19):5-11.

McCombs JS, Luo M, Johnstone BM, and Shi L (2000a). The Use of Conventional
Antipsychotic Medications for Patients with Schizophrenia in a Medicaid Population:
Therapeutic and Cost Outcomes Over 2 Years. Value in Health. 3(3):222-231.

McCombs JS, Nichol MB, Johnstone BM (2000b) Psychotic Drug Use Patterns and the
Cost of Treating Schizophrenia. Psychiatric Services. 51(4):525-527.

McCombs JS, Shi L, Croghan TW, and Stimmel, GL (2003). Access to Drug Therapy
and Substitutions Between Alternative Antidepressants Following an Expansion of the
California Medicaid Formulary. Health Policy. 65(3):301-311.

116
McKee M, Britton A, Black N, McPherson K, Sanderson C, and Bain C (1999). Methods
in Health Services Research: Interpreting the Evidence: Choosing Between Randomised
and Non-Randomized Studies. British Medical Journal. 319:312-315.

Merriam-Webster Dictionary. Available at: http://www.merriam-webster.com
/dictionary/prognosis. Accessed August 6, 2007.

Miettinen OS (1976). Stratification by a Multivariate Confounder Score. American
Journal of Epidemiology. 104:609-620.

Motheral BR, Cox ER, Mager D, Henderson R, and Martinez R (2001). Express Scripts
Prescription Drug Atlas, 2001: A Study of Geographic Variation in the Use of
Prescription Drugs, Express Scripts, Inc., January 2002.

Narrow WE, Rae DS, Robins LE, and Regier DA (2002). Revised Prevalence Estimates
of Mental Disorders in the United States. Archives of General Psychiatry. 59:115-123.

Neyman, J (1923). On the Application of Probability Theory to Agricultural Experiments.
Essay on Principles. Section 9, translated in Statistical Science, (with discussion). 5(4):
465–480, 1990.

Norquist GS, Regier DA, and Rupp A (1996). Estimates of the cost of treating people
with schizophrenia: contributions of data from epidemiologic surveys, in Handbook of
Mental Health Economics and Health Policy: Schizophrenia. Vol 1. Edited by
Moscarelli M, Rupp A, Sartorius N. New York, Wiley.

Pagano M, Gauvreau K. Principles of Biostatistics. Duxbury Press; 2 edition (March 9,
2000)

Peto R (1982). Statistical Aspects of Cancer Trials. Treatment of Cancer. K. E. Halnan
(ed.), 867-871. London: Chapman and Hall.

Pike MC, Anderson J, Day N (1979). Some insight into Miettinen’s Multivariate
Confounder Score Approach to Case-Control Study Analysis. Epidemiology in
Community Health. 33:104-106.

Popcock SJ, Assman SE, Enos LE, and Kasten LE (2002). Subgroup Analysis, Covariate
Adjustment and Baseline Comparisons in Clinical Trial Reporting: Current Practice and
Problems. Statistics in Medicine. 21(19):2917-2930.

Quandt R (1972). A New Approach to Estimating Switching Regressions. Journal of the
American Statistical Association. 67: 306-310.

Rosenbaum PR (1995). Observational Studies. Springer Verlag, New York.

117
Rosenbaum PR, and Rubin DB (1983). The Central Role of the Propensity Score I
Observational Studies for Causal Effects. Biometrika. 70(1):41-55.

Rosenbaum PR, Rubin DB, (1984). Reducing Bias in Observational Studies Using
Subclassification on the Propensity Score. Journal of American Statistical Association.
79(387):516-524.

Rothwell PM (1995). Can Overall Results of Clinical Trials Be Applied to All Patietns?
Lancet. 345:1616-1619.

Rothwell PM (2005). External Validity of Randomized Controlled Trials: “To Whom Do
the Results of this Trial Apply?” The Lancet. 365(9453):13-14.

Rubin DB (1973a). Matching to Remove Bias in Observational Studies. Biometrics. 29:
159-183.

Rubin DB (1973b). The Use of Matched Sampling and Regression Adjustments to
Remove Bias in Observational Studies. Biometrics. 29:185-203.

Rubin DB (1974). Estimating Causal Effects of Treatments in Randomized and
Nonrandomized Studies. Journal of Educational Psychology. 66(5):688-701.

Rubin DB (1977). Assignment to Treatment Group on the Basis of a Covariate. Journal
of Educational Statistics.2(1):1-26.

Rubin, D. B., (1978). Bayesian inference for causal effects: The Role of Randomization.
Annals of Statistics. 6:34–58.

Schmoor C, Olschewski M, and Schumacher M (1996). Randomized and Non-
Randomized Patients in Clinical Trails: Experiences with Comprehensive Cohort Studies.
Statistics in Medicine. 15(3):236-271.

Schwartz JS (1984). The Role of Professional Medical Societies in Reducing Practice
Variations. Health Affairs. 3(2):90-101.

Shepherd J, Cobbe SM, Ford I, Isles CG, Lorimer AR, Macfarlane PW, McKillop JH,
Packard CJ (1995). Prevention of Coronary Heart Disease with Pravastatin in Men with
Hypercholesterolemia. New England Journal of Medicine. 333:1301-1308.

Soumerai SB (2004). Benefits and Risks of Increasing Restriction on Access to Costly
Drugs in Medicaid. Health Affairs. 23(1):135-146.

Starfield B (2000). Is US Health Really the Best in the World? Journal of the American
Medical Association. 284:483-485.

118
Steg PG, López-Sendón J, Lopez de Sa R, Goodman SG, Gore JM, Anderson FA,
Himbert D, Allegrone j,Van de Werf F (2007). External Validity of Clinical Trials in
Acute Myocardial Infarction. Archives of Internal Medicine.167(1):68-73

Strauss D (1998). On Miettinen’s Multivariate Confounder Score. Journal of Clinical
Epidemiology. 51(3):233-236.

Stroup TS, McEvoy JP, Swartz MS, Byerly MJ, Glick ID, Simpson GM, Stevens MC,
and Lieberman JA (2003). The National Institute of Mental Health Clinical
Antipsychotic Trials of Intervention Effectiveness (CATIE) Project: Schizophrenia Trial
Design and Protocal Development. Schizophrenia Bulletin. 29(1): 15-31.

Tunis SR, Stryer DB, and Clancey CM (2003). Practical Clinical Trials: Increasing the
Value of Clinical Research for Decision Making in Clinical and Health Policy. Journal
of American Medical Association. 290(12):1624-1632.

VanderKlaauw W (2002). A Regression-Discontinuity Evaluation of the Effect of
Financial Aid Offers on College Enrollment. International Economic Review.
43(4):1249-1287.

Weiden PJ, Olfson M (1995) Cost of relapse in schizophrenia, Schizophrenia Bulletin.
21(3):419–429.

Wennberg J, Wennberg D (2000). Dartmouth Atlas of Health Care in Michigan. Detroit,
MI: Blue Cross, Blue Shield of Michigan and Blue Care Network. Available at:
http://www.bcbsm.com/atlas/overview.shtml. Accessed August 10, 2007.

Zarin DA, Yound JL, and West JC (2005). Challenges to Evidence-Based Medicine: A
Comparison of Patients and Treatments in Randomized Controlled Trials with Patients
and Treatment in a Practice Research Network. Social Psychiatry and Psychiatric
Epidemiology. 40(1):27-35.

National Alliance of Mental Illnesses. Available at:
http://www.nami.org/Template.cfm?Section=By_Illness&Template=/TaggedPage/Tagge
dPageDisplay.cfm&TPLID=54&ContentID=23036 Access August 8, 2007.

Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition (DSM-IV).
Available at: http://allpsych.com/disorders/dsm.html. Accessed on: August 8, 2007.

119
APPENDIX
Table A.1: Baseline Descriptive Statistics
Covariate Olanzapine
(n=38,804)
Risperidone
(n=37,560)
Quetiapine
(n=11,413)
Low
TAPs
(n=26,395)
Medium
TAPs
(n=56,901)
High
TAPs
(n=48,785)
SCHIZOPHRENIA
TYPE

SIMPLE** 0.36 0.36 0.22 0.61 0.13 0.58
DISORG** 0.37 0.31 0.40 0.15 0.03 0.23
CATA** 0.13 0.10 0.15 0.09 0.03 0.12
PARAS** 9.28 8.40 8.75 4.89 1.32 9.04
ACUTE** 0.29 0.31 0.24 0.17 0.07 0.26
LATENT** 0.03 0.06 0.03 0.07 0.01 0.04
RESID** 0.69 0.73 0.52 0.79 0.17 0.90
SAFF** 8.17 7.18 9.58 3.78 1.50 5.23
OTHNOS** 5.80 5.59 4.70 4.94 1.16 6.36
UNSPEC** 74.87 76.96 75.42 84.50 95.58 77.23
AGE
AGE_CAT18** 6.10 6.64 6.62 4.94 2.63 3.79
AGE_CAT25** 14.81 14.63 15.41 13.40 9.62 14.22
AGE_CAT35** 25.92 23.30 27.57 23.41 09.33 24.49
AGE_CAT45** 23.22 20.20 25.76 21.15 20.46 20.26
AGE_CAT55** 12.84 11.33 12.21 13.93 16.74 12.05
AGE_CAT65** 17.11 23.90 12.42 23.16 31.22 25.19
MALE** 46.92 44.21 41.07 45.88 25.83 45.84
RACE/
ETHNICITY

WHITE* * 50.13 52.53 54.51 53.45 46.41 49.94
BLACK* 14.09 14.36 12.91 13.90 16.04 17.15
HISPANIC** 6.19 6.52 7.25 4.53 9.50 4.04
ASIAN** 6.90 4.88 5.81 3.89 3.55 3.65
OTHER** 22.69 21.70 19.53 24.23 24.50 25.22
COMORBIDITIES
ACC** 2.18 2.02 2.87 2.10 2.77 2.05
ANX** 10.03 9.51 12.02 8.85 9.56 8.62
ARRTDX** 4.46 4.78 4.28 4.49 7.26 5.07
BIP** 12.60 10.29 16.12 6.79 2.76 7.47
BLOOD** 6.74 6.38 6.41 6.08 12.54 6.36
CARDIO** 16.73 16.09 17.11 16.84 59.63 18.36
CHILD** 0.47 0.64 0.69 0.72 0.09 0.26
CONGE** 1.70 1.78 1.88 2.28 3.18 1.71
DEMEN** 3.30 5.60 3.15 6.79 1.40 5.53
DEP** 58.24 54.90 66.86 39.23 64.20 40.42
DIGES** 17.42 16.69 19.16 19.14 35.16 17.20
DIABDX** 8.90 9.94 10.25 9.98 18.23 9.81
ENDO** 10.58 10.51 11.49 9.83 16.66 10.36
EPSDX** 0.83 1.04 1.29 0.77 0.47 1.18
GENTI** 17.15 16.60 18.86 17.59 33.83 17.51
GLAUC** 1.01 1.30 1.24 1.33 2.17 1.27
HBP** 14.52 14.40 14.00 14.20 24.79 15.56
HIV** 0.16 0.12 0.16 0.08 0.33 0.12

120
Table A.1: Baseline Descriptive Statistics (Continued)
Covariate Olanzapine
(n=38,804)
Risperidone
(n=37,560)
Quetiapine
(n=11,413)
Low
TAP
(n=26,395)
Med
TAP
(n=56,901)
High
TAP
(n=48,785)
HLIPID** 13.25 11.82 14.59 10.37 16.80 9.70
INFEC** 96.31 97.18 97.35 97.50 97.88 98.09
INJUR** 18.59 18.16 20.68 19.36 27.06 19.04
MANIC** 0.71 0.68 0.98 0.43 0.18 0.50
MUSCL** 28.51 27.22 31.21 29.58 47.60 26.51
NEOPL** 4.35 4.31 4.52 5.04 12.66 4.49
NERV** 30.57 33.03 32.26 35.20 43.87 33.97
OCIRC** 14.20 16.78 13.14 15.36 25.76 17.07
OTHPSY** 4.11 4.80 4.65 3.50 1.85 4.71
PERI** 0.21 0.23 0.23 0.20 0.42 0.24
PERSO** 1.18 1.23 1.47 1.46 0.64 1.44
PSTG** 0.78 0.80 0.90 0.71 1.92 0.89
RESPI** 26.11 25.70 27.76 30.16 43.28 28.14
SABUSE** 6.85 5.23 9.05 4.41 4.36 4.22
SEXD 0.14 0.11 0.11 0.10 0.11 0.09
SKIN** 14.24 15.46 14.15 17.90 18.96 16.87
SUICDX** 3.59 3.37 4.39 2.72 1.34 3.46
CONCOMITANT
DRUG USE

AD** 51.55 48.05 60.57 33.95 55.33 34.32
HYPNT** 5.64 4.79 6.44 2.22 4.51 2.29
/SEIZU** 15.01 15.01 22.41 15.09 14.80 12.56
ARRH** 7.52 8.50 7.46 7.06 11.97 7.51
APARK ** 1.45 1.57 2.07 1.13 1.15 1.21
ANXIE** 4.75 4.31 5.78 4.22 6.51 3.48
MOOD** 20.83 19.13 24.52 13.09 4.76 15.22
LIPID ** 9.22 8.20 10.21 7.05 9.22 6.35
EPS1** 22.64 25.41 22.19 20.56 9.48 44.74
EPS2* 4.17 3.93 5.20 4.19 5.96 3.34
DIABRX** 7.49 8.70 8.64 8.41 15.95 7.73
DIABDX** 18.93 18.16 19.85 16.78 29.89 15.17
PTHER** 10.44 13.04 7.39 19.72 7.22 21.81
PRIOR
ANTIPSYCHOTIC
DRUG USE

OLA** 72.29 10.60 23.42 4.33 1.84 6.39
RIS** 12.50 71.84 18.81 5.32 1.90 6.72
QUE** 3.33 3.28 66.27 1.28 0.59 1.80
OAP** 2.28 2.14 3.57 1.61 0.36 2.12
TAP** 21.29 21.66 17.87 81.51 72.98 81.11
EPISODE TYPE
RESTART** 59.97 62.19 45.78 75.81 88.65 72.53
SWT** 11.75 10.91 16.96 4.87 1.65 5.79
AUG 9.76 9.54 17.11 10.83 4.25 11.30
LATE_SWT** 18.53 17.36 20.14 8.49 5.45 10.38

121
Table A.1: Baseline Descriptive Statistics (Continued)
Covariate Olanzapine
(n=38,804)
Risperidone
(n=37,560)
Quetiapine
(n=11,413)
Low
TAP
(n=26,395)
Med
TAP
(n=56,901)
High
TAP
(n=48,785)
PRIOR HEALTH
SERVICE
UTILIZATION

LTC** 9.59 16.12 7.86 10.39 7.89 13.74
AHOSP** 4.51 4.61 5.07 4.60 10.77 5.23
PHOSP** 4.16 3.69 5.66 0.82 0.43 1.80
REHAB** 0.29 0.30 0.26 0.31 0.45 0.19
CMHC** 15.62 37.56 61.08 10.40 6.63 14.07
PSYCH** 2.10 2.98 1.25 4.09 1.65 4.21
* p<0.05; **p<0.01
122
Table A.2: The PPS Model Considering Risperidone as the CT
Covariate Point Estimate 95% Wald Confidence Limits
SIMPLE 1.501 1.034 2.180
PARAS 1.134 1.036 1.242
LATENT 1.659 0.653 4.214
OTHNOS 1.251 1.128 1.387
AGE_CAT35 1.087 1.005 1.175
AGE_CAT45 1.246 1.149 1.352
AGE_CAT55 1.378 1.251 1.519
AGE_CAT65 1.074 0.975 1.183
MALE 1.147 1.089 1.209
BLACK 0.692 0.639 0.750
HISP 0.688 0.613 0.773
ASAIN 0.781 0.686 0.890
OTHER_RACE 0.927 0.870 0.987
ARRTDX 0.909 0.803 1.030
BIP 0.839 0.765 0.921
BLOOD 1.073 0.964 1.195
CARDIO 0.757 0.641 0.894
CHILD 1.474 1.101 1.975
DEMEN 1.440 1.301 1.595
DEP 0.770 0.690 0.858
DIADX 0.905 0.811 1.010
ENDO 1.086 0.996 1.184
HBP 1.254 1.053 1.494
HIV 0.421 0.150 1.186
HLIPID 1.070 0.985 1.163
INFEC 1.152 0.981 1.352
INJUR 0.889 0.829 0.954
MUSCL 0.892 0.837 0.950
NERV 1.044 0.987 1.105
OCIRC 0.922 0.855 0.995
PERSO 0.848 0.670 1.073
PSTG 0.755 0.537 1.060
SABUSE 0.664 0.581 0.758
SKIN 1.135 1.057 1.219
SUICDX 0.765 0.655 0.895
AD 1.147 1.028 1.281
HYPNT 0.829 0.729 0.942
SEIZU 1.133 1.057 1.216
ANXIE 0.767 0.666 0.883
MOOD 1.133 1.060 1.211
EPS1 1.331 1.253 1.413
EPS2 0.884 0.772 1.012
DIABRX 0.917 0.818 1.028
DIABDX 0.900 0.840 0.965
PTHER 1.079 0.998 1.168
OLA 0.802 0.732 0.879
RIS 0.909 0.853 0.969
QUE 0.778 0.671 0.902
OAP 1.154 0.984 1.353
TAP 1.150 1.068 1.239
SWT 2.189 1.981 2.419

123
Table A.2: The PPS Model Considering Risperidone as the CT (Continued)
Covariate Point Estimate 95% Wald Confidence Limits
LATE_SWT 1.418 1.312 1.533
AUG 1.714 1.543 1.904
LTC 1.780 1.643 1.929
AHOSP 0.850 0.742 0.974
PHOSP 0.917 0.790 1.064
CMHC 0.917 0.861 0.976
PSYCH 1.300 1.132 1.493

Table A.3: The PPS Model Considering Quetiapine as the CT
Covariate Point Estimate 95% Wald Confidence Limits
DISORG 1.622 0.834 3.157
PARAS 1.210 1.030 1.421
AGE_CAT25 0.902 0.785 1.035
AGE_CAT55 1.297 1.126 1.494
MALE 1.080 0.980 1.191
BLACK 0.689 0.590 0.804
HISP 0.678 0.554 0.831
ASAIN 0.782 0.629 0.972
ACC 0.555 0.393 0.784
ARRTDX 1.228 0.982 1.535
BIP 0.869 0.755 1.000
DEMEN 1.813 1.432 2.295
DEP_PRE 0.909 0.822 1.006
DIADX 0.865 0.734 1.019
GLAUC 0.668 0.410 1.090
HLIPID 1.233 0.985 1.545
MANIC 1.298 0.798 2.113
MUSCL 0.844 0.755 0.942
SABUSE 0.848 0.709 1.015
SKIN 1.191 1.040 1.365
SUICDX 0.728 0.561 0.944
APARK 0.829 0.599 1.147
LIPID 0.853 0.657 1.109
EPS1 1.232 1.100 1.378
QUE 0.864 0.774 0.964
SWT 2.490 2.155 2.876
LATE_SWT 1.627 1.410 1.879
AUG 2.195 1.901 2.535
LTC 1.852 1.577 2.174

124
Table A.4: The PPS Model Considering Low Potency TAPs as the CT
Covariate Point Estimate 95% Wald Confidence Limits
SIMPLE 1.894 1.317 2.722
PARAS 1.091 0.933 1.275
ACUTE 2.248 1.117 4.522
RESID 1.613 1.161 2.241
SAFF 0.883 0.727 1.072
OTHNOS 1.359 1.179 1.567
AGE_CAT35 1.237 1.110 1.378
AGE_CAT45 1.480 1.325 1.653
AGE_CAT55 1.408 1.242 1.596
AGE_CAT65 0.901 0.793 1.022
MALE 1.129 1.052 1.212
BLACK 0.543 0.485 0.609
HISP 0.463 0.375 0.571
ASAIN 0.352 0.274 0.451
OTHER_RACE 0.763 0.702 0.830
BIP 0.662 0.563 0.777
BLOOD 1.207 1.039 1.403
CARDIO 0.923 0.830 1.026
CHILD 1.443 1.018 2.045
CONGE 1.252 1.013 1.547
DEMEN 1.972 1.758 2.213
DEP 0.634 0.531 0.757
DIGES 0.902 0.820 0.993
ENDO 1.078 0.954 1.218
GENTI 0.851 0.767 0.945
GLAUC 0.810 0.571 1.150
INJUR 0.855 0.776 0.941
MUSCL 0.874 0.802 0.952
NEOPL 0.840 0.706 1.000
NERV 1.045 0.970 1.126
OCIRC 0.904 0.812 1.006
PERI 3.051 1.455 6.401
PSTG 0.194 0.077 0.489
RESPI 0.844 0.776 0.917
SABUSE 0.478 0.378 0.604
SKIN 1.456 1.333 1.591
SUICDX 0.703 0.538 0.918
AD 1.282 1.068 1.540
SEIZU 1.244 1.135 1.364
ARRH 0.884 0.764 1.021
ANXIE 0.554 0.440 0.698
MOOD 1.200 1.080 1.334
EPS1 1.326 1.219 1.443
EPS2 0.894 0.741 1.078
DIABDX 0.838 0.757 0.928
OLA 0.745 0.618 0.899
RIS 0.685 0.580 0.809
QUE 0.560 0.382 0.821
OAP 0.431 0.308 0.604
TAP 0.769 0.705 0.839
SWT 1.232 1.055 1.439
125
Table A.4: The PPS Model Considering Low Potency TAPs as the CT (Continued)
Covariate Point Estimate 95% Wald Confidence Limits
LATE_SWT 0.753 0.657 0.864
LTC 1.225 1.090 1.377
AHOSP 0.720 0.584 0.889
PHOSP 0.743 0.460 1.202
CMHC 1.100 0.972 1.243
PSYCH 1.323 1.127 1.552

126
Table A.5: The PPS Model Considering Medium Potency TAPs as the CT
Covariate Point Estimate 95% Wald Confidence Limits
SIMPLE 4.916 2.679 9.019
CATA 3.761 1.040 13.605
PARAS 2.285 1.818 2.871
RESID 4.047 2.355 6.954
SAFF 1.375 1.069 1.770
OTHNOS 1.960 1.514 2.537
AGE_CAT25 1.363 0.933 1.990
AGE_CAT35 1.715 1.194 2.464
AGE_CAT45 2.156 1.502 3.096
AGE_CAT55 2.366 1.638 3.417
AGE_CAT65 1.517 1.051 2.188
MALE 1.147 1.043 1.261
BLACK 0.817 0.723 0.923
HISP 0.420 0.336 0.525
ASAIN 0.744 0.581 0.952
BIP 1.131 0.898 1.425
BLOOD 0.730 0.607 0.877
DEMEN 1.664 1.227 2.256
DEP 2.164 1.814 2.582
DIGES 0.650 0.582 0.725
ENDO 0.760 0.655 0.883
GENTI 0.812 0.727 0.907
GLAUC 0.703 0.478 1.033
HBP 1.121 1.003 1.253
HIV 0.422 0.133 1.337
INJUR 0.782 0.697 0.878
MANIC 1.765 0.839 3.714
MUSCL 0.710 0.644 0.783
NEOPL 0.618 0.515 0.741
NERV 0.892 0.813 0.978
OCIRC 0.823 0.729 0.930
OTHPSY 1.407 1.054 1.879
PSTG 0.387 0.199 0.755
RESPI 0.863 0.784 0.951
SABUSE 0.612 0.477 0.786
SKIN 1.121 0.993 1.266
AD 0.780 0.660 0.921
HYPNT 0.725 0.559 0.941
SEIZU 0.781 0.681 0.896
MOOD 1.162 0.974 1.386
LIPID 0.887 0.768 1.026
EPS1 1.981 1.765 2.223
EPS2 0.741 0.594 0.923
DIABRX 0.886 0.775 1.012
DIABDX 0.709 0.636 0.791
PTHER 1.113 0.962 1.289
TAP 2.010 1.760 2.296
SWT 1.505 1.194 1.897
AUG 0.803 0.663 0.973
CMHC 1.385 1.189 1.612

127
Table A.6: The PPS Model Considering High Potency TAPs as the CT
Covariate Point Estimate 95% Wald Confidence Limits
SIMPLE 1.281 0.945 1.737
DISORG 1.438 0.914 2.264
CATA 2.030 1.114 3.699
PARAS 1.169 1.074 1.272
RESID 1.493 1.189 1.875
OTHNOS 1.280 1.163 1.408
AGE_CAT25 1.493 1.260 1.770
AGE_CAT35 1.671 1.418 1.970
AGE_CAT45 1.979 1.676 2.336
AGE_CAT55 2.136 1.797 2.539
AGE_CAT65 1.541 1.295 1.834
MALE 1.126 1.068 1.187
BLACK 0.688 0.639 0.741
HISP 0.557 0.477 0.649
ASAIN 0.566 0.481 0.666
OTHER_RACE 0.846 0.796 0.900
ACC 1.111 0.915 1.348
BIP 0.865 0.780 0.959
CARDIO 0.890 0.827 0.958
CHILD 1.884 1.230 2.886
DEMEN 1.129 1.009 1.262
DEP 0.814 0.771 0.859
DIGES 0.936 0.870 1.007
DIADX 1.094 0.980 1.221
ENDO 1.091 0.999 1.191
EPSDX 1.414 1.145 1.747
GENTI 0.933 0.866 1.005
HIV 0.539 0.193 1.504
INFEC 1.220 1.010 1.474
INJUR 0.799 0.741 0.860
MUSCL 0.893 0.837 0.953
OCIRC 0.852 0.788 0.922
OTHPSY 0.886 0.778 1.009
PERI 0.519 0.224 1.205
PERSO 0.840 0.664 1.062
PSTG 0.557 0.372 0.833
SABUSE 0.635 0.545 0.740
SEXDE 0.421 0.129 1.372
SKIN 1.233 1.151 1.320
SUICDX 0.872 0.743 1.023
HYPNT 0.649 0.527 0.800
SEIZU 1.108 1.027 1.195
APARK 0.809 0.631 1.037
ANXIE 0.616 0.517 0.734
EPS1 1.442 1.365 1.524
EPS2 0.726 0.620 0.851
DIABRX 0.881 0.780 0.995
OLA 0.807 0.714 0.912
RIS 0.870 0.775 0.977
QUE 0.792 0.640 0.979
OAP 0.816 0.680 0.980
128
Table A.6: The PPS Model Considering High Potency TAPs as the CT (Continued)
Covariate Point Estimate 95% Wald Confidence Limits
TAP 0.911 0.852 0.974
SWT 1.104 0.986 1.235
LATE_SWT 0.877 0.802 0.960
AUG 0.945 0.853 1.047
LTC 1.327 1.220 1.442
AHOSP 0.857 0.749 0.982
PHOSP 0.676 0.539 0.848
CMHC 1.282 1.188 1.383
PSYCH 1.227 1.087 1.385

129
Figure A.1: Histogram of Risperidone PPS by Treatment
0 0. 05 0. 1 0. 15 0. 2 0. 25 0. 3 0. 35 0. 4 0. 45 0. 5 0. 55 0. 6 0. 65 0. 7 0. 75 0. 8 0. 85 0. 9 0. 95 1
0
2000
4000
6000
8000
10000
C
o
u
n
t
ri s_pps

0
5000
10000
15000
20000
C
o
u
n
t
0
5000
10000
15000
20000
C
o
u
n
t
0
5000
10000
15000
20000
C
o
u
n
t
0
5000
10000
15000
20000
C
o
u
n
t
0 0. 05 0. 1 0. 15 0. 2 0. 25 0. 3 0. 35 0. 4 0. 45 0. 5 0. 55 0. 6 0. 65 0. 7 0. 75 0. 8 0. 85 0. 9 0. 95 1
0
5000
10000
15000
20000
C
o
u
n
t
ri s_pps

Figure A.2: Refined Breakdown of Episodes by Treatment Type When Utilizing
Risperidone as the CT
34,929
35,223
10,094
25,636
54,226
47,488
Olanzapine
Risperidone
Quetiapine
Low Potency TAPs
Medium Potency TAPs
High Potency TAPs

130
Figure A.3: Histogram of Quetiapine PPS by Treatment
0 0. 05 0. 1 0. 15 0. 2 0. 25 0. 3 0. 35 0. 4 0. 45 0. 5 0. 55 0. 6 0. 65 0. 7 0. 75 0. 8 0. 85 0. 9 0. 95 1
0
1000
2000
3000
C
o
u
n
t
que_pps

0
5000
10000
15000
20000
25000
C
o
u
n
t
0
5000
10000
15000
20000
25000
C
o
u
n
t
0
5000
10000
15000
20000
25000
C
o
u
n
t
0
5000
10000
15000
20000
25000
C
o
u
n
t
0 0. 05 0. 1 0. 15 0. 2 0. 25 0. 3 0. 35 0. 4 0. 45 0. 5 0. 55 0. 6 0. 65 0. 7 0. 75 0. 8 0. 85 0. 9 0. 95 1
0
5000
10000
15000
20000
25000
C
o
u
n
t
que_pps

Figure A.4: Refined Breakdown of Episodes by Treatment Type When Utilizing
Quetiapine as the CT
36,263
35,427
10,079
24,930
50,078
46,837
Olanzapine
Risperidone
Quetiapine
Low Potency TAPs
Medium Potency TAPs
High Potency TAPs

131
Figure A.5: Histogram of Low Potency TAPs PPS by Treatment
0 0. 05 0. 1 0. 15 0. 2 0. 25 0. 3 0. 35 0. 4 0. 45 0. 5 0. 55 0. 6 0. 65 0. 7 0. 75 0. 8 0. 85 0. 9 0. 95 1
0
2000
4000
6000
C
o
u
n
t
l ow _pps
0
5000
10000
15000
20000 C
o
u
n
t
0
5000
10000
15000
20000 C
o
u
n
t
0
5000
10000
15000
20000 C
o
u
n
t
0
5000
10000
15000
20000 C
o
u
n
t
0 0. 05 0. 1 0. 15 0. 2 0. 25 0. 3 0. 35 0. 4 0. 45 0. 5 0. 55 0. 6 0. 65 0. 7 0. 75 0. 8 0. 85 0. 9 0. 95 1
0
5000
10000
15000
20000 C
o
u
n
t
l ow _pps

Figure A.6: Refined Breakdown of Episodes by Treatment Type When Utilizing
Low Potency TAPs as the CT
35,239
34,239
9,494
24,569
50,797
45,861
Olanzapine
Risperidone
Quetiapine
Low Potency TAPs
Medium Potency TAPs
High Potency TAPs

132
Figure A.7: Histogram of Medium Potency TAPs PPS by Treatment
0 0. 05 0. 1 0. 15 0. 2 0. 25 0. 3 0. 35 0. 4 0. 45 0. 5 0. 55 0. 6 0. 65 0. 7 0. 75 0. 8 0. 85 0. 9 0. 95 1
0
10000
20000
30000
C
o
u
n
t
m ed_pps

0
5000
10000
15000
20000 C
o
u
n
t
0
5000
10000
15000
20000 C
o
u
n
t
0
5000
10000
15000
20000 C
o
u
n
t
0
5000
10000
15000
20000 C
o
u
n
t
0 0. 05 0. 1 0. 15 0. 2 0. 25 0. 3 0. 35 0. 4 0. 45 0. 5 0. 55 0. 6 0. 65 0. 7 0. 75 0. 8 0. 85 0. 9 0. 95 1
0
5000
10000
15000
20000 C
o
u
n
t
m ed_pps

Figure A.8: Refined Breakdown of Episodes by Treatment Type When Utilizing
Medium Potency TAPs as the CT
37,150
36,084
10,948
25,365
56,284
44,921
Olanzapine
Risperidone
Quetiapine
Low Potency TAPs
Medium Potency TAPs
High Potency TAPs

133
Figure A.9: Histogram of High Potency TAPs PPS by Treatment
0 0. 05 0. 1 0. 15 0. 2 0. 25 0. 3 0. 35 0. 4 0. 45 0. 5 0. 55 0. 6 0. 65 0. 7 0. 75 0. 8 0. 85 0. 9 0. 95 1
0
5000
10000
15000
C
o
u
n
t
hi gh_pps

0
10000
20000
30000
C
o
u
n
t
0
10000
20000
30000
C
o
u
n
t
0
10000
20000
30000
C
o
u
n
t
0
10000
20000
30000
C
o
u
n
t
0 0. 05 0. 1 0. 15 0. 2 0. 25 0. 3 0. 35 0. 4 0. 45 0. 5 0. 55 0. 6 0. 65 0. 7 0. 75 0. 8 0. 85 0. 9 0. 95 1
0
10000
20000
30000
C
o
u
n
t
hi gh_pps

Figure A.10: Refined Breakdown of Episodes by Treatment Type When Utilizing
High Potency TAPs as the CT

25,623
26,967
6,987
20,175
32,457
38,968
Olanzapine
Risperidone
Quetiapine
Low Potency TAPs
Medium Potency TAPs
High Potency TAPs

134
Table A.7: Cutpoints for Risperidone Quintiles
Risperidone Quintile Lower Bound Upper Bound
1 0.10 0.15
2 0.15 0.18
3 0.18 0.23
4 0.23 0.31
5 0.31 0.50

Table A.8: Cutpoints for Quetiapine Quintiles
Quetiapine Quintile Lower Bound Upper Bound
1 0.10 0.13
2 0.13 0.17
3 0.17 0.23
4 0.23 0.29
5 0.29 0.50

Table A.9: Cutpoints for Low Potency TAPs Quintiles
Low Potency TAPs
Quintile
Lower Bound Upper Bound
1 0.05 0.10
2 0.10 0.13
3 0.13 0.17
4 0.17 0.22
5 0.22 0.40

Table A.10: Cutpoints for Medium Potency TAPs Quintiles
Medium Potency TAPs
Quintile
Lower Bound Upper Bound
1 0.00 0.01
2 0.01 0.02
3 0.02 0.04
4 0.04 0.06
5 0.06 0.20

Table A.11: Cutpoints for High Potency TAPs Quintiles
High Potency TAPs
Quintile
Lower Bound Upper Bound
1 0.10 0.13
2 0.13 0.15
3 0.15 0.17
4 0.17 0.21
5 0.21 0.30
135
Table A.12a: Descriptive Statistics for Olanzapine Quintile 1
Covariate Success
(N=814)
No Success
(N=6002)
SCHIZOPHRENIA TYPE
SIMPLE 0.49 0.32
DISORG 0.25 0.13
CATA 0.37 0.12
PARAS** 7.62 4.77
ACUTE 0.25 0.33
LATENT 0.00 0.03
RESID 0.37 0.15
SAFF 8.11 7.33
OTHNOS 3.44 2.37
UNSPEC** 79.12 84.46
AGE
AGE_CAT18 11.79 9.96
AGE_CAT25 17.08 16.48
AGE_CAT35 24.32 26.44
AGE_CAT45 24.20 24.48
AGE_CAT55 10.93 10.55
AGE_CAT65 11.67 12.10
MALE 40.54 39.07
RACE/ETHNICITY
WHITE 38.33 34.99
BLACK 17.57 20.08
HISPANIC 10.69 10.45
ASIAN 13.39 13.45
OTHER 20.02 21.04
COMORBIDITIES
ACC 1.84 2.50
ANX 13.02 12.51
ARRTDX 4.05 4.35
BIP 16.71 16.08
BLOOD 6.51 6.86
CARDIO 17.48 16.95
CHILD 0.25 0.20
CONGE 1.60 1.77
DEMEN 0.61 0.53
DEP 69.66 68.24
DIGES 18.18 18.83
DIABDX 10.32 10.11
ENDO 9.51 9.21
EPSDX 0.25 0.30
GENTI 20.64 20.39
GLAUC 0.98 1.15
HBP 15.11 15.51
HIV 0.00 0.07
HLIPID 15.23 13.31
INFEC 96.56 97.45
INJUR 21.25 20.96
MANIC 0.98 0.85
MUSCL 36.61 35.77
NEOPL 4.05 4.32
NERV 27.76 27.41
OCIRC 14.25 14.21
136
Table A.12a: Descriptive Statistics for Olanzapine Quintile 1 (Continued)
Covariate Success
N=814
No Success
N=6002
OTHPSY 3.81 3.30
PERI 0.25 0.27
PERSO 0.74 0.95
PSTG 1.11 1.38
RESPI 27.89 30.16
SABUSE 7.99 7.58
SEXD 0.12 0.12
SKIN 10.20 10.85
SUICDX 3.44 3.42
CONCOMITANT DRUG USE
AD 60.93 60.26
HYPNT 7.62 9.06
SEIZU 14.50 14.80
ARRH* 7.13 6.00
APARK 1.11 0.67
ANXIE 6.51 7.10
MOOD 14.37 15.29
LIPID 11.18 9.31
EPS1 7.00 7.68
EPS2 4.79 4.77
DIABRX 8.85 8.81
DIABDX 22.24 22.38
PTHER 4.30 4.58
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 86.12 85.84
RIS 5.28 5.05
QUE 2.77 4.05
OAP 0.49 0.23
TAP 3.81 4.45
EPISODE TYPE
RESTART 92.75 92.70
SWT 0.61 0.57
AUG 1.35 1.22
LATE_SWT 5.28 5.51
PRIOR HEALTH SERVICE
UTILIZATION

LTC 1.23 0.62
AHOSP 6.02 5.38
PHOSP 5.16 4.72
REHAB 0.25 0.37
CMHC** 63.64 59.50
PSYCH** 0.61 0.28
* p<0.05; **p<0.01

137
Table A.12b: Descriptive Statistics for Olanzapine Quintile 2
Covariate Success
N=955
No Success
N=5863
SCHIZOPHRENIA TYPE
SIMPLE 0.42 0.22
DISORG 0.21 0.24
CATA 0.21 0.14
PARAS 7.43 6.77
ACUTE 0.52 0.29
LATENT 0.00 0.00
RESID 0.31 0.31
SAFF 8.06 7.42
OTHNOS 3.04 3.87
UNSPEC 79.79 80.74
AGE
AGE_CAT18 6.49 5.94
AGE_CAT25 15.08 15.16
AGE_CAT35 24.40 25.82
AGE_CAT45* 25.86 22.65
AGE_CAT55 12.46 14.36
AGE_CAT65 15.71 16.07
MALE 45.45 45.68
RACE/ETHNICITY
WHITE 47.23 49.33
BLACK 12.04 10.06
HISP* 5.24 4.37
ASIAN 9.32 8.77
OTHER 26.18 27.48
COMORBIDITIES
ACC 1.47 1.18
ANX 7.43 8.03
ARRTDX 3.66 3.60
BIP 13.40 12.20
BLOOD 4.50 5.66
CARDIO 18.22 16.34
CHILD 0.52 0.65
CONGE 1.15 1.60
DEMEN* 1.78 0.90
DEP 60.73 62.10
DIGES** 12.88 16.97
DIABDX 8.59 8.39
ENDO 9.01 9.65
EPSDX 0.84 0.51
GENTI 14.66 16.63
GLAUC 0.52 1.11
HBP 15.60 14.45
HIV* 0.00 0.05
HLIPID 13.30 15.13
INFEC 96.13 96.15
INJUR 16.02 15.83
MANIC 0.94 0.65
MUSCL* 23.77 27.56
NEOPL 3.77 4.50
NERV 28.18 26.28
OCIRC 13.19 12.57
138
Table A.12b: Descriptive Statistics for Olanzapine Quintile 2 (Continued)
Covariate Success
N=955
No Success
N=5863
OTHPSY 2.20 2.88
PERI 0.00 0.10
PERSO 0.84 1.11
PSTG 0.84 0.44
RESPI 24.40 23.42
SABUSE 4.19 4.90
SEXD 0.10 0.03
SKIN 12.57 11.60
SUICDX 2.41 2.80
CONCOMITANT DRUG USE
AD 54.14 56.08
HYPNT* 6.18 4.61
SEIZU 14.03 14.33
ARRH 6.91 7.04
APARK 0.94 1.18
ANXIE 3.56 3.87
MOOD 20.00 17.57
LIPID 9.42 10.81
EPS1 10.79 10.42
EPS2 3.87 3.99
DIABRX 8.17 7.47
DIABDX 19.58 19.09
PTHER 7.23 5.95
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 83.87 83.59
RIS 6.07 5.42
QUE* 3.66 2.46
OAP* 0.73 0.31
TAP 7.02 6.23
EPISODE TYPE
RESTART 85.76 86.83
SWT 1.15 1.16
AUG 2.83 2.17
LATE_SWT 10.26 9.84
PRIOR HEALTH SERVICE
UTILIZATION

LTC** 3.87 1.62
AHOSP 3.87 3.29
PHOSP 2.93 3.07
REHAB 0.42 0.36
CMHC* 52.98 49.53
PSYCH** 1.15 0.43
* p<0.05; **p<0.01

139
Table A.12c: Descriptive Statistics for Olanzapine Quintile 3
Covariate Success
N=1406
No Success
N=5408
SCHIZOPHRENIA TYPE
SIMPLE 0.28 0.28
DISORG 0.28 0.54
CATA 0.14 0.15
PARAS 12.23 11.58
ACUTE 0.14 0.26
LATENT 0.07 0.02
RESID 0.92 0.74
SAFF 8.82 9.26
OTHNOS 6.61 6.62
UNSPEC 70.48 70.56
AGE
AGE_CAT18* 3.27 4.60
AGE_CAT25 14.65 14.02
AGE_CAT35 24.18 25.26
AGE_CAT45 22.05 23.80
AGE_CAT55 14.86 15.38
AGE_CAT65* 20.98 16.94
MALE 51.56 52.46
RACE/ETHNICITY
WHITE 54.48 54.60
BLACK 9.96 10.45
HISP 3.77 4.20
ASIAN** 7.04 5.25
OTHER 24.75 25.50
COMORBIDITIES
ACC 1.78 1.31
ANX 8.68 8.27
ARRTDX 4.77 3.85
BIP 9.39 10.63
BLOOD 5.62 6.01
CARDIO 16.43 15.64
CHILD 0.43 0.67
CONGE 1.07 1.74
DEMEN 3.20 2.68
DEP* 52.84 52.72
DIGES 14.51 15.29
DIABDX 8.18 7.67
ENDO 8.82 10.17
EPSDX 1.00 0.85
GENTI 13.16 14.83
GLAUC* 0.50 1.13
HBP 13.80 13.39
HIV 0.07 0.00
HLIPID 14.30 14.79
INFEC 93.60 64.88
INJUR 14.86 15.24
MANIC 0.50 0.52
MUSCL 23.68 23.61
NEOPL 4.98 4.18
NERV 26.66 30.40
140
Table A.12c: Descriptive Statistics for Olanzapine Quintile 3 (Continued)
Covariate Success
N=1406
No Success
N=5408
OCIRC 14.15 12.32
OTHPSY 3.34 3.66
PERI* 0.36 0.11
PERSO 1.07 1.07
PSTG 0.21 0.57
RESPI 21.91 22.73
SABUSE 4.20 3.83
SEXD 0.00 0.06
SKIN 13.73 14.83
SUICDX 2.77 2.96
CONCOMITANT DRUG USE
AD 47.51 46.09
HYPNT 4.13 3.83
SEIZU 14.74 13.37
ARRH 7.33 7.47
APARK ** 2.13 1.20
ANXIE 3.84 3.68
MOOD 21.05 22.15
LIPID 10.38 10.63
EPS1 18.92 19.93
EPS2 3.49 3.20
DIABRX* 7.89 6.25
DIABDX 18.56 17.83
PTHER 8.39 9.08
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 76.53 75.09
RIS 10.81 10.61
QUE 3.63 4.12
OAP* 2.06 1.33
TAP 12.02 12.85
EPISODE TYPE
RESTART 67.35 66.79
SWT 5.05 4.46
AUG 5.41 5.82
LATE_SWT 22.19 22.93
PRIOR HEALTH SERVICE
UTILIZATION

LTC** 9.03 6.23
AHOSP 3.06 3.11
PHOSP 2.49 3.11
REHAB 0.14 0.13
CMHC 42.60 45.34
PSYCH 1.35 1.11
* p<0.05; **p<0.01

141
Table A.12d: Descriptive Statistics for Olanzapine Quintile 4
Covariate Success
N=2027
No Success
N=4786
SCHIZOPHRENIA TYPE
SIMPLE 0.20 0.38
DISORG 0.59 0.50
CATA 0.10 0.06
PARAS 11.45 11.20
ACUTE 0.10 0.27
LATENT 0.00 0.00
RESID 0.69 0.94
SAFF 7.84 8.52
OTHNOS 6.56 6.79
UNSPEC 72.47 71.33
AGE
AGE_CAT18 2.66 2.63
AGE_CAT25 14.50 13.54
AGE_CAT35 26.34 25.16
AGE_CAT45 22.35 21.79
AGE_CAT55** 11.94 14.19
AGE_CAT65 22.30 22.69
MALE 49.28 50.08
RACE/ETHNICITY
WHITE 59.40 59.24
BLACK 10.31 9.67
HISP 3.35 3.22
ASIAN 3.90 4.45
OTHER 23.04 23.42
COMORBIDITIES
ACC 1.78 1.50
ANX 7.70 8.46
ARRTDX 4.39 4.64
BIP 9.62 9.92
BLOOD 6.36 6.96
CARDIO 15.15 16.38
CHILD 0.49 0.42
CONGE 2.02 1.67
DEMEN* 5.23 4.10
DEP 53.92 51.80
DIGES 16.48 16.21
DIABDX 8.78 8.21
ENDO 11.69 9.78
EPSDX 1.23 1.30
GENTI 16.03 14.54
GLAUC 1.28 1.00
HBP 13.12 13.60
HIV 0.00 0.00
HLIPID 13.32 12.85
INFEC 96.10 95.51
INJUR 17.76 16.46
MANIC 0.59 0.69
MUSCL 24.47 23.94
NEOPL 3.45 4.33
NERV 33.45 32.72
142
Table A.12d: Descriptive Statistics for Olanzapine Quintile 4 (Continued)
Covariate Success
N=2027
No Success
N=4786
OCIRC 14.11 15.19
OTHPSY 4.49 4.18
PERI 0.20 0.15
PERSO 1.53 1.34
PSTG 0.39 0.40
RESPI 24.91 23.59
SABUSE 3.06 3.57
SEXD 0.05 0.04
SKIN 15.89 16.05
SUICDX 3.55 3.36
CONCOMITANT DRUG USE
AD 47.46 45.80
HYPNT 3.45 4.47
SEIZU 15.93 14.73
ARRH 7.75 8.94
APARK 2.66 2.28
ANXIE 2.91 3.61
MOOD 23.14 23.32
LIPID 9.03 8.65
EPS1 31.57 29.25
EPS2* 3.11 4.12
DIABRX 6.96 7.19
DIABDX 16.92 18.07
PTHER 15.15 13.48
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 59.36 58.95
RIS 21.41 20.77
QUE 3.65 4.70
OAP 4.31 3.36
TAP 32.41 30.44
EPISODE TYPE
RESTART 29.26 28.77
SWT 16.63 16.51
AUG* 17.51 15.46
LATE_SWT* 36.61 39.26
PRIOR HEALTH SERVICE
UTILIZATION

LTC 17.91 17.63
AHOSP 3.50 3.26
PHOSP* 2.07 2.34
REHAB 0.10 0.38
CMHC 32.56 33.99
PSYCH 2.71 2.30
* p<0.05; **p<0.01

143
Table A.12e: Descriptive Statistics for Olanzapine Quintile 5
Covariate Success
N=2938
No Success
N=3878
SCHIZOPHRENIA TYPE
SIMPLE 0.61 0.52
DISORG 0.61 0.54
CATA 0.17 0.15
PARAS 13.96 15.45
ACUTE 0.31 0.26
LATENT 0.00 0.00
RESID 1.77 1.68
SAFF 9.16 9.44
OTHNOS 11.81 11.09
UNSPEC 61.61 60.88
AGE
AGE_CAT18 0.71 1.03
AGE_CAT25** 10.62 12.58
AGE_CAT35 26.24 24.99
AGE_CAT45 22.60 23.57
AGE_CAT55** 16.68 13.33
AGE_CAT65 23.14 24.50
MALE 55.11 53.33
RACE/ETHNICITY
WHITE 66.98 66.58
BLACK 5.17 5.23
HISP 1.36 1.55
ASIAN** 1.12 2.42
OTHER 25.36 24.21
COMORBIDITIES
ACC** 0.58 1.19
ANX 4.80 5.83
ARRTDX 4.94 5.13
BIP 7.01 7.19
BLOOD 7.35 6.16
CARDIO 14.91 15.68
CHILD 0.61 0.57
CONGE 1.50 1.19
DEMEN 8.17 9.10
DEP** 42.55 45.72
DIGES 16.07 15.34
DIABDX 7.05 8.17
ENDO 11.06 10.50
EPSDX 1.23 1.47
GENTI 13.10 12.84
GLAUC 1.12 1.03
HBP 12.90 13.02
HIV 0.00 0.03
HLIPID 11.84 11.78
INFEC* 96.38 95.85
INJUR* 13.65 15.70
MANIC* 0.14 0.41
MUSCL 20.05 20.58
NEOPL 4.05 3.89
NERV 34.79 34.06

144
Table A.12e: Descriptive Statistics for Olanzapine Quintile 5 (Continued)
Covariate Success
N=2938
No Success
N=3878
OCIRC 14.23 15.21
OTHPSY 4.97 5.49
PERI 0.07 0.21
PERSO 1.43 1.42
PSTG 0.34 0.26
RESPI 22.23 23.26
SABUSE 1.23 1.39
SEXD 0.03 0.03
SKIN** 21.14 18.62
SUICDX** 1.80 2.76
CONCOMITANT DRUG USE
AD** 39.21 42.11
HYPNT 2.18 2.19
SEIZU 15.21 14.36
ARRH 9.73 9.49
APARK 1.67 2.40
ANXIE 1.57 1.88
MOOD 29.88 28.98
LIPID 7.86 8.82
EPS1 55.24 53.79
EPS2 3.10 2.68
DIABRX** 4.77 6.52
DIABDX 15.04 15.01
PTHER 22.29 21.20
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA** 47.82 50.98
RIS 24.13 25.89
QUE 2.11 2.66
OAP 6.06 7.04
TAP 62.19 60.06
EPISODE TYPE
RESTART 4.15 5.18
SWT 42.75 39.84
AUG 27.71 29.42
LATE_SWT 25.39 25.55
PRIOR HEALTH SERVICE
UTILIZATION

LTC* 23.59 25.19
AHOSP 1.87 2.55
PHOSP 1.16 1.68
REHAB 0.27 0.15
CMHC** 23.96 27.23
PSYCH 6.40 5.62
* p<0.05; **p<0.01

145
Table A.13a: Descriptive Statistics for Risperidone Quintile 1
Covariate Success
N=820
No Success
N=6350
SCHIZOPHRENIA TYPE
SIMPLE 0.12 0.14
DISORG 0.00 0.43
CATA 0.12 0.09
PARAS* 7.20 5.26
ACUTE 0.61 0.28
LATENT 0.00 0.02
RESID 0.61 0.52
SAFF 7.20 7.80
OTHNOS 2.44 2.06
UNSPEC 81.71 83.40
AGE
AGE_CAT18 9.15 9.87
AGE_CAT25 17.80 16.90
AGE_CAT35 24.39 26.27
AGE_CAT45 20.48 19.78
AGE_CAT55 9.51 9.39
AGE_CAT65 18.66 17.80
MALE 32.80 31.51
RACE/ETHNICITY
WHITE 30.24 30.52
BLACK 24.88 26.52
HISP 12.56 14.06
ASIAN 11.22 9.70
OTHER 21.10 19.20
COMORBIDITIES
ACC 2.20 2.60
ANX* 11.71 14.60
ARRTDX 6.10 5.81
BIP 15.85 13.62
BLOOD 7.44 7.29
CARDIO 23.41 22.20
CHILD 0.24 0.08
CONGE 1.46 2.35
DEMEN 1.22 0.71
DEP 72.07 70.20
DIGES 18.54 19.84
DIABDX 15.12 14.19
ENDO 11.34 10.96
EPSDX 0.37 0.79
GENTI 21.10 22.22
GLAUC 2.07 1.56
HBP 19.02 18.63
HIV 0.24 0.17
HLIPID 12.20 13.87
INFEC 95.85 95.73
INJUR 11.06 23.54
MANIC 1.22 0.88
MUSCL 36.71 37.73
NEOPL 4.63 4.69
NERV* 27.56 31.15
146
Table A.13a: Descriptive Statistics for Risperidone Quintile 1 (Continued)
Covariate Success
N=820
No Success
N=6350
OCIRC 18.05 17.45
OTHPSY 4.27 3.43
PERI 0.49 0.22
PERSO 1.83 1.40
PSTG 1.71 1.24
RESPI 30.49 30.24
SABUSE 10.24 10.14
SEXD 0.00 0.13
SKIN 12.07 12.19
SUICDX 5.12 4.77
CONCOMITANT DRUG USE
AD 61.34 56.65
HYPNT 8.54 8.71
SEIZU 11.71 12.22
ARRH 9.15 7.39
APARK 0.98 0.66
ANXIE* 6.10 8.60
MOOD 13.66 13.75
LIPID 9.02 9.31
EPS1 11.10 9.45
EPS2 5.24 5.53
DIABRX 12.32 11.94
DIABDX 24.02 25.10
PTHER 4.88 5.42
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 8.17 7.87
RIS 85.49 84.85
QUE 3.05 2.93
OAP 0.24 0.28
TAP 5.00 4.54
EPISODE TYPE
RESTART 90.24 89.72
SWT 1.22 0.85
AUG 2.07 1.84
LATE_SWT 6.46 7.59
PRIOR HEALTH SERVICE
UTILIZATION

LTC 1.95 0.79
AHOSP 6.71 6.47
PHOSP 6.59 6.06
REHAB 0.37 0.47
CMHC 59.88 56.79
PSYCH 0.37 0.57
* p<0.05; **p<0.01

147
Table A.13b: Descriptive Statistics for Risperidone Quintile 2
Covariate Success
N=1180
No Success
N=5988
SCHIZOPHRENIA TYPE
SIMPLE 0.08 0.08
DISORG 0.17 0.28
CATA 0.25 0.10
PARAS* 7.88 6.31
ACUTE 0.34 0.30
LATENT 0.00 0.00
RESID 0.68 0.68
SAFF 8.22 7.16
OTHNOS 2.88 3.37
UNSPEC 79.49 81.70
AGE
AGE_CAT18 9.24 8.90
AGE_CAT25 16.02 16.00
AGE_CAT35 24.49 24.80
AGE_CAT45 19.92 19.77
AGE_CAT55* 9.75 11.84
AGE_CAT65 20.59 18.69
MALE 40.59 40.66
RACE/ETHNICITY
WHITE 48.90 48.36
BLACK 13.39 14.04
HISP 6.69 5.78
ASIAN 6.36 6.21
OTHER 24.66 25.60
COMORBIDITIES
ACC 1.61 2.10
ANX 9.24 8.75
ARRTDX 3.39 4.14
BIP 12.71 11.91
BLOOD 5.42 5.74
CARDIO 17.46 16.98
CHILD 0.42 0.38
CONGE 2.12 1.64
DEMEN 1.95 1.60
DEP 61.44 60.12
DIGES 16.78 16.12
DIABDX 9.58 9.59
ENDO 9.32 9.79
EPSDX 0.76 0.72
GENTI 17.71 16.90
GLAUC 1.02 1.49
HBP 14.58 14.48
HIV 0.00 0.05
HLIPID 12.37 13.26
INFEC 96.10 96.58
INJUR 18.47 16.35
MANIC 0.34 0.63
MUSCL 27.80 28.24
NEOPL 3.90 4.59
NERV 28.73 30.11
148
Table A.13b: Descriptive Statistics for Risperidone Quintile 2 (Continued)
Covariate Success
N=1180
No Success
N=5988
OCIRC 14.75 13.99
OTHPSY 4.07 3.12
PERI 0.25 0.20
PERSO 1.02 1.12
PSTG 0.34 0.40
RESPI 25.08 25.07
SABUSE 3.22 4.19
SEXD 0.17 0.17
SKIN 12.20 12.56
SUICDX 3.05 2.54
CONCOMITANT DRUG USE
AD 59.90 53.39
HYPNT 5.51 5.11
SEIZU 13.56 12.96
ARRH 7.03 7.80
APARK 1.19 1.19
ANXIE 4.49 3.69
MOOD 18.81 16.77
LIPID 8.81 9.52
EPS1* 18.56 16.18
EPS2 2.97 3.72
DIABRX* 8.31 8.55
DIABDX 16.10 18.52
PTHER 8.90 8.20
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA* 9.07 7.40
RIS 84.15 82.60
QUE 2.88 2.87
OAP 0.76 0.40
TAP 6.36 6.53
EPISODE TYPE
RESTART 83.14 83.10
SWT** 3.14 1.84
AUG 3.39 3.21
LATE_SWT 10.34 11.86
PRIOR HEALTH SERVICE
UTILIZATION

LTC 3.90 2.91
AHOSP* 5.00 3.71
PHOSP 3.90 3.51
REHAB 0.42 0.30
CMHC 43.90 44.89
PSYCH 1.36 0.97
* p<0.05; **p<0.01

149
Table A.13c: Descriptive Statistics for Risperidone Quintile 3
Covariate Success
N=1546
No Success
N=5624
SCHIZOPHRENIA TYPE
SIMPLE 0.26 0.21
DISORG 0.39 0.28
CATA 0.13 0.09
PARAS 9.31 9.64
ACUTE 0.52 0.34
LATENT 0.00 0.05
RESID 1.16 0.68
SAFF 7.44 7.70
OTHNOS 6.66 6.19
UNSPEC 74.13 74.82
AGE
AGE_CAT18 7.76 7.06
AGE_CAT25 15.78 14.85
AGE_CAT35 22.83 23.24
AGE_CAT45** 19.47 22.83
AGE_CAT55 11.71 11.56
AGE_CAT65* 22.90 20.47
MALE 49.61 50.37
RACE/ETHNICITY
WHITE 57.83 56.99
BLACK 12.10 10.97
HISP 3.55 3.93
ASIAN 3.23 3.86
OTHER 23.29 24.25
COMORBIDITIES
ACC 1.81 1.37
ANX 7.44 7.38
ARRTDX 4.01 3.73
BIP 8.34 8.52
BLOOD 5.37 5.85
CARDIO 15.46 15.26
CHILD 0.91 0.82
CONGE 1.62 1.40
DEMEN 3.49 3.45
DEP 50.13 51.85
DIGES 15.27 15.61
DIABDX 8.21 8.48
ENDO 8.93 8.93
EPSDX 0.78 0.89
GENTI 14.10 14.51
GLAUC 1.42 1.19
HBP 13.32 12.84
HIV 0.00 0.00
HLIPID 11.90 12.25
INFEC 96.38 97.14
INJUR 15.78 14.86
MANIC 0.71 0.71
MUSCL 21.99 23.56
NEOPL 4.08 3.52
NERV* 32.41 29.61
150
Table A.13c: Descriptive Statistics for Risperidone Quintile 3 (Continued)
Covariate Success
N=1546
No Success
N=5624
OCIRC 15.20 14.17
OTHPSY 4.85 3.97
PERI 0.26 0.11
PERSO 1.03 0.94
PSTG 0.37 0.32
RESPI 22.87 23.61
SABUSE 2.93 3.43
SEXD 0.13 0.11
SKIN 13.00 14.24
SUICDX 2.85 2.36
CONCOMITANT DRUG USE
AD 44.89 46.32
HYPNT 2.98 3.79
SEIZU 13.58 15.38
ARRH 8.73 7.66
APARK 2.01 1.35
ANXIE 2.46 2.77
MOOD 19.34 19.58
LIPID 8.54 8.45
EPS1 26.46 25.59
EPS2 3.62 3.63
DIABRX 8.15 8.00
DIABDX 17.59 16.18
PTHER 12.48 11.86
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 10.28 10.67
RIS 76.91 75.96
QUE 3.43 3.70
OAP 1.62 1.32
TAP 12.68 12.66
EPISODE TYPE
RESTART 70.63 69.43
SWT 4.40 4.69
AUG 7.89 7.04
LATE_SWT 17.08 18.83
PRIOR HEALTH SERVICE
UTILIZATION

LTC** 13.07 10.42
AHOSP 3.10 3.18
PHOSP 2.59 2.95
REHAB 0.26 0.34
CMHC 33.70 36.29
PSYCH 2.26 1.81
* p<0.05; **p<0.01

151
Table A.13d: Descriptive Statistics for Risperidone Quintile 4
Covariate Success
N=2038
No Success
N=5130
SCHIZOPHRENIA TYPE
SIMPLE 0.54 0.43
DISORG 0.20 0.29
CATA 0.05 0.06
PARAS 9.57 10.41
ACUTE 0.29 0.27
LATENT 0.10 0.10
RESID 0.69 0.74
SAFF 6.97 6.35
OTHNOS 7.31 6.73
UNSPEC 74.29 74.62
AGE
AGE_CAT18 4.12 4.70
AGE_CAT25 13.40 12.48
AGE_CAT35 19.97 20.47
AGE_CAT45 19.53 18.27
AGE_CAT55 11.04 11.85
AGE_CAT65 31.94 32.24
MALE 49.17 49.43
RACE/ETHNICITY
WHITE 63.35 63.22
BLACK 9.13 9.61
HISP* 3.58 2.63
ASIAN 2.55 2.79
OTHER 21.39 21.75
COMORBIDITIES
ACC 1.72 1.68
ANX 6.13 6.16
ARRTDX 5.00 4.39
BIP 7.80 7.43
BLOOD 6.04 6.10
CARDIO 14.33 14.85
CHILD 0.64 1.13
CONGE 1.37 1.70
DEMEN 7.31 6.92
DEP* 46.81 46.12
DIGES 14.82 15.79
DIABDX 7.90 8.77
ENDO 11.48 10.92
EPSDX 1.47 1.21
GENTI 12.81 14.64
GLAUC 1.42 1.31
HBP 11.97 12.69
HIV 0.00 0.00
HLIPID 11.68 10.43
INFEC 98.38 97.97
INJUR 17.12 16.94
MANIC 0.64 0.43
MUSCL 22.33 21.72
NEOPL 4.86 4.29
NERV 35.43 35.09
152
Table A.13d: Descriptive Statistics for Risperidone Quintile 4 (Continued)
Covariate Success
N=2038
No Success
N=5130
OCIRC 18.16 18.05
OTHPSY 5.40 5.77
PERI 0.15 0.12
PERSO 0.74 1.13
PSTG 0.29 0.29
RESPI 23.80 23.78
SABUSE 2.36 2.12
SEXD 0.05 0.08
SKIN 17.47 16.67
SUICDX 1.91 2.42
CONCOMITANT DRUG USE
AD 41.66 41.05
HYPNT 2.75 2.75
SEIZU 16.05 15.32
ARRH 9.81 9.82
APARK 1.57 2.30
ANXIE* 2.55 1.72
MOOD 21.00 20.49
LIPID 7.80 7.31
EPS1 31.80 31.87
EPS2 2.94 3.04
DIABRX 7.21 7.82
DIABDX 15.90 15.15
PTHER 17.47 16.10
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 14.52 14.93
RIS 64.33 65.48
QUE 4.27 3.74
OAP* 4.22 3.26
TAP 27.38 25.52
EPISODE TYPE
RESTART 47.15 47.02
SWT 13.30 12.16
AUG 13.15 13.33
LATE_SWT 26.40 27.49
PRIOR HEALTH SERVICE
UTILIZATION

LTC 28.85 28.97
AHOSP 3.04 3.31
PHOSP* 2.36 2.18
REHAB 0.10 0.12
CMHC 25.22 25.77
PSYCH* 4.61 3.43
* p<0.05; **p<0.01

153
Table A.13e: Descriptive Statistics for Risperidone Quintile 5
Covariate Success
N=2807
No Success
N=4363
SCHIZOPHRENIA TYPE
SIMPLE 1.00 0.89
DISORG 0.39 0.34
CATA 0.07 0.14
PARAS 10.94 11.71
ACUTE 0.18 0.28
LATENT 0.18 0.11
RESID 1.03 1.12
SAFF 5.70 6.26
OTHNOS 10.15 10.20
UNSPEC 70.36 68.37
AGE
AGE_CAT18 1.82 2.20
AGE_CAT25* 9.62 11.48
AGE_CAT35 21.48 20.01
AGE_CAT45 21.87 20.40
AGE_CAT55** 15.32 13.06
AGE_CAT65** 29.89 32.84
MALE 53.79 53.38
RACE/ETHNICITY
WHITE 69.33 70.20
BLACK 7.09 6.14
HISP 2.24 1.83
ASIAN 1.07 1.28
OTHER 20.27 20.54
COMORBIDITIES
ACC 1.71 2.15
ANX 5.59 5.75
ARRTDX 4.28 4.72
BIP* 6.02 7.36
BLOOD 6.77 6.33
CARDIO 12.86 14.05
CHILD 1.18 0.94
CONGE 1.64 1.54
DEMEN 16.14 15.63
DEP 38.97 40.50
DIGES* 16.17 14.07
DIABDX 6.63 6.81
ENDO 10.90 11.37
EPSDX 1.53 1.74
GENTI 12.61 12.38
GLAUC 0.78 0.87
HBP 11.61 11.85
HIV 0.00 0.00
HLIPID 9.76 9.51
INFEC 99.04 98.62
INJUR* 13.25 15.29
MANIC 0.32 0.41
MUSCL 20.63 20.40
NEOPL 3.71 4.08
NERV 39.65 38.67
154
Table A.13e: Descriptive Statistics for Risperidone Quintile 5 (Continued)
Covariate Success
N=2807
No Success
N=4363
OCIRC 17.06 18.66
OTHPSY 6.27 6.99
PERI 0.11 0.23
PERSO 1.21 1.08
PSTG 0.25 0.23
RESPI 23.33 23.24
SABUSE 1.00 1.03
SEXD 0.11 0.05
SKIN 22.91 22.07
SUICDX 1.39 2.02
CONCOMITANT DRUG USE
AD 34.88 36.26
HYPNT* 2.60 1.81
SEIZU 20.16 19.98
ARRH 9.41 9.70
APARK* 2.07 2.89
ANXIE 1.10 1.35
MOOD 24.58 25.60
LIPID 6.70 6.72
EPS1 45.96 47.15
EPS2 3.35 2.68
DIABRX 5.56 5.89
DIABDX 13.22 13.00
PTHER 24.40 24.80
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA* 10.30 12.15
RIS** 46.03 49.28
QUE* 2.17 2.98
OAP 4.81 5.75
TAP* 63.13 60.53
EPISODE TYPE
RESTART** 13.47 17.21
SWT* 37.48 34.82
AUG 23.23 24.11
LATE_SWT 25.83 23.86
PRIOR HEALTH SERVICE
UTILIZATION

LTC 38.69 39.77
AHOSP* 2.35 3.25
PHOSP 1.10 1.42
REHAB 0.11 0.30
CMHC 17.74 16.94
PSYCH 7.59 7.95
* p<0.05; **p<0.01

155
Table A.14a: Descriptive Statistics for Quetiapine Quintile 1
Covariate Success
N=225
No Success
N=1792
SCHIZOPHRENIA TYPE
SIMPLE 0.00 0.22
DISORG 0.00 0.00
CATA 0.00 0.22
PARAS 4.44 5.08
ACUTE 0.44 0.11
LATENT 0.00 0.00
RESID 0.00 0.22
SAFF 9.33 7.59
OTHNOS 3.56 2.18
UNSPEC 82.22 84.38
AGE
AGE_CAT18 7.56 6.86
AGE_CAT25 20.00 18.81
AGE_CAT35 27.11 29.80
AGE_CAT45 29.89 28.91
AGE_CAT55 5.78 6.98
AGE_CAT65 10.67 8.65
MALE 32.89 29.24
RACE/ETHNICITY
WHITE 49.78 53.52
BLACK 12.44 12.89
HISP 8.44 5.36
ASIAN 9.33 10.38
OTHER 20.00 17.86
COMORBIDITIES
ACC 2.22 2.73
ANX 12.89 12.67
ARRTDX 2.67 1.84
BIP 21.78 19.92
BLOOD 5.78 4.52
CARDIO 18.22 14.62
CHILD 0.00 0.78
CONGE 2.67 1.40
DEMEN 0.00 0.11
DEP 70.67 74.44
DIGES 18.22 19.81
DIABDX 9.78 11.22
ENDO 12.44 11.72
EPSDX 0.89 0.84
GENTI 17.78 20.37
GLAUC 0.44 1.17
HBP 16.44 13.50
HIV 0.00 0.17
HLIPID 12.44 11.83
INFEC 96.00 97.49
INJUR** 14.67 22.21
MANIC 1.33 0.84
MUSCL 33.78 36.27
NEOPL 3.56 4.19
NERV 30.22 30.92
156
Table A.14a: Descriptive Statistics for Quetiapine Quintile 1 (Continued)
Covariate Success
N=225
No Success
N=1792
OCIRC 9.78 11.44
OTHPSY 5.33 3.24
PERI 0.00 0.17
PERSO 0.89 1.56
PSTG 0.44 1.17
RESPI 34.67 29.46
SABUSE 13.33 11.27
SEXD 0.00 0.17
SKIN 10.67 9.77
SUICDX 3.11 3.74
CONCOMITANT DRUG USE
AD 64.89 67.80
HYPNT 5.78 6.14
SEIZU 28.44 24.94
ARRH* 10.22 6.08
APARK 1.78 2.40
ANXIE 5.78 7.03
MOOD 24.00 19.70
LIPID 8.44 9.43
EPS1 6.22 7.98
EPS2 7.56 5.97
DIABRX 8.44 8.48
DIABDX 15.56 20.26
PTHER 5.33 4.41
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 6.22 5.92
RIS 4.89 5.30
QUE 92.00 91.52
OAP 1.33 0.73
TAP 4.89 5.58
EPISODE TYPE
RESTART 94.22 93.08
SWT 0.44 0.50
AUG 0.89 0.84
LATE_SWT 4.44 5.58
PRIOR HEALTH SERVICE
UTILIZATION

LTC 0.00 0.06
AHOSP* 7.11 3.96
PHOSP 5.33 4.35
REHAB 0.44 0.39
CMHC 71.56 67.47
PSYCH 0.00 0.84
* p<0.05; **p<0.01

157
Table A.14b: Descriptive Statistics for Quetiapine Quintile 2
Covariate Success
N=294
No Success
N=1724
SCHIZOPHRENIA TYPE
SIMPLE 0.00 0.00
DISORG 1.02 0.46
CATA 0.34 0.06
PARAS 6.80 7.19
ACUTE 0.34 0.23
LATENT 0.00 0.06
RESID 0.68 0.87
SAFF 10.54 9.86
OTHNOS 4.08 3.60
UNSPEC 76.19 66.35
AGE
AGE_CAT18 8.50 9.22
AGE_CAT25 14.29 12.76
AGE_CAT35 29.93 28.25
AGE_CAT45 22.45 23.84
AGE_CAT55 17.01 16.42
AGE_CAT65 7.82 9.51
MALE 43.88 44.26
RACE/ETHNICITY
WHITE 54.76 56.50
BLACK 8.84 11.19
HISP* 8.16 5.10
ASIAN 4.76 4.29
OTHER 23.47 22.91
COMORBIDITIES
ACC 3.06 2.32
ANX 11.90 12.70
ARRTDX 4.42 4.06
BIP 17.69 14.62
BLOOD 4.76 7.31
CARDIO 11.90 16.18
CHILD 0.68 0.70
CONGE 1.36 2.09
DEMEN 0.34 0.87
DEP 70.41 67.52
DIGES 18.71 19.66
DIABDX 10.54 8.41
ENDO 9.86 9.69
EPSDX 0.34 0.81
GENTI 18.71 18.50
GLAUC 1.36 1.33
HBP 10.20 14.21
HIV* 0.68 0.06
HLIPID 16.33 17.81
INFEC 96.60 96.52
INJUR 23.13 18.39
MANIC 1.36 1.04
MUSCL 26.87 28.02
NEOPL 4.08 4.41
NERV 30.61 30.86
158
Table A.14b: Descriptive Statistics for Quetiapine Quintile 2 (Continued)
Covariate Success
N=294
No Success
N=1724
OCIRC 11.93 13.69
OTHPSY 5.44 3.54
PERI 0.34 0.35
PERSO** 3.74 1.39
PSTG 0.34 0.93
RESPI 24.15 26.39
SABUSE 9.52 8.00
SEXD 0.00 0.00
SKIN 17.01 14.33
SUICDX 3.74 4.87
CONCOMITANT DRUG USE
AD 63.27 61.14
HYPNT 8.50 6.55
SEIZU 22.11 22.04
ARRH 6.12 6.96
APARK 0.68 1.28
ANXIE 6.80 5.74
MOOD 22.79 22.16
LIPID 9.86 12.65
EPS1 17.01 14.50
EPS2 7.14 5.34
DIABRX 8.84 7.83
DIABDX 21.09 20.48
PTHER 7.82 6.09
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 12.24 11.77
RIS* 13.61 9.05
QUE 79.59 79.93
OAP 2.72 2.03
TAP** 13.61 8.70
EPISODE TYPE
RESTART 75.85 75.81
SWT 2.04 1.91
AUG 2.04 2.90
LATE_SWT 20.07 19.37
PRIOR HEALTH SERVICE
UTILIZATION

LTC 2.04 1.51
AHOSP 5.39 2.72
PHOSP 6.80 5.22
REHAB 0.34 0.23
CMHC 64.63 61.14
PSYCH 1.70 0.81
* p<0.05; **p<0.01

159
Table A.14c: Descriptive Statistics for Quetiapine Quintile 3
Covariate Success
N=385
No Success
N=1628
SCHIZOPHRENIA TYPE
SIMPLE 0.78 0.31
DISORG 0.52 0.18
CATA 0.00 0.12
PARAS 8.31 7.62
ACUTE 0.26 0.37
LATENT 0.00 0.00
RESID 0.00 0.49
SAFF 12.47 10.07
OTHNOS 6.49 5.59
UNSPEC 71.17 75.25
AGE
AGE_CAT18 5.97 7.74
AGE_CAT25 15.58 16.83
AGE_CAT35 27.27 25.86
AGE_CAT45 25.19 25.31
AGE_CAT55 11.95 11.55
AGE_CAT65 14.03 12.71
MALE 42.08 41.58
RACE/ETHNICITY
WHITE 52.99 52.09
BLACK 12.73 14.99
HISP 5.97 8.60
ASIAN 8.31 5.77
OTHER 20.00 18.55
COMORBIDITIES
ACC 1.30 2.21
ANX 12.47 12.53
ARRTDX 4.42 4.42
BIP 17.66 18.18
BLOOD 4.94 6.45
CARDIO 15.32 17.14
CHILD 1.04 0.55
CONGE 2.60 1.84
DEMEN 3.64 2.21
DEP* 73.25 67.63
DIGES 19.48 19.29
DIABDX 10.65 11.92
ENDO 10.39 12.59
EPSDX* 2.86 1.35
GENTI 18.70 18.98
GLAUC 0.52 1.23
HBP 13.51 14.43
HIV 0.26 0.18
HLIPID 13.77 15.66
INFEC 98.18 97.30
INJUR 17.92 19.90
MANIC 1.04 1.29
MUSCL** 37.14 29.67
NEOPL 4.42 5.10
NERV 35.32 31.51
160
Table A.14c: Descriptive Statistics for Quetiapine Quintile 3 (Continued)
Covariate Success
N=385
No Success
N=1628
OCIRC 14.03 12.16
OTHPSY 5.97 5.65
PERI 0.52 0.18
PERSO 1.82 1.47
PSTG 1.30 0.86
RESPI 29.09 28.01
SABUSE 10.39 9.95
SEXD 0.52 0.18
SKIN 10.39 13.14
SUICDX 4.94 4.55
CONCOMITANT DRUG USE
AD** 67.79 59.28
HYPNT 7.53 7.06
SEIZU 21.82 22.85
ARRH 5.97 7.86
APARK 2.34 2.27
ANXIE 6.75 5.96
MOOD 29.31 26.41
LIPID 11.17 10.26
EPS1 19.22 17.81
EPS2 5.45 4.73
DIABRX 9.61 9.77
DIABDX 19.48 18.61
PTHER 7.01 7.00
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 25.97 27.52
RIS 18.44 20.82
QUE 56.10 55.22
OAP 6.23 4.67
TAP 15.84 16.89
EPISODE TYPE
RESTART 22.08 21.74
SWT 13.25 12.53
AUG 15.32 16.77
LATE_SWT 49.35 48.96
PRIOR HEALTH SERVICE
UTILIZATION

LTC 7.79 7.43
AHOSP 6.23 5.71
PHOSP 7.27 7.68
REHAB 0.00 0.12
CMHC 61.56 61.43
PSYCH 1.30 1.04
* p<0.05; **p<0.01

161
Table A.14d: Descriptive Statistics for Quetiapine Quintile 4
Covariate Success
N=538
No Success
N=1469
SCHIZOPHRENIA TYPE
SIMPLE 0.19 0.14
DISORG 0.56 0.14
CATA 0.00 0.20
PARAS 10.97 10.69
ACUTE 0.56 0.34
LATENT 0.00 0.00
RESID 0.74 0.48
SAFF 11.52 11.84
OTHNOS 5.95 6.26
UNSPEC 69.52 69.91
AGE
AGE_CAT18 5.02 7.15
AGE_CAT25 17.66 17.77
AGE_CAT35 28.44 27.84
AGE_CAT45 25.65 23.28
AGE_CAT55 9.48 11.64
AGE_CAT65 13.75 12.32
MALE 47.03 45.95
RACE/ETHNICITY
WHITE 60.04 59.95
BLACK 11.15 9.73
HISP 3.90 4.22
ASIAN 3.72 5.72
OTHER 21.19 21.38
COMORBIDITIES
ACC 0.19 0.82
ANX 9.67 9.87
ARRTDX 4.28 5.17
BIP 15.06 16.07
BLOOD 5.39 5.99
CARDIO 14.87 15.32
CHILD 0.37 1.09
CONGE 1.30 1.84
DEMEN 2.97 1.84
DEP* 63.01 62.83
DIGES 16.17 19.06
DIABDX 9.85 8.44
ENDO 10.97 12.46
EPSDX 0.93 0.82
GENTI 14.13 17.22
GLAUC 0.74 0.48
HBP 12.45 12.59
HIV 0.19 1.14
HLIPID 15.43 13.48
INFEC 96.65 97.28
INJUR 16.36 18.38
MANIC 1.49 1.02
MUSCL 25.84 26.82
NEOPL 3.16 5.04
NERV 28.07 30.02
162
Table A.14d: Descriptive Statistics for Quetiapine Quintile 4 (Continued)
Covariate Success
N=538
No Success
N=1469
OCIRC 11.90 13.07
OTHPSY 3.53 4.36
PERI 0.00 0.27
PERSO 1.30 1.29
PSTG 0.56 0.54
RESPI 27.14 27.23
SABUSE 6.13 7.76
SEXD 0.00 0.14
SKIN 13.57 14.70
SUICDX 2.23 2.86
CONCOMITANT DRUG USE
AD 56.69 57.59
HYPNT* 4.46 7.22
SEIZU 22.49 20.63
ARRH* 5.20 8.44
APARK * 3.35 1.77
ANXIE 4.28 4.70
MOOD 29.00 26.28
LIPID 10.97 9.60
EPS1 30.48 31.25
EPS2 4.09 5.85
DIABRX 8.36 6.94
DIABDX 19.33 19.20
PTHER 9.53 11.52
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 40.71 42.34
RIS 31.97 30.02
QUE 53.35 49.22
OAP 5.20 6.60
TAP 27.88 27.09
EPISODE TYPE
RESTART 5.76 4.97
SWT 30.30 31.65
AUG 40.33 38.05
LATE_SWT 23.61 35.32
PRIOR HEALTH SERVICE
UTILIZATION

LTC 8.36 6.54
AHOSP 3.72 3.81
PHOSP* 5.02 5.65
REHAB 0.19 0.07
CMHC 57.81 59.70
PSYCH 1.30 1.70
* p<0.05; **p<0.01

163
Table A.14e: Descriptive Statistics for Quetiapine Quintile 5
Covariate Success
N=724
No Success
N=1300
SCHIZOPHRENIA TYPE
SIMPLE 0.28 0.46
DISORG 0.69 1.38
CATA 0.14 0.23
PARAS 16.71 17.00
ACUTE 0.00 0.15
LATENT 0.00 0.08
RESID 0.83 0.69
SAFF 7.87 9.31
OTHNOS 6.08 7.08
UNSPEC 67.40 63.62
AGE
AGE_CAT18 2.90 2.92
AGE_CAT25 8.01 8.85
AGE_CAT35 24.03 24.31
AGE_CAT45 22.10 24.46
AGE_CAT55 21.27 18.92
AGE_CAT65 21.69 20.54
MALE 51.52 54.15
RACE/ETHNICITY
WHITE 68.09 67.54
BLACK 4.14 2.77
HISP 2.07 2.38
ASIAN 2.07 2.31
OTHER 23.62 25.00
COMORBIDITIES
ACC 0.83 0.31
ANX 7.87 6.92
ARRTDX 7.60 6.23
BIP 6.63 7.46
BLOOD 8.29 6.08
CARDIO 17.13 15.46
CHILD 0.97 0.62
CONGE 2.76 1.69
DEMEN 10.64 8.92
DEP 51.24 55.69
DIGES 17.68 15.54
DIABDX 7.32 7.85
ENDO 10.91 10.69
EPSDX 2.35 2.46
GENTI 17.13 15.31
GLAUC 0.69 0.23
HBP 14.64 13.00
HIV 0.00 0.15
HLIPID 16.85 17.46
INFEC 98.48 97.85
INJUR 17.54 16.00
MANIC 0.41 1.08
MUSCL 20.99 21.08
NEOPL 4.42 3.46
NERV 37.57 34.15
164
Table A.14e: Descriptive Statistics for Quetiapine Quintile 5 (Continued)
Covariate Success
N=724
No Success
N=1300
OCIRC 15.33 15.62
OTHPSY 4.42 5.77
PERI 0.14 0.00
PERSO 0.83 1.23
PSTG 0.14 0.15
RESPI 23.62 23.23
SABUSE 2.07 3.54
SEXD 0.14 0.08
SKIN 22.79 20.00
SUICDX 1.38 0.69
CONCOMITANT DRUG USE
AD* 46.55 51.23
HYPNT 6.08 5.85
SEIZU 18.92 19.85
ARRH* 11.33 8.38
APARK 2.07 3.15
ANXIE 3.59 4.23
MOOD 25.41 27.77
LIPID 10.64 10.77
EPS1 49.96 49.31
EPS2 4.28 4.69
DIABRX 6.77 8.54
DIABDX 20.17 19.46
PTHER 10.91 12.38
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 38.81 40.31
RIS 33.84 36.69
QUE* 35.08 39.77
OAP 5.39 6.15
TAP 37.57 37.62
EPISODE TYPE
RESTART 2.49 2.62
SWT 47.93 47.00
AUG 35.64 37.85
LATE_SWT 13.95 12.54
PRIOR HEALTH SERVICE
UTILIZATION

LTC* 26.52 23.69
AHOSP 4.97 4.31
PHOSP 2.76 3.85
REHAB 0.55 0.15
CMHC 45.44 19.31
PSYCH 2.62 2.00
* p<0.05; **p<0.01

165
Table A.15a: Descriptive Statistics for Low Potency TAPs Quintile 1
Covariate Success
N=381
No Success
N=4533
SCHIZOPHRENIA TYPE
SIMPLE 0.26 0.13
DISORG 0.00 0.11
CATA 0.26 0.07
PARAS 2.89 3.68
ACUTE 0.00 0.07
LATENT 0.00 0.09
RESID 0.79 0.29
SAFF 3.94 4.90
OTHNOS 2.62 2.25
UNSPEC 89.24 88.42
AGE
AGE_CAT18* 2.89 5.21
AGE_CAT25 13.12 15.93
AGE_CAT35 19.69 22.41
AGE_CAT45 17.85 16.99
AGE_CAT55 11.29 11.45
AGE_CAT65** 35.17 28.02
MALE 36.22 35.67
RACE/ETHNICITY
WHITE 29.40 25.19
BLACK 23.88 26.03
HISP 8.92 11.78
ASIAN 9.71 12.38
OTHER 28.08 24.62
COMORBIDITIES
ACC 2.10 2.98
ANX 13.91 15.05
ARRTDX 6.30 5.98
BIP 10.50 10.90
BLOOD 5.77 7.30
CARDIO** 35.43 26.72
CHILD 0.52 0.22
CONGE 1.05 1.48
DEMEN 0.79 0.44
DEP 57.48 56.74
DIGES 25.72 27.42
DIABDX 12.60 13.32
ENDO 11.81 12.02
EPSDX 0.52 0.71
GENTI 25.20 29.27
GLAUC 2.67 1.84
HBP** 29.13 22.79
HIV 0.00 0.15
HLIPID 15.75 13.17
INFEC 98.69 98.19
INJUR 25.20 28.37
MANIC 0.52 0.68
MUSCL 40.42 45.25
NEOPL 9.19 7.41
NERV 35.17 37.88
166
Table A.15a: Descriptive Statistics for Low Potency TAPs Quintile 1 (Continued)
Covariate Success
N=381
No Success
N=4533
OCIRC 24.67 20.54
OTHPSY 4.46 3.75
PERI 0.00 0.15
PERSO 1.57 1.63
PSTG 0.24 0.26
RESPI 43.04 15.27
SABUSE 11.55 9.71
SEXD 0.15 0.52
SKIN 14.96 13.08
SUICDX 4.46 5.10
CONCOMITANT DRUG USE
AD 47.32 49.61
HYPNT 3.09 2.62
SEIZU 10.76 10.26
ARRH 11.55 9.82
APARK* 2.10 0.84
ANXIE 8.66 8.69
MOOD 9.71 11.71
LIPID* 12.07 8.98
EPS1 16.27 14.32
EPS2 5.25 6.82
DIABRX 10.76 11.10
DIABDX 26.25 25.97
PTHER 24.67 21.51
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 7.61 5.49
RIS 8.14 7.21
QUE 2.89 2.49
OAP 3.41 3.46
TAP 87.14 85.07
EPISODE TYPE
RESTART 66.67 71.37
SWT* 5.77 3.66
AUG 13.39 12.16
LATE_SWT 14.17 12.82
PRIOR HEALTH SERVICE
UTILIZATION

LTC* 8.92 5.01
AHOSP 8.92 2.82
PHOSP 1.05 1.54
REHAB 0.79 0.38
CMHC 14.44 12.62
PSYCH 2.36 2.67
* p<0.05; **p<0.01

167
Table A.15b: Descriptive Statistics for Low Potency TAPs Quintile 2
Covariate Success
N=540
No Success
N=4373
SCHIZOPHRENIA TYPE
SIMPLE 0.00 0.18
DISORG 0.19 0.16
CATA 0.00 0.05
PARAS 4.26 3.80
ACUTE 0.00 0.11
LATENT 0.19 0.02
RESID 0.56 0.23
SAFF* 5.74 3.82
OTHNOS 2.78 2.72
UNSPEC 86.30 88.91
AGE
AGE_CAT18 5.19 4.92
AGE_CAT25 14.07 12.65
AGE_CAT35 19.26 20.63
AGE_CAT45 15.19 16.30
AGE_CAT55 12.59 13.22
AGE_CAT65 33.70 32.29
MALE 38.70 38.10
RACE/ETHNICITY
WHITE 43.15 39.10
BLACK 18.89 19.41
HISP 4.63 4.99
ASIAN 2.41 2.20
OTHER 34.30 30.93
COMORBIDITIES
ACC 2.04 2.33
ANX 6.67 7.45
ARRTDX 4.44 5.08
BIP 7.41 7.16
BLOOD 4.26 5.83
CARDIO 17.59 20.31
CHILD 0.19 0.30
CONGE 1.30 1.88
DEMEN 1.48 1.01
DEP 45.19 44.68
DIGES 17.96 20.54
DIABDX 11.30 11.89
ENDO 7.96 9.70
EPSDX 0.19 0.64
GENTI 19.44 19.83
GLAUC 2.22 1.74
HBP 15.00 17.04
HIV 0.00 0.09
HLIPID 12.22 11.85
INFEC 96.11 97.51
INJUR 22.41 21.63
MANIC 0.00 0.39
MUSCL** 28.52 34.07
NEOPL 4.81 6.11
NERV 33.94 35.00
168
Table A.15b: Descriptive Statistics for Low Potency TAPs Quintile 2 (Continued)
Covariate Success
N=540
No Success
N=4373
OCIRC 16.67 17.78
OTHPSY 2.78 2.95
PERI 0.19 0.02
PERSO 1.45 1.30
PSTG* 0.37 0.05
RESPI 33.52 34.78
SABUSE 2.41 3.38
SEXD 0.00 0.11
SKIN 12.22 13.83
SUICDX 2.41 2.22
CONCOMITANT DRUG USE
AD 39.81 38.71
HYPNT 2.22 2.52
SEIZU 11.30 11.00
ARRH 9.06 8.89
APARK 1.30 1.07
ANXIE 2.96 3.54
MOOD 9.44 10.75
LIPID 8.70 8.64
EPS1 17.41 14.86
EPS2 5.74 4.87
DIABRX 10.37 10.88
DIABDX 19.44 21.91
PTHER 17.78 16.99
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 3.70 4.25
RIS 5.93 5.37
QUE 1.67 1.05
OAP 2.22 2.08
TAP 86.30 83.42
EPISODE TYPE
RESTART 73.70 77.29
SWT 4.44 4.32
AUG* 12.41 9.54
LATE_SWT 9.44 8.85
PRIOR HEALTH SERVICE
UTILIZATION

LTC** 9.26 6.06
AHOSP 4.63 4.21
PHOSP 0.74 0.66
REHAB 0.19 0.39
CMHC 10.93 10.06
PSYCH 2.22 2.20
* p<0.05; **p<0.01

169
Table A.15c: Descriptive Statistics for Low Potency TAPs Quintile 3
Covariate Success
N=721
No Success
N=4193
SCHIZOPHRENIA TYPE
SIMPLE 0.28 0.29
DISORG 0.00 0.12
CATA 0.28 0.14
PARAS 4.02 5.25
ACUTE 0.00 0.12
LATENT 0.14 0.07
RESID 0.55 0.55
SAFF 3.47 3.29
OTHNOS 4.02 3.17
UNSPEC 87.24 87.00
AGE
AGE_CAT18 5.83 5.03
AGE_CAT25 12.48 13.43
AGE_CAT35 21.78 21.13
AGE_CAT45 16.50 19.48
AGE_CAT55 14.29 14.09
AGE_CAT65 29.13 16.83
MALE 42.44 45.00
RACE/ETHNICITY
WHITE 56.17 56.67
BLACK 11.51 12.95
HISP 2.08 2.00
ASIAN 0.72 0.42
OTHER 29.82 27.67
COMORBIDITIES
ACC 1.39 1.55
ANX 5.13 4.60
ARRTDX 3.19 3.67
BIP 5.96 6.11
BLOOD 5.69 4.60
CARDIO 14.56 15.19
CHILD* 0.97 0.33
CONGE 1.66 1.65
DEMEN 3.05 2.22
DEP* 37.86 37.63
DIGES 15.95 16.60
DIABDX 10.26 9.42
ENDO 9.57 7.99
EPSDX 0.69 0.57
GENTI 16.09 14.19
GLAUC 1.39 0.93
HBP 12.21 12.74
HIV 0.00 0.05
HLIPID 8.88 9.40
INFEC 97.78 97.38
INJUR 17.75 16.48
MANIC 0.28 0.33
MUSCL 24.83 25.11
NEOPL 4.58 4.34
NERV 32.73 32.98
170
Table A.15c: Descriptive Statistics for Low Potency TAPs Quintile 3 (Continued)
Covariate Success
N=721
No Success
N=4193
OCIRC 15.53 15.34
OTHPSY 3.88 3.10
PERI 0.00 0.10
PERSO 1.11 0.95
PSTG 0.00 0.00
RESPI 27.60 26.26
SABUSE 0.69 1.65
SEXD 0.00 0.07
SKIN 16.50 14.69
SUICDX 1.25 1.24
CONCOMITANT DRUG USE
AD 33.84 33.22
HYPNT 1.11 2.05
SEIZU 10.96 12.64
ARRH 5.83 6.27
APARK 1.53 1.41
ANXIE 1.39 1.38
MOOD 11.23 11.95
LIPID 5.41 6.49
EPS1* 15.26 19.17
EPS2 3.33 3.58
DIABRX 8.04 9.11
DIABDX 13.87 14.64
PTHER 18.86 17.12
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 3.61 6.84
RIS* 2.91 4.94
QUE 0.42 0.91
OAP* 0.55 1.29
TAP* 85.44 81.85
EPISODE TYPE
RESTART 78.64 78.34
SWT 5.41 4.15
AUG 7.63 9.90
LATE_SWT 8.32 7.61
PRIOR HEALTH SERVICE
UTILIZATION

LTC 11.93 10.64
AHOSP 2.22 2.17
PHOSP 0.28 0.48
REHAB 0.42 0.31
CMHC 9.85 9.02
PSYCH 2.91 2.72
* p<0.05; **p<0.01

171
Table A.15d: Descriptive Statistics for Low Potency TAPs Quintile 4
Covariate Success
N=940
No Success
N=3976
SCHIZOPHRENIA TYPE
SIMPLE 1.06 0.65
DISORG 0.32 0.20
CATA 0.00 0.08
PARAS 6.28 5.86
ACUTE 0.00 0.18
LATENT 0.11 0.10
RESID* 1.28 0.60
SAFF 3.51 3.62
OTHNOS 5.21 5.16
UNSPEC 82.23 83.55
AGE
AGE_CAT18 5.64 4.68
AGE_CAT25 12.45 13.20
AGE_CAT35 25.11 26.38
AGE_CAT45 23.72 23.16
AGE_CAT55 14.79 15.64
AGE_CAT65 18.30 16.93
MALE 52.87 53.14
RACE/ETHNICITY
WHITE 68.94 70.57
BLACK 7.55 6.41
HISP 0.55 1.06
ASIAN 0.21 0.05
OTHER 22.23 22.41
COMORBIDITIES
ACC 1.49 1.28
ANX 5.00 4.00
ARRTDX 3.30 3.42
BIP 3.72 4.07
BLOOD 4.15 4.95
CARDIO 10.11 10.66
CHILD 0.74 0.80
CONGE 2.13 2.06
DEMEN 4.79 5.33
DEP 28.94 30.53
DIGES 12.45 12.78
DIABDX 7.45 8.63
ENDO 7.77 7.87
EPSDX 0.53 0.83
GENTI 8.94 10.31
GLAUC 0.21 0.63
HBP 8.30 9.05
HIV 0.11 0.08
HLIPID 7.98 9.46
INFEC 96.49 96.96
INJUR 14.15 12.93
MANIC 0.11 0.15
MUSCL 17.87 19.74
NEOPL 2.87 3.14
NERV 31.70 32.24
172
Table A.15d: Descriptive Statistics for Low Potency TAPs Quintile 4 (Continued)
Covariate Success
N=940
No Success
N=3976
OCIRC 11.70 10.79
OTHPSY 2.45 2.72
PERI 0.11 0.05
PERSO 0.64 1.36
PSTG 0.00 0.00
RESPI 18.72 20.42
SABUSE 0.11 0.35
SEXD 0.00 0.08
SKIN 17.66 17.76
SUICDX 1.38 0.96
CONCOMITANT DRUG USE
AD 26.28 27.79
HYPNT 1.60 1.56
SEIZU 18.30 15.74
ARRH 5.21 5.36
APARK 0.96 1.13
ANXIE 0.96 0.63
MOOD 14.57 12.47
LIPID 5.64 6.09
EPS1 21.28 21.91
EPS2 2.45 2.59
DIABRX 7.34 7.17
DIABDX 10.64 9.98
PTHER 15.53 17.98
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 2.98 3.40
RIS 4.57 4.70
QUE 0.32 0.45
OAP 0.43 0.30
TAP 79.15 80.73
EPISODE TYPE
RESTART 80.64 78.45
SWT 4.79 5.38
AUG 10.11 10.34
LATE_SWT* 4.47 5.84
PRIOR HEALTH SERVICE
UTILIZATION

LTC 13.09 11.44
AHOSP 1.17 1.84
PHOSP 0.32 0.28
REHAB 0.21 0.20
CMHC 8.30 8.98
PSYCH 4.15 3.87
* p<0.05; **p<0.01

173
Table A.15e: Descriptive Statistics for Low Potency TAPs Quintile 5
Covariate Success
N=1440
No Success
N=3472
SCHIZOPHRENIA TYPE
SIMPLE 1.60 1.47
DISORG 0.28 0.14
CATA 0.00 0.12
PARAS 6.25 6.39
ACUTE 0.69 0.40
LATENT 0.00 0.03
RESID 1.74 2.22
SAFF 1.88 2.25
OTHNOS 10.76 11.18
UNSPEC 76.81 75.81
AGE
AGE_CAT18 4.24 4.29
AGE_CAT25 9.93 9.65
AGE_CAT35 24.40 26.94
AGE_CAT45 30.97 30.21
AGE_CAT55 17.57 16.85
AGE_CAT65** 10.35 14.60
MALE 61.46 59.53
RACE/ETHNICITY
WHITE 80.90 81.51
BLACK* 3.82 2.65
HISP 0.56 0.20
ASIAN 0.00 0.06
OTHER 14.72 15.58
COMORBIDITIES
ACC 1.74 1.24
ANX 4.17 4.55
ARRTDX 4.03 3.05
BIP 1.94 2.48
BLOOD 6.81 5.50
CARDIO 7.08 8.21
CHILD 1.94 1.79
CONGE 4.31 3.40
DEMEN 22.36 20.22
DEP 18.61 19.50
DIGES* 14.03 11.49
DIABDX 5.63 6.48
ENDO 9.03 8.99
EPSDX 1.46 0.86
GENTI 8.68 7.95
GLAUC 0.35 0.32
HBP 5.90 6.71
HIV 0.00 0.03
HLIPID 7.22 7.57
INFEC 97.64 96.89
INJUR 11.03 10.14
MANIC 0.35 0.14
MUSCL** 20.83 16.10
NEOPL 1.94 2.25
NERV 35.49 34.97
174
Table A.15e: Descriptive Statistics for Low Potency TAPs Quintile 5 (Continued)
Covariate Success
N=1440
No Success
N=3472
OCIRC 9.24 11.35
OTHPSY* 3.19 4.41
PERI 0.42 0.35
PERSO 1.74 1.38
PSTG 0.00 0.03
RESPI 17.57 16.07
SABUSE 0.00 0.20
SEXD 0.00 0.09
SKIN 29.10 28.51
SUICDX 0.63 0.89
CONCOMITANT DRUG USE
AD 16.81 18.00
HYPNT 1.32 1.58
SEIZU 23.65 23.68
ARRH 3.54 3.92
APARK* 0.69 1.35
ANXIE 0.35 0.43
MOOD 16.81 17.37
LIPID 4.86 5.21
EPS1 33.82 32.69
EPS2 2.29 2.19
DIABRX 4.17 4.52
DIABDX 9.03 7.95
PTHER 21.53 21.28
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 2.36 2.88
RIS 3.47 3.25
QUE 0.07 0.26
OAP 0.14 0.17
TAP** 72.64 76.58
EPISODE TYPE
RESTART** 80.14 76.64
SWT 6.11 6.97
AUG* 10.14 12.24
LATE_SWT 3.61 4.15
PRIOR HEALTH SERVICE
UTILIZATION

LTC 16.46 18.63
AHOSP 0.76 1.15
PHOSP 0.14 0.20
REHAB 0.07 0.12
CMHC 8.68 9.88
PSYCH 7.57 7.57
* p<0.05; **p<0.01

175
Table A.16a: Descriptive Statistics for Medium Potency TAPs Quintile 1
Covariate Success
N=58
No Success
N=11199
SCHIZOPHRENIA TYPE
SIMPLE 0.00 0.00
DISORG 0.00 0.01
CATA 0.00 0.00
PARAS 0.00 0.04
ACUTE 0.00 0.07
LATENT 0.00 0.00
RESID 0.00 0.00
SAFF 3.45 0.24
OTHNOS 1.72 0.17
UNSPEC 94.83 99.46
AGE
AGE_CAT18 1.72 6.82
AGE_CAT25 12.07 15.22
AGE_CAT35 24.14 19.47
AGE_CAT45 17.24 15.25
AGE_CAT55 10.34 11.24
AGE_CAT65 34.48 31.99
MALE 15.52 17.81
RACE/ETHNICITY
WHITE 41.38 37.21
BLACK 15.52 15.08
HISP 25.86 25.13
ASIAN 3.45 4.13
OTHER 13.79 18.45
COMORBIDITIES
ACC 5.17 4.86
ANX 29.31 10.30
ARRTDX 18.97 12.09
BIP 5.17 1.17
BLOOD 32.76 29.89
CARDIO 51.72 35.44
CHILD 0.00 0.06
CONGE 8.62 5.52
DEMEN 0.00 0.54
DEP 79.31 36.99
DIGES 72.41 64.41
DIABDX 31.03 26.90
ENDO 31.03 34.23
EPSDX 0.00 0.65
GENTI 62.07 60.23
GLAUC 3.45 4.30
HBP 36.21 28.73
HIV 0.00 0.96
HLIPID 24.14 20.71
INFEC 100.00 98.30
INJUR 58.62 45.66
MANIC 0.00 0.03
MUSCL 72.41 66.51
NEOPL 37.99 29.53
NERV 53.45 57.05
176
Table A.16a: Descriptive Statistics for Medium Potency TAPs Quintile 1
(Continued)
Covariate Success
N=58
No Success
N=11199
OCIRC 44.83 41.52
OTHPSY 5.17 1.37
PERI 0.00 1.17
PERSO 0.00 0.36
PSTG 6.90 7.90
RESPI 74.14 60.41
SABUSE 25.86 6.97
SEXD 0.00 0.13
SKIN 18.97 24.56
SUICDX 8.62 1.29
CONCOMITANT DRUG USE
AD 63.79 30.91
HYPNT 8.62 8.05
SEIZU 29.31 22.03
ARRH 13.79 13.72
APARK 1.72 1.22
ANXIE 18.97 7.17
MOOD 3.45 2.22
LIPID 13.79 15.01
EPS1 5.17 1.78
EPS2 10.34 10.21
DIABRX 21.14 25.86
DIABDX 46.55 44.83
PTHER 5.17 2.97
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 5.17 1.05
RIS 1.72 1.06
QUE 1.72 0.40
OAP 0.00 0.14
TAP 70.69 44.91
EPISODE TYPE
RESTART 87.93 91.70
SWT 0.00 0.21
AUG 8.62 5.51
LATE_SWT 3.45 2.57
PRIOR HEALTH SERVICE
UTILIZATION

LTC 8.62 8.84
AHOSP 20.69 23.28
PHOSP 3.45 0.26
REHAB 1.72 0.79
CMHC 8.62 3.54
PSYCH 3.45 1.00
* p<0.05; **p<0.01

177
Table A.16b: Descriptive Statistics for Medium Potency TAPs Quintile 2
Covariate Success
N=161
No Success
N=11096
SCHIZOPHRENIA TYPE
SIMPLE 0.00 0.01
DISORG 0.00 0.02
CATA 0.00 0.01
PARAS 0.00 0.11
ACUTE 0.00 0.05
LATENT 0.00 0.00
RESID 0.00 0.01
SAFF 0.62 0.43
OTHNOS 0.00 0.31
UNSPEC 99.38 99.06
AGE
AGE_CAT18 2.48 2.43
AGE_CAT25 4.35 10.14
AGE_CAT35 19.26 19.62
AGE_CAT45 20.50 17.48
AGE_CAT55 15.53 14.21
AGE_CAT65 37.89 36.08
MALE 22.98 20.33
RACE/ETHNICITY
WHITE 42.86 45.15
BLACK 20.50 15.48
HISP 10.56 12.34
ASIAN 1.83 3.98
OTHER 21.74 23.65
COMORBIDITIES
ACC 4.97 3.29
ANX 9.32 10.18
ARRTDX 6.83 8.92
BIP 3.11 1.34
BLOOD 16.32 22.98
CARDIO 39.13 32.83
CHILD 0.00 0.05
CONGE 3.73 4.06
DEMEN 2.48 0.91
DEP 73.29 54.34
DIGES 50.93 46.75
DIABDX 17.39 21.43
ENDO 27.33 21.35
EPSDX 1.24 0.51
GENTI 39.75 43.26
GLAUC 5.59 2.79
HBP 32.92 26.89
HIV 0.62 0.36
HLIPID 16.15 19.09
INFEC 98.14 97.80
INJUR 40.99 33.27
MANIC 0.00 0.05
MUSCL 68.32 57.69
NEOPL 22.36 17.30
NERV 52.80 51.28
178
Table A.16b: Descriptive Statistics for Medium Potency TAPs Quintile 2
(Continued)
Covariate Success
N=161
No Success
N=11096
OCIRC 33.54 32.86
OTHPSY 3.73 1.65
PERI 0.62 0.43
PERSO 1.86 0.55
PSTG 1.86 1.33
RESPI 55.28 50.86
SABUSE 6.83 5.18
SEXD 0.00 0.06
SKIN 32.30 21.83
SUICDX 1.86 1.01
CONCOMITANT DRUG USE
AD 64.60 45.18
HYPNT 5.59 5.89
SEIZU 18.01 17.61
ARRH 10.56 13.77
APARK 1.86 1.31
ANXIE 6.89 7.48
MOOD 4.97 2.95
LIPID 13.04 14.19
EPS1 5.59 3.27
EPS2 6.21 7.24
DIABRX 16.77 18.29
DIABDX 44.72 36.72
PTHER 8.70 4.10
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 2.48 1.16
RIS 1.24 1.35
QUE 0.62 0.35
OAP 0.00 0.19
TAP 70.81 63.79
EPISODE TYPE
RESTART 85.09 90.84
SWT 0.62 0.62
AUG 4.97 3.22
LATE_SWT 9.32 5.32
PRIOR HEALTH SERVICE
UTILIZATION

LTC 11.18 9.08
AHOSP 17.39 13.73
PHOSP 0.62 0.37
REHAB 1.24 0.50
CMHC* 1.24 4.70
PSYCH 1.24 1.25
* p<0.05; **p<0.01

179
Table A.16c: Descriptive Statistics for Medium Potency TAPs Quintile 3
Covariate Success
N=329
No Success
N=10927
SCHIZOPHRENIA TYPE
SIMPLE 0.00 0.02
DISORG 0.00 0.02
CATA 0.00 0.01
PARAS 0.91 0.25
ACUTE 0.00 0.06
LATENT 0.00 0.02
RESID 0.00 0.03
SAFF 1.52 0.60
OTHNOS 0.30 0.34
UNSPEC 97.26 98.65
AGE
AGE_CAT18 1.52 1.54
AGE_CAT25 10.03 7.92
AGE_CAT35 14.59 18.77
AGE_CAT45 22.80 19.74
AGE_CAT55 18.54 16.24
AGE_CAT65 35.52 35.80
MALE 21.88 23.57
RACE/ETHNICITY
WHITE 72.86 47.76
BLACK 23.40 17.12
HISP 5.78 7.07
ASIAN 3.65 3.81
OTHER 24.32 24.23
COMORBIDITIES
ACC 2.13 2.56
ANX 12.16 10.26
ARRTDX 6.99 4.15
BIP 3.04 2.28
BLOOD 8.51 9.22
CARDIO 33.43 30.50
CHILD 0.00 0.10
CONGE 4.26 3.09
DEMEN 0.61 1.20
DEP 76.60 67.10
DIGES 35.87 34.26
DIABDX 17.63 18.07
ENDO 13.07 13.98
EPSDX 0.61 0.43
GENTI 31.91 31.75
GLAUC 2.74 2.04
HBP 26.44 25.28
HIV 0.00 0.21
HLIPID 20.06 17.02
INFEC 97.26 97.60
INJUR 24.62 25.99
MANIC 0.30 0.14
MUSCL 56.23 50.52
NEOPL 10.94 10.32
NERV 44.68 44.71
180
Table A.16c: Descriptive Statistics for Medium Potency TAPs Quintile 3
(Continued)
Covariate Success
N=329
No Success
N=10927
OCIRC 28.27 26.37
OTHPSY 0.91 1.64
PERI 0.00 0.22
PERSO 2.56 0.70
PSTG 0.61 0.29
RESPI 45.59 43.11
SABUSE 4.86 4.03
SEXD 0.00 0.14
SKIN 18.54 19.04
SUICDX 2.74 1.17
CONCOMITANT DRUG USE
AD 67.17 56.50
HYPNT 3.95 4.24
SEIZU 14.29 14.62
ARRH 14.59 12.46
APARK 1.82 1.27
ANXIE 9.42 7.22
MOOD 4.26 3.64
LIPID 14.29 12.93
EPS1 5.17 4.93
EPS2 7.90 6.03
DIABRX 15.81 16.24
DIABDX 33.74 31.97
PTHER 8.81 5.34
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 2.43 1.51
RIS 1.50 2.74
QUE 0.91 0.44
OAP 0.30 0.26
TAP 77.20 76.28
EPISODE TYPE
RESTART 88.45 89.76
SWT 1.22 1.06
AUG 4.56 3.94
LATE_SWT 5.78 5.24
PRIOR HEALTH SERVICE
UTILIZATION

LTC 6.99 9.15
AHOSP 8.21 8.70
PHOSP 0.61 0.38
REHAB 0.61 0.42
CMHC 5.47 4.82
PSYCH 2.13 1.65
* p<0.05; **p<0.01

181
Table A.16d: Descriptive Statistics for Medium Potency TAPs Quintile 4
Covariate Success
N=595
No Success
N=10656
SCHIZOPHRENIA TYPE
SIMPLE 0.00 0.03
DISORG 0.00 0.06
CATA 0.00 0.00
PARAS 1.01 0.69
ACUTE 0.00 0.05
LATENT 0.00 0.03
RESID 0.00 0.06
SAFF 0.50 1.53
OTHNOS 1.18 0.72
UNSPEC 97.31 96.84
AGE
AGE_CAT18 2.02 1.34
AGE_CAT25 6.05 6.95
AGE_CAT35 19.39 17.90
AGE_CAT45 22.18 21.02
AGE_CAT55 17.65 19.29
AGE_CAT65 32.77 33.49
MALE 27.39 27.75
RACE/ETHNICITY
WHITE 46.72 48.55
BLACK 19.50 17.91
HISP 3.03 2.62
ASIAN 4.54 3.64
OTHER 26.22 27.28
COMORBIDITIES
ACC 2.52 1.84
ANX 6.55 8.65
ARRTDX 4.54 5.11
BIP 2.35 2.74
BLOOD 4.37 5.31
CARDIO 26.22 28.14
CHILD 0.17 0.13
CONGE 0.84 2.05
DEMEN 1.18 1.58
DEP 76..81 77.83
DIGES 17.98 21.54
DIABDX 14.96 15.06
ENDO 9.35 8.57
EPSDX 0.50 0.34
GENTI 22.86 22.11
GLAUC 0.67 1.22
HBP 22.86 23.94
HIV 0.17 0.12
HLIPID 15.29 15.49
INFEC 98.66 97.18
INJUR 19.50 19.95
MANIC 0.00 0.15
MUSCL 40.84 41.19
NEOPL 3.36 4.91
NERV 38.66 39.00
182
Table A.16d: Descriptive Statistics for Medium Potency TAPs Quintile 4
(Continued)
Covariate Success
N=595
No Success
N=10656
OCIRC 18.15 19.32
OTHPSY 2.35 1.81
PERI 0.00 0.21
PERSO 0.03 0.60
PSTG 0.00 0.12
RESPI 36.30 36.65
SABUSE 1.34 3.26
SEXD 0.17 0.10
SKIN 16.47 16.23
SUICDX 1.34 1.34
CONCOMITANT DRUG USE
AD 66.55 67.69
HYPNT 2.35 3.06
SEIZU 10.42 11.76
ARRH 10.08 11.55
APARK 1.34 1.18
ANXIE 4.20 5.71
MOOD 5.88 4.81
LIPID 11.76 11.62
EPS1 9.41 7.98
EPS2 4.71 4.08
DIABRX 15.46 14.40
DIABDX 23.53 24.13
PTHER 8.74 7.45
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 1.68 1.71
RIS 2.52 1.55
QUE 0.00 0.59
OAP 0.50 0.38
TAP 82.18 85.09
EPISODE TYPE
RESTART 87.90 89.05
SWT 2.18 1.24
AUG 4.37 4.69
LATE_SWT 5.55 5.02
PRIOR HEALTH SERVICE
UTILIZATION

LTC 8.07 7.58
AHOSP 3.87 5.51
PHOSP 0.17 0.38
REHAB 0.17 0.35
CMHC 4.54 6.16
PSYCH 2.35 1.59
* p<0.05; **p<0.01

183
Table A.16e: Descriptive Statistics for Medium Potency TAPs Quintile 5
Covariate Success
N=1039
No Success
N=10230
SCHIZOPHRENIA TYPE
SIMPLE 0.19 0.30
DISORG 0.00 0.08
CATA 0.10 0.10
PARAS 6.58 3.47
ACUTE 0.10 0.14
LATENT 0.00 0.00
RESID 0.97 0.31
SAFF 5.81 3.99
OTHNOS 4.55 3.21
UNSPEC 81.70 88.41
AGE
AGE_CAT18 1.06 1.07
AGE_CAT25 9.39 7.78
AGE_CAT35 22.94 19.85
AGE_CAT45 27.11 27.79
AGE_CAT55 21.97 22.86
AGE_CAT65 17.52 20.65
MALE 41.14 37.73
RACE/ETHNICITY
WHITE 56.44 52.60
BLACK 10.75 14.98
HISP 1.65 0.79
ASIAN 2.03 2.86
OTHER 29.14 28.77
COMORBIDITIES
ACC 1.16 1.35
ANX 7.26 8.49
ARRTDX 2.71 3.44
BIP 6.58 5.70
BLOOD 2.23 2.52
CARDIO 18.78 22.20
CHILD 0.00 0.13
CONGE 1.06 1.37
DEMEN 2.81 2.63
DEP 73.57 84.19
DIGES 10.55 10.47
DIABDX 9.58 10.41
ENDO 5.32 5.03
EPSDX 0.29 0.39
GENTI 12.49 13.16
GLAUC 0.29 0.60
HBP 16.94 19.93
HIV 0.10 0.02
HLIPID 11.23 12.19
INFEC 96.13 96.91
INJUR 10.04 11.40
MANIC 0.48 0.48
MUSCL 18.20 24.38
NEOPL 1.65 1.83
NERV 27.30 28.65
184
Table A.16e: Descriptive Statistics for Medium Potency TAPs Quintile 5
(Continued)
Covariate Success
N=1039
No Success
N=10230
OCIRC 8.62 9.94
OTHPSY 2.23 2.61
PERI 0.10 0.13
PERSO 0.77 0.96
PSTG 0.00 0.03
RESPI 22.75 25.67
SABUSE 1.94 2.54
SEXD 0.10 0.15
SKIN 11.62 13.51
SUICDX 1.36 1.76
CONCOMITANT DRUG USE
AD 63.54 75.89
HYPNT 1.94 1.56
SEIZU 8.42 8.53
ARRH 8.62 8.86
APARK 0.77 0.79
ANXIE 3.68 5.19
MOOD 9.58 9.24
LIPID 7.45 8.61
EPS1 33.01 24.56
EPS2 1.65 2.51
DIABRX 8.71 10.48
DIABDX 10.36 13.23
PTHER 14.62 14.43
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 3.12 4.07
RIS 3.51 3.68
QUE 1.16 1.04
OAP 0.87 0.68
TAP 92.16 93.42
EPISODE TYPE
RESTART 81.70 83.85
SWT 5.23 4.05
AUG 6.78 6.30
LATE_SWT 6.29 5.81
PRIOR HEALTH SERVICE
UTILIZATION

LTC 4.84 5.11
AHOSP 3.00 3.22
PHOSP 0.68 0.64
REHAB 0.29 0.21
CMHC 17.42 11.79
PSYCH 2.32 2.53
* p<0.05; **p<0.01

185
Table A.17a: Descriptive Statistics for High Potency TAPs Quintile 1
Covariate Success
N=899
No Success
N=6895
SCHIZOPHRENIA TYPE
SIMPLE 0.11 0.39
DISORG 0.00 0.06
CATA 0.11 0.01
PARAS** 8.12 4.80
ACUTE 0.33 0.20
LATENT 0.11 0.06
RESID 0.11 0.19
SAFF 5.01 4.89
OTHNOS 3.45 3.07
UNSPEC** 82.65 86.32
AGE
AGE_CAT18 6.45 5.50
AGE_CAT25 16.02 15.40
AGE_CAT35 23.58 22.07
AGE_CAT45 12.46 15.26
AGE_CAT55 9.23 7.98
AGE_CAT65 32.26 33.79
MALE 37.04 37.43
RACE/ETHNICITY
WHITE 38.71 36.87
BLACK 23.38 21.69
HISP 5.66 6.79
ASIAN 5.56 4.79
OTHER 27.25 29.31
COMORBIDITIES
ACC 1.78 2.41
ANX 9.79 9.51
ARRTDX 6.45 5.98
BIP 7.79 7.83
BLOOD 6.12 6.35
CARDIO 21.58 24.31
CHILD 0.00 0.04
CONGE 1.67 1.94
DEMEN 5.45 4.90
DEP 53.28 50.49
DIGES 17.69 20.00
DIABDX 11.57 10.73
ENDO 8.79 10.40
EPSDX 0.56 0.67
GENTI 20.80 21.36
GLAUC 0.78 1.35
HBP 19.24 20.26
HIV 0.00 0.01
HLIPID 11.46 10.63
INFEC 97.55 97.51
INJUR 23.92 23.58
MANIC 0.22 0.45
MUSCL* 29.14 33.20
NEOPL 4.67 5.11
NERV 35.82 35.50
186
Table A.17a: Descriptive Statistics for High Potency TAPs Quintile 1 (Continued)
Covariate Success
N=899
No Success
N=6895
OCIRC 20.02 22.02
OTHPSY 5.01 5.69
PERI 0.00 0.06
PERSO 1.22 1.57
PSTG 0.56 0.35
RESPI 30.03 30.89
SABUSE 5.45 4.79
SEXD 0.11 0.06
SKIN 15.13 14.04
SUICDX 4.23 3.96
CONCOMITANT DRUG USE
AD 45.05 42.51
HYPNT 2.67 2.49
SEIZU 11.90 11.18
ARRH 7.90 9.22
APARK 1.78 1.54
ANXIE 3.78 3.35
MOOD 14.57 13.60
LIPID 7.01 7.34
EPS1 28.92 26.82
EPS2 4.89 4.48
DIABRX 11.12 9.96
DIABDX* 15.02 18.25
PTHER 20.80 21.00
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 7.79 6.69
RIS 7.90 6.87
QUE 1.89 2.15
OAP 2.89 2.42
TAP 81.75 79.32
EPISODE TYPE
RESTART 71.19 72.46
SWT* 6.67 4.83
AUG 11.79 11.39
LATE_SWT 10.34 11.33
PRIOR HEALTH SERVICE
UTILIZATION

LTC 12.01 10.41
AHOSP 7.01 6.18
PHOSP 2.67 2.16
REHAB 0.22 0.19
CMHC 11.68 9.67
PSYCH 2.34 3.09
* p<0.05; **p<0.01

187
Table A.17b: Descriptive Statistics for High Potency TAPs Quintile 2
Covariate Success
N=1091
No Success
N=6703
SCHIZOPHRENIA TYPE
SIMPLE 0.64 0.46
DISORG 0.18 0.10
CATA 0.00 0.03
PARAS 6.60 6.21
ACUTE 0.18 0.16
LATENT 0.00 0.06
RESID 0.46 0.34
SAFF 5.41 4.70
OTHNOS 3.67 3.61
UNSPEC 82.86 84.32
AGE
AGE_CAT18 2.47 2.43
AGE_CAT25 16.59 14.96
AGE_CAT35 24.84 29.96
AGE_CAT45 15.77 17.23
AGE_CAT55 8.43 9.74
AGE_CAT65 31.90 31.67
MALE 40.88 40.98
RACE/ETHNICITY
WHITE 47.48 45.82
BLACK 18.61 19.93
HISP 2.75 3.10
ASIAN 1.92 2.42
OTHER 29.24 28.73
COMORBIDITIES
ACC 2.38 1.78
ANX 7.06 6.25
ARRTDX 5.13 4.61
BIP 7.33 7.46
BLOOD 5.96 6.12
CARDIO 18.24 18.95
CHILD 0.09 0.13
CONGE 1.28 1.40
DEMEN 4.95 4.91
DEP 41.43 41.77
DIGES 16.32 15.77
DIABDX 10.17 9.29
ENDO 9.81 9.37
EPSDX 0.70 1.10
GENTI 16.53 14.48
GLAUC 1.01 1.21
HBP 15.12 16.14
HIV 0.00 0.06
HLIPID* 7.70 9.73
INFEC 97.89 97.93
INJUR 18.61 17.90
MANIC 0.64 0.63
MUSCL 26.95 25.65
NEOPL 4.67 4.21
NERV 34.28 33.43
188
Table A.17b: Descriptive Statistics for High Potency TAPs Quintile 2 (Continued)
Covariate Success
N=1091
No Success
N=6703
OCIRC 17.42 16.56
OTHPSY 4.77 4.13
PERI 0.09 0.04
PERSO 1.19 1.33
PSTG 0.27 0.15
RESPI 27.86 25.88
SABUSE 3.48 3.04
SEXD 0.00 0.03
SKIN 14.48 14.96
SUICDX 3.12 2.46
CONCOMITANT DRUG USE
AD 35.01 35.30
HYPNT 1.01 1.21
SEIZU 11.82 10.52
ARRH 7.88 8.44
APARK 1.74 1.28
ANXIE 2.11 1.55
MOOD 13.23 13.26
LIPID 5.13 6.47
EPS1 33.64 33.30
EPS2 2.84 2.73
DIABRX 7.61 7.70
DIABDX 14.39 15.60
PTHER 19.34 20.48
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 4.86 5.88
RIS 6.51 5.86
QUE 1.37 1.52
OAP 2.20 2.22
TAP 82.40 81.13
EPISODE TYPE
RESTART 74.24 74.64
SWT 4.58 5.27
AUG 9.99 10.03
LATE_SWT 11.18 10.07
PRIOR HEALTH SERVICE
UTILIZATION

LTC 13.47 12.79
AHOSP 3.67 4.09
PHOSP* 0.64 1.49
REHAB 0.09 0.07
CMHC 9.99 9.92
PSYCH 3.76 3.37
* p<0.05; **p<0.01

189
Table A.17c: Descriptive Statistics for High Potency TAPs Quintile 3
Covariate Success
N=1227
No Success
N=6516
SCHIZOPHRENIA TYPE
SIMPLE 0.90 0.58
DISORG 0.00 0.23
CATA 0.00 0.03
PARAS 7.58 7.52
ACUTE 0.16 0.20
LATENT 0.08 0.00
RESID 0.81 0.51
SAFF 4.73 4.80
OTHNOS 6.03 5.02
UNSPEC 79.71 81.11
AGE
AGE_CAT18 1.30 1.70
AGE_CAT25 16.22 15.93
AGE_CAT35 26.41 26.69
AGE_CAT45 20.21 18.55
AGE_CAT55 12.88 11.96
AGE_CAT65 22.98 25.17
MALE 48.41 48.17
RACE/ETHNICITY
WHITE 53.63 55.00
BLACK 17.11 16.57
HISP 1.06 1.32
ASIAN 1.39 1.34
OTHER 26.81 25.77
COMORBIDITIES
ACC 1.87 1.89
ANX 4.65 5.05
ARRTDX 4.48 4.13
BIP 5.95 7.04
BLOOD 5.38 4.94
CARDIO 14.26 14.26
CHILD 0.16 0.20
CONGE 1.22 1.24
DEMEN 6.52 6.11
DEP 37.65 36.40
DIGES 12.55 13.83
DIABDX 7.82 9.04
ENDO 8.56 8.89
EPSDX 1.14 0.78
GENTI 13.45 13.64
GLAUC 1.14 1.00
HBP 11.65 12.08
HIV 0.00 0.00
HLIPID 8.56 8.62
INFEC 97.72 98.20
INJUR 14.43 14.47
MANIC 0.65 0.45
MUSCL 19.89 21.06
NEOPL 3.67 4.01
NERV 31.05 31.17
190
Table A.17c: Descriptive Statistics for High Potency TAPs Quintile 3 (Continued)
Covariate Success
N=1227
No Success
N=6516
OCIRC 14.51 13.69
OTHPSY 3.75 3.50
PERI 0.00 0.00
PERSO 1.14 0.97
PSTG 0.00 0.06
RESPI 24.45 24.40
SABUSE 1.87 1.70
SEXD 0.00 0.00
SKIN 15.24 16.02
SUICDX 1.87 2.23
CONCOMITANT DRUG USE
AD 31.62 30.91
HYPNT 0.81 0.81
SEIZU 11.90 12.28
ARRH 5.46 6.68
APARK 1.06 1.09
ANXIE 0.90 1.00
MOOD 14.59 15.12
LIPID 5.79 5.46
EPS1 42.87 43.86
EPS2 1.79 2.06
DIABRX 5.62 6.69
DIABDX 13.37 13.21
PTHER 21.11 20.50
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 5.54 5.51
RIS 6.76 6.35
QUE 1.14 1.34
OAP 1.79 1.76
TAP 80.68 79.70
EPISODE TYPE
RESTART 74.16 74.94
SWT 5.62 5.48
AUG 10.35 10.36
LATE_SWT 9.86 9.22
PRIOR HEALTH SERVICE
UTILIZATION

LTC 16.87 16.61
AHOSP 2.77 2.90
PHOSP 0.98 0.97
REHAB 0.08 0.17
CMHC 10.43 10.85
PSYCH 2.93 3.94
* p<0.05; **p<0.01

191
Table A.17d: Descriptive Statistics for High Potency TAPs Quintile 4
Covariate Success
N=1534
No Success
N=6310
SCHIZOPHRENIA TYPE
SIMPLE 0.72 0.74
DISORG 0.39 0.17
CATA 0.07 0.11
PARAS 11.67 11.63
ACUTE 0.33 0.29
LATENT 0.07 0.06
RESID 0.91 0.95
SAFF 4.76 5.71
OTHNOS 7.56 7.48
UNSPEC 73.53 72.85
AGE
AGE_CAT18* 0.26 0.79
AGE_CAT25 12.65 13.06
AGE_CAT35 26.92 26.72
AGE_CAT45 25.16 24.64
AGE_CAT55 14.28 14.23
AGE_CAT65 20.73 20.55
MALE 52.80 53.20
RACE/ETHNICITY
WHITE 59.45 61.90
BLACK 12.39 11.77
HISP 0.72 0.87
ASIAN 0.72 0.62
OTHER 26.73 24.83
COMORBIDITIES
ACC 1.11 1.25
ANX 3.32 4.28
ARRTDX 3.52 3.12
BIP 6.13 5.58
BLOOD 5.28 4.56
CARDIO 12.58 11.49
CHILD 0.20 0.25
CONGE 1.11 1.22
DEMEN 6.13 6.70
DEP 28.23 28.65
DIGES 12.91 11.77
DIABDX 8.41 8.23
ENDO 9.97 8.81
EPSDX 1.17 1.13
GENTI 13.49 11.92
GLAUC 1.11 1.33
HBP 11.08 9.81
HIV 0.00 0.00
HLIPID 8.60 7.75
INFEC 98.76 98.19
INJUR 9.71 9.94
MANIC 0.20 0.24
MUSCL 18.71 16.83
NEOPL 3.32 3.23
NERV 28.23 30.71
192
Table A.17d: Descriptive Statistics for High Potency TAPs Quintile 4 (Continued)
Covariate Success
N=1534
No Success
N=6310
OCIRC 11.73 11.16
OTHPSY 3.19 3.14
PERI 0.00 0.00
PERSO 0.65 0.87
PSTG 0.00 0.03
RESPI 22.75 22.69
SABUSE 0.52 0.89
SEXD 0.00 0.00
SKIN 17.67 18.37
SUICDX 1.43 1.70
CONCOMITANT DRUG USE
AD 243.71 24.29
HYPNT 0.59 0.35
SEIZU 11.93 12.61
ARRH 5.54 6.15
APARK 0.59 0.76
ANXIE 0.46 0.60
MOOD 15.91 16.59
LIPID 4.90 5.15
EPS1 58.34 57.34
EPS2 1.56 1.54
DIABRX 6.13 5.63
DIABDX 12.91 11.84
PTHER 20.53 21.39
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 5.15 5.25
RIS 5.93 6.51
QUE 1.30 1.27
OAP 2.09 1.90
TAP 80.57 92.73
EPISODE TYPE
RESTART 74.58 74.18
SWT 5.74 6.15
AUG 7.69 8.05
LATE_SWT 11.99 11.62
PRIOR HEALTH SERVICE
UTILIZATION

LTC 16.69 18.13
AHOSP 2.02 2.14
PHOSP 0.46 0.90
REHAB 0.26 0.19
CMHC 15.32 14.29
PSYCH 5.80 4.66
* p<0.05; **p<0.01

193
Table A.17e: Descriptive Statistics for High Potency TAPs Quintile 5
Covariate Success
N=1873
No Success
N=5920
SCHIZOPHRENIA TYPE
SIMPLE 0.91 1.03
DISORG 0.69 0.63
CATA 0.37 0.29
PARAS 16.44 17.75
ACUTE 0.32 0.17
LATENT 0.00 0.02
RESID 2.46 2.47
SAFF 5.13 5.51
OTHNOS 13.99 14.12
UNSPEC 59.69 58.02
AGE
AGE_CAT18 0.21 0.35
AGE_CAT25 9.72 11.08
AGE_CAT35 24.51 26.37
AGE_CAT45 32.41 30.32
AGE_CAT55 20.29 19.56
AGE_CAT65 12.87 12.31
MALE 61.24 61.69
RACE/ETHNICITY
WHITE 72.72 72.20
BLACK 6.78 6.72
HISP 0.37 0.24
ASIAN 0.21 0.17
OTHER 19.91 20.68
COMORBIDITIES
ACC 1.33 1.35
ANX 2.83 3.16
ARRTDX 3.04 2.89
BIP 4.27 4.36
BLOOD 5.29 4.26
CARDIO 9.88 8.51
CHILD 0.80 0.69
CONGE 1.17 1.17
DEMEN 6.62 6.72
DEP 18.05 20.17
DIGES* 12.01 10.30
DIABDX 8.06 7.79
ENDO 9.61 9.88
EPSDX 2.30 2.60
GENTI 9.08 8.61
GLAUC 0.96 0.79
HBP 8.54 7.52
HIV 0.00 0.02
HLIPID* 9.66 7.94
INFEC 98.77 98.78
INJUR 7.10 7.01
MANIC 0.11 0.19
MUSCL 14.58 14.27
NEOPL 3.90 3.33
NERV 30.86 29.51
194
Table A.17e: Descriptive Statistics for High Potency TAPs Quintile 5 (Continued)
Covariate Success
N=1873
No Success
N=5920
OCIRC 7.37 7.74
OTHPSY 2.51 2.47
PERI 0.00 0.00
PERSO 0.64 0.61
PSTG 0.11 0.00
RESPI 22.80 21.89
SABUSE 0.53 0.56
SEXD 0.00 0.00
SKIN 22.64 21.45
SUICDX 0.64 1.15
CONCOMITANT DRUG USE
AD* 15.54 17.69
HYPNT 0.11 0.20
SEIZU 15.86 15.56
ARRH 5.77 5.24
APARK 0.32 0.52
ANXIE 0.21 0.15
MOOD 16.82 16.57
LIPID 6.03 5.24
EPS1 79.02 79.92
EPS2 0.64 0.81
DIABRX 4.32 4.32
DIABDX 10.95 11.30
PTHER 21.52 21.62
PRIOR ANTIPSYCHOTIC
DRUG USE

OLA 4.48 4.86
RIS* 4.27 5.51
QUE 0.96 0.95
OAP 1.28 1.64
TAP 82.81 83.97
EPISODE TYPE
RESTART 76.03 74.00
SWT 6.78 7.92
AUG 10.46 11.44
LATE_SWT 6.73 6.64
PRIOR HEALTH SERVICE
UTILIZATION

LTC 16.66 17.15
AHOSP 1.66 1.76
PHOSP 0.11 0.29
REHAB 0.11 0.05
CMHC 23.01 23.36
PSYCH 6.94 6.52
* p<0.05; **p<0.01

195
Figure A.11: Crude Treatment Effect by Risperidone Quintile
0
50
100
150
200
250
300
350
400
450
500
1 2 3 4 5
Risperidone Quintiles
Average Time to All Cause Discontinuation
Olanzapine
Risperidone
Quetiapine
Low TAP
Med TAP
High TAP

Figure A.12: Crude Treatment Effect by Quetiapine Quintile
0
50
100
150
200
250
300
350
400
450
500
1 2 3 4 5
Quetiapine Quintiles
Average Time to All Cause Discontinuation
Olanzapine
Risperidone
Quetiapine
Low TAP
Med TAP
High TAP

196
Figure A.13: Crude Treatment Effect by Low Potency TAPs Quintile
0
50
100
150
200
250
300
350
400
450
500
1 2 3 4 5
Low TAPs Quintiles
Average Time to All Cause Discontinuation
Olanzapine
Risperidone
Quetiapine
Low TAP
Med TAP
High TAP

Figure A.14: Crude Treatment Effect by Medium Potency TAPs Quintile
0
50
100
150
200
250
300
350
400
450
500
1 2 3 4 5
Medium TAPs Quintiles
Average Time to All Cause Discontinuation
Olanzapine
Risperidone
Quetiapine
Low TAP
Med TAP
High TAP

197
Figure A.15: Crude Treatment Effect by High Potency TAPs Quintile
0
50
100
150
200
250
300
350
400
450
500
1 2 3 4 5
High TAPs Quintiles
Average Time to All Cause Discontinuation
Olanzapine
Risperidone
Quetiapine
Low TAP
Med TAP
High TAP

Figure A.16a: Adjusted Treatment Effect Utilizing Olanzapine as the CT
Olanzapine vs Quetiapine

0 0.5 1 1.5 2
Ola Quintile 1
Ola Quintile 2
Ola Quintile 3
Ola Quintile 4
Ola Quintile 5
Log Odds Ratio
Olanzapine Better Quetiapine Better
198
Figure A.16b: Adjusted Treatment Effect Utilizing Olanzapine as the CT
Olanzapine vs Low Potency TAPs

Figure A.16c: Adjusted Treatment Effect Utilizing Olanzapine as the CT
Olanzapine vs High Potency TAPs

0 0.5 1 1.5 2
Ola Quintile 1
Ola Quintile 2
Ola Quintile 3
Ola Quintile 4
Ola Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Ola Quintile 1
Ola Quintile 2
Ola Quintile 3
Ola Quintile 4
Ola Quintile 5
Log Odds Ratio
Olanzapine Better Low TAPs Better
Olanzapine Better High TAPs Better
199
Figure A.17a: Adjusted Treatment Effect Utilizing Risperidone as the CT
Risperidone vs Quetiapine

Figure A.17b: Adjusted Treatment Effect Utilizing Risperidone as the CT
Risperidone vs Low Potency TAPs

0 0.5 1 1.5 2
Ris Quintile 1
Ris Quintile 2
Ris Quintile 3
Ris Quintile 4
Ris Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Ris Quintile 1
Ris Quintile 2
Ris Quintile 3
Ris Quintile 4
Ris Quintile 5
Log Odds Ratio
Risperidone Better Quetiapine Better
Risperidone Better Low TAPs Better
200
Figure A.17c: Adjusted Treatment Effect Utilizing Risperidone as the CT
Risperidone vs Medium Potency TAPs

Figure A.17d: Adjusted Treatment Effect Utilizing Risperidone as the CT
Risperidone vs High Potency TAPs

0 0.5 1 1.5 2
Ris Quintile 1
Ris Quintile 2
Ris Quintile 3
Ris Quintile 4
Ris Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Ris Quintile 1
Ris Quintile 2
Ris Quintile 3
Ris Quintile 4
Ris Quintile 5
Log Odds Ratio
Risperidone Better Medium TAPs Better
Risperidone Better High TAPs Better
201
Figure A.18a: Adjusted Treatment Effect Utilizing Quetiapine as the CT
Quetiapine vs Olanzapine

Figure A.18b: Adjusted Treatment Effect Utilizing Quetiapine as the CT
Quetiapine vs Risperidone

0 0.5 1 1.5 2
Que Quintile 1
Que Quintile 2
Que Quintile 3
Que Quintile 4
Que Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Que Quintile 1
Que Quintile 2
Que Quintile 3
Que Quintile 4
Que Quintile 5
Log Odds Ratio
Quetiapine Better Olanzapine Better
Quetiapine Better Risperidone Better
202
Figure A.18c: Adjusted Treatment Effect Utilizing Quetiapine as the CT
Quetiapine vs Low Potency TAPs

Figure A.18d: Adjusted Treatment Effect Utilizing Quetiapine as the CT
Quetiapine vs Medium Potency TAPs

0 0.5 1 1.5 2
Que Quintile 1
Que Quintile 2
Que Quintile 3
Que Quintile 4
Que Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Que Quintile 1
Que Quintile 2
Que Quintile 3
Que Quintile 4
Que Quintile 5
Log Odds Ratio
Quetiapine Better Low TAPs Better
Quetiapine Better Medium TAPs Better
203
Figure A.18e: Adjusted Treatment Effect Utilizing Quetiapine as the CT
Quetiapine vs High Potency TAPs

Figure A.19a: Adjusted Treatment Effect Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Olanzapine

0 0.5 1 1.5 2
Que Quintile 1
Que Quintile 2
Que Quintile 3
Que Quintile 4
Que Quintile 5
Log Odds Ratio
0 1 2 3 4
Low Quintile 1
Low Quintile 2
Low Quintile 3
Low Quintile 4
Low Quintile 5
Log Odds Ratio
Quetiapine Better High TAPs Better
Low TAPs Better Olanzapine
Better
204
Figure A.19b: Adjusted Treatment Effect Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Quetiapine

Figure A.19c: Adjusted Treatment Effect Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Medium Potency TAPs

0 1 2 3 4
Low Quintile 1
Low Quintile 2
Low Quintile 3
Low Quintile 4
Low Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Low Quintile 1
Low Quintile 2
Low Quintile 3
Low Quintile 4
Low Quintile 5
Log Odds Ratio
Low TAPs Better Quetiapine Better
Low TAPs Better Medium TAPs Better
205
Figure A.19d: Adjusted Treatment Effect Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs High Potency TAPs

Figure A.20a: Adjusted Treatment Effect Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Olanzapine

0 0.5 1 1.5 2
Low Quintile 1
Low Quintile 2
Low Quintile 3
Low Quintile 4
Low Quintile 5
Log Odds Ratio
0 10 20 30 40
Med Quintile 1
Med Quintile 2
Med Quintile 3
Med Quintile 4
Med Quintile 5
Log Odds Ratio
Low TAPs Better High TAPs Better
Olanzpine Better
206
Figure A.20b: Adjusted Treatment Effect Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Risperidone

Figure A.20c: Adjusted Treatment Effect Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Quetiapine

0 20 40 60
Med Quintile 1
Med Quintile 2
Med Quintile 3
Med Quintile 4
Med Quintile 5
Log Odds Ratio
0 20 40 60
Med Quintile 1
Med Quintile 2
Med Quintile 3
Med Quintile 4
Med Quintile 5
Log Odds Ratio
Quetiapine Better
Risperidone Better
207
Figure A.20d: Adjusted Treatment Effect Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Low Potency TAPs

Figure A.20e: Adjusted Treatment Effect Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs High Potency TAPs

0 5 10 15 20
Med Quintile 1
Med Quintile 2
Med Quintile 3
Med Quintile 4
Med Quintile 5
Log Odds Ratio
0 5 10 15 20
Med Quintile 1
Med Quintile 2
Med Quintile 3
Med Quintile 4
Med Quintile 5
Log Odds Ratio
High TAPs Better
Low TAPs Better
208
Figure A.21a: Adjusted Treatment Effect Utilizing High Potency TAPs as the CT
High Potency TAPs vs Olanzapine

Figure A.21b: Adjusted Treatment Effect Utilizing High Potency TAPs as the CT
High Potency TAPs vs Risperidone

0 1 2 3
High Quintile 1
High Quintile 2
High Quintile 3
High Quintile 4
High Quintile 5
Log Odds Ratio
0 1 2 3
High Quintile 1
High Quintile 2
High Quintile 3
High Quintile 4
High Quintile 5
Log Odds Ratio
High TAPs Better Olanzapine Better
High TAPs Better Risperidone Better
209
Figure A.21c: Adjusted Treatment Effect Utilizing High Potency TAPs as the CT
High Potency TAPs vs Quetiapine

Figure A.21d: Adjusted Treatment Effect Utilizing High Potency TAPs as the CT
High Potency TAPs vs Low Potency TAPs

0 1 2 3
High Quintile 1
High Quintile 2
High Quintile 3
High Quintile 4
High Quintile 5
Log Odds Ratio
0 1 2 3
High Quintile 1
High Quintile 2
High Quintile 3
High Quintile 4
High Quintile 5
Log Odds Ratio
High TAPs Better Quetiapine Better
High TAPs Better Low TAPs Better
210
Figure A.21e: Adjusted Treatment Effect Utilizing High Potency TAPs as the CT
High Potency TAPs vs Medium Potency TAPs

Figure A.22a: Adjusted Treatment Effect with PS (Prior) Utilizing Olanzapine as the CT
Olanzapine vs Risperidone

0 1 2 3
High Quintile 1
High Quintile 2
High Quintile 3
High Quintile 4
High Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Ola Quintile 1
Ola Quintile 2
Ola Quintile 3
Ola Quintile 4
Ola Quintile 5
Log Odds Ratio
High TAPs Better Medium TAPs Better
Olanzapine Better Risperidone Better
211
Figure A.22b: Adjusted Treatment Effect with PS (Prior) Utilizing Olanzapine as the CT
Olanzapine vs Quetiapine

Figure A.22c: Adjusted Treatment Effect with PS (Prior) Utilizing Olanzapine as the CT
Olanzapine vs Low Potency TAPs

0 0.5 1 1.5 2
Ola Quintile 1
Ola Quintile 2
Ola Quintile 3
Ola Quintile 4
Ola Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Ola Quintile 1
Ola Quintile 2
Ola Quintile 3
Ola Quintile 4
Ola Quintile 5
Log Odds Ratio
Olanzapine Better Quetiapine Better
Olanzapine Better Low TAPs Better
212
Figure A.22d: Adjusted Treatment Effect with PS (Prior) Utilizing Olanzapine as the CT
Olanzapine vs Medium Potency TAPs

Figure A.22e: Adjusted Treatment Effect with PS (Prior) Utilizing Olanzapine as the CT
Olanzapine vs High Potency TAPs

0 0.5 1 1.5 2
Ola Quintile 1
Ola Quintile 2
Ola Quintile 3
Ola Quintile 4
Ola Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Ola Quintile 1
Ola Quintile 2
Ola Quintile 3
Ola Quintile 4
Ola Quintile 5
Log Odds Ratio
Olanzapine Better High TAPs Better
Olanzapine Better Medium TAPs Better
213
Figure A.23a: Adjusted Treatment Effect with PS (Prior) Utilizing Risperidone as the CT
Risperidone vs Olanzapine

Figure A.23b: Adjusted Treatment Effect with PS (Prior) Utilizing Risperidone as the CT
Risperidone vs Quetiapine

0 0.5 1 1.5 2
Ris Quintile 1
Ris Quintile 2
Ris Quintile 3
Ris Quintile 4
Ris Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Ris Quintile 1
Ris Quintile 2
Ris Quintile 3
Ris Quintile 4
Ris Quintile 5
Log Odds Ratio
Ridperidone Better Olanzapine Better
Risperidone Better Quetiapine Better
214
Figure A.23c: Adjusted Treatment Effect with PS (Prior) Utilizing Risperidone as the CT
Risperidone vs Low Potency TAPs

Figure A.23d: Adjusted Treatment Effect with PS (Prior) Utilizing Risperidone as the CT
Risperidone vs Medium Potency TAPs

0 0.5 1 1.5 2
Ris Quintile 1
Ris Quintile 2
Ris Quintile 3
Ris Quintile 4
Ris Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Ris Quintile 1
Ris Quintile 2
Ris Quintile 3
Ris Quintile 4
Ris Quintile 5
Log Odds Ratio
Risperidone Better Low TAPs Better
Risperidone Better Medium TAPs Better
215
Figure A.23e: Adjusted Treatment Effect with PS (Prior) Utilizing Risperidone as the CT
Risperidone vs High Potency TAPs

Figure A.24a: Adjusted Treatment Effect with PS (Prior) Utilizing Quetiapine as the CT
Quetiapine vs Olanzapine

0 0.5 1 1.5 2
Ris Quintile 1
Ris Quintile 2
Ris Quintile 3
Ris Quintile 4
Ris Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Que Quintile 1
Que Quintile 2
Que Quintile 3
Que Quintile 4
Que Quintile 5
Log Odds Ratio
Risperidone Better Medium TAPs Better
Quetiapine Better Olanzapine Better
216
Figure A.24b: Adjusted Treatment Effect with PS (Prior) Utilizing Quetiapine as the CT
Quetiapine vs Risperidone

Figure A.24c: Adjusted Treatment Effect with PS (Prior) Utilizing Quetiapine as the CT
Quetiapine vs Low Potency TAPs

0 0.5 1 1.5 2
Que Quintile 1
Que Quintile 2
Que Quintile 3
Que Quintile 4
Que Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Que Quintile 1
Que Quintile 2
Que Quintile 3
Que Quintile 4
Que Quintile 5
Log Odds Ratio
Quetiapine Better Risperidone Better
Quetiapine Better Low TAPs Better
217
Figure A.24d: Adjusted Treatment Effect with PS (Prior) Utilizing Quetiapine as the CT
Quetiapine vs Medium Potency TAPs

Figure A.24e: Adjusted Treatment Effect with PS (Prior) Utilizing Quetiapine as the CT
Quetiapine vs High Potency TAPs

0 0.5 1 1.5 2
Que Quintile 1
Que Quintile 2
Que Quintile 3
Que Quintile 4
Que Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Que Quintile 1
Que Quintile 2
Que Quintile 3
Que Quintile 4
Que Quintile 5
Log Odds Ratio
Quetiapine Better Medium TAPs Better
Quetiapine Better High TAPs Better
218
Figure A.25a: Adjusted Treatment Effect with PS (Prior) Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Olanzapine

Figure A.25b: Adjusted Treatment Effect with PS (Prior) Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Risperidone

0 1 2 3 4
Low Quintile 1
Low Quintile 2
Low Quintile 3
Low Quintile 4
Low Quintile 5
Log Odds Ratio
0 1 2 3 4
Low Quintile 1
Low Quintile 2
Low Quintile 3
Low Quintile 4
Low Quintile 5
Log Odds Ratio
Low TAPs Better Olanzapine
Better
Low TAPs Better Risperidone Better
219
Figure A.25c: Adjusted Treatment Effect with PS (Prior) Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Quetiapine

Figure A.25d: Adjusted Treatment Effect with PS (Prior) Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Medium Potency TAPs

0 1 2 3 4
Low Quintile 1
Low Quintile 2
Low Quintile 3
Low Quintile 4
Low Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Low Quintile 1
Low Quintile 2
Low Quintile 3
Low Quintile 4
Low Quintile 5
Log Odds Ratio
Low TAPs Better Quetiapine Better
Low TAPs Better Medium TAPs Better
220
Figure A.25e: Adjusted Treatment Effect with PS (Prior) Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs High Potency TAPs

Figure A.26a: Adjusted Treatment Effect with PS (Prior) Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Olanzapine

0 0.5 1 1.5 2
Low Quintile 1
Low Quintile 2
Low Quintile 3
Low Quintile 4
Low Quintile 5
Log Odds Ratio
0 10 20 30 40
Med Quintile 1
Med Quintile 2
Med Quintile 3
Med Quintile 4
Med Quintile 5
Log Odds Ratio
Low TAPs Better High TAPs Better
Olanzapine Better
221
Figure A.26b: Adjusted Treatment Effect with PS (Prior) Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Risperidone

Figure A.26c: Adjusted Treatment Effect with PS (Prior) Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Quetiapine

0 20 40 60
Med Quintile 1
Med Quintile 2
Med Quintile 3
Med Quintile 4
Med Quintile 5
Log Odds Ratio
0 20 40 60
Med Quintile 1
Med Quintile 2
Med Quintile 3
Med Quintile 4
Med Quintile 5
Log Odds Ratio
Risperidone Better
Quetiapine Better
222
Figure A.26d: Adjusted Treatment Effect with PS (Prior) Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Low Potency TAPs

Figure A.26e: Adjusted Treatment Effect with PS (Prior) Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs High Potency TAPs

-5 5 15 25
Med Quintile 1
Med Quintile 2
Med Quintile 3
Med Quintile 4
Med Quintile 5
Log Odds Ratio
0 5 10 15 20
Med Quintile 1
Med Quintile 2
Med Quintile 3
Med Quintile 4
Med Quintile 5
Log Odds Ratio
Medium TAPs Better Low TAPs Better
High TAPs Better
223
Figure A.27a: Adjusted Treatment Effect with PS (Prior) Utilizing High Potency TAPs as the CT
High Potency TAPs vs Olanzapine

Figure A.27b: Adjusted Treatment Effect with PS (Prior) Utilizing High Potency TAPs as the CT
High Potency TAPs vs Risperidone

0 1 2 3
High Quintile 1
High Quintile 2
High Quintile 3
High Quintile 4
High Quintile 5
Log Odds Ratio
0 1 2 3
High Quintile 1
High Quintile 2
High Quintile 3
High Quintile 4
High Quintile 5
Log Odds Ratio
High TAPs Better Olanzapine Better
High TAPs Better Risperidone Better
224
Figure A.27c: Adjusted Treatment Effect with PS (Prior) Utilizing High Potency TAPs as the CT
High Potency TAPs vs Quetiapine

Figure A.27d: Adjusted Treatment Effect with PS (Prior) Utilizing High Potency TAPs as the CT
High Potency TAPs vs Low Potency TAPs

0 1 2 3
High Quintile 1
High Quintile 2
High Quintile 3
High Quintile 4
High Quintile 5
Log Odds Ratio
0 1 2 3
High Quintile 1
High Quintile 2
High Quintile 3
High Quintile 4
High Quintile 5
Log Odds Ratio
High TAPs Better Quetiapine Better
High TAPs Better Low TAPs Better
225
Figure A.27e: Adjusted Treatment Effect with PS (Prior) Utilizing High Potency TAPs as the CT
High Potency TAPs vs Medium Potency TAPs

Figure A.28a: Adjusted Treatment Effect with PS (Post) Utilizing Olanzapine as the CT
Olanzapine vs Risperidone

0 1 2 3
High Quintile 1
High Quintile 2
High Quintile 3
High Quintile 4
High Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Ola Quintile 1
Ola Quintile 2
Ola Quintile 3
Ola Quintile 4
Ola Quintile 5
Log Odds Ratio
High TAPs Better Medium TAPs Better
Olanzapine Better Risperidone Better
226

Figure A.28b: Adjusted Treatment Effect with PS (Post) Utilizing Olanzapine as the CT
Olanzapine vs Quetiapine

Figure A.28c: Adjusted Treatment Effect with PS (Post) Utilizing Olanzapine as the CT
Olanzapine vs Low Potency TAPs

0 0.5 1 1.5 2
Ola Quintile 1
Ola Quintile 2
Ola Quintile 3
Ola Quintile 4
Ola Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Ola Quintile 1
Ola Quintile 2
Ola Quintile 3
Ola Quintile 4
Ola Quintile 5
Log Odds Ratio
Olanzapine Better Quetiapine Better
Olanzapine Better Low TAPs Better
227

Figure A.28d: Adjusted Treatment Effect with PS (Post) Utilizing Olanzapine as the CT
Olanzapine vs Medium Potency TAPs

Figure A.28e: Adjusted Treatment Effect with PS (Post) Utilizing Olanzapine as the CT
Olanzapine vs High Potency TAPs

0 0.5 1 1.5 2
Ola Quintile 1
Ola Quintile 2
Ola Quintile 3
Ola Quintile 4
Ola Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Ola Quintile 1
Ola Quintile 2
Ola Quintile 3
Ola Quintile 4
Ola Quintile 5
Log Odds Ratio
Olanzapine Better Medium TAPs Better
Olanzapine Better High TAPs Better
228
Figure A.29a: Adjusted Treatment Effect with PS (Post) Utilizing Risperidone as the CT
Risperidone vs Olanzapine

Figure A.29b: Adjusted Treatment Effect with PS (Post) Utilizing Risperidone as the CT
Risperidone vs Quetiapine

0 0.5 1 1.5 2
Ris Quintile 1
Ris Quintile 2
Ris Quintile 3
Ris Quintile 4
Ris Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Ris Quintile 1
Ris Quintile 2
Ris Quintile 3
Ris Quintile 4
Ris Quintile 5
Log Odds Ratio
Risperidone Better Olanzapine Better
Risperidone Better Quetiapine Better
229
Figure A29c: Adjusted Treatment Effect with PS (Post) Utilizing Risperidone as the CT
Risperidone vs Low Potency TAPs

Figure A.29d: Adjusted Treatment Effect with PS (Post) Utilizing Risperidone as the CT
Risperidone vs Medium Potency TAPs

0 0.5 1 1.5 2
Ris Quintile 1
Ris Quintile 2
Ris Quintile 3
Ris Quintile 4
Ris Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Ris Quintile 1
Ris Quintile 2
Ris Quintile 3
Ris Quintile 4
Ris Quintile 5
Log Odds Ratio
Risperidone Better Low TAPs Better
Risperidone Better Medium TAPs Better
230
Figure A.29e: Adjusted Treatment Effect with PS (Post) Utilizing Risperidone as the CT
Risperidone vs High Potency TAPs

Figure A.30a: Adjusted Treatment Effect with PS (Post) Utilizing Quetiapine as the CT
Quetiapine vs Olanzapine

0 0.5 1 1.5 2
Ris Quintile 1
Ris Quintile 2
Ris Quintile 3
Ris Quintile 4
Ris Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Que Quintile 1
Que Quintile 2
Que Quintile 3
Que Quintile 4
Que Quintile 5
Log Odds Ratio
Risperidone Better High TAPs Better
Quetiapine Better Olanzapine Better
231
Figure A.30b: Adjusted Treatment Effect with PS (Post) Utilizing Quetiapine as the CT
Quetiapine vs Risperidone

Figure A.30c: Adjusted Treatment Effect with PS (Post) Utilizing Quetiapine as the CT
Quetiapine vs Low Potency TAPs

0 0.5 1 1.5 2
Que Quintile 1
Que Quintile 2
Que Quintile 3
Que Quintile 4
Que Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Que Quintile 1
Que Quintile 2
Que Quintile 3
Que Quintile 4
Que Quintile 5
Log Odds Ratio
Quetiapine Better Risperidone Better
Quetiapine Better Low TAPs Better
232
Figure A.30d: Adjusted Treatment Effect with PS (Post) Utilizing Quetiapine as the CT
Quetiapine vs Medium Potency TAPs

Figure A.30e: Adjusted Treatment Effect with PS (Post) Utilizing Quetiapine as the CT
Quetiapine vs High Potency TAPs

0 0.5 1 1.5 2
Que Quintile 1
Que Quintile 2
Que Quintile 3
Que Quintile 4
Que Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Que Quintile 1
Que Quintile 2
Que Quintile 3
Que Quintile 4
Que Quintile 5
Log Odds Ratio
Quetiapine Better Medium TAPs Better
Quetiapine Better High TAPs Better
233
Figure A.31a: Adjusted Treatment Effect with PS (Post) Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Olanzapine

Figure A.31b: Adjusted Treatment Effect with PS (Post) Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Risperidone

0 1 2 3 4
Low Quintile 1
Low Quintile 2
Low Quintile 3
Low Quintile 4
Low Quintile 5
Log Odds Ratio
0 1 2 3 4
Low Quintile 1
Low Quintile 2
Low Quintile 3
Low Quintile 4
Low Quintile 5
Log Odds Ratio
Low TAPs Better Olanzapine
Better
Low TAPs Better Risperidone Better
234

Figure A.31c: Adjusted Treatment Effect with PS (Post) Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Quetiapine

Figure A.31d: Adjusted Treatment Effect with PS (Post) Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs Medium Potency TAPs

0 1 2 3 4
Low Quintile 1
Low Quintile 2
Low Quintile 3
Low Quintile 4
Low Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Low Quintile 1
Low Quintile 2
Low Quintile 3
Low Quintile 4
Low Quintile 5
Log Odds Ratio
Low TAPs Better Quetiapine Better
Low TAPs Better Medium TAPs Better
235

Figure A.31e: Adjusted Treatment Effect with PS (Post) Utilizing Low Potency TAPs as the CT
Low Potency TAPs vs High Potency TAPs

Figure A.32a: Adjusted Treatment Effect with PS (Post) Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Olanzapine

0 20 40
Med Quintile 1
Med Quintile 2
Med Quintile 3
Med Quintile 4
Med Quintile 5
Log Odds Ratio
0 0.5 1 1.5 2
Low Quintile 1
Low Quintile 2
Low Quintile 3
Low Quintile 4
Low Quintile 5
Log Odds Ratio
Low TAPs Better High TAPs Better
Olanzapine Better
236

Figure A.32b: Adjusted Treatment Effect with PS (Post) Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Risperidone

Figure A.32c: Adjusted Treatment Effect with PS (Post) Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Quetiapine

0 20 40 60
Med Quintile 1
Med Quintile 2
Med Quintile 3
Med Quintile 4
Med Quintile 5
Log Odds Ratio
0 20 40 60
Med Quintile 1
Med Quintile 2
Med Quintile 3
Med Quintile 4
Med Quintile 5
Log Odds Ratio
Risperidone Better
Quetiapine Better
237

Figure A.32d: Adjusted Treatment Effect with PS (Post) Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs Low Potency TAPs

Figure A.32e: Adjusted Treatment Effect with PS (Post) Utilizing Medium Potency TAPs as the CT
Medium Potency TAPs vs High Potency TAPs

-5 5 15 25
Med Quintile 1
Med Quintile 2
Med Quintile 3
Med Quintile 4
Med Quintile 5
Log Odds Ratio
0 5 10 15 20
Med Quintile 1
Med Quintile 2
Med Quintile 3
Med Quintile 4
Med Quintile 5
Log Odds Ratio
Medium TAPs Better Low TAPs Better
High TAPs Better
238

Figure A.33a: Adjusted Treatment Effect with PS (Post) Utilizing High Potency TAPs as the CT
High Potency TAPs vs Olanzapine

Figure A.33b: Adjusted Treatment Effect with PS (Post) Utilizing High Potency TAPs as the CT
High Potency TAPs vs Risperidone

0 1 2 3
High Quintile 1
High Quintile 2
High Quintile 3
High Quintile 4
High Quintile 5
Log Odds Ratio
0 1 2 3
High Quintile 1
High Quintile 2
High Quintile 3
High Quintile 4
High Quintile 5
Log Odds Ratio
High TAPs Better Olanzapine Better
High TAPs Better Risperidone Better
239

Figure A.33c: Adjusted Treatment Effect with PS (Post) Utilizing High Potency TAPs as the CT
High Potency TAPs vs Quetiapine

Figure A.33d: Adjusted Treatment Effect with PS (Post) Utilizing High Potency TAPs as the CT
High Potency TAPs vs Low Potency TAPs

0 1 2 3
High Quintile 1
High Quintile 2
High Quintile 3
High Quintile 4
High Quintile 5
Log Odds Ratio
0 1 2 3
High Quintile 1
High Quintile 2
High Quintile 3
High Quintile 4
High Quintile 5
Log Odds Ratio
High TAPs Better Quetiapine Better
High TAPs Better Low TAPs Better
240

Figure A.33e: Adjusted Treatment Effect with PS (Post) Utilizing High Potency TAPs as the CT
High Potency TAPs vs Medium Potency TAPs

0 1 2 3
High Quintile 1
High Quintile 2
High Quintile 3
High Quintile 4
High Quintile 5
Log Odds Ratio
High TAPs Better Medium TAPs Better

Abstract (if available)

Abstract The foundation of this dissertation is built upon the belief that treatment effects are often heterogeneous. Thus, different patients experience different outcomes on the same medication. The existence of such heterogeneity gives indication that the current clinical evidence may not be appropriate. This becomes increasingly evident when heterogeneous treatment effects (HTE) prove to be qualitative. Thus basing clinical decisions on averages could have implications on the patients' well-being, the cost of healthcare to society, and the availability of medications in the marketplace.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Effects of a formulary expansion on the use of atypical antipsychotics and health care services by patients with schizophrenia in the California Medicaid Program

PDF

Assessment of the impact of second-generation antipscyhotics in Medi-Cal patients with bipolar disorder using panel data fixed effect models

PDF

Three essays on comparative economic and clinical outcomes of antipsychotics using retrospective claims data

PDF

An evaluation of risk predictors for disease progression in Veteran Affairs patients with chronic hepatitis C

PDF

Selected papers on the evaluation of healthcare costs of prematurity and necrotizing enterocolitis using large retrospective databases

PDF

Healthcare provider recommendations: a panacea to improving influenza vaccination rates?

PDF

The impact of Patient-Centered Medical Home on a managed Medicaid plan

PDF

Investigating a physiological pathway for the effect of guided imagery on insulin resistance

Asset Metadata

Creator Stafkey-Mailey, Dana Renée (author)

Core Title The development of the prognostic propensity score: an introduction to a method to identify optimal treatment according to individual tailoring variables when heterogeneous treatment effects are ...

School School of Pharmacy

Degree Doctor of Philosophy

Degree Program Pharmaceutical Economics

Publication Date 12/05/2008

Defense Date 10/06/2008

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag antipsychotics,heterogeneous treatment effects,OAI-PMH Harvest,observational study,personalized medicine,prognosis,schizophrenia,stratification,subclassification

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor McCombs, Jeffrey S. (committee chair), Ahn, Jeonghoon (committee member), Ridder, Geert (committee member)

Creator Email stafkey@usc.edu,stafkey_mailey@yahoo.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-m1874

Unique identifier UC1324521

Identifier etd-StafkeyMailey-2479 (filename),usctheses-m40 (legacy collection record id),usctheses-c127-138238 (legacy record id),usctheses-m1874 (legacy record id)

Legacy Identifier etd-StafkeyMailey-2479.pdf

Dmrecord 138238

Document Type Dissertation

Rights Stafkey-Mailey, Dana Renée

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Repository Name Libraries, University of Southern California

Repository Location Los Angeles, California

Repository Email cisadmin@lib.usc.edu