Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
Compliance study of second-generation antipsychotics on patients with schizophrenia
(USC Thesis Other)
Compliance study of second-generation antipsychotics on patients with schizophrenia
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
COMPLIANCE STUDY OF SECOND-GENERATION ANTIPSYCHOTICS ON PATIENTS WITH SCHIZOPHRENIA by Lei Chen A Dissertation Presented to the FACULTY OF THE GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (PHARMACEUTICAL ECONOMICS AND POLICY) December 2005 Copyright 2005 Lei Chen Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. UMI Number: 3220096 Copyright 2005 by Chen, Lei All rights reserved. INFORMATION TO USERS The quality of this reproduction is dependent upon the quality of the copy submitted. Broken or indistinct print, colored or poor quality illustrations and photographs, print bleed-through, substandard margins, and improper alignment can adversely affect reproduction. In the unlikely event that the author did not send a complete manuscript and there are missing pages, these will be noted. Also, if unauthorized copyright material had to be removed, a note will indicate the deletion. ® UMI UMI Microform 3220096 Copyright 2006 by ProQuest Information and Learning Company. All rights reserved. This microform edition is protected against unauthorized copying under Title 17, United States Code. ProQuest Information and Learning Company 300 North Zeeb Road P.O. Box 1346 Ann Arbor, Ml 48106-1346 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. DEDICATION To my parents, my husband, Bing Liu, and my son, Bryan Liu Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. ACKNOWLEDGEMENTS I wish to express my sincere thanks and appreciation to my mentor, Dr. Jeffrey S. McCombs, for his invaluable advice, personal warmth, and unfailing support in every aspect during my graduate study, for giving me the opportunity to work on numerous projects that helped develop my research and organization skills. His positive attitude and belief in unlimited personal growth has been constant inspiration to me. I can think of no one who cares and helps students so much like him. I am very grateful to Dr. Joel W. Hay, for his critical spirit and resourceful mind, for making me realize and explore my potentials. He assisted me to have my first paper published. He always gave me stimulating and instructive suggestions whenever I had a discussion with him about my research. I would also like to convey my heartfelt appreciation to Dr. Geert Ridder, for his expertise in econometric/statistical modeling. As his student taking all of his econometrics courses, and as an empirical researcher to apply those models in the real-world studies, I realized how wonderful and rewarding to have him guide my learning of the most difficult part of empirical research. My thanks are also extended to Tina Bogan and Shaunna Thomas, who have made my life much easier by their excellent administrative and logistic work. I especially enjoy working with Jinhee Park as my colleague and friend, whose cheerful personality and endurance in face of difficulty make tedious work a pleasant journey. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Most of all, I am greatly indebted to my husband, who stands by me all the time, especially sharing my frustration, stress, and confusion. His belief in me, sweet support, and enormous patience helped me through the toughest moments in life. I am tempted to mention my naughty son, for the happiness he brings to the family, and for his noncompliant of my limited time spent with him. I wish to extend a special word of gratitude to my nanny, for her loving care of my son. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. TABLE OF CONTENTS Dedication ii Acknowledgements iii List of Tables vii List of Figures viii Abstract ix Chapter 1: Introduction 1 Chapter 2: Disease Background 3 2.1 Disease characteristics 3 2.2 Social and economic burden 5 2.3 General treatment guidelines 6 2.4 Typical antipsychotics 8 2.5 Atypical antipsychotics 10 2.6 Treatment guidelines for antipsychotic use 14 Chapter 3: Research Question And Significance 16 3.1 Research question 16 3.2 Limitation of previously published literature 18 Chapter 4: Data 20 4.1 Data source 20 4.2 Episode of drug therapy 22 4.3 Episode selection period 26 4.4 Inclusion and exclusion criterion 27 4.5 Final selection criterion 27 4.6 Episode follow-up 28 4.7 Multiple episodes vs. one episode per patient 28 4.8 Definition of patient compliance with drug therapy 30 4.9 Construct of covariates 30 Chapter 5: Methodology 35 5.1 Theoretical framework 35 5.2 Strategy 1 - baseline model with unconfoundedness assumption 43 5.3 Strategy 2 - identification of treatment effect adjusting for selection bias 44 5.4 Strategy 3 - panel data estimation 52 v Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chapter 6: Results 56 6.1 Descriptive statistics 56 6.2 Baseline model of compliance 58 6.3 Compliance model with treatment selection adjustment 61 6.4 Results of panel data model 69 Chapter 7: Discussion And Conclusion 73 7.1 Summary 73 7.2 Limitations of study 76 7.3 Conclusion 78 Bibliography 79 v i Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. LIST OF TABLES 6.1 Descriptive statistics of baseline characteristics by drug 57 6.2 Model parameter 59 6.3 Parameter estimate from baseline model (TAP as reference) 59 6.4 Results from baseline model by type of episode (TAP as reference) 60 6.5 Parameter estimates from baseline model (olanzapine as reference) 61 6.6 Results from mother logit (TAP ass reference) 63 6.7 Results from outcome equation with selection bias terms 65 6.8 Confidence interval of outcomes (bootstrap with 500 reps) 68 6.9 Parameter estimate from panel data model (TAP as reference) 70 6.10 Results from panel data model by type of episode (TAP as reference) 71 6.11 Results from panel data model (olanzapine as reference) 72 v ii Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. LIST OF FIGURES 4.1 Treatment episodes for drug A 23 4.2 Sorted episodes for a patient 23 4.3a Real restarter 24 4.3b Late switcher 24 4.4a Switching episode with shorter gap 25 4.4b Switching episode with crossover 25 4.5 Augmenting episode 26 4.6 Combo therapy 27 4.7 Distribution of number of mono-restart episodes in the study period 29 v iii Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. ABSTRACT This study compared treatment adherence for patients with schizophrenia across alternative atypical antipsychotics (olanzapine, risperidone, and quetiapine), using typical antipsychotics as the comparison treatment. Treatment adherence was measured by time to all-cause discontinuation after the initiation of drug therapy. The MediCal dataset containing a 100% sample of patients treated for schizophrenia during the period 1994 to 2003 was used to create an analytic file consisting of all drug treatment episodes initiated by the patient. The study focused on episodes initiated between January 1999 and March 2001 in which previously treated patients restarted drug therapy using only one medication. A Cox proportional hazards model was estimated based on the first restart episode of each patient initiated in the study period. The Cox proportional hazard model was then re-estimated using a latent index framework to adjust for treatment selection bias. Finally, a panel data, fixed-effect estimation procedure was employed to account for omitted variable/heterogeneity of the patient across alternative medications using data for patients with at least two treatment episodes. All three estimation approaches provided consistent results. Patients are more compliant with atypical antipsychotics than with typical antipsychotics. Among atypical antipsychotics, patients treated with quetiapine tend to be less compliant than those treated with olanzapine or risperidone, while those treated with olanzapine and risperidone displayed similar compliance outcomes. In addition, the baseline Cox model and the panel data model compared the compliance in ‘returning’ patients, who initiated ix Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. the therapy with the same medication as used in their most recent treatment attempt; and ‘switching’ patients, who initiated the therapy with an alternative antipsychotic medication. Patients treated with atypical antipsychotics are less likely to discontinue the therapy compared to typical antipsychotics in both ‘returning’ and ‘switching’ patients. However, this differential effect was larger in ‘switching’ patients than in ‘returning’ patients. This implies that typical antipsychotics may work well for some patients who take medication episodically, but that switching between conventional antipsychotics may not be beneficial as switching to atypical antipsychotics. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. CHAPTER 1: INTRODUCTION Schizophrenia is a chronic and devastating mental disorder that imposes enormous financial and emotional burdens on the society. Patients with schizophrenia used to be treated with typical antipsychotics. However, their side effects frequently resulted in patients discontinuing medication, which often led to increased costs to treat symptom relapse and rehospitalization. Atypical antipsychotics were introduced in late 1980’s. With an improved side effect profile and increased efficacy in controlling symptoms, atypical antipsychotics have largely substituted for typical antipsychotics as first-line treatment of schizophrenia. At present, three atypical antipsychotics have been widely used for several years, including risperidone (1994), olanzapine (1996), and quetiapine (1997). Though they belong to one drug class, they are very different from each other pharmacologically and clinically. On the other hand, typical antipsychotics have not been abandoned entirely and still play a significant role in treating schizophrenia. How these drugs compare to each other in real-world clinical practice is an interesting policy question, especially in light of significant differences in drug use pattern. According to treatment guidelines, patients with schizophrenia should be treated with antipsychotics permanently as long as the medication is effective in controlling the symptoms and can be well tolerated. Therefore, the drug use adherence can be used to differentiate between drugs. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. The challenge in estimating treatment effects with causal interpretation from observational (real world) data is the nonrandom treatment selection process that could potentially bias the research findings. The purpose of this study is to apply advanced econometric methods to compare treatment compliance among four types of antipsychotics while controlling for treatment selection bias. The primary outcome measure used in this study is time to termination of medication use. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. CHAPTER 2: DISEASE BACKGROUND 2.1. Disease characteristics Schizophrenia is a condition of the brain/mind characterized by an array of unusual internal experiences, socially inappropriate behavior and reduced participation in ordinary social and occupational activities (Torrey 1991). Its characteristic symptoms include: positive symptoms (delusions, hallucinations); disorganization (disorganized speech & thought, and disorganized behavior) and inappropriate affect (expression of emotions that are incompatible with the content of speech); negative symptoms (affective flattening, poverty of thought and speech, anhedonia, lack of motivation); cognitive dysfunction (deficits in attention, memory and executive functioning, and lack of insight) (American Psychiatric Association 1997, Goff et al. 2001). Approximately 1.0% of the adult population in U.S. suffers from schizophrenia in any given year (Keith et al. 1991, American Psychiatric Association 1997, Narrow et al. 2002). The illness typically appears during adolescence or early adulthood, affecting men and women with equal frequency (McGlashan 1988). The disease is characterized by three repeated phases that merge into one another without absolute, clear boundaries between them: acute phase, stabilization phase, stable phase. Wide variations in disease progression are seen among individuals and within an individual over time. While some individuals are free of further episodes for longer period of time, the majority displays frequent exacerbations and remissions, and a small proportion remaining chronically severely psychotic. Complete remission (i.e., a return to full premorbid functioning) is not common in this disorder. 3 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Schizophrenia is associated with substantial psychiatric comorbidity. The lifetime risk of depression in patients with schizophrenia is 81% compared with a risk of 7% to 25% in the general population, and depression in chronic schizophrenia has been related to a risk of relapse and suicide. Obsessive compulsive (OC) symptoms are estimated to occur in 10% to 25% patients. It has been suggested that patients with OC symptoms are more severely disabled, have longer hospital stay, and are less treatment responsive (Green et al. 2003). Approximately 40-50% of patients with schizophrenia have a lifetime substance use disorder (Blanchard et al. 2000). Comorbid substance abuse is associated with clinical exacerbations, poorer overall functioning, larger doses of typical antipsychotic therapy, treatment non-compliance, increased rate of relapse and hospitalization, and higher rates of homeless and unemployment (Green et al. 2003, Buckley 1998). Nearly 50% of patients with schizophrenia suffer from at least one comorbid medical condition (Goldman 1999). Reduced exercise, poor nutrition, increased cigarette smoking, substance abuse, and obesity may contribute to increased risks for cardiovascular and pulmonary diseases, HIV and hepatitis B & C, as well as diabetes mellitus (Green et al. 2003). However, patients with schizophrenia are often unable to give a complete or reliable account of their medical illnesses and their negative symptoms serve as an interfering factor with treatment as well. Unsurprisingly, their comorbid conditions often go unrecognized, untreated or undertreated (Goff et al. 2001). Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. The overall rate of mortality among patients with schizophrenia is nearly double that expected in general population (Allebeck et al. 1986). Nearly 50% of patients attempt suicide, and 10% die by suicide (Tsuang et al. 1999). The highest risk of suicide was seen in 15-44 year age group, who were almost 30 times more likely to die by suicide than their peers (Anderson et al. 1991). 2.2 Social and economic burden The care of patients with schizophrenia involve a wide range of health and social welfare services, such as hospital care, community health care, social care services (social workers), non-professional care givers (usually close relatives), private sector providers under contract to the public sector, and voluntary organizations (Knapp 1997). Medical resource utilization can be greatly affected by the severity of the disease and the comorbid conditions of the patient. Although only 1% of adults have the disease, the treatment expenditures account for more than 2.5% of all health care expenditures (Rupp & Keith 1993, Regier & Narrow 1993). The direct cost of treating schizophrenia to the United States was estimated at 19 billion in 1991 (Wyatt et al. 1995). Public funds (federal, state, local) finance the majority (64%) of treatment costs for schizophrenia (Rupp & Keith 1993). Medicaid is a major payer in funding public mental health by the late 1990s (Hogan 1999). The annual direct cost of caring for patients with schizophrenia was estimated between $15,000 and $19,600 per patient in the Massachusetts Medicaid program (Dickey et al. 1996), $26,000 in the Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. California Medicaid program (McCombs et al. 2000a), and $13,650 for ambulatory patients and $95,000 for institutionalized patients in Michigan, Kentucky, Alabama and Georgia states (Lyu et al. 2001). Productivity loss and family burden account for major indirect cost of schizophrenia. The indirect cost was estimated at $46 billion in 1991(Wyatt et al. 1995). The unemployment rate of patients with schizophrenia can reach 70-80% (Davies & . Drummond 1990, McCreadie 1992, Attkisson et al. 1992). The value of time contributed by relatives for care of the mentally ill in 1985 was estimated at $2.5 billion (McGuire 1991). Nearly 10% of the totally and permanently disabled population is composed of people suffering from schizophrenia. (Rupp & Keith 1993). About one-third of homeless single adults suffer from severe mental illness, largely schizophrenia (Attkisson et al. 1992). 2.3 General treatment guidelines The development of a treatment plan for an individual with schizophrenia requires the consideration of cross-sectional (e.g., current clinical status) and longitudinal issues (e.g., frequency, severity, treatments, and consequences of past episodes). Whenever feasible, treatment planning should use an integrated approach with appropriate pharmacologic, psychotherapeutic, psychosocial, and rehabilitative interventions. Many patients require continuous care over the course of their lives, with no limits as to duration of treatment. Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In addition to treating patients directly, the psychiatrist frequently functions as a collaborator, consultant, and/or supervisor with other mental health professionals in a team approach. At present, there is no cure for schizophrenia. However, treatment can decrease morbidity and mortality associated with the disorder. The general goals of treatment are to decrease the frequency, severity, and psychosocial consequences of acute psychotic episodes and to maximize psychosocial functioning between episodes. The specific goals of treatment depend on the phase of the illness and other specific characteristics of the patient. In addition, many patients with schizophrenia need specific community, supportive, and rehabilitative services to address the impairments in role function associated with their disorder. Treatment with antipsychotic agents has proven to be the most effective form of therapy in most patients (Kane & Marder 1993). Antipsychotics are used for the treatment of acute episodes, the prevention of future episodes, and improvement of symptoms between episodes. In addition, various ancillary medications have been used to enhance the therapeutic efficacy of antipsychotic medications and to treat residual symptoms. These medications include lithium, benzodiazepines, and anticonvulsants. Patients with schizophrenia may receive their care in a variety of settings. Hospitalization is used to facilitate rapid resolution of acute symptoms, and is usually indicated for patients who pose a serious threat of harm to themselves or others, or have general 7 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. medical or psychiatric problems that make outpatient treatment unsafe or ineffective. Nursing homes are suitable for some geriatric or medically disabled chronic patients. Day treatment programs are used to provide ongoing supportive care for marginally adjusted schizophrenic patients in the later part of the stabilization phase and the stable phase of illness. Supporting housing is a psychosocial support program widely used for patients who do not live with their families and would benefit from some supervision in their living arrangements. 2.4 Typical antipsychotics Since the introduction of chlorpromazine in 1953, typical antipsychotics have dramatically altered the lives of patients with schizophrenia, and allowed most chronically institutionalized patients to live in the community (Goff et al. 2001). Pharmacologically, typical antipsychotics block D2-receptors in a nonspecific manner throughout the brain. Though the positive symptoms of illness are relieved through D2- receptor blockade, the lack of receptor selectivity often causes intolerable side effects. The neurologic side effects are the most troublesome. About 30% - 60% of the patients develop extrapyramidal symptoms (EPS) (Weiden 1986), which often require maintenance antiparkinson drug therapy during prolonged antipsychotic treatment (Gelenberg 1987). Tardive dyskinesia (TD), an irreversible movement disorder, occurs at a rate of 5% per year of exposure (Jeste & Caligiuri 1993). Elderly patients are much more vulnerable, exhibiting TD at a rate of 30% after 1 year and 60% after 2 years of typical antipsychotic use (Jeste et al. 1995). Other frequent side effects of typical 8 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. antipsychotics include sedation, anticholinergic and antiadrenergic effects (including dry mouth, blurred vision, constipation, tachycardia, and urinary retention), weight gain and disturbances in sexual function (American Psychiatry Association 1997). Therapeutic effect of typical antipsychotics is limited as well. Thirty percent of patients remain unresponsive to these agents (Kane & Marder, 1993), requiring continuous hospitalization or other restrictive environments to provide close monitoring for safety and self-care (Goff et al. 2001). Negative symptom and cognitive dysfunction cannot be effectively controlled using typical antipsychotics (Dixon et al. 1995, Goff et al. 2001). As a result, patients are usually unable to return to work or have satisfactory interpersonal relationships; no more than 30% of patients were able to hold part- or full-time jobs (Mulkern & Manderscheid 1989). It is not surprising that patients with schizophrenia do not always take their drugs as prescribed. Studies have shown that 24% to 63% of outpatients with schizophrenia take less than the prescribed dosage, and this is even true among inpatients (Lindstrom & Bingefors 2000). Discontinuation is frequent. Approximately 40% patients stop taking their antipsychotic medication within one year, and about 75% stop taking the medication within two years (Perkins 1999). A study from Slovenia showed that 63% of first-episode patients with schizophrenia who were taking depot medications drop out during the first year of treatment (Novak-Grubic & Tavcar 1999). Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Using Medi-Cal data from 1987-1996, McCombs et al. (1999) documented that in California, more than 24% of patients with schizophrenia did not use any antipsychotic medication over a one-year period. For the patients who used an antipsychotic medication, over 24% of treated patients delayed the therapy for more than 30 days, and for those patients with no delay in treatment, over 47% switched or augmented their initial antipsychotic medication. The mean duration of uninterrupted therapy was 142 days, and only 11.6% of treated patients achieved uninterrupted antipsychotic drug therapy for one year. Similar patterns over 1 year were confirmed in other states’ Medicaid programs (Lyu et al., 2001). In economic terms, it was estimated that in California, delays in therapy and switching or augmentation of initial antipsychotic drug therapy in the first year were both associated with more than $9,000 in additional health care cost (McCombs et al., 2000a). Over two years, delayed therapy were associated with increased costs of $12,285, changes in therapy were associated with higher total direct cost of $17,644 (McCombs et al., 2000b). With Wisconsin Medicaid data, irregular use of antipsychotics was also shown to significantly increase the hospital use and cost (Svarstad et al., 2001). 2.5 Atypical antipsychotics The development of atypical antipsychotics represents a significant advance in the treatment of patients with schizophrenia. Clozapine was the first atypical agent approved by US Food and Drug Administration (FDA) in 1989. Subsequently, risperidone was 10 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. approved in 1994, olanzapine in i996, quetiapine in late 1997, ziprasidone in 2001, and aripiprazole in 2002. Pharmocologically, atypical agents differ from typical psychotics in their ‘limbic- specific’ D2 binding and high ratio of 5-HT2 binding to D2 binding, which are associated with decreasing risk of EPS and potentially improving negative symptoms (Worrel et al. 2000). In general, atypical agents are characterized by (a) a decreased or absent risk of EPS or other movement disorders at doses that produce antipsychotic effects, (b) minimally elevated prolactin concentrations that was associated with disturbances in sexual function, and (c) significant reduction in both positive and negative symptoms of schizophrenia (Kinon & Lieberman 1996, Lindenmayer et al. 1994). Atypical antipsychotic agents do have distinct therapeutic and side effect profiles that distinguish them from one another. The following comparison is focused largely on clozapine, risperidone, olanzapine and quetiapine, due to the limited use of other newer antipsychotics in the data used in this study. Risperidone 6mg/day was superior to haloperidol; olanzapine is at least equal to haloperidol for reducing positive symptoms and has been more effective than haloperidol for reducing negative symptoms; however, quetiapine was not shown to be better for the management of positive or negative symptoms when compared to typical agents (Worrel et al. 2000). Similar findings were also reported in a meta-analysis of data from randomized controlled trials (Leucht et al. 1999). Clozapine remains the only drug with 11 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. documented efficacy for the management of treatment-resistant schizophrenia (Kane 1988). It has also been reported to reduce violence, aggregation, and suicide in schizophrenia (Glazer & Dickson 1998, Meltzer 1998). But due to clozapine’s life- threatening hematologic complication in about 1% of patients, its use is recommended after sequential failed trials with newer atypical and conventional antipsychotics (McEvoy et al. 1999). Studies have shown that atypical antipsychotics as a group appear to be superior to typical agents in improving cognitive function. However, they seem to target at different domains. Clozapine improves attention and verbal fluency, as well as some types of executive function; whereas, results of its effects on working memory and secondary verbal and spatial memory were inconclusive. Risperidone has consistent positive effects on working memory, executive functioning, and attention; while improvement in verbal learning and memory was inconsistent. Olanzapine improves verbal learning and memory, verbal fluency, and executive function; but not attention, working memory, or visual learning and memory (Meltzer & McGurk 1999). Though atypical antipsychotics largely reduced the risk of EPS and other movement disorders, the use of risperidone still poses dose dependent EPS liability, and the risk of EPS rises when doses above approximately 6 mg/day are prescribed, with no additional benefit accruing. (Caroff et al. 2002). 12 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. As with conventional antipsychotics, risperidone produces dose-related increases in serum prolactin levels in men and women, but such elevation was not shown to be linked to the emergence of possible prolactin-related side effects, such as galactorrhea (secretion of liquid from nipple) and menstrual cycle change (Kleinber et al. 1999). In comparison, clozapine does not increase prolactin at doses within the clinical range; olanzapine does not increase prolactin above normal at standard dosing levels (Casey 1996). More prominent problems and impediments with the use of atypical antipsychotics are metabolic disorders, such as obesity, dyslipidemia, and type II diabetes. Clozapine appears to show the greatest weight gain, followed by olanzapine, quetiapine, risperidone, and ziprasidone in a meta-analysis comparing short-term use (10 weeks) of atypical antipsychotics (Allison et al. 1999). Long-term follow-up studies showed that the largest weight gains are associated with clozapine and olanzapine, and the smallest with quetiapine and ziprasidone. Risperidone is associated with modest weight changes that are not dose related (Nasrallah 2003). Problems with serum lipid levels have also been reported for clozapine and olanzapine, but not risperidone (Ghaeli et al. 1996, Gaulin et al. 1999, Dursun et al. 1999, Henderson et al. 2000, Koro et al. 2002a, Osser et al. 1999). Although schizophrenia itself is associated with an increased risk of diabetes (Mukherjee et al. 1996), the evidence suggested that clozapine and olanzapine are associated with highest risk of diabetes (Koller E et al. 2001, Roller EA et al. 2002, Roller EA et al. 2003, Henderson et al. 2002, Lindenmayer et al. 2003, Fuller et al. 2003, Roro et al. 2002b, Sernyak et al. 2002, 13 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Gianfrancesco et al. 2002, Gianfrancesco et al. 2003, Caro et al. 2002). Reports are relatively limited with quetiapine and ziprasidone due to their relatively limited use. 2.6 Treatment guidelines for antipsychotic use The advent of second-generation antipsychotic agents has rapidly and dramatically changed clinical practice. At present, clinical experts strongly prefer the atypical antipsychotics to typical antipsychotics as the first-line treatment for schizophrenia in most clinical situations because they are better tolerated, equal in treating positive symptoms and equal or better in treating negative symptoms. Typical antipsychotics have only three indications: (1) for stable patients who have had a good response to them without major side effects, (2) for patients who require intramuscular administration of depot medication (not yet available for the atypicals), and (3) for the acute management of aggression/violence in some patients, especially those needing depot medication. Clozapine should be used for patients who have not responded to sequential trials of newer atypical and conventional antipsychotics (McEvoy et al. 1999). In the maintenance phase of treatment, physicians should select medication, dose, and route of administration most likely to enhance adherence and reduce side effects. A minimum of 12-24 months of antipsychotic use is recommended for first episode patients who have gone into remission after the acute psychotic symptom has resolved. Longer- term treatment (up to lifetime) is recommended for patients when diagnosis of schizophrenia is clearly established. Continuous treatment is preferred over intermittent 14 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. use of drugs. Monotherapy is highly recommended except for the most treatment- refractory patients (Miller et al. 1999, McEvoy et al. 1999, Lehman et al. 1998). Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. CHAPTER 3: RESEARCH QUESTION AND SIGNIFICANCE 3.1 Research question Currently, there are 4 major drugs used in the battle against schizophrenia: risperidone (R), olanzapine (O), quetiapine (Q), and typical antipsychotics (TAP). As summarized in the literature review, these treatment options offer different pharmacological and clinical advantages. How these differences in clinical profile are transformed into differences in drug use adherence in real clinical setting is a very interesting, yet under-explored issue. The purpose of this study is to apply advanced econometric methods to conduct a head- to-head comparison among these 4 treatment options in terms of patient’s adherence to drug therapy as measured by time to medication discontinuation. Head-to-head comparisons across different antipsychotic treatment are crucial to the effective management of patients with schizophrenia. According to the Texas Medication Algorithm Project (Miller 1999), comparison studies among atypical antipsychotics are largely lacking. Facing so many drug choices, physicians have difficulty predicting patient response and level of adherence to a particular antipsychotic medication based solely on their own clinical practice, and can benefit from feedback information about their prescribing patterns and subsequent use of intensive healthcare services of their patients (McEvoy et al. 1999). However, results from randomized clinical trials (RCT) may not help guide clinical practice as these studies are conducted in highly structured practice settings. In addition, RCT are not well suited for the studies in which there are Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. more than two treatment options to compare, which requires much larger sample size and is usually not feasible financially. With large real world data, observational studies comparing multiple treatment options appear to be the best source in providing knowledge of medication selection and predicting the likelihood of successful treatment. This study focuses on drug use adherence, measured by time to discontinuation, as its primary outcome for a number of reasons. First, a survey on physician and patient attitudes toward antipsychotic therapy revealed that psychiatrists identified compliance as the major challenge in treating schizophrenia, adverse effects and lack of efficacy were cited by patients as the major reasons for drug discontinuation (Hellewell & Cantillon 1998). But more importantly, medication adherence is directly related to other clinical and economic outcomes. Drug nonadherence was reported to be the best predictor of symptom relapse (Ayuso-Gutierrez et al. 1997). With each relapse, schizophrenia becomes increasingly resistant to treatment (Sheitman et al. 1997, Weiden et al. 1996), constituting one of the most costly aspects in treating patients with schizophrenia (Weiden et al. 1995, Almond et al. 2004). Without surprise, interruption of antipsychoitc treatment is associated with increases in rates of hospitalization and treatment costs (Svarstad et al. 2001, Gilmer et al. 2004). Third, drug adherence is also an indicator of drug treatment effectiveness. An effective agent with a more tolerable side-effect profile would be expected to have longer duration in use than those that were switched or terminated soon after the initiation of the treatment. In literature, time to medication discontinuation has been recognized as an index measure of treatment effectiveness, 17 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. capturing the judgments of both patients and clinicians on the medication effectiveness, safety and tolerability (Stroup et al. 2003). 3.2 Limitation of previously published literature The available published head-to-head outcome studies focused on olanzapine and risperidone, probably due to the later introduction and limited use of other atypical antipsychotics. Two large RCTs comparing the efficacy and safety of olanzpine and risperidone showed similar compliance rates, though they both reported more cases of weight gain in olanzapine (Tran et al. 1997, Conley et al. 2001) One meta-analysis (Santarlasci & Messori 2003) compared dropout rate between olanzapine and risperidone from the published clinical trials. It was found that the risk of treatment discontinuation was significantly greater for risperidone than for olanzapine (42% vs. 33%, respectively). However, the result was based on the statistical method that didn’t adjust for the timing of discontinuation. In addition, this meta-analysis bears the same limitations as RCT in that adherence to drug therapy is artificially inflated in a highly controlled and structured clinical practice setting. Three observational studies that examined drug use patterns (Zhao 2002, Gibson et al. 2004, Rascati et al. 2003) found that patients treated with olanzapine were more compliant than risperidone treated patients. However, two of these studies used descriptive statistics without adequate adjustment of baseline difference in patient 18 with permission of the copyright owner. Further reproduction prohibited without permission. characteristics (Gibson et al. 2004, Rascati et al. 2003). Only Zhao’s study (2002) used multivariate analysis, adjusting for observed patient demographics, a proxy indicator for disease severity, and the presence of comorbid medical conditions. However, there are two major limitations with the method used in this study. First, duration in drug use is non-negative and skewed toward zero, but the author used simple OLS regression estimation technique, which required normal distribution of outcome variable. Second, no attempt was made to adjust for unobserved factors in drug selection that may be correlated with the compliance. This may cause correlation between drug choice and error term, thus biasing the estimates. In summary, although several studies presented certain important data regarding compliance, they are restricted in a number of ways. First, they only looked at olanzapine and risperidone. A more relevant comparison should include quetiapine which has been available since 1997. Typical antipsychotics should also be included to serve as a baseline comparator in such head-to-head comparisons. Second, evidence on adherence derived by clinical trials is quite limited in its generalizability to real world practice. Research using real-world, long-term follow-up data would provide a more accurate assessment of compliance. Third, previous observational studies didn’t use an appropriate statistical method to address the issue of treatment selection bias. An improved methodology framework is needed in this type of research. 19 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. CHAPTER 4: DATA 4.1 Data source The fee-for-service California Medicaid (MediCal) dataset from 1994 to 2003 was used to identify patients who used any type of antipsychotics during this 10-year period. The patients with schizophrenia were then identified if the diagnosis of schizophrenia (International Classification of Diseases 9th Revision, ICD-9 codes: 295.0 - 295.9) was recorded on at least 1 paid claim over the entire data span. MediCal finances a wide range of health care services for the poor and the disabled population as long as they remain eligible for the program. The dataset contains patient level demographic information, such as date of birth, gender, race, county, as well as claims for all covered services paid on behalf of the recipients. Medical claims include type and date of service, amount billed, amount paid, and units (days) of service. Prescription drug claims identify a specific product by National Drug Code (NDC), and consist of records of date of prescription filling, dose, strength, and day-of-supply. The advantage of using MediCal data is that it provides rich and detailed information regarding health care resource utilization and drug use profile over time. However, this data is not without problems, which must be addressed. 20 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Payments incurred by dually eligible elderly or disabled recipients for both Medicare and Medicaid programs are under reported, because Medicare becomes the primary payer for these patients. To correct this problem for cost of institutional care, the days of service were multiplied by the average per diem cost to derive the payment for hospitalization, skilled nursing facility (SNF), and intermediate care facility (ICF). A per diem price of $1,053 per day was used for hospital care and $270 per day of care in SNF and ICF (Hiehle & Cline 2002). For simplicity and consistency, this method was applied to all patients regardless of age or eligibility status. For ambulatory services covered under Part B of Medicare, the amount paid by MediCal was inflated on the basis of the Medicare part B deductible (100$) and copayment rate (20%). Actual MediCal expenditures for outpatient services were used for patients covered by MediCal only and for services not covered by Medicare part B. All reported non-institutional expenditures were adjusted to 2003 dollars using MediCal-specific fee schedule adjustments. Eligibility data for our sample was not available, creating a potential problem that a gap in coverage would be counted as discontinuation of treatment. Elderly, disabled or blind patients were assumed to be eligible for MediCal program permanently. To ensure continued eligibility for patients who were not elderly, disabled or blind, we excluded those patients with zero total treatment cost in any three consecutive months during the defined screening period (with more details later). The logic behind this screening method is that schizophrenia is a severe mental disease that requires continuous treatment, making it unlikely that such a patient can avoid any form of therapy for three months. This method has been used in previous studies based on MediCal data 21 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (McCombs et al. 1999, McCombs et al. 2000a, McCombs et al. 2000b, Lyu et al. 2001). However, this method excludes patients who may completely withdraw from treatment due to the homelessness and patients incarcerated by the criminal justice system. 4.2 Episodes of drug therapy In the raw data, up to 12 different antipsychotics were identified for every patient, and as many as 90 refills were recorded for each drug. All treatment episodes related to a particular antipsychotic medication were extracted from its prescription records. As illustrated in Figure 4.1, a treatment episode of an antipsychotic medication was defined as a continuous use of a medication up to a gap of >15 days between the end of reported day-of-supply and the next prescription refill. The start date of an episode was the “index date”. The “end date” of an episode was calculated as the last prescription date related to the episode plus the reported day-of-supply for that refill. The duration of an episode is the length between “end date” and “index date”. A gap less than 15 days was allowed within a treatment episode. Depending on the number of greater-than-15-day gaps, a patient may have multiple treatment episodes on one medication (e.g. A l, A 2...). Once this process was applied to each antipsychotics used by a patient (e.g. A l, A 2..., B l, B 2..., C l, C 2..., ...), all treatment episodes generated in the entire data were sorted in ascending order of their index dates, starting from the first observed episode (Figure 4.2). As shown in the graph, a patient can have a series of treatment attempts on the same and different antipsychotics, which are related. 22 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Figure 4.1: Treatment episodes for drug A G ap> 15 3 4 I Al a . 1 A 2 T h e sta rt o f each p rescrip tio n fill ^ Index d ate - H ' E n d date ----------------- R e p o rted d ay s o f su p p ly fo r each p rescrip tio n G ap in rep o rted day o f su p p ly g re a te r th an 15 A n e p iso d e o f d ru g therapy Figure 4.2: Sorted episodes for a patient 1994 2003 T im e Depending on when and which medication was initiated, four types of treatment episodes can be abstracted, demonstrated in the details as the following: (1) ‘First’ episode was defined as the very first episode that could be identified in the entire data, for example, episode Al in Figure 4.2; 23 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (2) A ‘restarter episode’ was defined when a patient initiated an episode of drug treatment after a break of greater than 15 days following prior episode. Restart episode can be further categorized as ‘real restarter’, defined as returning on the same drug as used in prior period (Figure 4.3a), and ‘late switcher’, defined as starting the use of a different antipsychotic medication after the break in therapy (Figure 4.3b); Figure 4.3a: Real restarter O lan zap in e G ■ / ip > = 15 days O lan zap in e (real restarter) Figure 4.3b: Late switcher O lan zap in e G / ap> = 15 days R isp erid o n e (late sw itch er) (3) A ‘switching episode’ is defined when a different drug episode was started after the termination of the prior drug episode, either with a short gap (<15 days) in prescription (Figure 4.4a) or allowing for a period of treatment crossover of less than 30 days (McEvoy et al. 1999) (Figure 4.4b); (4) An ‘augmenter episode’ was defined when the patient started a second drug episode while remaining on active therapy on the previous drug episode. The overlapping period of augmenter episode should be longer than 30 days (Figure 4.5). 24 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Figure 4.4a: Switching episode with shorter gap O lan zap in e G ap < = 1 5 days / \ R isp erid o n e (sw itch er) Figure 4.4b: Switching episode with crossover O lan zap in e C / •ossover < = 3 0 days R isp erid o n e (sw itch er) Figure 4.5: Augmenting episode O lan zap in e Q u etiap in e (au g m en ter) Finally, depending on how many drugs were prescribed, we could also define mono therapy if only one antipsychotics was prescribed, or combo-therapy if more than one antipsychotics were prescribed at the same time (Figure 4.6). Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Figure 4.6: Combo-therapy O lan zap in e Q u etiap in e 4.3 Episode selection period While all treatment episodes in the period 1994-2003 were used to classify the type of episodes, only those initiated between January 1999 and March 2001 were included in this study for three reasons. First, it avoided the period when adjustment in clinical practice was observed after the restriction on prescription of risperidone, olanzapine and quetiapine was lifted in October 1997 (McCombs et al. 2004). Second, early use of quetiapine may have been underdosed since its approval in October 1997. In order to give it an equal footing in its comparison to risperidone and olanzapine, the first one-year of quetiapine availability was not included. Third, ziprasidone was approved in March 2001, so treatment episodes initiated after that were excluded to avoid any adjustment in treatment pattern that may have resulted based on ziprasidone’s introduction. 4.4 Inclusion and exclusion criteria Not all episodes initiated in the defined period were used. Episodes that satisfied any one of the following criteria was excluded: 26 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (1) Episodes initiated when the patient was younger than 18; (2) Episodes initiated when the patient was older than 100, due to possible reporting error in date of birth; (2) Episodes with less than 6-month of data available prior to the episode index date; (3) The first observed episode of drug therapy for each patient in the data was excluded due to the uncertainty about information of prior treatment episodes. Patients who satisfied one of the following conditions were excluded: (1) Patients who ever used clozapine in the entire data span were excluded, because they are usually more severe, thus representing a very different sample from those who can be treated with other types of antipsychotics; (2) Patients with zero total treatment cost in any three consecutive months during the ‘screening’ period, but not elderly, disabled or blind. The ‘screening period’ varied between patients. The ‘end-date’ of the screening period was the date of last available claim record. The start-date of the screening period was set at the index date of the prior episode or 6 month prior to the episode’s index date, whichever was longer. This insures that a minimum of prior 6 month of data and the information of prior episode are included in the study. 4.5 Final selection criteria A total of 208,915 episodes of antipsychotic therapy were identified for analysis, of which 144,827 were restart episodes (70%). Among restart episodes, 10,251 (7.08%) 27 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. episodes were initiated using combination therapy. This made it natural to focus on mono-restart treatment episode as the unit of analysis for conducting head-to-head comparison of drug performance. Moreover, attribution of treatment effect and statistical adjustment for treatment selection bias are much more straightforward using restart episodes. For example, restart episodes constitute a clear point of reinitiating a therapy after a break in treatment and the physician has 4 treatment options from which to select. Conversely, drug selection options for switching and augmenting episodes are restricted, because physicians have to choose among antipsychotics not used in the immediate prior episode. In addition, augmenting episodes overlap with the previous treatment episode, thus the treatment outcomes achieved in each episode are highly correlated. 4.6 Episode follow-up Each selected episode of a patient during January 1999 and March 2001 was then followed on to the last recorded claim in the entire data of that patient. In other words, the follow-up periods vary between patients and extended beyond March 2001. Accordingly, the date of last recorded claim is defined as “last date”. 4.7 Multiple episodes vs. one episode per patient In the period from January 1999 to March 2001, there were 134,576 mono-restart episodes, which corresponded to 52,121 patients. The descriptive statistics showed that 37.3% of the patients had only one restart episode in this period; the remaining 62.7% of 28 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. patients had multiple mono-restart episodes. The distribution in detail was shown in Figure 4.7. Figure 4.7: Distribution of number of mono-restart episodes in the study period ;= 40 0s r 30 c 20 | 10 £ o 37.3 93 ftfi 14.93 14 3 y.bJ —i ■ i — —I — I —■ —i 2 3 Number of episode 4 more than 4 The fact that we have multiple spells of drug use for majority of the patients raises the question of how to use the data. Answer to this question is also relevant to what statistical method is appropriate. There are two ways to handle this situation. One way is to use just one episode for each patient in the study period, for example, the first episode, as if we only observed one treatment spell for everyone. This sampling scheme corresponds to the cross-sectional study design. The other way is to use multiple episodes for each patient: a panel data design. Panel data sampling is similar to asking each patient about their two most recent treatment episodes, in analogy to asking workers about their two most recent employment records in human labor economic studies. In this study, both sampling schemes were used and corresponding statistical methods will be proposed in the method section below. 29 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. 4.8 Definition of patient compliance with drug therapy Patient compliance with a particular antipsychotic treatment is measured by time to discontinuation of that medication following the episode index date. If the “end date” of a selected drug episode for a patient is before the “last date” of that patient, time to discontinuation was determined by the duration of the episode, calculated as the difference between the “end date” and “index date” of that episode. If otherwise, time to discontinuation is right censored due to loss of follow-up, and time to follow-up was determined by the difference between the “last date” of the patient and the “index date” of that episode. According the data structure, only the last episode in the entire data could be possibly censored. Among all the 134,576 mono-restart episodes that were initiated between January 1999 and March 2001, 2.1% of all mono-restart episodes were found to be censored. 4.9 Construction of covariates An exhaustive list of covariates was created to reflect baseline characteristics of a patient before a restart episode was initiated, as well as the information regarding the scenarios of drug selection, such as time of prescription, drug use history, and drug diffusion rate. (1) Patient’s demographic characteristics: age at the initiation of a treatment episode, gender, race, and disability status were derived directly from the data. Urban residency was based on county of residence and population density in 1990. The county of residence variable was recorded only once with the first identified claim record. We 30 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. assumed that patients did not move during the study period, which may result in possible mismatching if the patient moved from one county to another. (2) A range of variables related to patient’s health status in prior 6 months: • Mental disorder profile documenting 9 types of mental diseases identified based on ICD-9 code: organic psychotic disease, bipolar disorder, depression, anxiety, dementia, substance abuse, personality disorder, and other mental disorder. • Dichotomous variables denoting psychotropic drug use, such as mood stabilizers, antidepressants, anti-seizure medications, and depot (long-acting) antipsychotics. These variables serve as another indicator of comorbid mental disorders. • Dichotomous variables denoting the existence of diabetes, hyperlipidemia, arrthymia, and EPS. The first three conditions were identified based on either ICD-9 codes or related drug use, while EPS was determined solely by ICD-9 code (333.00-333.99, E853.xx, E854.xx). These conditions are of particular interest, because they capture physicians’ cautiousness in prescribing atypical antipsychotics based on the literature that document the association of atypical antipsychotic use with these conditions. For example, physicians may be less likely to prescribe olanzapine to diabetic patients, or resperidone to patients with EPS diagnoses, etc. • A count variable indicating the number of medical conditions other than diabetes, hyperlipidemia, arrthymia and EPS. It is expected that patients with larger number of comorbid medical diseases are more ill. • Dichotomous variable denoting the prior health care use in acute/psychiatric hospital, or long-term care service. Patients with prior hospital admission may be 31 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. more severely ill than patients without a recent history of these events. Those receiving long-term care are generally old patients with complex comorbid diseases. • Dichotomous variable denoting if a patient attempted suicide in prior 6 months based on ICD-9 codes in the range E950-959, 959.9 or 300.9. Patients with prior suicidal attempt may be more severely ill than other patients. • Total treatment cost in prior 6 months. This variable is used here as a summary variable for the overall health status of the patient. Because the distribution of total costs is highly right skewed, we divided the values into 5 quintiles. (3) Patient’s compliance propensity with antipsychotic therapy. The patient’s propensity of compliance with the drug therapy was measured by two variables. One variable is the number of uninterrupted days of antipsychotic therapy achieved during their previous treatment attempt. The other variable is days without antipsychotic therapy (gap) before the index date of a restart episode. The second variable also captured the factor that may be related to the long break in therapy, for example, patient thought they don’t need to be treated. (4) Prior antipsychotic drug use. Four dichotomous variables were created to indicate the type of antipsychotics used in prior 6 months: olanzapine, risperidone, quetiapine, and typical antipsychotics. Physicians usually consider the patient’s drug use history when determining which drug to prescribe. If the patient responded well to one particular antipsychotic medication, the physician would probably restart the patient on the same drug. If not, he might initiate treatment with an alternative medication. 32 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (5) Medication diffusion rate, measured by number of prescribed pills/per patient for each type of antipsychotics in each county over time. Physicians prescribing patterns may be correlated with some systematic geographic factors, such as their awareness of new treatment due to local level education or advertising, or penetration rate of managed care in a county that may have impact on physicians’ prescribing habit. It is expected that the drug with higher diffusion rate is more likely to be prescribed. This method was used in a study that compared the rate of hospital admission between typical and atypical antipsychotics on patients with schizophrenia in Medicaid programs (Salkever et al. 2004). The total prescribed pills of each of antipsychotics in a county during a calendar quarter were tabulated from our data, then divided by the total number of patients with schizophrenia in that county; the calculated numbers were then attached to the treatment episode occurred in that calendar quarter using the same drug. In the study period, there were 9 calendar quarters, 57 counties, and 4 types of antipsychotics. So altogether, there were 4*57*9=2352 different values. (6) Time trend. Quarter dummy variables based on the episode index date were used to capture global historical factors such as changes in economic activity, employment, and social welfare policy that vary over time and may affect prescribing decisions. (7) How long atypical antipsychotics were on market. Time since FDA approval measures the level of maturity in atypical antipsychotic use, and indirectly reflects the ‘perceived’ drug attributes by the physicians that change over time. For example, more and better evidence accumulates as more patients used a specific drug and more clinical studies became available in the literature. Therefore, what physicians thought about individual antipsychotic medication might change over time and affect their drug 33 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. selection decision. The difference in month between the date of drug introduction and the index date of an episode was calculated for three types of atypical antipsychotics, and then the arithmetic average of the three values was taken. (8) Type of restart episode. A dichotomous variable was created to differentiate two types of restart episodes: real restarter (same_drug=l), late switcher (same_drug=0). Our data showed that 77.10% of the restart episodes used the same drug as prescribed before, the remaining used different antipsychotics, called Tate switcher’. 34 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. CHAPTER 5: METHODOLOGY Evaluating the effectiveness of alternative interventions is a central objective in social science and medicine. However, the problem of potential selection bias arises in most studies, especially when randomization is not feasible and the researchers have to evaluate alternative treatments using observational data. Individuals observed to receive treatment A often possess different characteristics from individuals receiving treatment B, resulting in unbalanced comparison groups. When unbalance comes from unobserved characteristics, advanced statistics/econometric methods are required to derive unbiased treatment effect estimates. 5.1 Theoretical framework Counterfactual and treatment effect Suppose we have a random sample of size N from a large population. Let D be an indicator of observed treatment exposure, with Dj = 1 if treated, and £). = 0 if control. Let Xj be a vector of observed characteristics of each unit. Let Yf° denote the outcome for unit i under control, and Y - the outcome under treatment. Each agent is observed in only one state at a time, so that either Y° or Y' is observed for each person, but never both. Therefore, Y° and Y' are called ‘counterfactuals’ (Rubin 1974). The framework asserts that individuals have potential outcomes in all states, even though they can be 35 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. observed in only one state. The response Y actually observed is the one that would be seen under the exposure actually received, formalized as Y = Y'D + (1 - D)Y". The treatment effect for unit i is defined as A, =Y- - Y ° , which is never observed. The average treatment effect (ATE) is defined as the expected gain from participating in the program for a randomly chosen individual in the population, or the mean treatment effect over everyone in the targeted population. ATE conditional on X = x can be expressed as (1) Aat" :(x) = E(A\ X = x). Then we can obtain unconditional estimates by integrating (x) over the distribution o f X, (2) & An' : = E(A) = E(Y] - T°). A conceptually different parameter is the effect of treatment on the treated (TT). This is the average gain from treatment for those that actually select into the treatment: (3) An (x,D = 1) - i?[A\X = x,D -\\. The unconditional TT is defined as (4) A7 7 ' (D = 1) = E(Yl - Y° | D = 1) . Imbens and Angrist (1994) introduced a third treatment effect, local average treatment effect (LATE) that estimated the expected outcome gain for those induced to receive treatment through a change in the instrument from Z = 1 to Z - 0. The variable Z is assumed to affect the treatment decision, but not to affect the outcomes T1 and T°, thus defined as instrumental variable (IV). LATE is expressed as 36 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (5) E(Y' -Y° \ D(z) = I,D(z') = 0). LATE is a robust estimation of treatment effect without imposing strong distributional assumptions. The estimation of LATE maintains an assumption of monotonicity. In the binary-treatment-instrument case, the assumption requires that the instrument must affect the treatment assignment of all individuals in the same direction and thus in a monotone fashion; to put it formally, 7’ (Z, = 1) > 7’ (Z, = 0) or Tj (Z, = 1) < 7’ (Z, = 0) for all i. LATE has two problems: (a) it is defined by the instrument, and thus different instruments define different average treatment effects for the same group of individuals eligible to receive the treatment; (b) it is an average treatment effect for a subset of individuals that is inherently unobservable no matter what the instrument (Winship and Morgan, 1999). There are several ways to identify the treatment effect under different assumptions, as presented in the following section. Estimation under randomization of treatment If the treatment is assigned randomly, it means (6) Y°,Y] Y D . This implies mean independence (7) E(Y°\D) = E(Y°), (8) E(Y'\D) = E(Y'). 37 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Hence, (9) Aati; = A1 1 ' = E(Y | D = 1) - E(Y \ D = 0). Random assignment to treatment groups insures that ATE and TT are the same, and can be estimated by the difference of the mean outcomes between treatment group and control group. Estimation under unconfoundedness assumption Randomization is not feasible in many situations, such as those relevant to our study. However, if we assume that X contains all variables that affect both the treatment selection and the outcome, we get the following (10) D 1 ( Y \ Y ° ) \ X . It means that conditional on X, treatment is assigned randomly. Then (11) E (Y '-Y ° \D = l,X = x) = E(Yl -Y ° \X = x), (12) Aate (x) = An (x). But generally AAlh * A11, because the averaging is over different populations. The unconfoundedness assumption applies to the situation where the imbalance between two groups is only due to observable factors. One of the simple approaches is to run multivariate regression with carefully developed control variables in order to remove the correlation between the treatment variable and the error term (see equation 13). Differential treatment effect is identified by the direct estimation of the parameter 38 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. associated with a treatment dummy (or several dummies depending on how many drugs are compared), where it takes on a value of 1 if the patient receives the treatment, and 0 if the patient doesn’t. For example, a regression specified as (13) Y = X/3 + y* D + s where D is the treatment dummy. Because D and c are uncorrelated after X is included, y would be unbiased estimation of treatment effect. Latent index model Regardless of the care taken to fully document the differences between treatment groups, there may still exist some unobservable factors that drive treatment selection and are correlated with the treatment outcome, for example, disease severity. Failure to control for these factors can lead to biased estimates of treatment effects using simple multivariate regression. The term ‘bias’ refers to the consequence that the research will incorrectly attribute a differential effect in outcome to the treatment received, when in reality it was due to the initial unobservable differences in characteristics between the two groups. To put it econometrically, the treatment variable becomes endogenous when it is correlated with error term that includes unobservable factors. A multivariate regression with an endogenous control variable will always result in erroneous conclusion. Hence advanced modeling is warranted. One popular approach for dealing with unobserved selection bias was introduced by Heckman (1976, 1979). This approach specifies a latent index model that connects the 39 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. rule of treatment assignment with the potential treatment outcomes. The rule is based on the expected net utility from receiving the treatment: individuals participate in a program if net utility is non-negative and do not participate if net utility is negative. This approach assumes a parametric error distribution (for example, normal distribution) and allows for dependence between the errors in the outcome and choice equation. While convenient in implementation, this approach has been criticized for its reliance on distributional assumptions and lack of robustness to departures from normality. Though more robust approaches without strong distributional assumptions have been proposed, for example, LATE, “they typically estimate only one treatment parameter and are quite limited in the range of policy questions they can answer”, according to Heckman and Vytlacil (2000b). Further, Vytlacil (1999) demonstrated that the monotonicity and independence assumptions imposed in LATE analysis are actually equivalent to those required to specify a nonparametric latent index model. Later on, Heckman developed the marginal treatment effect (MTE) and united the recent treatment effect literature with the classical selection bias literature, and obtained simple closed- form expressions for four treatment parameters of interest in parametric models: ATE, TT, LATE, and MTE. Their work was motivated by the observation that “despite the recent advances in flexible estimation of selection models, simple two-step correction procedures continue to dominate applied work on this topic” (Heckman 2000). Heckman’s work was presented as followings. Consider a model of potential outcomes: (14) Yx =XJ3' +U], (15) Y° = X P * + U \ 40 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (16) d * =ze + uD . The first two equations denote outcome equations in two possible ‘states’. Each agent is observed in only one state. Let D denote the observed treatment decision, where D{Z) = 1 denotes receipt of treatment and D(Z) = 1 denotes noreceipt. The variable D* is a latent variable that generates D(Z) according to a threshold, for example, (17) D(Z) = l[D*(Z)>o\=\[z9 + UD >o] where 1 [ a ] is the indicator function which takes the value 1 if the event A is true and the value 0 if otherwise. This model has been called the ‘switching regression model’. ATE conditional on X = x can be expressed as (18) ATE(x) = E (A \X = x) = x (0 ] - 0 ° ) . Then we can obtain unconditional estimates by integrating (x) over the distribution of X, (19) ATE = E(A) = \ATE(X)dF{X) ta — '^A T E (xj) n 7 = 1 where n is the sample size. TT conditional on X = x is expressed as (20) 7T (x,z,D(z) = \) = E [b \X = x,Z = z, D ( z ) = l] = *(/?' -/? ° ) + £ ( f / '- U ° \ U D > - z 0 ,X = x,Z = z ) = *(/?' - 0 ° ) + E(Ul - U° | U D > -z0) where the third equality follows from the assumption that (UD U lU°) is independent of (.X ,Z ). As with ATE, we can obtain an unconditional estimate by integrating over the joint distribution of X and Z for those who actually receive treatment. Letting n, be the number of observations with D, = 1, TT can be approximated as follows: 41 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (21) TT = E (A | D(Z) = 1) = JTT(X,Z,D(Z) = \)dF(*X,Z \ D(Z) = 1) ■£>(*,) = 1). n, The LATE parameter is defined as (22) Z ^ rJ E(D(z) = 0,D(z,) = l , ^ = ^) = ^(A|T»(z) = 0,J D(z') = l,A = x) = x(0'-J3°) + E(U1- U ° \ - z '0 < U D < - z 0 ,X = x = x (p 1 -J3°) + E(U] -U ° \ - z 9 < U D <-zO) where the third equality follows from the assumption that (UD UX U°) is independent of (.X ,Z ). The unconditional version of LATE is defined as (23) £(A | D(z) - 0, D(z ) = 1) = [LATE(D(z) = 0,D{z) = \,Z)dF{X) = - f jLA TE(D(z) = 0 ,D (z) = 1 , x,). n ,=i The marginal treatment effect (MTE) is the treatment effect for individuals with a given value of U D, (24) MTE(x,un) - E (A \X = x,UD = uD) = x(fi] -J3°) + E(U' - U° \ U I} = uD,X = x) = x(j3' - p ° ) + E(Ul - U ° \ U D =uD) where the third equality follows from the assumption that (UD U lU°) is independent of (X,Z ) . Evaluation of the MTE parameter at low values of u " averages the outcome gain for those with unobservables making them least likely to participate, while evaluation of the MTE parameter at high value of un is the gain for those individuals with 42 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. unobservables which make them most likely to participate. The unconditional version of MTE can be written as (25) MTE(ud)= ^M T E (X ,u °)d F (x )^-Y jMTE(xl,uD). n / = 1 The MTE parameter can also be expressed as the limit form of the LATE parameter, (26) lim LATE(x,D(z) = 0,D(z =1) :0->:0 = x(/3' -/? ° )+ lim E(U' -U ° \-z'9 < U D < -z 6 ,X = x z0->z'e = x(J3' - /?°) + E(U' - U-0 | U n = -z'O) = MTE{x,-z &). Thus MTE parameter measures the average gain in outcomes for those individuals who are just indifferent to the receipt of treatment when zO index is fixed at the value - a " . 5.2 Strategy 1 - baseline model with unconfoundedness assumption This study has made significant effort to create an exhaustive list of independent variables that are thought to correlate with both treatment selection and treatment outcome in order to reduce or remove the endogeneity of drug variable. In strategy 1, only the first episode of each patient during the sampling period was used. A multivariate Cox proportional hazard model was estimated with the specification as the following, (27) 2 (0 = 20 (0 exp(X/3 + y * D) where 20(O , is the baseline hazard rate of drug discontinuation that changes over time; 43 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. D, is a vector of dummy variables denoting the drug used as initial therapy; X , is a vector of personal characteristics at the baseline thought to affect the duration of a treatment episode, including: patient’s demographic characteristics, patients’ health status in prior 6 months, patients’s compliance propensity with antipsychotic therapy, sam edrug variable, interaction terms between same_drug variable and different types of antipsychotics (typical antipsychotics as reference group). This model is estimated with partial likelihood without knowing baseline hazard. For this model, the effect of the treatment is measured by the estimate of y . Under unconfoundedness assumption, the assignment D is exogenous, hence the estimated y would be unbiased. 5.3 Strategy 2 - estimation of treatment effect adjusting for selection bias The treatment effect literature discussed above is focused dominantly on binary choice, which involves only two treatment options. In polychotomous choice setting, Lee’s selection model (Lee 1982, 1983) is still the main framework. However, most of the selection models in the literature deal with linear model with normality assumption, for example, the studies of log transformed cost. Cox proportional hazard model is nonlinear. Interestingly, through a transformation, nonlinear Cox model can be transformed to a linear model with type I extreme value distribution error term. Least-square regression can then be used to estimate the coefficients of explanatory variables if the data are not heavily censored (Keifer, 1988). This is the approach taken in this study. In order to 44 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. simplify the estimation, the first episode for each patient during the sampling period was use. Linear transformation The linear transformation of Cox proportional hazard model takes the form (Kiefer, 1989) (28) - \ n h \ t ) = X/3 + u This transformation involves the calculation of integrated baseline hazard rate nonparametrically. Specifically, all the completed episodes of therapy in the sample of size n are ordered from the shortest to the longest, t] <t2 <... < tk . Let h - be the number of completed spells of drug therapy f ., for j = 1 ,...K . Let nj be the number of spells not completed before duration (people at risk). The baseline hazard, A° (I j), is the probability of completing a spell at duration 1 2, conditional on the spell’s reaching duration t ., as expressed by 1° ( |; ) = hj /«■. Cumulative hazard rate (CHR) can then be computed as A( ) (/j) = ^ A° (l J) .Note, the longer the duration of the therapy is, the larger '-j the CHR. Correspondingly, the larger the CHR is, the smaller the - In A0 (/). In our data, censored patients were used to compute baseline hazard rate as included in the denominator of the equation, but were dropped in the following steps, because the least- square regression require non-censored data. With 2% censored sample, the result wouldn’t be affected much. 45 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. After this transformation, Lee’s two-stage model was specified as follows: (29) U *sj = Vsj + rjsi, (30) Z ), = s, iff U], > MaxU\,, k = 1,2, 3,4 (k * s ), (31) Ys = -lnA °v = Xff +us , where s=l, 2, 3, 4. Equation (29) and (30) relate the rule of assigning patients to treatment to the potential treatment outcome through random utility maximization, f/* is the expected utility associated with choice 5 for individual i. It is comprised of a deterministic component Vsi and a random component rjs, s=l,2,3,4 indexes four drugs, i=l ...N indexes individual patients. In equation (31), one potential outcome associated with each choice s for individual i is specified, where u is a random variable with type I extreme value distribution. Discrete choice model Multinomial logit model (MNL) is most widely used to model polychotomous choice. The assumption of the model is that the error terms are independently and identically distributed (iid) with the type I extreme-value distribution. Under this assumption, the model demonstrates a property called ‘irrelevance of independent alternative’ (IIA). According to this assumption, the four drug choices being evaluated in this study are independent to each other. That is to say the covariates that influence the drug choice are fully captured by the observed variables, there is no unobserved confounders left in the error term that cause ss to be correlated. However, this assumption is not likely held in 46 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. our study. At the time when atypical agents are recommended as first-line therapy, patients who are still treated with typical antipsychotics should respond well to them without side effects. On the other hand, in the presence of some complicating clinical problems, such as insomnia, dysphoria, suicidal behavior, comorbid substance abuse, cognitive problems, compulsive water drinking, atypicals are strongly preferred over typicals (McEvoy et al. 1999). Put in another way, patients using typical and atypical agents are quite different, while patients using alternative atypical agents may share some unobserved similarities, which are absorbed into error terms es. In order to approximate the correlation between drug choices, a more flexible model was adopted for this study. McFadden (1975) has shown that any model that specifies choice probabilities, including models that do not exhibit IIA, can be approximated with logit models. Let P* = f (zjn;zjn’ ,sn) be the “true” model, where zm is observed data relating to alternative i as faced by decisionmaker n , zjn is observed data relating to alternatives other than i faced by decisionmaker n , and sn is a vector of characteristics of the decisionmaker. Define Wjn = log(/>*) and evaluate logit probabilities based on Win: ew » e'og,;: P: Pin=- L in y e Wi» y ^ y p * 2-1 jeJ „ "/W y ./« This equation shows that any choice model can, with an appropriate choice of Wjn, be put into the logit form, giving rise to the term “mother logit”. All that is required is that additional variables be added to representative utility, in particular, variables that related to alternatives other than the one for which the representative utility is designated. While 47 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. the resulting model might not be consistent with utility maximization, a logit model can always be used to fit whatever the choice probability there is (Train K, 1986). Therefore a variable indicating how long atypical antipsychotics were on market was included in the logit model as described in the data section. Calculate selection correction terms Following Lee’s notation, (32) s x = MaxU*s - tjs (k & s) , (33) I = s, iff ex <Vs. The marginal probability of choosing s can be expressed as (34) prob(D = s) = prob(ex <VX ) = Fx (Vs) = cxp(X/3x) / exp(X/?k). k The selectivity bias terms for each choice, As, were derived by taking expectation of ux conditional on truncated distribution of s x, that is E(ux \ s x < Vx). A crucial property used in deriving the conditional truncated mean is the linearity of ux conditional on ex. Because us is type I extreme value distribution, sx can be transformed to that distribution using a strictly increasing function J = G~'F, where G is a cumulative function of type I extreme value distribution. As s is specified to have the distribution function F{s ) , s =J(e) = G~'F(e) will be a random variable of type I extreme value distribution. In this way, us can be expressed as a linear equation of s* x, 48 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. (35) us = (J(ss ) - j U j ) + vv = (s*s - ftj ) + v, where vv and s] are independent, ju, = E(s*). Then selectivity bias term for the choice s is (36) E(s] | g; < °J F(K) where (37) ft(W s))= C U f j i e ^ d s l . •1 / ( —CO) The cumulative probability function and density function of type I extreme value distribution are as the follows, (38) G(x) = ex p (-ex p (-x )), (39) g(x) = exp(-x) * ex p (-ex p (-x )). The inverse function of G(x) can then be expressed as (40) G ' = - In - ln(x). So the predicted probability from mother logit model, F(VS), was plugged in G 1 , and * * * * the result became the upper bound in the integration ju( J(VS)) = s s g , {ss )dss . J-c o Outcome regression with correction terms Each drug-specific outcome equation was run using data for those patients who chose that drug. The dependent variable is the negative log transformed baseline CHR. In each 49 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. outcome equation, four correction terms were included to approximate correlation of drug selection, t ! a .i F ( K ) Note the inverse relationship between y and the duration of therapy. Therefore, positive parameters in the model would be associated with decrease in duration. OLS regression with robust standard errors (White 1980) was computed to account for heteroskedasiticity of errors. Confidence intervals of estimated outcome After estimation of (s=l,2,3,4), the entire sample was used to derive the expected outcome for each drug with j)v = Xfts . Bootstrap estimation procedures with 500 replications were performed to calculate the confidence intervals of the expected outcomes (Efron 1986). To make the results more interpretable, the original data on days of therapy was linearly projected to the corresponding cumulative hazard rate on the entire sample: (42)t = ro+r i *y = r 0+ ri*F°(t). The calculated ys was then converted to cumulative hazard rate with exp(-j)) and plugged back to equation W to get ts. The same bootstrap procedures were performed to derive the confidence interval of days on therapy for each drug. 50 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Model specification and discussion of exclusion restriction In estimation of latent index model, if the variables used to model the probability of receiving the treatment are the same as those used in the outcome equation, the derived selection bias terms tend to be highly correlated with the observed variables in the outcome equation, resulting in a severe multicollinearity problem. The solution is to identify instrumental variable (IVs) that affect the drug selection but are not correlated to the outcome. In this section, we focus on the rationing behind the choice of covariates in selection and outcome equations respectively, and explain why some variables can be used as IV. Physicians were the main decision-makers. When making the treatment decision, physicians are assumed to take into consideration or be affected by the following factors: • Patient’s demographic characteristics • The patient’s health status in prior 6 month, including a list of proxy variables, such as mental disorder profile, diagnoses of diabetes, arrthymia, hyperlipidemia or EPS, a count of other medical conditions, prior health care use in acute/psychiatric hospital or long-term care service, suicidal attempt, total treatment cost in prior 6 months • Antipsychotic use in prior 6 months. Physicians usually refer to patients’ drug use history when making a prescription decision, and would probably maintain the same drug therapy as before unless patients has not respond well in past treatment attempt 51 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. • Patient’s propensity for compliance with antipsychotic therapy • Medication diffusion rate • Time trend • How long atypical antipsychotics were on market, as a proxy of drug attribute Most of the variables listed above also affect the treatment outcome, such as patient’ demographic characteristics, and health status in prior 6 months. However, some variables are much more related to drug selection than to the treatment outcome. Antipsychotics do not have lasting effect, therefore, which antipsychotic medication was used in the prior 6 month was not included in the outcome equation of treatment compliance. Medication diffusion rate, time trend, and how long atypical antipsychotics were on market represent exogenous factors affecting drug choice and unrelated to the patient’s clinical response to the treatment. Therefore, they are included in the selection equation and used as IVs. 5.4 Strategy 3 - panel data estimation The primary motivation for using panel data is to solve the problem of omitted variables. For example, patients have different genetic backgrounds that determine how a patient may respond to particular antipsychotics. If we assume that this factor is correlated with the drug selection, then explicit control of this variable could largely remove the selection bias. This assumption is equivalent to the unconfoundedness assumption, with the form (43) D 1.Y \ X ,v l 52 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. where v, is individual-specific and time-invariant factor that could correlate with X. The limitation of this assumption is that this factor maybe not able to capture all the factors that affect the drug choice. The following is adopted from Lancaster (1990). Allowing v, enter Cox model multiplicatively, the model can be written as (44) AtJ (0 = A0 (t)v, exp(x'/7) where xf/ vary between persons and between spells for the same person, / =1,2; / = 1, ...N. Suppose that for person i, we observed two uncensored spells, which are independently distributed, given v,, then the partial likelihood function for individual i is (15) L _ (cxphyyj))1 ^- exp(x;,/?) + exp(x;2/?) where di = 1 if the first spell is shorter, zero otherwise. Equation (44) can be written as (46) Z ,= * f '( I - * ,) 1 "* ' where = --------------------------- l + exp((x,2 - x n) P) Now the partial likelihood function corresponds to a binary regression model with logit form for the event probability and with regressor vector given by xi2 - x n for person /. The existence of a solution for ft depends on there being variation in x over spells. 53 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Partial likelihood procedure successfully eliminate person-specific effects at a price of eliminating the effect of observed regressors that do not vary between spells for the same person, for example, gender and race. The model makes two assumptions. First, spells are assumed to be independent conditional on v, and the regressors. This assumption rules out lagged duration dependence. Second, the model assumes that no other unobservable factors exist that may determine the drug selection after controlling for stable individual characteristics v,. This argument is plausible. Clinically, the choice of antipsychotics is mostly driven by whether the patient could tolerate the side effect of the medication, which is likely determined by the genetic make-up of the patient. Horowitz and Lee (2004) proposed a new method to handle panel duration data with varying number of spells and dependent censoring. Though this method represents an advance in analyzing duration data with independent censoring, there are a number of drawbacks of this model that prevent us from using it. First, the model has to be estimated with multivariate kernel technique, which requires an enormous computational burden. Second, the model assumes a pure renewal process with no other states in between the two spells. But in our data, there exist gaps between treatment episodes with gaps as long as hundreds days being quite common. Third, Horowitz and Lee did not come up with a method about how to make use of all the spells efficiently. Therefore, our study would rather cling to a simplified panel data method, which is easier to understand and implement. 54 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In order to use panel data method, patients with at least two complete episodes were included and only the first two episodes were used. For patients who initiated only one restart treatment episode during January 1999 and March 2001, we searched the period after March 2001 in order to identify one more episode (excluding zipresidone and combo-therapy). According to Chamberlain (1985), “it is valid to select the first two spells of everyone, provided that the sampling interval is long enough so that (essentially) everyone has two complete spells.” 77% of the patients had sufficient data with which to identify 2 complete episodes of drug therapy. 55 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. CHAPTER 6: RESULTS 6.1 Descriptive Statistics During sampling period (January 1999 to March 2001), there were 51,042 uncensored restart episodes that met the selection criterion for this study. The descriptive statistics for the demographics, comorbiditiy profile, and prior health care utilization for each drug group were presented in Table 6.1. F or Chi-square tests were conducted to test whether these values are the same across four drug groups. Patients tend to use the same type of antipsychotics as prescribed in prior 6 months. For example, for patients initiating a restart episode with typical antipsychotics, 65.11% used TAP in prior 6 months, in contrast, only 7.01% used olanzapine, 4.72% used risperidone, and 1.76% used quetiapine. This pattern held true across all 4 types of antipsychotics, but to a lesser extent with quetiapine group, which received FDA approval in October 1997. Atypical antipsychotics were more likely to be prescribed to very young patients (<25 years), while typical antipsychotics to patients older than 45. Risperidone was most likely used in oldest population (> 65years), whereas olanzapine in patients aged 25-35, and quetiapine in patients 35-45. Patients who were disabled, black, or belong to other race category tend to use typical antipsychotics, while patients who were white, Hispanic or Asian were more likely to receive atypical antipsychotics. Quetiapine was least used in patients residing in urban areas. Olanzapine was most used in male patients. In general, quetiapine prescription was associated with more severe patients, as indicated by higher percentage in prior psychotropic drug use, other mental disorder diagnosis, acute and 56 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. psychiatric hospitalizations, suicidal attempts, and total treatment cost in prior 6 months. Patients with diabetes were least likely to get olanzapine, reflecting the physicians’ awareness of olanzapine’s higher risk in metabolic system effect. The use of atypical antipsychotics increased over time with the most prominent increase seen for quetiapine. Table 6.1: Descriptive statistics of baseline characteristics by drug TAP (N = 2 0 3 7 0 ) O lanzapine ( N = 16 0 6 9 ) Risperidone (N = 1 1 7 4 8 ) Quetiapine (N = 2 8 5 5 ) Dem ographics A g e in years 4 7 .1 8 * * * 4 4 .4 3 *** 4 5 .8 4 * * * 4 4 .2 0 * * * A g e b y ca teg o ry (col.% ) 18 < = A g e < 2 5 3.08% *** 5.61% *** 6.14% *** 6 .34% *** 2 5 < = A g e < 3 5 13.30% *** 17.52% *** 16.33% *** 16.32% *** 3 5 < = a g e < 4 5 2 9 .3 3 % * * * 30.82% *** 2 8 .48% *** 3 1 .0 3 % * * * 4 5 < = a g e < 5 5 27 .9 7 % * * * 25 .5 6 % * * * 2 4 .63% *** 2 6 .7 3 % * * * 5 5 < = a g e < 6 5 15.45% *** 1 2 .6 8 % *** 12.41% *** 12.155% *** A g e > = 6 5 1 0 .8 6 % *** 7.82% *** 1 2 .0 1 % *** 7.43% *** R ace b y categ o ry (col.% ) W hite 4 9 .7 6 % * * * 50.96% *** 50.64% *** 55 .2 4 % * * * B lack 2 0 .4 0 % * * * 17.70% *** 18.20% *** 16.01% *** H ispanic 2.3 3 % * * * 3.09% *** 3.36% *** 3 .29% *** A sian 1.67% *** 2 .91% *** 2 .44% *** 1.72% *** O ther race 25 .9 3 % * * * 25.3 3 % * * * 25 .3 6 % * * * 23 .7 5 % * * * O ther dem ographic factors D isab led 91 .1 9 % * * * 88.32% *** 8 7 .54% *** 87 .74% *** Urban 7 6 .79% *** 78 .69% *** 7 8 .69% *** 74 .5 7 % * * * M ale 50.0 1 % * * * 54 .52% *** 4 9 .83% *** 4 7 .8 8 % * * * Drug use in prior 6 m onths T yp ica l an tip sy ch o tics 65.1 1 % * * * 16.29% *** 14.69% *** 19.82% *** O lanzapine 7 .01% *** 64 .35% *** 8.61% *** 19.86% *** R isp eridone 4.7 2 % * * * 6 .76% *** 62 .7 9 % * * * 12.78% *** Q uetiap in e 1.76% *** 2.23% *** 1.95% *** 4 8 .0 9 % * * * D ep ot 3.71% *** 3.35% *** 2.4 1 % * * * 4 .31% *** M o o d S tab ilizer 19.04% *** 25 .3 6 % * * * 23 .2 5 % * * * 30.7 9 % * * * A n tid ep ressant 37 .9 0 % * * * 44 .6 6 % * * * 4 3 .9 3 % * * * 51.4 9 % * * * A n tiseizu re 13.24% *** 12.71% *** 14.14% *** 18.35% *** Com pliance propensity in history D ays on an tip sy ch o tic therapy 2 6 5 .0 3 * * * 2 5 3 .8 4 * * * 2 7 1 .4 5 * * * 2 5 9 .1 8 * * * G ap in an tip sy ch o tic therapy 2 4 3 .7 0 * 2 5 9 .7 5 * 2 6 0 .4 9 * 2 4 7 .7 3 * Diagnosis o f other m ental disorders in prior 6 months O rganic p sy ch o tic disorder 13.22% *** 16.51% *** 16.85% *** 17.58% *** B ipolar 8.52% *** 12.71% *** 10.97% *** 15.97% *** D ep ression 4.8 6 % * * * 6.37% *** 6 .15% *** 8.48% *** A n x iety 7.81% *** 9.55% *** 8.83% *** 11.38% *** Su bstance abuse 6 .43% *** 7.95% *** 6 .70% *** 9.00% *** O ther a ffectiv e disorder 0.76% *** 1.14% *** 0.7 7 % * * * 1.33% *** Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 6.1: Continued D em en tia 3 .00% *** 2 .15% *** 3.2 8 % * * * 2 .80% *** O ther m ental disorder 12.29% *** 13.32% *** 14.51% *** 16.85% *** Diagnosis o f medical conditions in prior 6 m onths D ia b etes 2 3 .5 0 % * * * 2 0 .0 1 % *** 2 1 .7 0 % * * * 2 2 .2 8 % * * * H yperlipidem ia 1 1 .440/0*** 10.80% *** 10.08% *** 10.51% *** A rrythm ia 9 .92% *** 8.98% *** 9.8 1 % * * * 10.30% *** EPS d ia g n o sis 0 .70% *** 0 .77% *** 0.7 2 % * * * 1.47% *** N u m ber o f other illn e sses 3 .0 1 * * * 2 8 3 * * * 2 9 2 *** 3 .0 8 * * * Health care utilization in prior 6 m onths A cu te h osp italization 5.66% ** 5.10% ** 5.77% ** 6.44% ** P sychiatric h ospitalization 5.77% *** 7.31% *** 7.0 7 % * * * ' 8.83% *** L ong-term care 7.17% *** 6.17% *** 9.1 2 % * * * 6.80% *** A ttem pted su icid e 3 .06% *** 4 .08% *** 3 .80% *** 5.60% *** T otal treatm ent co st 6 0 3 8 .2 2 * * * 6 6 9 2 .3 1 * * * 7 4 0 1 .3 3 * * * 7 8 7 0 .9 7 * * * Time trend dum m ies (row.% ) Q1 99 4 7 .0 1 % * * * 29 .1 7 % * * * 2 0 .3 6 % * * * 3 .46% *** Q 2 99 4 3 .2 0 % * * * 3 0 .46% *** 2 1 .6 3 % * * * 4.7 1 % * * * Q 3 99 39 .9 5 % * * * 3 1 .11% *** 2 3 .4 3 % * * * 5.51% *** Q 4 99 3 8 .56% *** 3 1 .80% *** 2 3 .7 4 % * * * 5.90% *** Q1 0 0 37 .3 2 % * * * 3 2 .17% *** 2 4 .7 0 % * * * 5.81% *** Q 2 0 0 32 .0 8 % * * * 3 4 .43% *** 2 5 .5 8 % * * * 7.92% *** Q 3 0 0 30.13% **^ 34.6 4 % * * * 2 6 .2 1 % * * * 9.02% *** Q 4 0 0 2 6 .6 4 % * * * 36.0 1 % * * * 2 7 .8 1 % * * * 9.54% *** Q1 01 2 6 .0 3 % * * * 36.6 4 % * * * 2 6 .9 9 % * * * 10.35% *** * P < 0.05, ** P < 0 .0 1 , ***P< 0.001 6.2 Baseline model of compliance There were 51,042 uncensored and 1,107 censored restart episodes included in the baseline model of compliance. The model was first estimated with typical antipsychotics as reference group, specified as the following: A(t) = A0(t)exp(X/3 + cc*SD + r] * 0 + y2* R + y,*Q + y4 * 0 * S D + y5*R*SD + y6*Q*SD ) where X is a vector o f patient’s characteristics at the baseline, including demographics, patient’s health status in prior 6 months, and patient’s compliance history; SD is a acronym for Same Drug variable. 58 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 6.2 summarized the coefficients that pertain to each study drug when used as a real restarter (SD=1) or as a delayed switcher episode (SD=0). Table 6.2: Model parameters t a p rr TA P ls Or r o L S Rr r Rl s Qr r Ql s SD=1 a — a — a — a — 0 = 1 , SD=0 — -- Y\ Y\ - -- - -- R=l, SD=0 - -- - ' -- Yi Yi - -- Q = l, SD=0 — -- -- -- - - Yz Y 3 0 = 1 , SD=1 - -- r* -- - - - — R = l, SD=1 — - - - Ys — — — Q =1,SD =1 — - - - ~ - . Y 6 -- R R : real restarter, L S: late sw itch er, SD : sam e d ru g The parameter estimates from the model are shown in table 6.3. First, The parameter for the SameDrug variable was negative and significant. It indicates that patients receiving typical antipsychotics were less likely to terminate the therapy if starting with the same Table 6.3: Parameter estimates from baseline model (TAP as reference) V a r ia b le s P a r a m e te r e stim a te s (* * * ) O lan zap in e -0 .4 7 0 6 R isperidone -0 .4 3 9 3 Q uetiapine -0 .3 5 0 5 Sam e drug -0 .4 0 4 3 0 1 a n za p in e*sam e drug 0 .3 7 1 5 R isp erid on e*sam e drug 0 .3 5 2 6 Q uetiap in e*sam e_d ru g 0.3441 * P < 0.05, * *P < 0.01, ***P< 0.001 antipsychotic medication as used in the prior treatment episode than switching to an alternative drug. Second, a significant interaction effect exists between type of episode and the drug used, documenting distinctive differential effect of atypical vs. typical antipsychotics between real restarters and late switchers. 59 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. The differential effects of atypical antipsychotics in comparison to typical antipsychotics were summarized separately for real restarters and late switchers in table 6.4. In both real restarters and late switchers, patients using atypical antipsychotics are less likely to discontinue the therapy compared to typical antipsychotics. Moreover, this differential effect was much larger in late switchers than in real restarters. For example, returning patients receiving olanzapine were 10% less likely to terminate the treatment than returning patients receiving typical antipsychotics, while switching patients receiving olanzapine were 37% less likely to terminate the treatment than switching patients receiving a typical antipsychotic medication. These results imply that typical antipsychotics may work well in some patients who use the medication periodically, while the superior treatment effects of atypical antipsychotics are more evident in late switchers. Table 6.4: Results from baseline model by type of episodes (TAP as reference) Drug use Real restart (T A P real restart as ref.) Late switcher (T A P late sw itch er as ref.) Param eter Odds ratio Param eter Odds ratio O lanzapine -0 .0 9 9 1 * * * 0 .9 0 6 * * * -0 .4 7 0 6 * * * 0 .6 2 5 * * * R isp eridone -0 .0 8 6 7 * * * 0 .9 1 7 * * * -0 .4 3 9 3 * * * 0 .6 4 4 * * * Q uetiapine -0 .0 0 6 5 * * * 0 .9 9 4 * * * -0 .3 5 0 5 * * * 0 .7 0 4 * * * * P < 0.05, * *P < 0.01, ***P< 0.001 In order to test if there exist differences in drug discontinuation among atypical antipsychotics, the model was rerun with olanzapine as the reference group: 2(t) = A0(t)exp(X/3 + a * SD + y{ *TAP + y2 * R + y3*Q + y4 * TAP * SD + y5* R* SD + y6*Q* SD) The parameter estimates from this model are shown in table 6.5. The results implied three points. First, patients using quetiapine were more likely to stop the treatment compared to 60 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. olanzapine ( y3 = 0.1201, p < 0.001), and this differential effect was same in late switchers and real restarters. Second, there is no difference in rate of discontinuation between olanzapine and risperidone, irrespective of type of restart episodes. Third, the rate of treatment discontinuation is the same for olanzapine in late switchers and real restarters, as indicated by the nonsignificant parameter for the same drug variable. Table 6.5: Parameter estimates from baseline model (olanzapine as reference) Variables Param eter Odds ratio R isp eridone 0 .0 3 1 3 1.032 Q uetiapine 0 . 1 2 0 1 *** 1 .128*** Sam e drug -0 .0 3 2 8 0 .9 6 8 R isp erid o n e* sa m e drug -0 .0 1 8 9 0.981 Q uetiap in e*sam e_d ru g -0 .0 2 7 4 0 .9 7 3 *. P < 0.05, * * P < 0 .0 1 , ***P < 0.001 6.3 Compliance model with treatment selection adjustment In this model, the non-linear Cox model was first transformed to a linear model, in which the dependent variable is negatively related to the duration of the drug therapy (see section 5.3). Then the adjustment of treatment selection was done through explicit modeling of drug choice with mother logit model. The predicted probability of drug choice obtained from mother logit was used to derive the selectivity bias terms that would enter the outcome equations. Next, the subsample of patients using one of the four medications was used to obtain unbiased parameter estimates for factors that affect the outcome associated with that medication through OLS regression. Altogether, there are four sets of such unbiased parameter estimates. Next, the total sample was used in each outcome equation to derive the expected outcome as if everyone were forced to take each 61 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. of the study medications. The original outcome (negative log-transformed cumulative hazard ratio) was then converted to days on therapy for ease of interpretation. Finally, bootstrap methods were used to produce the confidence intervals of the original and converted outcomes. While the main focus of the analysis is on the outcome equations, the estimated parameters from mother logit are presented as they provide insight into how the atypical antipsychotics were used by physician to treat patients. The results of the outcome estimation process are presented in the order of estimation, with emphasis on the grouped OLS and bootstrap results. Results from mother logit model All the baseline characteristics from the Table 6.1 were initially included in the mother logit model. Those factors that were not statistically significant were then combined into broader categories and dropped from the model. For example, acute and psychiatric hospitalizations were merge into a single hospitalization variable. Asian and Hispanic racial groups were combined with the ‘other race’ category. All of the mental disorders other than bipolar and dementia were added to other mental disorder category. The results of the final parsimonious model of drug selection are presented in Table 6.6 as odds ratio in comparison to typical antipsychotics. In general, atypical antipsychotics was more frequently used to treat very young patients (<25), patients with other mental disorders (for example, diagnosis of bipolar, or taking mood stabilizer), and patients with higher total treatment cost in prior 6 months. Typical 62 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. antipsychotics was more frequently used to treat black patients, patients with a better compliance history, prior depot use, prior nursing home care, or increased number of medical comorbidities. Male patients were more likely to receive olanzapine, while patients with diagnosis of diabetes or dementia, or prior use of an anti-seizure medication were less likely to use olanzapine. Patients living in urban areas, or with the diagnosis of other mental disorder were more likely to use risperidone. The likelihood of risperidone prescription decreased over time. Patients with prior EPS diagnosis were more likely to receive quetiapine. The type of antipsychotics used in prior 6 months was highly predictive of drug selection. For example, prior olanzapine use was associated with 26 folds increase in the likelihood of the patient returning on olanzapine. Similar results were found with risperidone and quetiapine. As expected, the likelihood of using a particular drug is positively correlated with the drug diffusion rate for that medication. Some results are counterintuitive, for instance, patients with more medical conditions tend to receive typical antipsychotics. Table 6.6: Results from mother logit (TAP as reference group) Olanzapine Risperidone Quetiapine Dem ographics 2 5 < = a g e < 3 5 0 .9 2 1 5 0 .7 6 8 0 * * 0 .8 2 1 4 3 5 < = a g e < 4 5 0 .7 7 7 1 * * 0 .6 5 4 2 * * * 0 .8 0 9 3 4 5 < = a g e < 5 5 0 .7 6 0 0 * * * 0 .6 4 7 8 * * * 0 .7 7 2 1 * 5 5 < = a g e < 6 5 0 .7 2 6 8 * * * 0 .6 2 5 4 * * * 0 .7 4 9 5 * A g e > = 6 5 0 .6 6 7 9 * * * 0 .7 4 9 4 * * 0 .6 4 2 6 * * D isab led 0 .9 2 9 7 0 .9 2 7 3 0 .9 2 6 0 B lack 0 .9 1 0 4 * 0 .9 6 0 2 0 .7 7 4 7 * * * O ther races 0 .9991 1 .0596 0 .9 5 2 3 Urban 1.0555 1 .1 649** 1.0210 M ale 1 .1 1 7 7 * * * 0 .9 6 2 6 0 .9 1 6 0 Drug use in prior 6 m onths T y p ica ls 0 .0 8 5 4 * * * 0 .0 8 8 3 * * * 0 .0 9 8 6 * * * O lanzapine 2 4 .8 1 2 0 * * * 1.3 5 2 6 * * * 2 .5 4 6 2 * * * R isperidone 1 .3 8 4 0 * * * 3 1 .4 8 1 1 * * * 1 .9 8 3 0 * * * 63 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 6.6: Continued Q uetiapine 1 .0 0 9 0 0 .9 0 6 6 4 3 .7 2 6 2 * * * D E P O T 0 .7 9 6 8 * * 0 .6 3 1 1 * * * 0 .6 2 5 1 * * * M o o d stab ilizer 1.0892* 0 .9 8 7 6 1 .3 1 4 8 * * * A n tid ep ressant 1 .0 6 0 6 1.0621 1 .3 0 9 7 * * * A nti seizu re m ed ication 0 .8 8 5 0 * * 0 .9 9 7 3 0 .9 8 3 9 Com pliance propensity in history Gap in an tip sy ch o tic treatm ent 0 .9 9 9 9 0 .9 9 9 9 * * 0 .9 9 9 7 * * * D ays on a n tip sy ch o tic therapy 0 .9 9 9 9 * * 0 .9 9 9 9 * 0 .9 9 9 6 * * * Diagnosis o f other m ental disorders in prior 6 m onths B ipolar 1 3 3 2 9 *** 1.1273* 1 .2631** D em en tia 0 .7 6 7 1 * * 0 .9 5 8 0 1 .0 0 3 6 O ther m ental disorders 1 .0399 1.0970* 1.0751 Diagnosis o f m edical conditions in prior 6 months D iabetes 0 .8 3 9 7 * * * 0 .9 0 9 6 * 0 .9 5 2 2 H yperlipidem ia 0 .9 3 3 3 0 .9 4 2 3 0 .8 5 7 8 A rrthym ia 0 .9 7 3 3 0 .9 5 3 7 1 .0749 E P S d ia g n o sis 1 .2 4 6 8 1 .1 8 2 0 1 .6 9 4 7 * N u m ber o f other illn ess 0 .9 4 2 9 * * * 0 .9 6 4 1 * * * 0 .9 4 1 2 * * * Health care utilization in prior 6 months H ospitalization 0 .8 6 1 0 * * 0 .9 1 7 7 0 .8 3 3 3 * L ong-term care 0 .7 4 4 7 * * * 0 .9 1 4 8 0 .7 7 9 5 * Su icidal attem pt 1 .0886 1 .0 8 2 7 1 .1 8 4 4 T otal co st in 2 0 -4 0 % quintile 1 .4 0 2 1 * * * 1 .2 4 3 2 * * * 1 .4 5 8 5 * * * T otal co st in 4 0 -6 0 % quintile 1 .7 3 3 5 * * * 1 .5 0 8 5 * * * 2 .0 2 8 1 * * * T otal co st in 6 0 -8 0 % quintile 2 .1 3 9 0 * * * 1 .7 1 1 0 * * * 2 .1 7 5 9 * * * T otal co st in 8 0 -1 0 0 % quintile 2 .0 3 2 1 * * * 1 .5 7 0 6 * * * 2 .1 6 3 4 * * * Drug diffusion rate by county & time T yp ical a n tip sych otics 0 .9 9 2 0 * * * 0 .9 9 0 1 * * * 0 .9 9 0 6 * * O lan zap in e 1 .0 1 1 9 * * * 0 .9 9 4 4 0 .9 9 2 2 R isp eridone 0 .9951 1.0 1 5 4 * * * 1.0031 Q uetiapine 0 .9 9 6 4 1 .0019 1 .0 3 6 5 * * * Time trend dum m ies 2 nd quarter, 1999 0 .8 9 5 0 0 .8 4 2 5 * 0 .8841 3 rd quarter, 1999 0 .8 6 7 0 0 .7 6 0 8 * 0 .7 0 5 8 4 th quarter, 1999 0 .7 5 9 1 0 .6 0 7 9 * * 0 .5 7 1 2 1 st quarter, 2 0 0 0 0 .7 1 4 0 0 .5 1 4 0 * * 0 .4 5 1 2 * 2nd quarter, 2 0 0 0 0 .8 2 9 6 0 .4 6 1 9 * 0 .5 0 7 5 3 r d quarter, 2 0 0 0 0 .7 2 3 2 0 .3 9 7 8 * 0 .4 3 6 6 4'irquarter, 2 0 0 0 0 .7 8 5 0 0 .3 7 8 8 * 0 .4 0 9 7 1 st quarter, 2001 0 .6 8 3 6 0 .2 8 3 7 * * 0 .3 2 2 9 Average market month 1 .0306 1 .0 6 1 4 * * 1.0625* *P < 0.05, * *P < 0.01, ***P<0.001 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Results from outcome equations with selection bias terms The dependent variable in the outcome equation is the negative log-transformed baseline cumulative hazard ratio. For each category of drugs (s=1,2,3,4), the subsample that chose s was used to run the corresponding outcome equation. Four sets of parameters were derived and the results were shown in Table 6.7. Note that the negative parameters were associated with longer duration according the model specification. The parameter estimates were different across 4 drugs. In most cases, the signs were the same, but the magnitudes of the effect were different. Table 6.7: Results from outcome equations with selection bias terms Typical use (2 0 3 7 0 ) Olanzapine (1 6 0 6 9 ) Risperidone (1 1 7 4 8 ) Quetiapine (2 8 5 5 ) Same drug -0 .5 5 0 2 * * * -0 .0 1 1 4 -0 .0 0 9 7 -0 .0 9 0 6 Dem ographics 2 5 < = a g e < 3 5 0 .0 2 8 9 0 .0 0 4 0 -0 .0 6 6 0 0 .0 7 7 9 3 5 < = a g e < 4 5 -0 .0 1 8 7 -0 .0 3 4 6 -0 .0 4 4 0 0 .0 4 5 1 4 5 < = a g e < 5 5 -0 .0 3 6 6 -0 .0 5 0 9 -0 .1 0 1 7 * 0 .0 3 0 0 5 5 < = a g e < 6 5 -0 .1 1 0 8 * -0 .1 0 6 9 * -0 .0 8 8 8 0 .0 0 5 2 A g e > = 6 5 -0 .1 1 6 1 * -0 .1 6 6 9 * * * -0.0321 -0 .0 5 3 8 D isab led -0 .0 9 5 1 * * 0 .0 2 9 2 0 .0 2 4 3 0 .0 1 5 6 B lack 0 .0 2 9 2 0 .1 2 9 9 * * * 0.1 0 0 1 * * * 0 .0 8 0 3 H ispanic -0 .0 3 3 2 0 .0 9 0 0 * 0 .0 2 1 9 -0 .1 2 8 7 O ther races -0 .0 2 9 7 0.0122 -0 .0 1 4 5 -0 .0 3 0 3 Urban -0 .0 6 4 0 * * 0 .0 0 1 8 0 .0 3 3 7 0 .0 3 7 5 M ale -0 .1 0 6 0 * * * 0 .0 1 5 4 -0 .0 0 1 4 0 .0 0 6 0 O ther psychotropic drug use in prior 6 months D ep o t 0 .0 5 1 9 0 .0 2 5 9 0 .0 5 4 6 -0 .0 7 6 5 M o o d stab ilizer -0 .0 5 4 7 * -0 .0 1 6 7 -0 .0 7 6 0 * * 0 .0 4 5 6 A n tid ep ressant 0 .0 4 6 9 * 0 .0 1 3 4 -0 .0 1 0 9 -0 .0 1 7 2 A n ti-seizu re 0 .0 6 0 1 * -0 .0 4 1 9 -0.0501 0 .0311 C o m p lian ce prop en sity in history Gap in an tip sy ch o tic therapy 0.0 0 0 1 * * * 0.0000 0.0000 0.0000 D ays on an tip sy ch o tic therapy -0 .0 0 0 7 * * * -0 .0 0 0 6 * * * -0 .0 0 0 6 * * * -0 .0 0 0 4 * * * Diagnosis o f other m ental Disorders in prior 6 months O rganic p sy ch o tic d isea se -0 .0 4 3 9 0 .0 8 1 9 * * * 0 .0 4 5 0 0 .0 9 3 4 B ipolar 0 .0 7 3 4 * 0 .0 9 0 9 * * * 0 .1 5 4 0 * * * -0 .0 4 4 6 D ep ression 0 .0 5 4 4 0 .0 4 5 3 0 .0 1 6 4 0 .1 7 2 4 * A n x iety 0 .0 7 9 0 0 .0 3 8 5 -0 .0 1 1 8 0 .0 8 3 1 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Table 6.7: Continued Su bstance abuse 0 .1 2 1 4 * * 0 .1 7 6 3 * * * 0 .1 5 2 3 * * * 0 .1 3 1 1 * O ther a ffectiv e disorder 0 .2 1 7 9 0 .0581 0 .0 2 8 0 0 .0 3 9 8 P erson ality disorder 0 .1 1 0 7 0 .3 9 6 5 * * * 0 .0 6 0 0 -0 .0 7 6 6 D em en tia -0 .2 6 2 5 * * * -0 .1 4 8 6 * -0 .1 2 0 4 * -0 .3 2 4 5 * O ther m ental disorders -0 .0 0 4 7 0 .0 0 7 0 0 .0 5 4 0 -0 .0 4 0 1 Diagnosis o f M edical conditions in prior 6 months D ia b etes -0 .0 3 3 9 0 .0 2 0 7 0 .0 1 4 3 0 .0 5 9 6 H y p erlip id em ia -0 .0 7 4 6 * * -0 .0 6 0 1 * -0 .0 8 5 2 * * -0 .0 4 3 3 A rrthym ia 0 .0 4 3 9 -0 .0 2 6 5 -0 .0 5 7 4 -0 .0 9 8 5 E P S d iag n o sis -0 .1 1 3 4 -0 .1 7 4 3 -0 .1 0 5 4 -0 .0 2 2 0 N u m b er o f other illn ess 0 .0 4 3 0 * * * 0 .0 2 4 8 * * * 0 .0 2 7 2 * * * 0 .0 0 6 8 Health care utilization in prior 6 months A cu te h osp italization 0 .2 0 0 6 * * * 0 .0 6 4 8 0 .1 4 8 7 * * -0 .0 3 8 9 P sychiatric h osp italization 0 .0 2 0 8 0 .2 0 4 4 * * * 0 .2 1 9 5 * * * 0 .1 0 9 6 L ong-term care 0 .4 8 5 1 * * * -0 .0 3 3 9 0.1 0 0 2 * -0 .0 0 0 7 Su icidal attem pt 0 .0 0 2 9 -0 .0 0 1 4 0 .1 5 4 6 * 0 .1 2 8 6 T otal co st in 2 0 -4 0 % quintile -0 .0 4 7 9 * -0 .1 2 0 4 * * * -0 .1 4 6 7 * * * -0 .1 5 2 4 * T otal co st in 4 0 -6 0 % quintile -0 .0 5 4 6 -0 .2 2 8 9 * * * -0 .2 0 1 8 * * * -0 .2 6 2 4 * * T otal c o st in 6 0 -8 0 % quintile -0 .0 3 5 1 -0 .2 5 3 2 * * * -0 .2 2 8 6 * * * -0 .3 0 1 9 * * * T otal co st in 8 0 -1 0 0 % quintile 0 .0 6 6 9 -0 .3 0 3 3 * * * -0 .3 1 8 9 * * * -0 .3 4 0 0 * * * Selection correction terms For typical an tip sy ch o tics 0 .0341 0 .0 4 7 9 -0 .0 6 2 2 0 .1 1 7 2 For olanzapine 0 .1 7 3 1 * 0 .0 2 2 9 0 .0 3 2 5 0 .2 8 0 5 * For risperidone 0 .1 7 2 0 * 0 .1 5 9 3 * * -0 .1 7 5 7 * 0 .1 5 1 6 For quetiapine 0 .2 9 8 8 * * * 0 .1 6 3 1 * * 0 .0 7 1 6 0.1120 * P < 0 .0 5 , * *P < 0.01, ***P< 0.001 Treatment episodes initiated with the same drug (real restarter) was shown to have better compliance than delayed switching episodes, however, this impact was statistically significant only in typical antipsychotics group. This is consistent with the finding from baseline compliance model, which also showed that patients receiving typical antipsychoitcs were less likely to terminate the therapy if starting with the same antipsychotic medication as used in the prior treatment episode than switching to an alternative drug, while episode type has no impact on compliance among atypical antipsychotics. 66 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Across all four drugs, the diagnosis of dementia and longer duration of uninterrupted antipsychotic therapy in prior treatment attempt were associated with better compliance, whereas substance abuse was shown to be associated with poorer compliance. Across three atypical antipsychotics, higher treatment costs were associated with better compliance, which may indicate that atypical antipsychotics perform well in high cost patients. For typical antipsychotics, olanzapine and risperidone, a diagnosis of hyperlipidemia was linked to better compliance, whereas the diagnosis with bipolar disorder and larger number of other illnesses were found to correlate with poorer compliance. For typical antipsychotics and olanzapine, age greater than 55 was associated with better compliance. For typical antipsychotic and risperidone, prior use of mood stabilizer was linked to better complicance, while a prior acute hospital or nursing home stay was associated with poor compliance. For olanzapine and risperidone, black race and prior psychiatric hospital admission were associated with poor compliance. For patients using typical antipsychotics, urban residency, disablility and male gender were factors associated with better compliance; while the prior use of antidepressant and anti seizure medications, and a longer gap between antipsychotic treatment attempts were linked to poorer compliance. For patients using olanzapine, Hispanic race, the diagnosis of organic psychotic disease and personality disorders were related to poorer compliance. In risperidone group, age in 45-55 was associated with better compliance, while prior suicide attempts were associated with poorer compliance. Except the additional finding that the diagnosis of depression was associated with poorer compliance, the rest of the variables were not shown to affect compliance in quetiapine group. The selection terms were found to be significant, which implied the existence of selection bias in the data. 67 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. However, the classic interpretation of these terms in binary choice two-stage model was not applicable here. Confidence interval of outcomes After the unbiased parameters in the four outcome equations were obtained ( fls ), the expected outcome for each drug was derived using the entire sample y x = Xji3X . Bootstrap estimation procedures with 500 replications were performed to calculate the confidence intervals of the expected outcomes. The smaller the expected outcome with a medication, the more compliance patients are using that drug. In order to make the result more interpretable, the results from - In A(/) were transformed into days of therapy (/) (see section 5.3). The bootstrap results for - In A(t) and I are presented in Table 6.8. Four methods of bootstrap estimation were implemented to get the robust estimation. Table 6.8: Confidence interval of outcomes (bootstrap with 500 reps) T y p ic a ls O la n z a p in e R isp e r id o n e Q u e tia p in e Outcome: -In A (point estimate and confidence interval) 0 .8 0 4 4 0 .4 1 8 9 0 .4 9 9 7 0 .5 1 5 8 N orm al 0 .8 0 0 4 -0 .8 0 8 3 0 .4 1 6 1 -0 .4 2 1 7 0 .4 9 7 0 -0 .5 0 2 4 0 .5 1 3 5 -0 .5 1 8 1 P ercentile 0 .8 0 0 6 -0 .8 0 8 4 0 .4 1 6 2 -0 .4 2 1 5 0 .4 9 7 1 -0 .5 0 2 5 0 .5 1 3 5 -0 .5 1 8 0 B C 0 .8 0 0 7 -0 .8 0 8 7 0 .4 1 6 1 -0 .4 2 1 3 0 .4 9 7 1 -0 .5 0 2 4 0 .5 1 3 6 -0 .5 1 8 0 B C A 0 .8 0 0 7 -0 .8 0 8 7 0 .4 1 6 1 -0 .4 2 1 3 0 .4 9 7 1 -0 .5 0 2 4 0 .5 1 3 6 -0 .5 1 8 0 Outcome: days (point estimate and con fidence interval) 58 .1 5 113.48 9 7 .7 3 9 1 .8 8 N orm al 5 7 .5 -5 8 .8 1 1 2 .8 -1 1 4 .1 9 7 .2 -9 8 .3 9 1 .5 -9 2 .3 P ercentile 5 7 .5 -5 8 .7 1 1 2 .9 -1 1 4 .2 9 7 .2 -9 8 .2 9 1 .5 -9 2 .3 B C 5 7 .5 - 5 8 .8 1 1 2 .8 -1 1 4 .1 9 7 .2 - 9 8 .2 9 1 .4 - 9 2 .3 B C A 5 7 .5 -5 8 .8 112.8-114.1 9 7 .2 -9 8 .2 9 1 .4 -9 2 .3 BC: bias corrected; BCA: bias corrected and accelated 68 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Three findings are notable. First, the duration of treatment with typical antipsychotics was much shorter than that with atypical antipsychotics. Patients stayed with typical antipsychotics for 58 days, while 92-113 days with atypical antipsychotics. Second, estimated duration of therapy across three atypical antipsychotics were relative close in value. Third, the result suggested that olanzapine was slightly better than risperidone or quetiapine in treatment compliance. For example, patients taking olanzapine stayed on the therapy 16 days longer than risperidone, 21 days longer than quetiapine. 6.4 Results of panel data model The final method for investigating possible differences in patient’s compliance across alternative medications consisted of a panel data analysis. Patients with at least two complete episodes were included in the sample and the first two episodes were used (N=40444). Individual-specific and time-invariant unobserved factor enter the Cox model multiplicatively, therefore, the partial likelihood corresponds to logistic regression with difference in covariates across two spells as the control variables (see section 5.4). Two models were estimated. In the first model, the same set of variables as those included in the baseline compliance model were used. In the second simplified model, summary count variables were created for number of medical conditions, mental disorders and types of medications taken. Two models gave the similar results. Results of the simplified model are reported here. 69 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Several results are noteworthy with typical antipsychotics as reference group, as shown in Table 6.9. First, the parameter for the same drug variable was negative and significant. It indicates that patients receiving typical antipsychoitcs were less likely to terminate the therapy if starting with the same antipsychotic medication as used in the prior treatment episode than switching to an alternative drug. Though this finding is consistent with baseline model, the magnitude of the estimated effect is smaller than that in baseline model, for examaple, -0.2584 vs. -0.4043, or in terms of odds ratio, 0.772 vs. 0.667. Second, there existed significant interaction effect between type of episode and the drug used, implying that the estimated effect on compliance of atypical antipsychotics relative to typical antipsychotics was different in real restarters and late switchers. Table 6.9: Parameter estimates from panel data model (TAP as reference) Variables Param eter estim ates O lan zap in e -0 .6 4 9 6 * * * R isp eridone -0 .5 6 7 9 * * * Q uetiap in e -0 .5 2 2 4 * * * Sam e drug -0 .2 5 8 4 * * * O lan zap in e*sam e drug 0 .1 4 8 6 * * R isp erid on e*sam e drug 0 .2 2 6 6 * * * Q uetiapine* sam e drug 0 .1 5 0 2 * * P < 0.05, ** P < 0 .0 1 , ***P< 0.001 As in the baseline model, the differential effects in comparison to typical antipsychotics were presented separately for real restarters and late switchers in Table 6.10. The similar results to baseline model include: (1) in both real restarters and late switchers, patients using atypical antipsychotics are less likely to discontinue the therapy compared to typical antipsychotics; (2) this differential effect was larger in late switchers than in real restarters as shown by the magnitude of the odds ratios. The estimated odds ratios of discontinuing therapy with atypical antipsychotics relative to typical antipsychotics in 70 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. late switchers were close between baseline and panel data models, for example, 0.52-0.59 in panel data model vs. 0.63-0.70 in baseline model. However, the estimated odds ratios in returning patients were quite different between two models. For example, using atypical antipsychotics were 29-39% less likely to discontinue the therapy than returning patients using typical antipsychotics in panel data model, compared to an estimated difference of 1-10% in baseline model. These differences may imply that after controlling for patient-specific unobserved factor, the superior effect of typical antipsychotics in patients restarting therapy on a drug used previously found in baseline model was is not evident in the panel data results. For ease of comparison, the results from baseline model were attached to Table 5.10 at the bottom. Table 6.10: Results from panel data model by type of episodes (TAP as reference) Real restart (T A P real restart as ref.) Late switcher (T A P late sw itch er as ref.) Param eter Odds ratio Param eter Odds ratio Results from panel data model O lanzapine -0 .5 0 1 0 * * * 0 .6 0 6 * * * -0 .6 4 9 6 * * 0 .5 2 2 * * R isperidone -0 .3 4 1 3 * * * Q 7 ] J * * * -0 .5 6 7 9 * * * 0 .5 6 7 * * * Q uetiapine -0 .3 7 2 2 * * * 0 .6 8 9 * * * -0 .5 2 2 4 * 0 .5 9 3 * Results from baseline model O lanzapine -0 .0 9 9 1 * * * 0 .9 0 6 * * * -0 .4 7 0 6 * * * 0 .6 2 5 * * * R isperidone -0 .0 8 6 7 * * * 0 .9 1 7 * * * -0 .4 3 9 3 * * * 0 .6 4 4 * * * Q uetiapine -0 .0 0 6 5 * * * 0 .9 9 4 * * * -0 .3 5 0 5 * * * 0 .7 0 4 * * * * P < 0 .0 5 , * * P < 0 .0 1 , ***P<0.001 As before, we tested if differences exist in rate of drug discontinuation among atypical antipsychotics by rerunning the model with olanzapine as reference group (Table 6.11). Two similar results were noticed. First, patients using quetiapine were more likely to terminate the treatment compared to olanzapine, and this differential effect was the same in late switchers and real restarters. Second, there is no difference in rate of 71 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. discontinuation between olanzapine and risperidone, irrespective of type of episodes. However, the likelihood of treatment discontinuation is significantly smaller in real restarters than in late switchers for olanzapine. In contrast, the baseline model did not detect any significant difference in the performance of olanzapine by type of episode. Table 6.11: Results from panel data model (olanzapine as reference) P a r a m e te r O d d s R a tio R isp eridone 0 .0 8 1 8 1.085 Q uetiap in e 0 .1 2 7 2 * 1.136* Sam e drug -0 .1 0 9 8 * * 0 .8 9 6 * * R isp erid on e*sam e drug -0 .0 7 8 0 1.081 Q u etiap in e*sam e drug -0 .0 0 1 6 1.002 * P < 0 .0 5 , * * P < 0 .0 1 , ***P< 0.001 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. CHAPTER 7. DISCUSSION AND CONCLUSION 7.1 Summary The objective of this study was to compare treatment adherence for patients with schizopharenia across alternative antipsychotics (olanzapine, risperidone, and quetiapine), using typical antipsychotics as the comparison treatment. The treatment adherence was measured by time to discontinuation after the initiation of drug therapy; Non-adherence to antipsychotic medication is a major challenge in treating patients with schizophrenia. Poor compliance is related to symptom relapse and increased cost of treatment. Antipsychotic medication with superior effectiveness and more benign side effect profile is supposed to bring better compliance, thus improving the treatment outcomes. The MediCal dataset containing a 100% sample of patients treated for schizophrenia during the period of 1994 to 2003 was used to create an analytic file, which focused on the restart episodes initiated between January 1999 and March 2001 with only one medication (134,576 mono-restart episodes, corresponding to 52,121 patients). All the selected treatment episodes for a patient were followed up on to the last recorded claim of that patient. An exhaustive list of covariates was generated to control patients’ demographics, prior health care utilization and comorbidity profile, prior compliance history with antipsychotic therapy, and type of restart episodes (real restarters vs. late switchers). 73 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. In order to document the robustness of out results, three statistical models were applied. Cox proportional hazards model was estimated as the baseline compliance model based on the first episode of each patient in the study period. With the same study sample, the second analysis used Cox proportional hazard model applied in latent index framework to adjust for treatment selection bias. Finally, panel data fixed-effect estimation was conducted on those patients with at least two treatment episodes to account for omitted variable/heterogeneity of the patient. Three models provided similar results. First, patients are more compliant with atypical antipsychotics (olanzapine, risperidone, quetiapine) than with typical antipsychotics. This result is not at all surprising given the large clinical trial literature that documents the clinical superiority of atypical antipsychotics, especially with regard to their improved side effect profile. Second, among atypical antipsychotics, patients treated with quetiapine tend to be less compliant than those treated with olanzapine or risperidone, while those treated with olanzapine and risperidone are very close in their compliance. In addition, the first and the third model allowed us to compare the compliance in two different clinical scenarios, ‘returning’ patients (real restarters) who initiated the therapy with the same medication as used in their most recent treatment attempt; and ‘switching’ patients (late switcher), who initiated the therapy with an alternative antipsychotic medication. Three additional findings were revealed in these two distinct patient groups. 74 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. First, patients treated with atypical antipsychotics are less likely to discontinue the therapy compared to typical antipsychotics in both ‘returning’ and ‘switching’ patients. However, this differential effect was larger in ‘switching’ patients than in ‘returning’ patients. This implies that typical antipsychotics may work well in some patients, and the duration of therapy achieved by restarting these patients on typical antipsychotics is almost equivalent to the duration achieved by switching them to atypical antipsychotics. In clinical practice, this suggests that physicians can put their patients on the same type of typical antipsychotics if they repeatedly used this drug in the past, switching them to atypical antipsychotics may not bring better compliance. Second, the odds ratio of discontinuing atypical relative to typical antipsychotics in ‘switching’ patients was 0.63-0.70 from the baseline model, which is close to the estimates from panel data model, 0.52-0.59. However, the odds ratios in ‘returning’ patients were 0.91-0.99 from the baseline model, which is much higher than that'from the panel data model, 0.61-0.71. While panel data model and baseline model used similar covariates, unobserved difference in patient-specific factor was better accounted for in panel data model, producing less biased estimates. In other words, the finding from the baseline model about favorable treatment effect of typical antipsychotics in ‘returning’ patients may be overestimated. Third, among three atypical antipsychotics, olanzapine and risperidone achieve similar level of treatment compliance in patients, irrespective of treatment scenarios; patients treated with quetiapine were significantly less compliant than those with olanzapine, and 75 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. this differential effect is the same for ‘returning’ and ‘switching’ patients. Baseline and panel data models provide consistent results. Previous studies focused on head-to-head comparison between two treatment options, either typical antipsychotics vs. atypical antipsychotics, or one specific atypical vs. a second atypical medication. To our knowledge, this is the first study of head-to-head comparison of treatment compliance between three atypical antipsychotics and typical antipsychotics in patients with schizophrenia. This study also makes two methodological contributions to the literature. First, this is the first study to apply Cox proportional hazard model in latent index framework with polychotomous choice. Second, this is also the first study to apply panel data method on treatment compliance; previous studies overwhelmingly used the data cross-sectionally, despite the fact that the patients take antipsychotics episodically. 7.2 Limitations of study Limitation of the data' Analyses based on paid claims data have a host of limitations that must be considered in interpreting these results. First, clinical information are not available on paid claims, so we cannot control disease severity in drug selection and outcome equations. However, we created an exhaustive list of covariates to minimize if not eliminate this problem, including prior treatment cost, dichotomous variables for comorbidities, the use of other 76 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. psychotropic medication, and prior hospitalization. Second, the data of this study did not have eligibility data. However, most of the sample included in this study gained MediCal eligibility due to disability status, thus reducing the likelihood of intermittent eligibility. Additionally, three-month zero cost screening greatly minimized the problem for those without permanent eligibility. Three, the use of California Medicaid program data limit the generalizability of the study results on patients with schizophrenia in other states. Limitation of the statistical methods Three statistical methods were applied, but each has its weakness. While an exhaustive list of control variables were included in the baseline Cox proportional hazard model to minimize the selection bias (omitted variable) problem, full correction of selection bias is not guaranteed. When Cox proportional hazard model was re-estimated within a latent index framework, the identification of model depends heavily on the availability of good instrumental variables (I Vs) in the sense that I Vs are highly predictive of drug selection, but not correlated with the treatment outcome. Drug diffusion rates over time appear to be a good IV, as well as market year of drug introduction. The selection bias terms included in the outcome equation were derived from a mother logit model. There are two compelling reasons to use this approach. First, it is easy to estimate using available statistical software programs. Second, it was used to introduce the correlation between error terms among 4 drug choices, the violation of IIA. Theoretically, there are other models that 77 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. could handle this situation, such as multinomial probit or mixed logit (Train 2003). However, there is no software readily available to estimate these models. Finally, censored episodes were discarded to simplify the analysis. This was not a significant problem when only 2% of the study sample was right censored. However, if a data is heavily censored, this method may not be applicable. Fixed-effect panel data method allowed us to explicitly account for unobserved patient- specific factor. However, selection bias may not be fully corrected in this model. We assumed that stable unobserved patient-specific factor fully determined the drug selection. If there may exist some unobserved time-varying factors that affect the drug selection, failure to adjust for that can lead to biased conclusion. 7.3 Conclusion Despite the limitation of the study, we found that patients treated with second-generation antipsychotic medications consistently achieve better drug therapy outcomes than similar patients treated with conventional antipsychotics. Among atypical antipsychotics, olanzapine and risperidone appear to be associated with better compliance relative to quetiapine. These findings are robust across three methodologies used. 78 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. BIBLIOGRAPHY Allebeck P, Wistedt B. Mortality in schizophrenia: a ten-year follow-up based on the Stockholm County inpatient register. Arch Gen Psychiatry 1986; 43:650-653 Allison DB, Mentore JL, Heo M, Chandler LP, Cappelleri JC, Infante MC, Weiden PJ. Antipsychotic-induced weight gain: a comprehensive research synthesis. Am J Psychiatry 1999; 156:1686-1696 Almond S, Knapp M, Francois C, Toumi M, Brugha T. Replase in schizophrenia: costs, clinical outcomes and quality of life. Br J Psychiatry 2004; 184:346-351 American Psychiatric Association. Practice guidelines for the treatment of patients with schizophrenia. Am J Psychiatry 1997; 154(suppl. 4):l-63 Anderson C, Connelly J, Johnstone EC, Owens DG. Disabilities and circumstances of schizophrenic patients-a follow-up study. Cause of death. Br J Psychiatry 1991; 159(suppl. 13): 30-33 Attkisson C, Cook J, Karno M, Lehman A, McGlashan TH, Meltzer HY, O ’Connor M, Richardson D, Rosenblatt A, Wells K. Clinical Services Research. Schizophr Bull 1992; 18:561-626 Ayuso-Gutierrez JL, del Rio Vega JM. Factors influencing re-lapse in the long-term course of schizophrenia. Schizophr Res 1997; 28:199-206 Blanchard JJ, Brown SA, Horan WP, Sherwood AR. Substance use disorders in schizophrenia: review, integration, and a proposed model. Clin Psychol Rev 2000; 20(2):207-234 Buckley PF. Substance abuse in schizophrenia: a review. J Clin Psychiatry 1998; 59 (suppl 3):26-30 Caro JJ, Ward A, Levinton C, Robinson K. The risk of diabetes during olanzapine use compared with risperidone use: a retrospective database analysis. J Clin Psychiatry 2002; 63(12): 1135-1139 Caroff SN, Mann SC, Campbell EC, Sullivan KA. Movement disorders associated with atypical antipsychotic drugs. J Clin Psychiatry 2002; 63(suppl 4):12-19 Casey DE. Side effect profiles of new antipsychotic agents. J Clin psychiatry 1996; 57(suppl ll):40-45 79 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Chamberlain G: Heterogeneity, omitted variable bias, and duration dependence, in Longitudinal Analysis of Labor Market Data. Edited by Heckman JJ, Singer B. Cambridge: Cambridge University Press, 1983, pp 3-38 Conley RR, Mahmoud R. A randomized double-blind study of risperidone and olanazpine in the treatment of schizophrenia of schizoaffective disorders. Am J Psychiatry 2001; 158:765-774 Davies LM, Drummond MF. The economic burden of schizophrenia. Psychiatric Bulletin 1990; 14:522-525 Dickey B, Normand ST, Norton EC, Azeni H. Fisher W. Altaffer F. Managing the care of schizophrenia: lessons from a 4-year Massachusetts Medicaid study. Arch Gen Psychiatry 1996; 53(10):945-952 Dixon LB, Lehman AF, Levine J. Conventional antipsychotic medications for schizophrenia. Schizophr Bull 1995; 21:567-577 Dursun SM, Szemis A, Andrews H, Reveley MA. The effects of clozapine on levels of total cholesterol and related lipids in serum of pati,ments with schizophrenia: a prospective study. J Psychiatry Neurosci 1999; 24:453-455 Efron B, Tibshirani R. Bootstrap methods for standard errors, confidence intervals, and other measures of statistical accuracy. Statistical Sceince 1986; 1:54-77 Fuller MA, Shermock KM, Secic M, Grogg AL. Comparative study of the development of diabetes mellitus in patients taking risperidone and olanzapine. Pharamacotherapy 2003; 23(8): 1037-1043 Gaulin BD, Markowitz JS, Caley CF, Nesbitt LA, Dufresne RL. Clozapine-associated elevation in serum triglycerides. Am J Psychiatry 1999; 156:1270-1272 Gelenberg AJ. Treating extrapyramidal reactions: some current issues. J Clin Psychiatry 1987; 48(suppl):24-27 Ghaeli P, Dufresne RL. Serum triglyceride levels in patients treated with clozapine. Am J Health Syst Pharm 1996; 53:2079-2081 Gianfrancesco FD, Grogg AL, Mahmoud RA, Wang R, Nasrallah HA. Differential effects of risperidone, olanzapine, colzapine, and conventional antipsychotics on type 2 diabetes: findings from a large health plan database. J Clin Psychiatry 2002; 63:920-930 Gianfrancesco FD, White R, Wang R, Nasrallah HA. Antipsychotic-induced type 2 diabetes: evidence from a large health plan database. J Clin Psychoparmacology 2003; 23:328-335 80 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Gibson PJ, Damler R, Jackson EA, Wilder T, Ramsey JL. The impact of olanzapine, risperidone, or haloperidol on the cost o f schizophrenia care in a Medicaid population. Value Health 2004; 7(l):22-35 Gilmer TP, Dolder CR, Lacro JP, Folsom DP, Lindamer L, Garcia P, Jeste DV. Adherence to treatment with antipsychotic medication and health care costs among Medicaid beneficiaries with schizophrenia. Am J Psychiatry 2004; 161:692-699 Glazer WM, Dickson RA. Clozapine reduces violence and persistent aggression in schizophrenia. J Clin Psychiatry 1998; 59(suppl 3):8-14 Goff DC, Heckers S, Freudenreich O. Schizophrenia. Med ClinN Am 2001; 85(3):663- 689 Goldman LS. Medical illness in patients with schizophrenia. J Clin Psychiatry 1999; 60:10-15 Green AI, Canuso CM, Brenner MJ, Wojcik JD. Detection and management of comorbidity in patients with schizophrenia. Psychiatr Clin N Am 2003; 26:115-139 Heckman JJ. The common structure of statistical models of truncation, sample selection, and limited dependent variables and a simple estimator for such models. Ann Econ Soc Meas 1976, 5:475-492 Heckman JJ. Treatment selection as a specification error. Econometrica 1979; 47:153- 161 Heckman J, Tobias JL, Vytlacil E. Simple estimators for treatment parameters in a latent variable framework with an application to estimating the returns to schooling. NBER working paper 7950, 2000. http://www.nber.org/papers/w7950.pdf Heckman JJ, Vytlacil E: Local instrumental variables, in Nonlinear Statistical Inference: Essays in Honor O f Takeshi Amemiya. Edited by Hsiao C, Morimune K, Powell J. Cambridge: Cambridge University Press, 2000, Hellewell JSE, Cantillon M. Antipsychotic tolerability: the attitudes and perceptions of medical professionals, patients and caregivers towards the side effects of antipsychotic therapy. Presented at the Eleventh Congress of the European College of Neuropsychopharmacology, Paris, France, 1998 Oct 31-Nov 4. Abstract P.2.109 Henderson DC, Cagliero E, Gray C, Nasrallah RA, Hayden DL, Schoenfeld DA, Goff DC. Clozapine, diabetes mellitus, weight gain, and lipid abnormalities: a five-year naturalistic study. Am J Psychiatry 2000; 157:975-981 81 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Hiehle G, Cline M. California’s Medical Assistance Program, Annual Statistical Report, Calendar Year 2001. California Department of Health Services, Sacramento. 2002; pp 28 Hogan MF. Public-sector mental health care: new challenges. Health Affairs 1999; 18:106-111 Horowitz J, Lee S. Semiparametric estimation of a panel data proportional hazards model with fixed effects. J of Econometrics 2004; 119:155-198 Imbens G, Angrist J. Identification and estimation of local treatment effects. Econometrica 1994; 62:467-476 Jeste DV, Caligiuri MP, Paulsen JS, Heaton RK, Lacro JP, Harris MJ, Bailey A. Fell RL, McAdams LA. Risk of tardive dyskinesia in older patients. Arch Gen Psychiatry 1995; 52:756-765 Jeste DV, Caligiuri MP. Tardive dyskinesia. Schizophr Bull 1993; 19:303-316 Kane J, Honigfeld G, Singer J, Meltzer H. Clozapine for the treatment-resistent schizophrenic. A double-blind comparison with chlorpromazine. Arch Gen Psychiatry 1988;45:789-796 Kane JM, Marder SR. Psychopharmacologic treatment of schizophrenia. Schizophr Bull 1993; 19:287-302 Keith SJ, Regier DA, 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, Kiefer NM. Economic duration data and hazard functions. J Econ Lit 1988; 26:646-679 Kinon BJ, Lieberman JA. Mechanisms of action of atypical antipsychotic drugs: a critical analysis. Psychopharmacology 1996; 124:2-24 Kleinberg DL, Davis JM, de Coster R, Van Baelen B, Brecher M. Prolactin levels and adverse events in patients treated with risperidone. J Clin Psychopharmacol 1999; 19:57- 61 Knapp M. Costs of schizophrenia. Br J Psychiatry 1997; 171:509-518 Koller E, Schneider B, Bennett K, Dubitsky G. Clozapine-associated diabetes. Am J Med - 2001;111:716-23 82 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Koller EA, Cross JT, Doraiswamy PM, Schneider BS. Risperidone-asociated diabetes mellitus: a pharmacovigilance study. Pharmacotherapy 2003; 23:735-744 Koller EA, Doraiswamy PM. Olanzapine-associated diabetes mellitus. Pharmacotherapy 2002; 22:841-852 Koro CE, Fedder DO, L’ltalien GJ, Weiss S, Magder LS, Kreyenbuhl J, Revicki D, Buchanan RW. An assessment of the independent effects of olanzapine and risperidone exposure on the risk of hyperlipidemia in schizophrenia patients. Arch Gen Psychiatry 2002a; 59:1021-1026 Koro CE, Fedder DO, L’ltalien GJ, Weiss SS, Magder LS, Kreyenbuhl J, Revicki D, Buchanan RW. Assessment of independent effect of olanzapine and risperidoen on risk of diabetes among patients with schizophrenia: population based nested case-control study. BMJ 2002b; 325:243-245 Lancaster T: Limited information inference, in The Econometric Analysis of Transition Data. Edited by Lancaster T. Cambridge: Cambridge University Press, 1990, pp 268-269 Lee LF. Generalized econometric models with selectivity. Econometrica 1983; 51:507- 512 Lee LF. Some approaches to the correction of selectivity bias. Re Econ Stud 1982; 49:355-372 Lehman AF, Steinwachs DM, and the Co-Investigators of the PORT Project. Translating research into practice: the Shizophrenia Patient Outcomes Research Team (PORT) treatment recommendations. Schizophr Bull 1998; 24:1-10 Leucht S, Pitschel-Walz G, Abraham D, Kissling W Efficacy and extrapyramidal side- effects of the new antipsychotics olanzapine, quetiapine, risperidone, and sertindole compared to conventional antipsychotics and placebo. A meta-analysis of randomized controlled trials. Schizophr Res 1999; 35:51-68 Lindenmayer JP, Czobor P, Volavka J, Citrome L, Sheitman B, McEvoy JP, Cooper TB, Chakos M, Lieberman JA. Changes in glucose and cholesterol levels in patients with schizophrenia treated with typical or atypical antipsychotics. Am J Psychiatry 2003; 160:290-296 Lindenmayer JP, Grochowski S, Mabugut L. Clozapine effects on positive and negative symptoms: a six month trial in treatment -resistant schizophrenics. J Clin Pychopharmacol 1994; 14:201-204 Lindstrom E, Bingefors K. Patient compliance with drug therapy in schizophrenia: economic and clinical issues. Pharmacoeconomics 2000; 18:105-124 83 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Lyu RR, McCombs JS, Johnstone BM, Muse DN. Use of conventional antipsychotics and the cost o f treating schizophrenia. Health Care Financ Rev 2001; 23:83-99 McCombs JS, Nichol MB, Stimmel GL, Shi J, Smith RR. Use patterns for antipsychotic medications in Medicaid patients with schizophrenia. J Clin Psychiatry 1999; 60 (suppl 19):5-11 McCombs JS, Nichol NB, Johnstone BM, Stimmel GL, Shi J, Smith R. Antipsychotic drug use patterns and the cost of treating shcizophrenia. Psychiatr Services 2000a; 51:525-527 McCombs JS, Luo M, Johnstone BM, Shi L. The use of conventional antipsychotic medications for patients with schizophrenia in a Medicaid population: therapeutic and cost outcomes over 2 years. Value Health 2000b; 3:222-231 McCombs JS. Mulani P. Gibson PJ. Open access to innovative drugs: treatment substitutions or treatment expansion? Health Care Financ Rev 2004; 25:35-53 McCreadie, RG. The Nithsdale schizophrenia surveys. Soc Psych Psych Epid 1992; 27:40-45 McEvoy JP, Scheifler PL, Frances A. The expert consensus guideline series: treatment of schizophrenia 1999. J Clin Psychiatry 1999; 60 (suppl ll):3-80 McFadden D. On independence, structure and simultaneity in transportation demand analysis. Working paper No. 7511, Urban Travel Demand Forecasting Project, Institution of Transportation and Traffic Engineering, University of California, Berkeley. 1975 McGlashan TH. A selective review of recent North American long-term followup studies of schizophrenia. Schizophr Bull 1988; 14:515-542 McGuire TG. Measuring the economic costs of schizophrenia. Schizophr Bull 1991; 17:375-388 Meltzer HY, McGurk SR. The effects of clozapine, risperidone, and olanzapine on cognitive function in schizophrenia. Schizophr Bull 1999; 25:233-255 Meltzer HY. Suicide in schizophrenia: risk factors and clozapine treatment. J Clin Psychiatry 1998; 59(suppl 3): 15-20 Miller AL, Chiles JA, Chiles JK, Crismon ML, Rush AJ, Shon SP. The Texas Medication Algorithm Project (TMAP) schizophrenia algorithms. J Clin Psychiatry 1999; 60:649-657 Mukherjee S, Decina P, Bocola V, Saraceni F, Scapicchio PL. Diabetes mellitus in schizophrenic patients. Compr Psychiatry 1996; 37:68-73 84 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Mulkern VM, Manderscheid RW. Characteristics of community support program clients in 1980 and 1984. Hosp Community Psych 1989; 40:165-172 Narrow WE, Rae DS, Robins LE, Regier DA. Revised prevalence estimates of mental disorders in the United States. Arch Gen Psychiatry 2002; 59:115-123 Nasrallah H. A review of the effect of atypical antipsychotics on weight. Psychoneuroendocrinology 2003; 28 (suppl l):83-96 Novak-Grubic V, Tavcar R. Treatment adherence in first-episode schizophrenia. Psychiatric Services 1999; 50:970-971 Perkins DO. Adherence to antipsychotic medication. J Clin Psychiatry 1999; 60(suppl 21):25-30 Rascati KL, Johnsrud MT, Crismon ML, Lage MJ, Barber B. Olanzapine versus risperidone in the treatment o f schizophrenia: a comparison of costs among Texas Medicaik patients. Pharmacoeconomics 2003; 21:683-697 Regier DA, Narrow WE, Rae DS, Manderscheid RW, Locke BZ, Goodwin FK. The de facto US mental and addictive disorders service system. Arch Gen Psychiatry 1993; 50:85-94 Rubin DB. Estimating causal effects of treatments in randomized and nonrandomized studies. J Educ. Psychol 1974; 66:688-701 Rupp A, Keith S. The cost of schizophrenia: assessing the burden. Psychiat Clin N Am 1993; 16:413-431 Salkever DS, Slade EP, Karakus M. Palmer L, Russo PA. Estimation of antipsychotic effects on hospitalization risk in a naturalistic study with selection on unobservables. J Nerv Ment Dis 2004; 192:119-128 Santarlasci B, Messori A. Clinical trial response and dropout rates with olanzapine versus risperidone. Ann Pharmacother 2003; 37:556-563 Sernyak MJ, Leslie DL, Alarcon RD, Losonczy MF, Rosenheck R. Associateion of diabetes mellitus with use of atypical neuroleptics in the treatment of schizophrenia. Am J Psychiatry 2002; 159:561-566 Sheitman BB, Lee H, Strauss R, Lieberman JA. The evaluation and treatment of first- episode psychosis. Schizophr Bull 1997; 23:653-661 Stroup TS, McEvoy JP, Swartz MS, Byerly MJ, Glick ID, Canive JM, McGee MF, Simpson GM, Stevens MC, Lieberman JA. The Natioal Institute of mental Health 85 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Clinical Antipsychotic Trials of Intervention Effectiveness (CATIE) project: schizophrenia design and protocol development. Schizophre Bull 2003; 29:15-31 Svarstad BL, Shireman TI, Sweeney JK. Using drug claims data to assess the relationship of medication adherence with hospitalization and costs. Psychiatr Serv 2001; 52:805-811 Torrey EF. A viral-anatomical explanation of schizophrenia. Schizophr Bull 1991; 17:15- 18 Train K. Logit, in Qualitative Choice Analysis: Theory Econometrics, and An Application to Automobile Demand. Edited by Train K. Cambridge: MIT press, 1986, pp 21-22 Train K: Mixed logit, in Discrete Choice Methods With Simulation. Edited by Train K. Cambridge: Cambridge University Press, 2003, pp 138-154 Tran PV, Hamilton SH, kuntz AJ, Potvin JH, Andersen SW, Beasley C Jr, Tollefson GD. Double-blind comparison o f olanzapine versus risperidone in the treatment of schizophrenia and other psychotic disorders. J Clin Psychopharmacology 1997; 17:407- 418 Tsuang MT, Fleming JA, Simpson JC: Suicide and schizophrenia, in The Harvard Medical School guide to suicide assessment and intervention. Edited by Jacobs DG. San Francisco: Jossey-Bass, 1999, pp 287-299 Vytlacil, Edward. Independence, monotonicity,and latent variable models: an equivalence result. Working paper, University of Chicago, 1999 Weiden PJ, Olfson M. Cost of relapse in schizophrenia. Schizophr Bull 1995; 21 419-429 Weiden P, Aquila R, Standard J. Atypical antipsychotic drugs and long-term outcome in schizophrenia. J Clin Psychiatry 1996; 57:53-60 Weiden PJ, Shaw E, Man JJ. Cause of neuroleptic noncompliance. Psychiatr Ann 1986; 16:571-575 White H. A heteroskedasticity-consistent covariance matrix estimator and a direct test for heteroskedasticity. Econometrica 1980; 48:817-838 Winship C, Morgan SL. The estimation of causal effects from observational data. Annu Rev Sociol 1999; 25:659-706 Worrel JA, Marken PA, Beckman SE, Ruehter VL. Atypical antipsychotic agents: a critical review. Am J Health-Syst Ph 2000; 57:238-258 86 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission. Wyatt RJ, Henter I, Leary MC, Taylor E. An economic evaluation of shcizophrenia - 1991. Soc Psychiatry Psychiatr Epidemiol 1995; 30:196-205 Zhao Z. A retrospective economic evaluation of olanzapine versus risperidone in the treatment o f schizophrenia. Managed Care Interface 2002; 15(2):75-81 Reproduced with permission of the copyright owner. Further reproduction prohibited without permission.
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Controlling for biases from measurement errors in health outcomes research: A structural equation modeling approach
PDF
Effects of a formulary expansion of the use of SSRIs and health care services by depressed patients in the California Medicaid program
PDF
A methodology to identify high -risk patients with diabetes in the California Medicaid populations (Medi -Cal)
PDF
Assessing the cost implications of combined pharmacotherapy in the long term management of asthma: Theory and application of methods to control selection bias
PDF
A new paradigm to evaluate quality-adjusted life years (QALY) from secondary database: Transforming health status instrument scores to health preference
PDF
Cost analysis of three pharmacy counseling programs for diabetics in a health maintenance organization
PDF
Assessment of prognostic comorbidity in hospital outcomes research: Is there a role for outpatient pharmacy data?
PDF
Essays on regulation of public utilities and the provision of public goods
PDF
Three essays on comparative economic and clinical outcomes of antipsychotics using retrospective claims data
PDF
Dependence on foreign labor, quality of education and unemployment in the GCC countries: In search of solutions
PDF
Essays on organizational forms and performance in California hospitals
PDF
Basolateral transport of nucleosides in rabbit tracheal epithelial cells: An equilibrative nucleoside transporter outcome
PDF
Drug delivery to the posterior region of the eye: Characterization of organic cation transporters
PDF
Time dependent survival analysis of Kaiser Permanente/USC pharmacists' consultation intervention study
PDF
Hepatitis C in the post-interferon era: selected essays in health economics
PDF
The construct validity of the Health Utilities Index in patients with chronic respiratory disease in a managed care population
PDF
Conjunctival epithelial uptake of biodegradable nanoparticles: Mechanism, intracellular distribution, and absorption enhancement
PDF
Endogenous regulatory factors in the inhibition and down-regulation of immune responses
PDF
Design of a stealth liposome delivery system for a novel glycinamide ribonucleotide formyltransferase inhibitor
PDF
Drug efflux pumps in rabbit conjunctival epithelial cells
Asset Metadata
Creator
Chen, Lei
(author)
Core Title
Compliance study of second-generation antipsychotics on patients with schizophrenia
School
Graduate School
Degree
Doctor of Philosophy
Degree Program
Pharmaceutical Economics and Policy
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
economics, general,Health Sciences, Pharmacy,OAI-PMH Harvest
Language
English
Contributor
Digitized by ProQuest
(provenance)
Advisor
McCombs, Jeffrey S. (
committee chair
), Hay, Joel W. (
committee member
), Ridder, Geert (
committee member
)
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-c16-605265
Unique identifier
UC11341117
Identifier
3220096.pdf (filename),usctheses-c16-605265 (legacy record id)
Legacy Identifier
3220096.pdf
Dmrecord
605265
Document Type
Dissertation
Rights
Chen, Lei
Type
texts
Source
University of Southern California
(contributing entity),
University of Southern California Dissertations and Theses
(collection)
Access Conditions
The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the au...
Repository Name
University of Southern California Digital Library
Repository Location
USC Digital Library, University of Southern California, University Park Campus, Los Angeles, California 90089, USA
Tags
economics, general
Health Sciences, Pharmacy