Online Program

Instrumental Variable Methods for the Comparative Safety of Second-Generation Antipsychotic Medications

*Portia Yvonne Cornell, Harvard University 
Bruce Fireman, Kaiser Permanente Division of Research 
Vicki Fung, Mid-Atlantic Permanente Research Institute 
John Hsu, Harvard Medical School 
MaryBeth Landrum, Harvard Medical School 
Mary Price, Harvard Medical School 

Keywords: instrumental variables, mental health, comparative effectiveness, antipsychotic drugs,

Background: Clinical trial evidence indicates that second-generation antipsychotic medications (SGAs) have adverse metabolic outcomes, which increase cardiovascular risks and mortality. With growing use of SGAs, particularly in young patients, it is important to establish their safety in real world settings. Because a physician’s drug choice will be partly determined by patient characteristics that are unobserved in most data sets, traditional observational methods produce biased inferences.

Objective: We sought to refine a novel method for comparing the metabolic effects of SGAs in mentally ill patients using real world data.

Design: We use physician prescribing preference as an instrumental variable (IV) to predict SGA drug choice and arrive at unbiased estimates of the comparative safety between different SGAs with respect to body mass index (BMI). To address differences between physician panels that could act as confounders, we include in the instrument physician-level case-mix measures and stratify by physician specialty. We use two-stage least squares regression to estimate the local average effect of SGA choice on changes in BMI. We control for age, sex, payer, mental health diagnosis, and co-morbidities. Data are from electronic health records from a prepaid, integrated delivery system.

Population: 1,433 patients aged 20-35 who received a new prescription for one of five SGAs between 2008–2009, had no antipsychotic use in the previous 12 months, and for whom we had BMI measurements before and at least 90 days after the prescription fill. We restrict the analysis to physicians with four or more new SGA prescriptions in the study period.

Intervention: Prescription for one of aripiprazole, olanzapine, quetiapine, risperidone, and ziprasidone.

Outcome: The primary outcome measure is change in body-mass index (BMI) before and 90 days after initial prescription of the antipsychotic. Because this analysis uses observational data, hypotheses were generated post-data collection.

Findings: Preliminary analyses indicate that preference among SGAs for a new antipsychotic prescription is a strong predictor of treatment. A physician’s previous choice of aripiprazole, olanzapine, or risperidone for the plurality of previous prescriptions in the study period increases the probability that the next patient will receive a prescription for that drug by 20.1, 21.0, and 10.1 percentage points respectively, controlling for patient characteristics (p<0.05). We examine the balance of pre-prescription BMI, mental health diagnosis, and co-morbidities, and find that the instrument achieves better balance on these characteristics than actual treatment status. Thus far our estimates suggest no statistically significant differences in BMI effects between SGAs.

Conclusions: The physician preference IV represents a promising approach for assessing the comparative safety of existing therapies in real world settings. We will proceed to examine the performance of this instrument in different sub-groups (e.g., younger and elderly patients, and by mental health diagnosis) and for other outcomes including systolic blood pressure and LDL cholesterol. We will also test IV assumptions by comparing balance on these clinically relevant baseline characteristics. By testing the IV assumptions using a clinically rich database, this study will suggest analyses that can be applied to larger administrative databases, where these assumptions cannot be as carefully tested.