The study population was formed using administrative and healthcare claims data from three United States (US) sites: (1) Kaiser Permanente Health Plan (Kaiser Permanente) (Northern California; 2.9 million covered lives during the study period); (2) HealthCore Integrated Research Network (HealthCore) (14.5 million covered lives in health plans across the US) and (3) PharMetrics (43 million covered lives from 73 health plans across the US). These sites were chosen because they maintain extensive electronic databases, including outpatient pharmacy, and inpatient and outpatient encounter and claims records. Outpatient pharmacy data include date of prescription fill, drug name, dose per pill, number of pills dispensed, and days-supply. While the electronic format of data from each site chosen is not identical, the principal investigators of this study followed a common protocol for collecting study data, have extensive experience compiling data from these multiple sources, have developed the site-specific algorithms needed to construct a single, uniform analytic dataset, and have established data management and quality control checks. A limited data set (defined by the Health Insurance Portability and Accountability Act (HIPAA) of 1996) was used and all researchers had HIPAA-required business associate and data use agreements in place before conducting the research. This study was approved by the Kaiser Permanente Northern California Institutional Review Board (IRB) and deemed IRB exempt at HealthCore and PharMetrics.
Patients ≥ 18 years of age, newly exposed to SGAs between November 1, 2002 and March 31, 2005 were included in the cohort. This inception cohort consisted of all patients exposed to an SGA for at least 45 days and continuously enrolled for at least 3 months before and 6 months after the date of first prescription (index date) with no evidence of diabetes using all available historical data prior to the index date and no previous antipsychotic prescription (first generation or SGA) filled within 3 months of the index date.
Exclusion of Patients with Previous History of Diabetes Mellitus
Previous history of diabetes used all available historic data and was defined using the following criteria: (1) at least one emergency department visit with an International Classification of Diseases, 9th Revision (ICD-9) coded diagnosis indicative of diabetes (250.xx); or (2) at least two outpatient visits coded with diabetes (250.xx); or (3) at least one inpatient stay with an ICD-9 coded diagnosis indicative of diabetes (250.xx); or (4) filled at least one prescription for an oral anti-diabetic agent or insulin. Diagnosis codes or medication utilization were used as exclusion criteria to provide a conservative approach to excluding patients with a previous history of diabetes.
Classification of Antipsychotic Exposure
Data analysis accounted for drug switching and non-consistent use of drug treatment by categorizing person-time exposed to individual antipsychotic agents. Patients contributed exposed person-time to individual antipsychotics beginning 45 days after index date, continuing for the duration of the days-supply for an individual prescription. Person-time accumulated, as described previously, until the first of the following occurred: (1) patient switched to a different antipsychotic; (2) prescription filled for anti-diabetic pharmacotherapy; (3) disenrollment from health plan; (4) death; (5) end of study period. Exposed person-time categories were formed for aripiprazole, olanzapine, risperidone, ziprasidone, and quetiapine. Clozapine was not included due to insufficient numbers.
Daily dose prescribed for each agent was calculated by multiplying number of pills dispensed by dose per pill, divided by days supply for that prescription episode. Blinded to study outcome and exposure, the following decision rules were applied to tabulate dosing episodes to determine whether decision-rule modification was necessary:
1. If patient received consecutive prescriptions for the same dose of medication, but the prescription was filled within (+/-) 14 days of the claims start/end date, the gaps were bridged and patient was classified as continuously exposed to the prescribed/recorded dose.
2. If patient filled 2 or more prescriptions on the same day, with the same days-supply and same dose, the days-supplies were summed.
3. If patient filled 2 or more prescriptions on the same day, with the same dose but different days-supply recorded, the days-supplies were summed.
4. If patient filled 2 or more prescriptions on the same day, with different doses:
a. If the days-supply between prescriptions was within (+/-) 14 days, daily dose was summed to the maximum days-supply between the 2 prescriptions.
b. If the days-supply between prescriptions was greater than 14 days, dose was considered switched to the second prescription strength (after 14 days).
c. If the days-supplies were equal, daily dose was summed.
d. If the days-supplies were equal and there were ≥2 records on the same day (e.g. record one--100 mg/30 days, 200 mg/30 days; record two--100 mg/30 days, 200 mg/30 days), we summed the daily dose and divided by the number of records and summed the days-supply (e.g. 300 mg/60 days).
Using these decision rules, we obtained (blinded) clinical input to determine whether data/decision rules reflected expected clinical dosing distributions. The dosing distributions for each drug were reviewed by a team of clinical experts consisting of two clinicians (one clinician was the study sponsor lead; the second clinician was an outside academic consultant), and three PhD-level epidemiologists (two epidemiologists were from the contract research organization and the third represented the study sponsor). To determine dosing distributions, each expert team member was provided a distribution of the dosing for each drug in the study. Blinded to study outcome, the team collectively reviewed the distribution for each drug and determined--based on clinical and statistical judgment--that tertile categorization was appropriate because the distribution of outcomes was too sparse to allow more granular analyses. The team addressed and deleted extreme outliers in dosing (e.g., clinically implausible doses indicative of data entry errors).
Study Outcome (Newly Treated Diabetes)
Patients exposed to SGAs for at least 45 days during the period January 1, 2002 through March 31, 2005 were identified and followed for the outcome of treated diabetes. To ensure ample exposure for evaluation, individuals were followed from the 45th day after index date through the earliest of the following events: (1) prescription filled for anti-diabetic pharmacotherapy; (2) disenrollment from database/health plan; (3) death; (4) end of study period.
Current daily antipsychotic dose (categorized by tertiles for each drug; quantified in person-time) was calculated as a time-dependant covariate (i.e. dose could change over follow-up time). Newly treated diabetes mellitus was identified using pharmacy data to determine patients exposed to anti-diabetic therapies. Hazard ratios (HRs) for diabetes across dose tertiles for each SGA were calculated via Cox proportional hazards regression using the lowest dose tertile as the reference, and adjusting for age, sex, study site, year of cohort entry, history of antipsychotic use (>3 months prior to index), exposure to other pharmacotherapy (alpha blockers, beta blockers, statins, corticosteroids, fibrates, lithium, oral contraceptives). We fit separate models for each SGA. Tests for trend across dose within each SGA were performed by coding increasing dose as 1, 2, and 3 and treating as a continuous variable in the Cox regression model.