The study was a retrospective cohort study utilizing the Ingenix I3/LabRx claims dataset from 1/1/2003 through 12/31/2006. The dataset is a proprietary sample of individuals receiving health insurance benefits from United Health Care (UHC). UHC data include the inpatient, outpatient and prescription drug claims of more than 15 million of covered lives across the United States. The index date was the date of the first prescription claim for an atypical antipsychotic. Patients were followed for up to 1 year post-index. Because the dataset in this study was derived from an insurance claim database and the data conform to the Health Insurance Portability and Accountability Act of 1996 confidentiality requirements, the study did not require informed consent or institutional review board approval.
The study included outpatients aged 18-65 years with an ICD-9 code for bipolar disorder, manic, mixed or hypomanic (296.0x, 296.1, 296.4x, 6x, 7x, 8x). Eligible patients required at least 180 days or continuous health plan enrollment before, and 365 days after, the index date. Patients were included only if they were treated on a single atypical antipsychotic at index.
Patients were excluded from the analysis if they resided in a nursing home, hospice, or another type of long-term care facility, received mail-order prescriptions, or were diagnosed with a schizophrenia spectrum disorder (295.xx) during the pre- or post-index study period. Patients were also excluded if they used any atypical antipsychotic in the 180-day pre-index period, or had prescriptions for more than one atypical antipsychotic at index. Additionally, patients were also excluded if they were hospitalized within 7 days of their index antipsychotic prescription, in order to reduce treatment selection bias based on extreme agitation or instability.
Assessments and statistical analyses
The primary outcome of interest was the first psychiatric hospitalization in the follow-up period. Patients were censored for the following events: medical hospitalization, discontinuation of index antipsychotic (>15 days gap in coverage), or a prescription for a different antipsychotic during the follow-up period.
In order to control for treatment selection bias, we employed propensity score matching to construct comparison groups that shared similar demographic and clinical characteristics. Propensity score matching is a robust means of controlling for observed confounding in observational data . Propensity scores were calculated for each patient using logistic regression with independent variables of age, sex, region, pre-index diagnosis or treatment of diabetes or hyperlipidemia, pre-index psychiatric hospitalization, pre-index lipid or glucose laboratory claims, choice of pre-index mood stabilizer exposure and Charlson comorbidity index. The propensity score was the predicted probability of treatment calculated for each patient in the regression model. Patients in comparison treatment groups were matched 1:1 if their propensity scores were within 0.25 standard deviations of the logit of the propensity score. All analyses were conducted in propensity score-matched cohorts of the study sample.
The primary analysis used Cox proportional hazards regression to assess time-dependent risk of post-index psychiatric hospitalization with a pre-specified threshold for statistical significance of p < 0.05. Covariates for adjustment in the models included age, sex, diagnosis or treatment for diabetes or hyperlipidemia diagnosis, pre-index psychiatric hospitalization, pre-index lipid or glucose laboratory claims, choice of pre-index mood stabilizer and the Deyo Charlson comorbidity index . Intent-to-treat analysis was used for the cost analysis. Monthly treatment costs during the follow-up period were compared using generalized gamma regression controlling for pre-index costs in patients with positive post-index healthcare costs. First, we calculated the mean for each of the numeric covariates, and gave equal share of the categorical covariates, and then calculated the log mean of the fitted gamma distribution based on these covariate values and the parameter estimates and then exponentiated the log mean to get the cost in dollars. Gamma regressions were used to compare outcomes because gamma distribution is suggested by many as a close approximation of cost data. For example, Diehr and colleagues compared different methods to model healthcare cost data and concluded that, for understanding the effect of individual covariates on total costs, the gamma distribution might be preferred because it is a multiplicative model . Generalized gamma regression has been found to be a more robust estimator than traditional ordinary least squares regression in the analysis of healthcare expenditure data due to the distributional qualities of healthcare costs . Only patients with positive healthcare costs in the follow-up period were included in the analysis, which categorized costs into mental health (inpatient/ER and outpatient), medical (inpatient/ER and outpatient) and pharmacy (all medications used). We excluded patients with non-positive costs based on the assumption that patients taking medications were also receiving billable services and that the absence of such costs reflected aberrant data.
As a sensitivity analysis, we also replicated all multivariate regression analyses on the full unmatched samples.