We conducted a cohort study involving all British Columbia residents aged 65 years or more who filled a first-recorded (index) prescription for an oral antipsychotic medication between Jan. 1, 1996, and Dec. 31, 2004. To ensure a uniform 1-year eligibility period before filling the index prescription, all study subjects were required to have used at least 1 medical service and have filled at least 1 prescription in the two 6-month intervals before the index date. They could not have used an antipsychotic medication in the year before the index date. We restricted the analysis to include only new users of antipsychotic medications to guard against selection bias among prevalent users from early symptom emergence, drug intolerance or treatment failure.15
Patients with a diagnosis of cancer at the index date were excluded to avoid residual confounding introduced by selective prescribing of conventional antipsychotic medications (chlorpromazine, haloperidol) as antiemetics in the most serious cancer cases, because these patients are also more likely to die within 180 days independent of drug use.
Patients were identified in linked administrative data from the BC Ministry of Health that contained information on all physician services (Medical Services Plan), hospital admissions with up to 25 diagnostic codes and all dispensings of prescription drugs, independent of payor, recorded by the province-wide PharmaNet database. We further linked vital status information from the BC Vital Statistics Agency, the provincial vital statistics bureau. Underreporting and misclassification of data appear to be minimal because of the electronic data entry of all drug dispensings and because hospital diagnoses showed good specificity and completeness.16
Linkage, performed with the use of a personal health number unique to every BC resident, is considered complete among patients using the provincial health care system. All traceable personal identifiers were removed to protect patient confidentiality. The Institutional Review Board of the Brigham and Women's Hospital approved this study, and data-use agreements with the BC Ministry of Health were in place.
Atypical antipsychotic agents17
included in the analyses were risperidone (74.7% of all atypical agents dispensed), quetiapine (14.9%), olanzapine (10.1%) and clozapine (0.3%). Other antipsychotic medications were considered to be conventional17
and included loxapine (69.4% of all conventional antipsychotic drugs dispensed), haloperidol (11.0%), chlorpromazine (7.4%), trifluoperazine (5.0%), thioridazine (3.1%), pimozide (2.4%), promazine (2.4%), perphenazine (1.5%), fluphenazine (0.2%), mesoridazine (0.1%) and thiothixene (< 0.1%). We converted daily doses to chlorpromazine-equivalent milligrams using the midpoints of recommended ranges in geriatric prescribing guidelines.18
We used the median daily dose in the population as a cutoff to assess the effect of higher and lower doses.
The study outcome was death from any cause, as recorded by the BC Vital Statistics Agency. A set of potential confounders was measured based on health care utilization data within 6 months before the initiation of index drug use (index date). These confounders included sociodemographic characteristics (age, sex, race, nursing home residence), generic markers of comorbidity that have shown good validity in predicting death19
(hospital admission for any reason, number of physician visits, number of distinct prescription drugs excluding antipsychotic medications listed earlier and Charlson Comorbidity Index score20
), psychiatric morbidity (dementia, delirium, mood disorders, psychotic disorders and other psychiatric disorders), prior use of anticholinergic drugs and current use of anticholinergic drugs. We also identified the presence of conditions that are independent predictors of death and were related to antipsychotic drug use in earlier research, including arrhythmias (defined by the presence of a diagnosis of ventricular or other cardiac arrhythmia plus use of a group I–IV antiarrhythmia medication); diabetes (defined by the presence of a diabetes diagnosis plus use of antidiabetic medications); cerebrovascular disease (both cerebral hemorrhagic and ischemic events); congestive heart failure; acute myocardial infarction (MI); other evidence of ischemic heart disease, including angina (defined as the presence of a diagnosis of angina and nitroglycerin use), percutaneous coronary intervention or coronary artery bypass graft surgery; and other cardiovascular conditions (valvular disease, aneurysm or peripheral vascular disease).
We performed 3 types of statistical analyses: multivariable Cox regression analysis, propensity score analysis and instrumental variable estimation.
For the multivariable Cox regression analysis, we computed distributions of sociodemographic, clinical and utilization characteristics among the users of conventional and atypical antipsychotic drugs and then calculated the mortality during the first 180 days after initiation of either drug class. We chose a period of 180 days on the basis of the duration of trials in the FDA's repeat analysis (the trials lasted from 4 to 26 weeks, with a modal duration of 10 weeks).6
We constructed unadjusted and multivariable (controlling for calendar year and all covariates listed earlier) Cox proportional hazard models to estimate mortality ratios within 180 days after the start of antipsychotic drug use without censoring, analogous to an intention-to-treat analysis in randomized trials. Models of mortality in the first 39 days, in 40–79 days and in 80–180 days of drug use were also constructed. Adjusted models were run separately in strata defined by dementia and nursing home status. We also investigated whether a dose–response relation existed in adjusted models by separating the conventional drug users into 2 groups: those taking the median daily dose or less and those taking more than the median daily dose.
For the propensity score analysis, we developed Cox regression models adjusted for propensity scores21
for more efficient estimation.22,23
Propensity scores were derived from predicted probabilities estimated in logistic regression models of conventional versus atypical antipsychotic drug use. The final non-parsimonious propensity score model contained all covariates listed earlier and discriminated well between the type of drug used (c
statistic = 0.78). Cox regression models of mortality were stratified across tenths of the propensity score.
Finally, we used instrumental variable analysis to provide estimates that would remain unbiased even if important confounding variables were unmeasured.24–26
An instrumental variable is an observable factor related to treatment choice but unrelated to patient characteristics and outcomes. As in other recent work,27
we used as the instrument the prescribing physician's preference for conventional versus atypical antipsychotic medication (as indicated by their most recent new prescription of antipsychotic drug). Using 2-stage linear regression for the instrumental variable estimation and additional adjustment for measured patient characteristics, we calculated the risk difference of 180-day mortality between patients using conventional and those using atypical antipsychotic medications. Linear regression to estimate risk difference is valid in large samples such as ours.28
Because patient-level observations were clustered in physician practices, we performed robust calculations of standard errors of the regression parameters to account for the within-physician correlation of outcomes (Stata statistical software, version 9, StataCorp LP, College Station, Tex.).