We summarized categorical baseline characteristics as frequencies (and percentages) and compared across exposure groups with chi-square tests. We summarized continuous baseline characteristics as means (and SDs) and compared across exposure groups with independent t-tests. We reported the proportion of residents missing data for individual baseline characteristics.
We derived inverse probability of treatment weights from an estimated propensity score, which we derived by regressing exposure status on baseline covariates (see Appendix B for covariates), including medications dispensed in the year before cohort entry. In our study, the propensity score was the probability that a resident would be dispensed trazodone or zopiclone, conditional on their baseline characteristics.(28 (link)) We included missing values for categorical variables as an additional category. There were no missing values for continuous variables. We modeled the average treatment effect because we could foresee any cohort member potentially receiving either exposure drug and we wished to understand the average effect of treatment in the entire cohort.(29 (link)) Treatment weights were inspected for outlying values (greater than 50).(30 (link)) Crude (i.e., unweighted) and weighted cause-specific hazard ratios comparing outcome rates associated with zopiclone or trazodone use were derived from cause-specific hazards models, because we wanted to understand the association between our exposure and outcome rate in residents who had not yet had an outcome and were, therefore, at risk of having an outcome.(31 (link)) Weighted cause-specific hazards models were adjusted for all baseline characteristics for which there were statistically significant differences (i.e., p<.05) between exposure groups. We verified that hazard ratios did not vary over time, and we used robust standard errors to account for within-subject homogeneity in outcomes induced by weighting.(32 (link))Our primary analyses were based on an intention-to-treat principle (i.e., residents who were newly dispensed either exposure drug remained in the cohort even if they were dispensed the other exposure drug during the follow-up period); residents were followed until the first of the outcome of interest, death, or 180 days after index date. In secondary analyses, we censored residents who were dispensed the other exposure drug during the 180-day follow-up period (i.e., per protocol). We reported weighted incidence rates as the number of events per 100 person-years. Where numbers permitted, we planned to conduct subgroup analyses based on residents’ age, sex, dementia severity, and concurrent antipsychotic prescription. Analyses were conducted with SAS (version 9.4; Cary, NC) and STATA SE (version 13.0; StataCorp LP, College Station, TX).