Propensity score stratification for observational comparison of repeated binary outcomes

A two-stage longitudinal propensity adjustment is described for bias reduction in treatment effectiveness estimates in observational studies. The initial stage characterizes those who receive various ordinal doses of treatment in a model of time-varying propensity for treatment intensity. The second stage incorporates the propensity adjustment in longitudinal treatment effectiveness analyses of a binary outcome that are stratified by propensity quantile. Mantel-Haenszel pooled parameter estimates are then calculated as weighted means of quantile-specific estimates. A simulation study compares four approaches to quantile stratification and shows that the quintile stratification reduces more bias than when fewer strata are used with longitudinal data. Statistical power, type I error and coverage are also evaluated. A longitudinal, observational study of antidepressant effectiveness illustrates the approach.