Comparing statistical methods in estimating per-protocol effects to address sparse followup issue in pragmatic clinical trials with treatment non-adherence

Md. Belal Hossain

Keywords: sparse follow-up, non-adherence, per-protocol, pragmatic trials

Although the treatments are randomly assigned at baseline for pragmatic trials, subjects may deviate from the protocol because of switching to other treatments, loss to follow-up due to side-effects. An additional issue arises when there is an infrequent measurement frequency of post-randomization prognostic factors. Inability to adjust for non-adherence and unmeasured confounding due to irregular measurement frequency could bias the treatment effect, depending on the estimation method. The inverse probability of adherence weighting (IPAW) method has been shown to substantially reduce the per-protocol effect estimate’s bias after addressing the treatment non-adherence issue if those unmeasured longitudinal values are imputed more frequently. Previous applications of the IPAW method relied on the last observation carried forward (LOCF) approach for imputation, although LOCF is known for artificially reduced within-subject variability. In the present study, we propose using multiple imputations (MI) as an alternative, which harnesses other covariates’ predictive ability. These two methods’ performances have not been compared in the literature yet in the IPAW context in solving the data sparseness issues in the presence of treatment non-adherence. We simulated pragmatic trial data with varying amounts of sparse measurements on post-randomization covariates under different realistic clinical settings. We assessed the statistical properties of IPAW estimators when data imputed under LOCF versus MI approaches with the treatment non-adherence issue.