MODELLING RESPONSE WITH TIME-VARYING COMPLIANCE IN LONGITUDINAL DATA: A SIMULATION STUDY WITH TWO-STAGE FRAMEWORK USING COMPLIANCE REGRESSION RESIDUALS IN CLINICAL TRIALS

Abstract

Background: Clinical studies, such as randomized controlled trials, typically measure response and key event incidence throughout the follow-up period. Patients may skip certain assessment visits due to lack of efficacy or safety concerns. Missing data being a common problem in statistical literature, approaches to handle it may still result in biased knowledge discovery. Analysis and interpretation become problematic when missing data percentage is substantial. Additionally, compliance to planned treatment paradigm could also be a problem as patients might not adhere to prescribed treatment regimen. Twin consequences of non-compliance and missing data are rarely addressed simultaneously, even though numerous innovative techniques to handle non-compliance or missing response in randomized trials have been proposed.

Materials and Methods: This article attempted to address the missing response by deploying 2 stage modelling for analysing longitudinal response using time-varying compliance regression residual. Given the variation in longitudinal outcomes, accounting for the dependency between continuous response and treatment compliance can be informative, especially for imputing missing data. EM algorithm is used in this process and compared with/without 2 stage modelling. Simulation study is created with missing response and non-compliance to assess the effectiveness of proposed estimators in scenarios, including both continuous and binary treatment compliance.

Results: Method was applied on simulated data with varying correlation and multiple missing scenarios in both the cases. The results were compared using the absolute bias and mean squared error (MSE).

Conclusions: The MSE was smallest for the proposed method compared to without joint modelling and no imputation analysis, indicating better results with the proposed method.