Analysis of Treatment Effectiveness in Longitudinal Observational Data

Abstract

Assessing treatment effectiveness in longitudinal observational data is complicated as patients may change medications at any time. To illustrate, three general statistical strategies were utilized to assess treatment effectiveness in an observational schizophrenia study: ignoring, eliminating, and modeling the switching. Differential switching rates produced dramatic differences in estimates of treatment effectiveness across the strategies, with p-values ranging from nearly 0 to almost 1. Ignoring the treatment switching by utilizing intent-to-treat approaches resulted in treatment effect estimates of near zero. Various methods of eliminating the switching, such as epoch analyses and on-drug subset analyses, along with use of marginal structural models generated reasonably consistent non-zero treatment effect estimates. When analyzing longitudinal observational data, researchers must understand the options, key concepts and assumptions behind the various statistical methods available. Marginal structural models are a promising approach to estimation of causal treatment effects in such data.