Abstract

Causal inference attempts to uncover the structure of the data and eliminate all non-causative explanations for an observed association. The goal of most, if not all, statistical inference is to uncover causal relationships, but it is not in general possible to conclude causality from standard statistical inference procedures, merely that the observed association between two variables is not due to chance. The need for causal inference procedures is apparent in many fields, but is perhaps most pressing in the field of health research, where quantifying the efficacy of new therapies, or uncovering the etiology of diseases, is often rendered complicated due to difficulties inherent in observational studies. Even in experimental studies, partial compliance to treatment regimens can compromise a well-designed experiment. The complexity of models, and corresponding inference procedures, is heightened in the context of longitudinal studies, where time-dependent confounding may be present.

The statistical study of adaptive treatment strategies, also called dynamic treatment regimes, is an area that has grown to prominence over the last two decades. While sequentially randomized trials can provide high quality evidence for the efficacy of tailored sequences of treatments, the sample sizes needed to estimate heterogeneous effects that form the basis of treatment personalization are such that non-experimental data sources are commonly used to estimate adaptive treatment strategies. Thus, methods in causal inference have formed the basis for most estimation approaches in this area.

The aim of this proposed session is to showcase the application and development of statistical methods that address challenges and biases in observational data to help solve problems relating to personalization of care. The session would consist of four speakers, and cover the topics of missing data, model mis-specification, measurement error, and irregular observation times.