The lay abstract featured today (for the Tutorial in Biostatistics on Modern approaches for evaluating treatment effect heterogeneity from clinical trials and observational data by Ilya Lipkovich, David Svensson, Bohdana Ratitch, Alex Dmitrienko) is from Statistics in Medicine with the full article now available to read here.

Abstract

In this tutorial, the authors cover a broad range of statistical approaches supporting personalized, a.k.a. precision, medicine. Instead of traditional presentation of estimated treatment effects as averages over patient populations of interest, precision medicine presumes that treatment effect may vary depending on patient’s characteristics (genetics, demographics, disease history, etc.). Therefore, personalized medicine may suggest existence of subpopulations with superior or inferior treatment effects compared to those in the overall population, resulting in different optimal treatments for different patient subgroups based on predicted individual treatment benefits. Such predictions can be derived by data-driven methods designed to assess heterogeneity of treatment effects across patients based on data from randomized and observational clinical trials. Methods for evaluating heterogeneous treatment effects (HTEs) are especially challenging. First, individual treatment effects (differences in what the outcome would be on one treatment versus another for each patient) are not observable at an individual level, because each patient is typically assigned only to one of the candidate treatments in a clinical study. Secondly, it is often unknown which variables may explain heterogeneity of treatment effects and this knowledge may need to be derived from potentially high dimensional data. The first challenge is common for causal inference, the second-for machine learning. Thirdly, methods for evaluating HTEs need to address the multiple testing problem, which is particularly challenging in this analysis context and, if ignored, would lead to false discovery and lack of reproducibility. The authors summarize the main types of data-driven analysis methods, explain the evolution and key features of methods from each of these classes, as well as methods for inference about HTEs, based on over 200 references from the recent literature. The use of key selected methods is illustrated with data examples and available R packages, making the analysis accessible to a broad community of clinical statisticians and researchers. The authors stress that the task of a data-driven evaluation of HTEs remains challenging, and scientists should rely on comprehensive data evaluation using multiple methods and principled inferential procedures rather than on a single technique and (often misleading) visualizations when interpreting the results.