History-Adjusted Marginal Structural Models: Optimal Treatment Strategies

Maya L. Petersen, Division of Biostatistics, School of Public Health, University of California, Berkeley
Mark J. van der Laan, Division of Biostatistics, School of Public Health, University of California, Berkeley

Abstract

Much of clinical medicine involves choosing a future treatment plan that is expected to optimize a patient's long-term outcome, and modifying this treatment plan over time in response to changes in patient characteristics. However, dynamic treatment regimens, or decision rules for altering treatment in response to time-varying covariates, are rarely estimated based on observational data. In a companion paper, we introduced a generalization of Marginal Structural Models, named History-Adjusted Marginal Structural Models, that estimate modification of causal effects by time-varying covariates. Here, we illustrate how History-Adjusted Marginal Structural Models can be used to identify a specific type of optimal dynamic treatment regimen. Estimation and interpretation of this dynamic treatment regimen are illustrated using an example drawn from the treatment of HIV infection using antiretroviral drugs.