Purpose of review
Pharmacoepidemiologists are often interested in estimating the effects of dynamic treatment strategies, where treatments are modified based on patients' evolving characteristics. For such problems, appropriate control of both baseline and time-varying confounders is critical. Conventional methods that control confounding by including time-varying treatments and confounders in an outcome regression model may not have a causal interpretation, even when all baseline and time-varying confounders are measured. This problem occurs when time-varying confounders are, themselves, affected by past treatment. We review alternative analytic approaches that can produce valid inferences in the presence of such confounding. We focus on the parametric g-formula and inverse probability weighting of marginal structural models, two examples of Robins' g-methods.
Unlike standard outcome regression methods, the parametric g-formula and inverse probability weighting of marginal structural models can estimate the effects of dynamic treatment strategies and appropriately control for measured time-varying confounders affected by prior treatment. Few applications of g-methods exist in the pharmacoepidemiology literature, primarily due to the common use of administrative claims data, which typically lack detailed measurements of time-varying information, and the limited availability of or familiarity with tools to help perform the relatively complex analysis. These barriers may be overcome with the increasing availability of data sources containing more detailed time-varying information and more accessible learning tools and software.
With appropriate data and study design, g-methods can improve our ability to make causal inferences on dynamic treatment strategies from observational data in pharmacoepidemiology.