Estimating the treatment effect on the treated under time-dependent confounding in an application to the Swiss HIV Cohort Study. Â Journal of the Royal Statistical Society
When comparing time varying treatments in a non-randomized setting, one must often correct for time-dependent confounders that influence treatment choice over time and that are themselves influenced by treatment.
In the current work, Gran and colleagues present a new two-step procedure, based on additive hazard regression and linear increments models, for handling such confounding when estimating average treatment effects on the treated. The approach can also be used for mediation analysis.
The method is applied to data from the Swiss HIV Cohort Study, estimating the effect of antiretroviral treatment on time to acquired immune deficiency syndrome or death. Compared with other methods for estimating the average treatment effects on the treated the method proposed is easy to implement by using available software packages in R.
In sum, compared with other methods that estimate average total treatment effect under time-dependent confounding on a time-to-event outcome, the proposed method is a two-step approach, where each step has the benefit of being easy to implement by using two simple existing statistical software packages.
The outcome model also has the benefit of being a hazard regression model in the traditional sense, which typically is not so in the g-estimation approach. The authors claim that there are advantages in considering several approaches for handling the thorny issue of time-dependent confounding, and that the procedure that is described in this paper serves as a valuable addition.