Title: Estimating the treatment effect of the treated under time-dependent confounding - applied to simulated data and to data from the Swiss HIV Cohort Study

Abstract: The issue of time-dependent confounding is central in causal inference. When comparing treatments in a non-randomised setting, one will often need to correct for confounders that influence treatment choice over time, and that are themselves influenced by treatment. We present a new method for handling such confounding when estimating treatment effects of the treated. Using Farewell's linear increments model, we estimate what the values of possible confounding processes would have been, counterfactually, if treated individuals had not started treatment. These estimated counterfactuals are used to find a causally valid estimate of the treatment effect. The approach also yields a definition of a natural direct and indirect effect as functions of time. The validity of the approach is demonstrated in a simulation study. The method is also applied to data from the Swiss HIV Cohort Study, analysing the effect of antiretroviral treatment in HIV patients on time to AIDS or death. Compared to other methods that are used to adjust for time-dependent confounding our approach makes more assumptions about the covariate processes themselves rather than the treatment initiation. R code for carrying out the analyses is available.