Abstract
Time-varying confounding forms a pervasive problem in observational studies that attempt to assess the total effect of a time-varying exposure on a survival outcome. It is tempting to see prosperity in the use of routine survival models with time-varying covariates, but unfortunately the resulting estimates of the exposure effect are prone to bias when, as often, time-varying confounders are themselves affected by earlier exposures. This is because standard regression adjustment for such confounders eliminates indirect effects of early exposures that are mediated via those covariates, and in addition, may induce a so-called collider-stratification bias. Martinussen et al. (2011) demonstrated how a valid adjustment for time-varying confounding is attainable when effects are parameterized on the additive hazard scale. Because they focused on the special case of 2 exposures, the first of which is dichotomous and randomly assigned, we here extend their results to general time-varying exposures, as well to settings with unmeasured confounding but data on an instrumental variable. Relative to G-estimation for structural nested accelerated failure time models, the attraction of the proposed approach is that it naturally accommodates independent censoring without requiring an artificial recensoring procedure to maintain unbiased estimating equations. Relative to inverse probability weighting for marginal structural survival models, the advantages are that it tends to yield more stable inferences by avoiding inverse probability weighting, and that it can incorporate instrumental variables and effect modification by time-varying covariates.
This talk will be based on joint work with Torben Martinussen, Eric Tchetgen Tchetgen and David Zucker.