Abstract:
|
Much attention has been given to the identification of causal effects in the presence of unmeasured confounding. Many proposals require access to auxiliary variables (e.g. instrumental variables, negative controls) that satisfy strong, untestable assumptions. In this talk, I will instead show how one can derive a `bespoke instrument' from a measured confounder, in order to remove residual confounding bias. This strategy requires an external reference population who did not have access to the exposure, and a stability condition on the outcome-confounder association between populations. I will extend the identification results of Richardson & Tchetgen Tchetgen (2021), outline the semiparametric efficiency theory for a general bespoke instrumental variable model, and describe multiply robust estimators of the average treatment effect in the treated. The utility of the estimators is demonstrated in simulation studies and a data analysis.
|