How should one report inferences about a population parameter when it is not estimable from observable data without the imposition of strong, untestable assumptions?
This question is relevant to a a wide variety of settings in propsective randomized and observational studies. For example, in the evaluation of the causal effects of treatment
- in studies with potentially informative missing or censored data;
- on functional outcomes in studies with high mortality rates;
- in studies with non-compliance;
- in studies with uncontrolled treatment assignment.
In my research, I focus on frequentist and Bayesian approaches to evaluating the robustness of results to assumptions such as missing at random and no unmeasured confounding. I am interested in formal sensitivity analysis procedures as well as methods for constructing bounds on treatment effects. I favor the incorporation of auxiliary information into the analysis to help increase the plausibility of assumptions, narrow the range of the sensitivity analysis, or shrink the bounds.
