Applied Statistics Seminar on Tuesday (February 7, 2023) at 12 PM

Abstract:

The exchangeability of units between treatment groups is a key and typically untestable assumption for evaluating causal intervention effects in observational studies. Standard methods assuming exchangeability can yield biased treatment effect estimates if the assumption does not hold. Existing methods evaluate the sensitivity of treatment effect estimates to non-exchangeability due to unmeasured confounders only. In practice, non-exchangeability can occur for either unmeasured confounders or reverse causality. We propose an index of sensitivity to non-exchangeability (ISENSE) to measure the impact of non-exchangeability on treatment effect estimates. Unlike existing methods, ISENSE does not require imposing assumptions on the types, numbers, and distributions of unmeasured confounders, and it can handle both unmeasured confounders and reverse causality. ISENSE is a computationally inexpensive local sensitivity method based on a Taylor-series approximation to the non-exchangeability likelihood, evaluated at the parameter estimates under the exchangeability assumption. One can interpret ISENSE intuitively through the unit-free “MinNE” statistic values that capture the minimum non-exchangeability needed to cause important sensitivity. We evaluate ISENSE using simulation studies and illustrate its use with an example using administrative data from British Columbia, Canada.

Presenter: