Sensitivity analysis for observational studies: past, now and future

Abstract: Any observational study of causal relationships relies on untestable assumptions. Thus, the credibility of a study crucially depends on the extent that its assumptions can be defended. The purpose of this talk is to give a high-level overview of how the statistical literature on sensitivity analysis evolved from its inception in the debate about smoking and lung cancer in the 1950s. I will review the first sensitivity analysis by Cornfield in 1959, the model and methodology based on randomization inference proposed by Rosenbaum, and other related literature. A recent emerging trend is to formulate sensitivity analysis for observational studies as a stochastic programming problem, which provides a unified conceptual framework for many existing models and methods. I will review some recent progress in this area and highlight the variety of theoretical, methodological, and practical difficulties involved in this framework. Part of this talk is based on joint work with Yao Zhang and Tobias Freidling.

Mathematics

Audience: researchers in the topic

Fluids and Structures Seminar @ UEA