Agnostic Sensitivity Analysis

Christopher Schwarz (New York University)

Abstract: The threat of endogeneity is ubiquitous within applied empirical research. A `near Bayesian' method of sensitivity analysis is developed and implemented, overcoming a number of difficulties with existing approaches. The procedure targets the distribution of possible causal effects (DOPE) and summarizes the uncertainty of estimates to regressor-error dependencies. The procedure samples from the set of valid correlation matrices to generate the a priori distribution of causal effects under ignorance of the control function which would achieve conditional independence. This allows scholars to make probabilistic statements regarding the sensitivity of their results to arbitrary combinations of omitted variables, systematic measurement errors, selection biases, reciprocal causation, and certain SUTVA violations. Unlike existing approaches, the methodology naturally lends itself to the comparative robustness of studies and extends naturally to semi-parametric distributional regression models which allow for both heterogeneous treatment effects and inference beyond the mean.

View Poster in a New Tab