Sensitivity analyses play an important role in assessing the reliability of causal claims in observational data by revealing what qualities an unobservable confounder must have to substantively change our belief about an estimated causal effect. While numerous sens-itivity analyses exist, they remain underutilized in applied empirical work. Our focus is on overcoming the hurdles of widespread use. By elaborating on the familiar “omitted variable bias” framework, this paper proposes an intuitive and easy to use set of sensitiv-ity analysis tools for linear outcome models. In particular, we introduce a new measure of sensitivity of linear regression coefficients called the “robustness value”. The robust-ness value reveals the “minimum strength of association” (in terms of partial R2) that unobserved confounding would need to have with the treatment and with the outcome to change the research conclusions. It’s a simple to compute and easy to interpret measure that, we argue, should be rountinely reported in regression tables. We further provide visualizations that help researchers make more elaborate arguments about the sensitiv-ity of point estimates, t-values and confidence intervals, by benchmarking with observed covariates and contemplating worst-cases scenarios. These tools work for continuous or binary treatments, do not require assumptions on the treatment assignment mechanism nor on the distribution of the unobserved confounder, and the results can be used to assess the effect of multiple confounders, whether they influence the outcome linearly or not. We illustrate these methods with a running example that estimates the effect of exposure to violence in western Sudan on attitudes toward peace. Open-source software, sensemakr for R implements the methods and visualizations provided here, together with natural-language interpretations of the results.

Read more about the Annual Meeting here: [ Ссылка ]

свернуть