64.011 - Sensitivity Analysis: Quantifying the Discourse About Causal Inference

Session Type: Professional Development Course

Abstract

Participants will learn how to quantify concerns about causal inferences due to unobserved variables or populations. Participants will also learn how to calculate the correlations associated with an unobserved confounding variable or the amount of one’s sample that would have to be replaced to invalidate an inference.
The instructors will present a general framework for characterizing the robustness of inferences from randomized experiments or observational studies. Calculations for bivariate and multivariate analysis will be presented in SPSS, SAS, and Stata, with an excel spreadsheet for other applications. Additional topics include a typology of thresholds for making inferences, null hypotheses of non-zero effects, evaluating thresholds relative to characteristics of observed variables or populations, and extensions to non-linear models. The live format will be a mixture of presentation, individual exploration, and group work. The course is aimed at graduate students and professors who are comfortable with basic regression and multiple regression.
Topics covered in this course include:
Overview and logistics, introduction to the counterfactual and % bias to invalidate an inference.
Interpreting % bias to invalidate an inference.
Example of application to an observational study.
Example of application to a randomized experiment.
Extensions of % bias thinking, introduction to correlational framework, how regression works.
Application of correlational framework to observational study
Worked example for the impact threshold for a confounding variable.
Application of correlation approach to external validity
Conclusion Application of correlation approach to external validity

Sub Unit

• Professional Development and Training Committee

Instructor

• Kenneth A. Frank, Michigan State University