The Potential and Challenges of Current Methods for Sensitivity Analysis in Structural Equation Modeling

• In Event: Sensitivity Analysis in Structural Equation Modeling

Abstract

Structural Equation Modeling (SEM) is popular because it is very general and flexible. One way to obtain validity evidence for a theory supported by a SEM study is through a sensitivity analysis to omitted confounders, which evaluates the extent that the conclusions of the SEM study hold firmly if there were additional variables in the model. Therefore, sensitivity analysis is an important tool to probe the boundaries of the conclusions of a research study. The objective of this paper is to review and demonstrate a recently-proposed method for sensitivity analysis in SEM. The SEM sensitivity analysis method consists of specifying a phantom variable to represent a candidate omitted confounder and choosing paths from the phantom variable to the variables in the model that the researcher believes are possible. Then, the researcher selects a set of values (i.e., sensitivity parameters) to fix each path of the phantom variables to, and runs the model multiple times, fixing the paths of the phantom variable to a different value from the pre-selected set in each run. Finally, the researcher compares the coefficients and p-values from the group of results obtained from models with phantom variables to the original model results and identifies which path values of the phantom variable resulted in substantial changes.

Author

• Walter L. Leite, University of Florida