Regression Discontinuity: Examining Model Misspecification

Abstract

The Regression Discontinuity (RD) design looks similar to the non-equivalent group design, which uses analysis of covariance, but assumptions and advantages are much different. The major problem in analyzing data from the RD design is model misspecification. If the regression equation or statistical model does not reflect the data distribution, then biased estimates of the treatment effect will occur. For example, if the true pre-post relationship is curvilinear, but the regression equation only modeled linear regression effects, the treatment effects would be biased. However, a statistical approach is possible using a full model with all terms specified and then test restricted sub-models that omit individual parameters.