Measuring a Binary Response's Range of Influence in Logistic Regression

Abstract

This article proposes a conceptually simple class of diagnostics to determine the potential of a single binary response to affect the outcome of a logistic regression. We simply rerun the regression once for each observation after changing the response and calculate how much the statistics of interest have changed. We show how statistics of this type may be helpful in checking the robustness of a logistic regression and study their relationship to earlier proposed diagnostics.

KEY WORDS:

• Diagnostic
• Generalized linear model
• Leverage