An alternative method for logistic regression on contingency tables with zero cell counts

Abstract

This paper introduces an alternative method for solving a problem of non-convergence in logistic regression. The method does not require any special software to be developed. It simply involves modifying the data by replacing the zero count by 1 and doubling a corresponding non-zero count. The method is compared with that based on penalized likelihood suggested by Firth. Results show that the data modification method provides statistical significance of associations similar to Firth’s method while using standard logistic regression output.

Keywords

• zero cell count
• logistic regression
• data modification