Abnormally wide confidence intervals in logistic regression: interpretation of statistical program results1

This study describes the behavior of eight statistical programs (BMDP, EGRET, JMP, SAS, SPSS, STATA, STATISTIX and SYSTAT) when performing a logistic regression with a simulated data set that contains a numerical problem created by the presence of a cell value equal to zero. The programs respond in different ways to this problem. Most of them give a warning, although many simultaneously present incorrect results, among which are confidence intervals that tend toward infinity. Such results can mislead the user. Various guidelines are offered for detecting these problems in actual analyses, and users are reminded of the importance of critical interpretation of the results of statistical programs. ABSTRACT