Logistic regression is the most commonly used technique for multivariate data analysis in case-control studies, clinical trials, or studies in the social sciences. Major statistical packages like SAS, SYSTAT, BMDP, SPSS, AND EGRET, all provide logistic regression modules that rely on large-sample theory for their validity. The conclusions os small studies analyzed by these packages are therefore often discredited. An exact theory for logistic regression, that remains valid in small samples, although well-recognized in the statistical literature, has never been considered a practical option because it is very computationally intensive. Recent publications, by the principal investigator and co-investigator, of efficient numerical algorithms for exact logistic regression have changed the picture, placing exact logistic regression within reach of the working statistician. Phase I demonstrated that these new algorithms can be implemented on microcomputers, attached to major statistical packages, and do indeed provide new insights about binary data in small studies. The phase II goal is to develop powerful, reliable, easy-to-use, commercially viable software for exact logistic regression, that can either operate in stand-alone form through a spreadsheet interface, or be inserted into SAS, SYSTAT, and EGRET. The software will be available on micro, mini, and mainframe computers under many operating systems. It will handle both, unconditional logistic regression and conditional logistic regression for matched case-control studies.
