In many clinical or epidemiological studies, the objective is to relate a binary response variable to treatment, exposure levels and other covariates. Usually, this is done by using a logistic model with the data analysis being based on large sample assumptions. Frequently, however, the sample size is so small or the data structure so sparse as to invalidate these assumptions. For example, studies of cancer etiology, prevention or treatment often have a small number of subjects. In such cases, small sample or exact methods of data analysis need to be used. Hitherto, these methods, although well-recognized in statistical theory, tended to be extremely computationally intensive and were not considered to be a practical option by data analysts. The recent development by the principal investigator and consultant of a computationally efficient multivariate shift algorithm for these problems has, however, changed the picture and placed exact logistic regression within the reach of the working statistician. The main goal of this project is to do the preliminary work for developing a user-friendly and widely marketable software package for this algorithm. Specifically, this project will, (a) undertake a study of the feasibility of implementing this algorithm on mini and micro-computers, (b) explore the possibility of attaching the algorithm to existing statistical packages like SAS and EPIC, and (c) lay the groundwork for developing a convenient data manipulation interface for the algorithm.
