Citation

St. Laurent, R.T., Weissfeld, L.A., & Moulton, L.H. (1995).   Bartlett correction of binomial regression analyses of small samples.   Proceedings of the Twentieth Annual SAS Users Group International Converence, Cary, NC: SAS Institute, pp. 1235-1239.

Abstract

Many epidemiologic investigations are necessarily restricted in the number of subjects to be studied.   Limited numbers of cases or exposed subjects may be available, or high data collection costs may limit the size of the study.   Yet the need to analyze several confounding or effect modifying variables simultaneously is not necessarily diminished.   Thus for dichotomous response variables, the use of binomial regression models (usually the logistic model) is often the analytic approach of choice in such situations.   In this paper we describe potential inferential pitfalls when analyzing data sets which are effectively "small", and we explain and discuss several possible courses of remedial action.   The Bartlett correction to the likelihood ratio statistic is seen to be a particularly useful tool in this setting.