Stukel's Extended Logistic Regression Analysis with R

Objective: For a logistic regression model, the degree to which predicted probabilities agree with actual outcomes can be expressed as a classification table. Being crucial in model adequacy checking, such tables may be slightly different when the same data are modeled with different statistical packages. The underlying reason is that when classifying a set of binary data, if the observations used to fit the model are also used to estimate the classification error, the resulting error-count estimate is biased. In order to cope with this problem, SAS suggests an algorithm, whereas the software is not publicly available. R is a free downloadable programme which is particularly designed for statistical computation, including the logistic regression analysis. The purpose of this study is to present a new function in R which carries out an extended logistic regression analysis of a binary data from the construction of its reduced-biased classification table, to the inference of its model parameters by calling the lrm(.) function under the Design package where necessary. Material and Methods: The performance of ext.logreg(.) is evaluated in terms of the accuracy of estimates and computational cost. Results: From the results of two binary datasets, it is observed that ext.logreg(.) via R estimates the model parameters and constructs the unbiased classification table as accurate as SAS programme under PROC logistic function without losing the computational demand. Conclusion: The free downloadable ext.logreg(.) function can be seen as an alternative computational tool in the analysis of logistic regression when the validation of predicted probabilities is essential.