Confidence Intervals for the Predictions of Logistic Regression in the Presence and Absence of a Variance- Covariance Matrix

Confidence intervals for fitted values provide valuable information about the usefulness of regression models. Although such intervals can be easily calculated using standard statistical software for response variables that have normally distributed errors (e.g., in ordinary least-squares regression), it is more difficult to calculate them for response variables that have binomially distributed errors (e.g., logistic regression). Although a number of statistical packages provide confidence intervals for fitted values directly for logistic regression models, some commonly used packages do not (e.g., SPSS). In this article we outline a method of calculating these intervals simply by fitting a model after transforming variables. This technique is evaluated by comparing results with those obtained using a method that utilizes the variance-covariance matrix. Both techniques are described in detail and applied to simple and multiple logistic regression along with step by step instructions and software commands for SPSS Version 10.1.