t " THE PRACTICAL VALUE OF LOGISTIC REGRESSION

Logistic multiple regression using the method of maximum likelihood is now the method of choice for many regression-type problems involving binary, ordinal. or nominal dependent variables. Logistic regression does not require grouping of observations to obtain valid estimates of effects and of outcome probabilities, and it has been shown in the binary case to provide more accurate probability estimates than linear discriminant analysis when the assumptions of the latter (i.e., multivariate normality of predictor variables with common covariance matrix) are violated. Even when multivariate normality holds, logistic regression has been shown to yield probability estimates virtually as accurate as those obtained using discriminant analysis. The assumptions of the logistic regression model are for the most part straightforward and easy to verify. A general purpose SAS macro language program to verify the assumptions of the binary or ordinal model graphically will be discussed. Examples demonstrating the advantages of logistic regression for binary and ordinal dependent variables over other methods will also be presented.