Five hundred fifty-eight patients from two general medical/surgical ICUs were studied with respect to the relationship of mortality and length of stay to 18 treatment and condition variables. Two multiple logistic regression models were developed to estimate the probability of mortality. One was a model incorporating main effects only; the other allowed for the interaction of pairs of variables. For ICU mortality, the best predictive variables using the main effects model were coma, cardiopulmonary resuscitation (CPR) in the ICU, and shock. With the interactive model, the variables selected were shock plus coma, and CPR in the ICU.