In their editorial, Spiegelman and Hertzmark (1) recommend an easy method to estimate risk and prevalence ratios, and they include SAS macros for performing the calculations. The method uses maximum likelihood when the correct binomial model converges and a Poisson model with a robust variance estimator when the correct model fails to converge. We agree completely with using maximum likelihood estimators (MLEs) when the model converges. However, one can do better than the Poisson model when the correct model fails to converge. There are two deficiencies in the Poisson approximation. First, as the authors (1) mention, the estimates are not as efficient as those for the MLEs of the log-binomial model. Second, a point that the authors do not mention, estimates of probabilities obtained from the Poisson model can exceed 1 because the wrong model is being fit (2). In 2003, Deddens et al. (3) proposed using maximum likelihood even when the correct model fails to converge. Because of difficulties in computing MLEs when the model fails to converge, they proposed obtaining MLEs on a modified data set such that these MLEs could be made arbitrarily close to the MLEs for the original data set. Doing so retains the efficiency properties of maximum likelihood as well as assures that estimated probabilities will be between 0 and 1.