Purpose: Patient-level predictions from population-based predictive models are subject to uncertainty. The impact of this uncertainty can be described using probabilistic sensitivity analyses (PSA). The logistic normal distribution is frequently used to describe probability distributions in second order Monte Carlo simulations. However, its standard error (SE) depends upon whether the upper or lower 95% confidence limit of the probability is used in the transform, as the confidence interval (CI) is not symmetric around the mean. We describe a technique that avoids this transformation when using the output of logistic regressions as inputs to decision models. Methods: We created a two-stage decision support tool for severe sepsis that guides the use of drotrecogin alfa (activated). First, logistic regression models were developed to calculate patient-specific mortality with and without treatment These were then used as inputs to a seventy-five state Markov model. Patient-specific predictors included age, gender and twelve readily available clinical characteristics. We calculated the logit and SE for 28-day mortality for hypothetical patients with specific risk profiles, either receiving or not receiving treatment. We also calculated the resultant probabilities of death with their 95% confidence intervals. For each patient, we performed PSAs in two ways: using a log normal distribution to describe the logit and its SE, and also using the approach described by Doubilet et al creating a logistic normal distribution from the probabilities and their upper and lower 95% CIs. Results: For a series of hypothetical patients, SEs in the distributions of the differences in expected utility between the two strategies were smallest when using the lower 95% CIs for probabilities of death less than 50% and were the largest for probabilities of death greater than 50%. Our method using log normal distributions based on the logit of the regression model yielded intermediate SEs. Conclusions: Uncertainty in the results of PSAs may be under or overestimated when using logistic normal distributions to describe probabilistic inputs to decision models. Direct use of the logit and its SE through a log normal transformation simplifies calculations and eliminates these errors when performing PSAs with decision models that incorporate predictive logistic regressions.