LAUSR

3. The number of patients with viral infections is at any rate too low for far-reaching conclusions, as pointed out by Drs. Karakonstantis and Kalemaki who problematize the use of logistic regression, citing the paper by Moons et al. [3]. Moons et al. warn against overfitting in prediction models when a final prediction model is identified by selecting a smaller number of predictors from a larger set of potential predictors, and then quantify the predictive performance. This has been done to a very small extent in our analyses as a set of predictors were pre-selected on a substantial basis. We have tried to avoid overfitting by not doing a stepwise selection of predictors. One method to adjust for overfitting is also to decrease the level of significance for including terms in the model, e.g., from a nominal level of 0.05–0.01. In the model for NLR an interaction term was included with p = 0.005 and for MLR with p = 0.001. Nevertheless, we agree that overfitting may have occurred although we have tried to take precautions. For the small viral infection group the ORs are all very uncertain and unstable with large confidence intervals reflecting the low power for drawing inferences about this group. Any firm conclusions for this group cannot be drawn, and the results should be confirmed in future studies. The group was included for completeness in this study which is primary targeted at bacterial infections. Ideally we could have done some internal validation using resampling techniques, but this was not done. 4. Septicemia is, admittedly, an unclear term, and Drs. Karakonstantis and Kalemaki are right in pointing out that we should have used the term sepsis, defined by Bone et al. [4] with later revisions [5] which include bacteremia defined as positive blood culture. Patients with bacteremia without clinical signs of infection were not included. All patients were admitted for fever, and Dear Editors,