Trial of alcohol-based hand gel in critical care units.

patients in each unit were observed for 1 year in each branch of the crossover study. Twelve beds multiplied by 365 days is 4,380 bed-days; so they had 4,380 bed-days as a maximum (we do not know if the units were consistently fully occupied). Their reported infection rates are approximately 1 to 4 infections per 1,000 bed-days. This means that they observed approximately 4 to 16 infections over the entire year of the intervention for each arm of the study. This range represents a very small number of infections, and without getting into the details of the underlying Poisson regression model, the inherent variability on these numbers will be relatively high. So, what does this mean? It means that the data are very noisy, and the study is unlikely to be able to demonstrate an effect even if it is there. Rough calculations suggest that even if improvements in hand hygiene adherence were able to decrease infection rates by 50%, then this study would have only roughly a 20% chance of demonstrating the effect. Another issue important to this data set is whether the infections themselves are independent or whether they occurred in clusters (clumped in time). If they were clustered (which would mean that they were not statistically independent), then this analysis would be weakened even more, because ignoring the clustering would give a false sense of the amount of information contained in the data. If the infections are clustered in time, then the analysis is inappropriate. To the authors' credit, they do acknowledge that the study is "underpowered to detect small differences in rates of infection," but it may be underpowered to demonstrate larger differences, too. The reviewers of this article should have noted to the authors that this is a good article but requested that they leave out the comments on "detectable changes in the incidence of healthcareassociated infection," because there does not appear to be enough information to generate a reliable conclusion.