Reply to: Some concerns about diagnostic test accuracy for infections

To the editor

We appreciate Kataoka et. al.’s interest in our article and thank them for their comments.^1,2 They are correct in noting that the 24/7/365 nature of emergency medicine can pose challenges to interpretation of clinical research. The ideal population they describe is not obtainable in the vast majority of emergency departments (EDs) due to staffing challenges and expense associated with around the clock research coordinator coverage. We believe there are few, if any, observational studies conducted in the ED that provide true 24/7 enrollment. Absent patient informed consent, we are unable to provide information for the vast majority of data points in Table 1 as these were based on data collected from the patient or the treating physician via survey. We do believe that potential bias is mitigated somewhat by our enrolling on evening hours (until 11PM) and including some weekend coverage in our enrollment. We also believe the population is broadly representative of the older ED population seen in our department given that approximately 80% of our ED patient’s present between 7AM and 11PM. Also, overall in our ED patients 70 years of age and older, mean age is 76 years (vs. 74 in the study population), 57% are females (vs. 58%), and 12% are African American (vs. 17%).

Their second concern was the calculation of the confidence intervals (CIs) for the likelihood ratios in Table 3. The discrepancy described is due to use of different, but equally valid, asymptotic approximations in the calculation of the CIs. The asymptotic for our CI is on the ratio which is why our CIs are symmetric about the estimate.³ It is also why they can return negative values as an asymptotic normal approximation doesn’t require a strictly positive distribution. The estimates Kataoka et al reference use an asymptotic on the log scale and then back transform. As a result, these intervals will never go below 0 but they also are not symmetric about the estimate. These are simply two acceptable asymptotic approaches and regardless of which is used, the study conclusions are not substantially changed.