Confidence interval methods for antimicrobial resistance surveillance data

Background: Antimicrobial resistance (AMR) is one of the greatest global health challenges today, but burden assessment is hindered by uncertainty of AMR prevalence estimates. Geographical representation of AMR estimates typically pools data collected from several laboratories; however, these aggregations may introduce bias by not accounting for the heterogeneity of the population that each laboratory represents. Methods: We used AMR data from up to 381 laboratories in the United States from The Surveillance Network to evaluate methods for estimating uncertainty of AMR prevalence estimates. We constructed confidence intervals for the proportion of resistant isolates using (1) methods that account for the clustered structure of the data, and (2) standard methods that assume data independence. Using samples of the full dataset with increasing facility coverage levels, we examined how likely the estimated confidence intervals were to include the population mean. Results: Methods constructing 95% confidence intervals while accounting for possible within-cluster correlations (Survey and standard methods adjusted to employ cluster-robust errors), were more likely to include the sample mean than standard methods (Logit, Wilson score and Jeffreys interval) operating under the assumption of independence. While increased geographical coverage improved the probability of encompassing the mean for all methods, large samples still did not compensate for the bias introduced from the violation of the data independence assumption. Conclusion: General methods for estimating the confidence intervals of AMR rates that assume data are independent, are likely to produce biased results. When feasible, the clustered structure of the data and any possible intra-cluster variation should be accounted for when calculating confidence intervals around AMR estimates, in order to better capture the uncertainty of prevalence estimates.