Evaluating multiple imputation (MI) in the Active Bacterial Core surveillance system

The Active Bacterial Core surveillance (ABCs) is an active laboratory- and population-based surveillance system that monitors disease incidence from five invasive bacterial pathogens of public health importance. Monitoring disease incidence by race is important as incidence often varies by race, and assists in monitoring the Healthy People 2010 goal of eliminating health disparities. Some approaches to resolving missing data (such as single imputation) replace each missing value with a single value, which does not account for the uncertainty around the missing data and can lead to underestimation of the variance. Multiple imputation (MI) approaches replace each missing value with multiple values, allowing the variance estimation to account for the imputed missing data. Therefore, a multiple imputation evaluation was conducted to assess the reliability of multiply imputed race data in ABCs. We constructed 200 datasets using logistic regression by modeling missing data patterns, allowing for multiple plausible data points to provide for more accurate variance estimation. A frequentist evaluation counted the number of datasets with 95% confidence intervals (CIs) containing the true race-specific incidence, and 97% of the CIs contained the true disease incidence within each race category. A Bayesian evaluation calculated the average area within the normal approximation of the true distribution of race-specific disease incidence that contained the 95% CIs from the imputed data, and showed a median 94% for interval coverage. Multiple imputation approaches for missing race data result in statistically valid inferences that properly reflect uncertainty due to missing values.