Estimating variability for age-standardized hospital morbidity rates: a comparative study of automated methods for web-based atlas production using medical databases

Abstract

Accurate interval estimation of standardized hospital morbidity rates is essential for comparing health care utilization across geographical areas or different populations. There are typically repeated hospital admissions for individuals in the population in any specific time period; hence counts of admissions may not be well modelled by standard distributions for count data based on a Poisson assumption. This leads to empirical distributions of the counts with heavier tails than the Poisson. This paper complies and reviews various approaches for interval estimation of standardized hospital morbidity rates used at health agencies and examines their suitability under a broad range of conditions. The focus here is on approaches used for automated production of rates for developing atlases, using large-scale (e.g., country-wide) medical databases. We consider parametric models which incorporate such overdispersion, including the zero-inflated Poisson, negative binomial, zero-inflated negative binomial and Poisson-inverse-Gaussian distributions. Additionally, we consider the use of robust methods based on simple moment estimators for computing the mean and variance of the distribution of the counts, and Poisson-based methods currently utilized for some published rates when patient-level data are unavailable. A simulation study is conducted to compare the different approaches. Various confidence interval construction methods are also examined. Our results indicate that the moment approach as well as confidence intervals based on a log transformation of the rate provide more accurate inference for morbidity rates than the other methods. We recommend the use of the moment approach due to its simplicity of implementation. Some cautions regarding ad hoc approaches currently in use are also provided.