Random effects models for identifying the most harmful medication errors in a large, voluntary reporting database

Abstract

Medical errors are a major source of preventable morbidity, mortality and healthcare costs. Voluntary reporting systems are useful data sources that collect detailed information on the circumstances of medical errors occurring in hospitals. Identifying the characteristics of errors that frequently result in patient harm when they occur would allow investigators to prioritize among the many sources of potential errors and design targeted prevention strategies. In this paper, we use data from MEDMARX, a large anonymous and voluntary reporting system for medication errors, to identify the combinations of error characteristics that are more likely to result in harm. To this end, we consider a Bayesian hierarchical model with crossed random effects and a flexible specification of the random effects distribution. We then provide a ranking of the errors using optimal Bayesian ranking based on their probability of harm. The use of optimal Bayesian ranking accounts for the varying amount of uncertainty across the random effects estimates. Finally, we examine the sensitivity of results to different specifications of the random effects distributions. The utility of flexible random effects assumptions is illustrated by empirically comparing results under several choices. We found that errors caused by mistakes in reconciling a patient’s current medication list with the medications prescribed at hospital discharge have an estimated 10.5% probability of harm. These errors had the highest rate of harm of errors that occur during the prescribing stage of medication use. In addition, we found that the results are sensitive to the random effects distribution used in estimation. Thus, an approach that explores this sensitivity is important for accurately comparing the relative harm across errors.