Methods for Medication Error Report Databases
Our work is motivated by the United States Pharmacopoeia's MEDMARX ® database, a national, internet-accessible database of medication error reports. The primary aim of my thesis research will be to develop statistical methodology to extract useful information from medication error reporting systems such as MEDMARX. Identifying the types and root causes of error is crucial to implementing interventions to reduce error, but no general statistical framework exists for making inference on these variables.

As a means to that end we have begun by developing methodology to evaluate the causal continuumn hypothesis with respect to medication errors and quantify the evidence and uncertainty for this hypothesis in the MEDMARX database. The causal continuum hypothesis states that the causes and contributing factors of error that frequently lead to near-misses (errors that do not result in patient harm) are the same as the causes and contributing factors that frequently lead to adverse events (errors that do result in patient harm). This work is being conducted with Laura Morlock of the Department of Health Policy and Management and Francesca Dominici.

The definitions of the causes and contributing factors of medication errors from our manuscript Bayesian hierarchical models for analysis of medication errors may be found here. The manuscript is here.

We are now working on methods to identify the most harmful combinations of medication error characteristics. Although in the causal continuum analysis, we found that the prevalence of harm did not vary across most causes and contributing factors, there may still exist some combinations of causes and other error characteristics that frequently result in harm. Identifying these errors are important for setting intervention priorities across healthcare facilities.

Optimal Propensity Score Stratification
Tom Louis and I have been investigating the optimal way to choose strata in a stratified propensity score analysis of bivariate treatment effect as a function of the amount of imbalance in propensity between the two groups. We are considering a strata to be optimal if it produces an effect estimate with minimum MSE. As a comparison to the non-parametric stratification approach, we also consider the use of semi-parametric generalized additive models (GAMs) in the propensity score analysis.