| In recent years the medical community has given increasing attention to the standardization and improvement of healthcare quality and safety. In particular, medical errors have been shown to be a significant source of preventable morbidity, mortality and healthcare costs. Observational data, including voluntary reporting databases of medical errors, have been suggested as an important tool in quality and safety research. For example, we consider MEDMARX, a large, national, voluntary reporting system for medication errors. This dataset provides large amounts of data from many healthcare facilities and can be useful for studying medication error. In this thesis, we summarize selected background literature and consider three issues in the application of statistical methods to the analysis of observational data in health services research. First, we develop Bayesian hierarchical models for quantifying the evidence in MEDMARX for the causal continuum hypothesis, which states that the causes and contributing factors of errors that result in patient harm are similar to the causes and contributing factors of errors that do not result in patient harm. We also use our models to identify the causes and contributing factors that may be aberrant from the causal continuum hypothesis. We found that the data strongly support the causal continuum hypothesis for most causes and contributing factors. Next, we develop a Bayesian hierarchical model to rank the combinations of error type and error causes that have the highest log odds of harm in MEDMARX. We compare this model to an empirical Bayes approach that was proposed for ordering medication-event combinations in an FDA spontaneous reporting database. In each approach, we consider the sensitivity of results to the specification of the random effects distributions. We found that both approaches provide reasonable orderings of the combinations of error characteristics, but the Bayesian hierarchical model with optimal Bayesian ranking provides a more accurate characterization of the evidence. Finally, we consider applications of propensity score methodology in adjusting for bias due to confounders when estimating the effect of a binary treatment. We compare stratification and regression approaches via Monte Carlo simulation, and illustrate the approaches in an analysis of the effect of insurance plan choice on satisfaction with asthma care in an observational study of asthma patients. In the simulations we considered, generalized additive models with a smooth term for propensity score outperform stratification, even when the regression assumption of additivity is violated. |
