Accounting for misclassification errors in observational studies using administrative databases

点击翻译按钮获取中文摘要

Health services administrative databases have been widely used to assess risks associated with diseases and their treatments. Although these administrative databases are excellent sources of ready-to-use information on large number of patients at minimal costs, they are prone to measurement errors. Measurement errors can be due to case or exposure misclassification and can lead to erroneous results if not addressed properly in the study design and/or statistical analyses. My research program includes three projects, each addressing one aspect of measurement error.Project 1. Administrative databases are often used to estimate the prevalence of a disease without accounting for misclassification errors inherit in these databases. I will use a Bayesian approach to develop and assess methods of combining information available from administrative databases as well as other sources of information such as the Canadian Community Health Survey, to estimate the prevalence of a disease accounting for misclassification errors inherit in these data. This information can be combined in several ways to define either one, two or three case-definitions that may be assumed to be either conditionally independent or a correlation function may be modeled and results based on these assumptions compared.Project 2. In studying medication-related adverse events from administrative databases, the risk time-window defined as the period of time during which a patient could be at risk of medication-related adverse events is not exactly known. This project aims to study the effect of varying the risk time window values on the assessment of medication-related adverse events using simulated data.Project 3. Missing data are a major problem in medical research and can lead to biased results if not addressed properly. Multiple imputation techniques have been proposed to estimate the missing information based on observed values. This project aims to assess whether multiple imputation techniques can be used to infer the cause of death from administrative databases when only the date of death is known to the investigator.