Title: Handling missing data in propensity score analyses of electronic health record data
Authors: Elizabeth Williamson - London School of Hygiene and Tropical Medicine (United Kingdom) [presenting]
Abstract: Missing data are a ubiquitous problem in medical research. When the data being used to address the research question are not collected for the purpose of research, the extent of missing data tends to be greater. Further, the mechanisms by which data become missing in EHR are likely to be different from those operating in more traditional study designs. The focus will be on studies where the main aim is to address a question of causal inference, for example estimating the comparative effect of two drugs on a health outcome. Particularly in settings using routinely collected health data, such as electronic health record (EHR) data, propensity score analysis is frequently applied; this analysis framework raises additional issues related to missing data. The way in which missing data is handled in these studies can lead to imprecise estimates and/or bias. Various commonly used missing data methods will be discussed, and the plausibility of the assumptions underlying these methods will be explored in studies using data from electronic health records. It will demonstrate that some missing data methods commonly described as ad-hoc can produce valid statistical inference under specific assumptions which can sometimes be plausible in these settings.