Estimating the Observability of an Outcome from an Electronic Health Records Dataset Using External Data.

Abstract

One of the key limitations of electronic health records (EHR) data is that not all health care encounters are observed. The degree to which patient information is captured is referred to as observability. Poor observability, and in particular differential observability, can lead to biased estimates and inference. As such, understanding the degree of observability is important in EHR based studies. In this study, we propose using external data with known observability to assess the degree of overall observability in EHR. We also construct a test for differential observability in the target EHR dataset. Using principles from the transportability literature, we show that we can use a balancing score based weight to estimate the observability of our target outcome. We conduct a series of simulation experiments to understand the conditions under which dataset features must be required to obtain proper inference. To illustrate this, we consider hospital readmissions among patients with end stage renal disease as our outcome of interest. We use administrative claims data, where the outcome is fully observed, as our external data.