|
Abstract:
|
Electronic Health Records (EHRs) are an increasingly common data source in many clinical analyses. While their size and convenience makes them very appealing for research purposes, there are some biases that may arise in their use. Many of these biases arise from the fact that people tend to be sicker when they interact with a health system. We term this bias "Informed Presence". While conceptually a missing data problem, the emphasis is on the fact that what we observe generates the bias. Using standard data from our institution's EHR we illustrate a number of biases - both simple and more complex - that may arise in the analysis of EHR data. These include biases from imperfect phenotying algorithms, location of service, patient referral status, and number of encounters. Some of the biases are easily accounted for, while others require more careful consideration.
|
