Impact Of Look Back Period Length On Associations When Using Electronic Health Record Data For Epidemiologic Research

Introduction: Studies using electronic health record (EHR) data often have a limited number of years available for analysis. There is a trade off between the look back period length used to define baseline characteristics and follow up duration used to define outcomes. Objective: Quantify the impact of 6, 12, and 24 month look back periods on the association between diabetes (DM) and subsequent cardiovascular (CV) events using EHR data alone and in combination with Medicare claims. Methods: EHR data from an academic health system and a federally qualified health center from 2009-2014 were linked to Medicare claims data. Eligibility criteria were age ≥65 years, Durham County address, 24 months of continuous enrollment after first claim, EHR encounter in the 2011 index year, and no history of cardiovascular disease (CVD) in the 24 months prior to the index date (i.e., look back period). DM was defined using EHR ICD-9 codes, HbA1c ≥6.5%, or glucose lowering medication, and using claims based diagnosis codes or glucose lowering medication. The outcome was a major CV event (myocardial infarction, stroke, or cardiac procedure) defined by diagnosis or procedure codes. Hazard ratios (HR) compared time to the outcome between patients with and without DM. Results: In 5473 patients, mean age was 77 years, 67% were female and 28% were Black. The prevalence of DM using EHR data only increased with a longer look back period (6 months [19%]; 12 months [21%]; 24 months [23%]) but was less than the prevalence using all available data from EHRs and claims together (28%) (Table 1A). Shorter look back periods resulted in higher HRs (6 month HR=1.64) as compared to HRs from longer look back periods (24 month HR=1.41) using EHR data alone or all available data from the EHR and claims together (HR=1.43) (Table 1B). Conclusions: To avoid over estimating associations, studies of CVD using EHR data to identify baseline conditions may want to use 12-24 month look back periods in the absence of additional administrative data. This may also lead to a shorter follow-up period.