Background: Electronic health records (EHRs) can provide valuable data for outcomes research. However, unlike administrative claims databases, EHRs lack eligibility tables or a standard way to define the benefit coverage period, which could lead to underreporting of healthcare utilization or outcomes, and could result in surveillance bias. We tested the effect of using a proxy eligibility period (eligibility proxy) when estimating a range of health resource utilization and outcomes parameters under varying degrees of missing encounter data.
Methods: We applied an eligibility proxy to create a benchmark cohort of chronic obstructive pulmonary disease (COPD) patients with 12 months of follow-up, with the assumption of no missing encounter data. The benchmark cohort provided parameter estimates for comparison with 9,000 simulated datasets representing 10-90% of COPD patients (by 10th percentiles) with between 1 and 11 months of continuous missing data. Two analyses, one for datasets using an eligibility proxy and one for those without an eligibility proxy, were performed on the 9,000 datasets to assess estimator performance under increasing levels of missing data. Estimates for each study variable were compared with those from the benchmark dataset, and performance was evaluated using bias, percentage change, and root-mean-square error.
Results: The benchmark dataset contained 6,717 COPD patients, whereas the simulated datasets where the eligibility proxy was applied had between 671 and 6,045 patients depending on the percentage of missing data. Parameter estimates had better performance when an eligibility proxy based on the first and last month of observed activity was applied. This finding was consistent across a range of variables representing patient comorbidities, symptoms, outcomes, health resource utilization, and medications, regardless of the measures of performance used. Without the eligibility proxy, all evaluated parameters were consistently underestimated.
Conclusion: In a large COPD patient population, this study demonstrated that applying an eligibility proxy to EHR data based on the earliest and latest months of recorded activity minimized the impact of missing data in outcomes research and improved the accuracy of parameter estimates by reducing surveillance bias. This approach may address the problem of missing data in a wide range of EHR outcomes studies.