The Impact of Variable Days of Coverage on Estimating Diagnosis and Treatment Prevalence From Administrative Claims Data

OBJECTIVES: Closed administrative claims data are a mainstay in HEOR, and commonly used to estimate diagnosis or treatment prevalence for the US population. This estimation is complicated by persons with incomplete pharmacy/medical coverage, and to what extent such persons with incomplete coverage can/should be included when estimating prevalence remains a question. Using a simulation-based approach, this research assesses the impact of variable days of coverage on prevalence estimation.

METHODS: Using closed payor claims from Inovalon from 2016-2022 (data cohort) we identified the subset of persons with complete enrollment each year (analysis population) and randomly sampled from this group (sample cohort). The sample cohort was then matched 1:1 with randomly selected persons from the data cohort and the observed (often incomplete) matched coverage patterns imposed on the sample cohort. The resulting analysis cohort contained persons with variable pseudo coverage each year representative of the data cohort, but with underlying known claims for the full year.

RESULTS: When diagnosis/treatment prevalence was moderate to high in the population (e.g. ICD10 I21*, Acute MI, ~400/100,000), including incomplete enrollment (≥180 days medical) resulted in under-estimation of true prevalence compared to including only complete enrollment (median error 4.4% vs. 0.002%), with similar estimate variability. When population prevalence was extremely low (C82.52, Diffuse follicle center lymphoma, <0.1/100,000), including incomplete enrollment resulted in reduced under-estimation (1.7% vs. 3.9%) and variability. Other treatments/diagnoses produced similar results.

CONCLUSIONS: For common diagnoses/treatments, including persons with incomplete enrollment results in under-estimation of prevalence. However, for extremely rare diagnoses/treatments including persons with incomplete enrollment may result in slightly more accurate prevalence estimation.