Study Design
Immortal time considerations in linked data Amarnath Marthi* Amarnath Marthi Marthi Marthi Marthi Marthi Marthi Univeristy of North Carolina at Chapel Hill
Linking data sources, such as electronic health records (EHRs) to administrative claims, may address incomplete capture of exposures, covariates, and outcomes needed for robust epidemiological analyses. While offering advantages, linkages can introduce selection bias and misclassification. One less-studied threat to internal validity is immortal time bias resulting from linkage.
We discuss how immortal time can be introduced by misalignment of the start of follow-up relative to (1) the data source linkage period (the years of data being linked) and/or (2) the specific date when a person is linked based on the recording, in all data sources, of individual-level data elements required for linking (individual linkage date). We illustrate this bias by comparing 90-day mortality after claims-defined hip fracture in adults ≥65 from a random sample of Medicare fee-for-service (FFS) beneficiaries (random cohort) to FFS beneficiaries linked to University of North Carolina health system EHR (linked cohort).
In analyses ignoring the data source linkage period, 90-day mortality post hip fracture was 14.3% (95%CI 14.1-14.5) in the random cohort (N=138816) compared to 11.3% (95%CI 10.6-12.0) in the linked cohort (N=8600). After accounting for the data source linkage period (N=7865), this risk increased to 12.4% (95%CI 11.6-13.1). Among hip fractures that occur after the individual linkage date (N=6210), 90-day mortality rose to 14.6% (95%CI 13.7-15.4) in the linked cohort. Results were similar after using standardized mortality weighting to balance the linked and random cohorts on age, sex, and race. We show how linkage-related immortal time bias may cause these differences.
When designing, conducting and reviewing research in linked datasets, researchers should consider not only if the correct linkages were generated but also ensure the start of follow-up does not occur before the data source linkage period or the individual’s linkage date to avoid immortal time bias.
