In clinical safety and effectiveness research using secondary health databases, patient medical histories are typically assessed using fixed look-back approaches. Conventional applications of these approaches exclude patients who are not continuously enrolled in the database for the entire look-back period (e.g. one year), and data occurring outside this period is ignored. An alternate approach has been suggested which assesses all of the available data history, though concerns exist that results may be biased by systematic variation in the amount of available database across important study groups.
We used applied analyses as well as plasmode simulation methods to explore the application of short (1-year) and long (3-year) fixed look-backs and all-available data approaches in analyses of Medicare fee-for-service (FFS) claims data. We assessed the bias and efficiency of effect estimates when we used the different look-backs to 1) assess cohort eligibility and to 2) identify and adjust for confounders. In the applied analysis, we evaluated the effect of statin initiation (vs. non-use) on incidence of 1) cancer within six months (a negative control outcome we expected a priori to be null) and 2) all-cause mortality within two years. In the plasmode simulation, exposures (conceptually: statin initiation vs. non-initiation) and outcomes (conceptually: inpatient hospitalization) were simulated as a function of self-reported interview data obtained from the Medicare Current Beneficiary Survey (MCBS, which represented the true underlying confounder of exposure-outcome associations. We evaluated estimates after applying different look-back approaches in the linked claims data.
Compared to short fixed look-back approaches, all-available approaches selected cohorts with superior classification and produced less biased estimates. Compared to long fixed look-back approaches, all-available approaches selected more inclusive cohorts and produced more precise estimates. Though these studies were conducted in a fairly narrow (applied) setting, our findings provide real-world evidence that using all-available look-backs to classify patient histories is superior to fixed look-back approaches. Our findings provide context to investigators seeking to understand the mechanisms through which the different look-backs may produce different estimates.