Modelling multi-level survival data in multi-center epidemiological studies

Abstract

S21: When the Answer is “Big(ger) Data” in Environmental Epidemiology: What are the Questions?, Room 315, Floor 3, August 26, 2019, 1:30 PM - 3:00 PM Aim: Multi-center collaborations provide the opportunity for harmonized epidemiological analyses in large data sets that increase statistical power to detect associations. We assessed methodological approaches for the analysis of multi-level survival data within the framework of the ELAPSE project that pooled data from nine European cohorts to assess the association of long term residential exposure to air pollutants with mortality. Methods: After harmonizing individual and area-level variables between cohorts, we applied Cox proportional hazard models adjusting for these covariates. The data presented two levels of clustering: one corresponding to the initial cohort and one to the residential neighborhood level. We assessed five approaches to account for the first level: 1) not accounting for the cohort or using 2) dummies for cohorts, 3) strata per cohort, 4) a frailty term for cohort identification, 5) a random intercept per cohort under a mixed Cox. We further assessed second level clustering by applying 1) a random intercept per neighborhood and 2) a variance correction method -recognizing that approaches have different interpretations. We finally assessed the frailty vs the stratified approach for cohort adjustment through 1,000 simulations under nine scenarios for varying amount of heterogeneity between cohorts and pollutants’ effects. Results: Our results support that effect estimates derived from any of the approaches to account for different initial cohorts are stable except absence of any adjustment, while adjustment for the neighborhood clustering increased the estimates’ standard errors. Simulations confirmed the identical results between the stratified and frailty model except in the case of large heterogeneity between the underlying hazards and the pollutant effects that resulted in a difference of 0.0001 in the pollutant coefficient between methods. Conclusions: It is important to account for across cohort heterogeneity although the specific approach may be less important. We encourage investigators to consider study-specific conditions and objectives.