Indirect Versus Direct Adjustment; Limitation of Administrative Cohorts

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 Use of large administrative cohorts linked to national registries and environmental exposures has facilitated consistent population-based risk coefficients. However, despite their size advantage, the lack of person-level behavioural risk factors (e.g. smoking, diet) is an important limitation with the potential to bias risk estimates. In environmental epidemiology, indirect adjustment for unmeasured confounding has taken several forms. A recently proposed method by Shin et al. (2014) uses partitioned regression. This method does not attempt to estimate the missing risk factors directly from supplementary data, but rather estimates the association between the missing factors and the available factors contained in the survival model using a representative ancillary dataset. The advantage of this method is that adjustment is at the individual-level and can accommodate multiple missing risk factors simultaneously. This symposium presentation will, 1) describe the partitioned regression indirect adjustment methodology including modifications that incorporate time-varying exposure data and proportional weighting of datasets, 2) formally evaluate the method and representativeness of the ancillary dataset using Cox proportional hazard models, and 3) compare it to other indirect adjustment methods. As an example we apply the method to the relationship between fine particulate matter (PM2.5) and non-accidental mortality, but keep the discussion general to any exposure-disease outcome relationship. We use the 2001 Canadian Census Health and Environment Cohort (CanCHEC, N=2.4 million, 16-years follow-up) as our primary dataset, and the 2001 cycle of the Canadian Community Health Survey (CCHS, N=130,000) as the ancillary matching dataset. Our validation tests showed minimal adjustment bias (-1.2% to +2.3%), depending on the modifications applied, cause of death, and covariates in the model. Adjustment direction and magnitude were very similar (<0.5%) compared to equivalent models using a CCHS-mortality linked cohort. Discussion will focus on the generalizability of the validation tests and how to assess adjustment results using sensitivity tests.