Meta-analysis vs pooling

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 Large datasets, often combined across multiple cohorts, are increasingly important in environmental epidemiology to understand nonlinear dose response, effects at low exposure levels, and pollutant mixtures. One strategy is to combine cohorts in a pooled analysis. An alternative is cohort-specific analyses followed by meta-analysis. We discuss strengths and weaknesses of these two analytic approaches and explore implications in the ECHO Pathways consortium of three thousand children and a separate consortium of one million adults. Cohorts typically have distinct target populations, locations, and epochs. Meta-analysis automatically adjusts for cohort-stable characteristics, and it is straightforward to achieve the same effect in a pooled analysis with dummy variables. Harmonization presents a greater challenge, as confounders are often measured differently and/or missing across cohorts. Meta-analysis typically adjusts for the best confounder data available in each cohort. We describe a method to accomplish the same result in a pooled analysis with multiplicative interactions between non-harmonized variables and dummy variables for cohort. Additional insights come from restricting to confounders available in all cohorts and to cohorts with all key confounders. Pooled analysis may have greater precision if harmonized covariates are included without effect modification by cohort. A motivation for combining cohorts is to estimate nonlinear dose-response over a wider range of exposure levels than is available in a single cohort and to estimate effects of pollutant mixtures. In a pooled analysis, more flexible modeling approaches are available to identify these aspects of the health effect association than in a meta-analysis. Notably, this benefit persists if the pooled analysis includes adjustment for cohort and non-harmonized covariates. Meta-analysis and pooled analysis can adequately account for confounding and covariate harmonization in multi-cohort datasets. Advantages of pooled analysis include flexible estimation of nonlinear dose response including health effects at low exposure levels and estimation of health effects of pollutant mixtures.