Regresar

Within- and between-group regression for improving the robustness of causal claims in cross-sectional analysis

Abstract:

Background: A major objective of environmental epidemiology is to elucidate exposure-health outcome associations. To increase the variance of observed exposure concentrations, researchers recruit individuals from different geographic areas. The common analytical approach uses multilevel analysis to estimate individual-level associations adjusted for individual and area covariates. However, in cross-sectional data this approach does not differentiate between residual confounding at the individual level and at the area level. An approach allowing researchers to distinguish between within-group effects and between-group effects would improve the robustness of causal claims. Methods: We applied an extended multilevel approach to a large cross-sectional study aimed to elucidate the hypothesized link between drinking water pollution from perfluoroctanoic acid (PFOA) and plasma levels of C-reactive protein (CRP) or lymphocyte counts. Using within- and between-group regression of the individual PFOA serum concentrations, we partitioned the total effect into a within- and between-group effect by including the aggregated group average of the individual exposure concentrations as an additional pbkp_redictor variable. Results: For both biomarkers, we observed a strong overall association with PFOA blood levels. However, for lymphocyte counts the extended multilevel approach revealed the absence of a between-group effect, suggesting that most of the observed total effect was due to individual level confounding. In contrast, for CRP we found consistent between- and within-group effects, which corroborates the causal claim for the association between PFOA blood levels and CRP. Conclusion: Between- and within-group regression modelling augments cross-sectional analysis of epidemiological data by supporting the unmasking of non-causal associations arising from hidden confounding at different levels. In the application example presented in this paper, the approach suggested individual confounding as a probable explanation for the first observed association and strengthened the robustness of the causal claim for the second one.