Background: Analysts of epidemiologic data often contend with the problem of estimating the independent effects of many correlated exposures. General approaches include assessing each exposure separately, adjusting for some subset of other exposures, or assessing all exposures simultaneously in a single model such as semi-Bayes modeling. The optimal strategy remains uncertain, and it is unclear to what extent different reasonable approaches influence findings. We provide an empirical comparison of results from several modeling strategies.
Methods: In an occupational case-control study of lung cancer with 184 exposure substances, we implemented 6 modeling strategies to estimate odds ratios for each exposure-cancer association. These included one-exposure-at-a-time models with various confounder selection criteria (such as a priori selection or a change-in-the-estimate criterion) and semi-Bayes models, one version of which integrated information on previous evidence and chemical properties.
Results: While distributions of odds ratios were broadly similar across the 6 analytic strategies, there were some differences in point estimates and in substances manifesting statistically significant odds ratios, particularly between strategies with few or no occupational covariates and those with many. Semi-Bayes models produced fewer statistically significant odds ratios than other methods. A simple semi-Bayes model that shrank all the 184 estimates to a common mean yielded nearly identical results to one that integrated considerable prior information.
Conclusion: Different modeling strategies can lead to different results. Considering the conceptual and pragmatic difficulties of identifying confounders, these results suggest that it would be unwise to place uncritical reliance on any single strategy.