Hypothesis
Regression-based variable-selection methods can pick the true health-relevant exposures out of hundreds of correlated ones in an exposome analysis.
What failed
In simulations with 237 realistically correlated exposures and 0 to 25 true predictors, elastic net and sparse partial least squares averaged 76% sensitivity with a 44% false discovery proportion, GUESS and the deletion/substitution/addition algorithm 81% with 34%, and the environment-wide association study approach reached an 86% false discovery proportion; performance fell considerably when correlation between covariates was high.
Broken assumption
Regression-based selection can separate causal exposures from correlated ones at exposome scale.
Why
The authors attribute the limits to correlation between exposures, which none of the methods resolved well enough to distinguish true predictors from their correlates.
From the published paper: Agier et al. Environmental Health Perspectives, 2016. ↗Public domain