The Effect of Residual Confounding in the Analysis of Large Observational Data Sets

We review how residual confounding in large observational data sets can mistakenly lead researchers to declare a predictor variable statistically significant. As sample sizes become very large, standard errors decrease dramatically, causing unmeasured variables correlated with the predictor of interest to overwhelm the model and yield precise estimates of the wrong thing, the residual confounder. Three examples from environmental epidemiology are provided. Specification Curve Analysis, SCA, and Local Control Analysis, LCA, are offered to reveal model selection bias and true null effects.