Why You Should Not Use the ITCV With Robust Standard Errors (and What to Do Instead)

Abstract

The Impact Threshold of a Confounding Variable (ITCV) is commonly interpreted as the minimum product of correlations an omitted variable would need to have with both the outcome and predictor of interest to overturn a study’s conclusions. We show formally and confirm through Monte Carlo simulations that this interpretation breaks down in regression models with heteroskedasticity- or cluster-robust standard errors. We recommend instead a bootstrap-based sensitivity analysis derived from the sensemakr framework, which maintains close-to-nominal coverage across all conditions examined. We provide Stata and R implementations through our bootmakr package, offer practical guidance, and demonstrate the approach by reproducing a published study.