The magnitude and direction of collider bias for binary variables (Q2192289)

Let \(Y\) be some dependent variable in a statistical model and let \(X\) be some explanatory vector. If \(Z\) is some variable influenced by both \(Y\) and \(X\), then the bias that occurs in estimating the effect of \(X\) by including \(Z\) in some model \(Y = X + Z\) is called collider bias. That is, collider bias occurs when the effect of a treatment variable \(X\) is overestimated or underestimated when we attempt to control for some variable \(Z\) which is actually the cause of both \(X\) and \(Y\). Collider bias is common in blind application of statistical models. It's easy to come up with ad-hoc numerical examples and hypothetical experimental situations where collider bias can occur in either direction. The present paper by Nguyen, Dafoe, and Ogburn take this to the systematic level and explicitly derive formulas for collider bias in the case of two binary variables (contingency tables) that are marginally independent with conditioning on collider variable. They give explicit formulas for bias on the covariance, risk difference, and in some cases the odds ratio for various causal relations among the variables. I believe this to be an important contribution to the statistical literature. It gives a quantitative force against bias in linear modelling, and it can be used to estimate the bias that can occur in erroneous models.