Not everything is a confounder, so do not condition on all the things
We spend this week thinking about ‘collider bias’ and its consequences.
Collider bias provides a particularly illuminating framework for understanding why all the things you were told not to do in your last statistics class were bad. Examples, include selection on the dependent variable, conditioning on post-treatment outcomes, and generally pointing statistical models at data whose provenance you are unsure of.
Collider bias is a particular issue for policy students as much of their data comes from organizations, who collected it for reasons unrelated but often unexpectedly relevant to the purpose you want it for. In particular, administrative data is usually created in response to some event, for example a birth, doctor’s appointment, accident, or arrest. Colliders will be particulary troublesome when we try to answer some questions with this kind of non-randomly sampled data.
WhatIf ch. 8
Hernan et al. 2004 ‘A Structural Approach to Selection Bias’ Epidemiology.
Elwert and Winship (2014) ‘Endogenous selection bias: The problem of conditioning on a collider variable’ Annual Review of Sociology.