Controlling for Omitted Variable Bias in Deep Neural Networks
The paper argues that decorrelating predictions from confounders does not actually remove omitted-variable bias in neural nets.
The authors propose adding control-variable effects to deep models through a generalized additive setup, then refitting the final layer of a pre-trained network with cross-fitting and ridge penalization. They describe ways to orthogonalize effects and marginalize predictions over the covariate distribution, depending on the scientific or fairness target. In simulations, the method consistently estimates true effects while compared methods need more data or miss them. On confounded neuroimaging data, it restores performance close to a model trained without confounding. ArXiv · AI/CL/LG's note
The authors propose adding control-variable effects to deep models through a generalized additive setup, then refitting the final layer of a pre-trained network with cross-fitting and ridge penalization. They describe ways to orthogonalize effects and marginalize predictions over the covariate distribution, depending on the scientific or fairness target. In simulations, the method consistently estimates true effects while compared methods need more data or miss them. On confounded neuroimaging data, it restores performance close to a model trained without confounding. ArXiv · AI/CL/LG's note
score 5