Identifiable Estimation of Causal Concept Effects under Visual Latent Confounding

Estimating the causal effect of human-interpretable visual concepts on outcomes<br/>is essential for auditing classifiers and assessing bias in image datasets. However, existing estimators typically assume unconfoundedness, a condition rarely met in practice, as concept annotations are seldom exhaustive. We formalize the prob-<br/>lem of visual latent confounding, where unannotated factors manifest as high-dimensional visual signatures that jointly influence observed concepts and out-comes. We present UnCoVAEr (Unobserved Confounding Variational AutoEn-coder), a latent-variable model that learns identifiable confounder representations<br/>from images. By leveraging observed concepts and outcomes as auxiliary vari-<br/>ables, we prove that UnCoVAEr identifies representations sufficient for backdoor adjustment under standard assumptions. Empirically, UnCoVAEr achieves lower<br/>causal concept effect estimation bias on MorphoMNIST and CelebA benchmarks, outperforming feature-adjustment, counterfactual, and latent-variable baselines