colloquium

We propose a semiparametric framework for causal inference with right-censored survival outcomes and many weak invalid instruments, motivated by Mendelian randomization in biobank studies. Under a structural accelerated failure time model, we construct an augmented inverse probability of censoring weighted moment condition that uses both censored and uncensored observations. A heteroscedasticity condition on the treatment model ensures point identification of the causal effect despite censoring and instrument invalidity. For estimation and inference, we develop GEL-NOW (Generalized Empirical Likelihood with Non-Neyman Orthogonal and Weak moments). We use deep neural networks to estimate a diverging number of nuisance functions that enter the moments in a Neyman-orthogonal manner. A central challenge is that the conditional censoring distribution enters nonorthogonally, so its estimation contributes to the first-order asymptotic distribution of the causal effect estimator. We derive this distribution and explicitly account for the additional uncertainty in the asymptotic variance. We also develop a censoring-adjusted overidentification test that incorporates this variance component. Simulation studies and applications to UK Biobank data demonstrate the robustness and practical utility of the proposed framework.