This paper addresses the issue of selection bias in recommender systems, where user feedback data often exhibit missing not at random (MNAR) patterns. To tackle this, the authors propose a Doubly Calibrated Estimator (DCE) that improves upon existing doubly robust (DR) estimators by calibrating both the imputation and propensity models. The DCE introduces calibration experts that consider different logit distributions across users, ensuring more accurate and reliable estimates. The authors also develop a tri-level joint learning framework that simultaneously optimizes these calibration experts alongside prediction and imputation models. Extensive experiments on real-world datasets demonstrate the effectiveness of the DCE in reducing bias and variance, leading to better unbiased recommendation performance compared to existing DR estimators. The proposed method is orthogonal to existing DR estimators and can be seamlessly integrated with them, offering a significant improvement in debiased recommendation tasks.This paper addresses the issue of selection bias in recommender systems, where user feedback data often exhibit missing not at random (MNAR) patterns. To tackle this, the authors propose a Doubly Calibrated Estimator (DCE) that improves upon existing doubly robust (DR) estimators by calibrating both the imputation and propensity models. The DCE introduces calibration experts that consider different logit distributions across users, ensuring more accurate and reliable estimates. The authors also develop a tri-level joint learning framework that simultaneously optimizes these calibration experts alongside prediction and imputation models. Extensive experiments on real-world datasets demonstrate the effectiveness of the DCE in reducing bias and variance, leading to better unbiased recommendation performance compared to existing DR estimators. The proposed method is orthogonal to existing DR estimators and can be seamlessly integrated with them, offering a significant improvement in debiased recommendation tasks.