Variance-Optimal Off-Policy Evaluation with Conjunct Effect Modeling
The Variance-Optimal-CEM (VOCEM) estimator improves off-policy evaluation by interpolating between two existing methods: OffCEM, which uses stable cluster-level weights, and Doubly Robust (DR) estimation. VOCEM selects an optimal interpolation coefficient to minimize variance. This results in an estimator whose variance is no larger than either endpoint, offering greater stability and lower Mean Squared Error across various settings.