Computing the Shapley value exactly requires evaluating the characteristic function v(·) over all 2^n possible coalitions, which becomes computationally intractable as the number of players n grows beyond 20-25. This paper addresses the question of howto draw coalitions efficiently when exact computation is infeasible. We first establish a unified framework for Monte Carlo Shapley estimators based on a level-stratified representation of the Shapley formula, and show that fixing the coalition-size levelbefore sampling strictly dominates drawing it at random. We then propose a novel Hybrid Exact/Sampling estimator that exploits the binomial structure of coalition counts: levels with few coalitions are enumerated exactly at zero variance cost, while only the large middle levels are sampled. The hybrid estimator is provably unbiased with strictly lower variance than any purely stochastic method at equal cost. We characterise precisely when this advantage is largest: the hybrid dominateswhen the characteristic function is pseudo-continuous and individual marginal contri- butions are moderate relative to the range of v - conditions met by the broad class of functions used in inequality analysis, production theory, and machine learninginterpretability. A Monte Carlo experiment with n = 20 players conrms a 36-89% RMSE reduction relative to the state-of-the-art permutation sampler. A Neyman optimal-allocation variant is also derived, which further reduces variance when within-level dispersion is heterogeneous and the budget is large.