MOSTLY HARMLESS SIMULATIONS? USING MONTE CARLO STUDIES FOR ESTIMATOR SELECTION∗