We present new evidence on the randomization methods used in existing experiments, and new simulations comparing these methods. We find that many papers do not describe the randomization in detail, implying that better reporting is needed. Our simulations suggest that in samples of 300 or more, the different methods perform similarly. However, for very persistent outcome variables, and in smaller samples, pair-wise matching and stratification perform best and appear to dominate the rerandomization methods commonly used in practice. The simulations also point to specific recommendations for which variables to balance on, and for which controls to include in the ex post analysis.