We report on differences in sensitivity and false-positive rate across five methods of global normalization using resting-state fMRI data embedded with simulated activation. These methods were grand mean session scaling, proportional scaling, ANCOVA, a masking method, and an orthogonalization method. We found that global normalization by proportional scaling and ANCOVA decreased the sensitivity of the statistical analysis and induced artifactual deactivation even when the correlation between the global signal and the experimental paradigm was relatively low. The masking method and the orthogonalization method performed better from this perspective but are both restricted to certain experimental conditions. Based on the results of these simulations, we offer practical guidelines for the choice of global normalization method least likely to bias the experimental results.