We study the determinants of biases in subjective performance evaluations in an MTurk experiment to test the implications of a standard formal framework of rational subjective evaluations. In the experiment, subjects in the role of workers work on a real effort task. Subjects in the role of supervisors observe subsamples of the workers' output and assess their performance. We conduct 6 experimental treatments varying (i) whether workers' pay depends on the performance evaluation, (ii) whether supervisors are paid for the accuracy of their evaluations, and (iii) the precision of the information available to supervisors. In line with the predictions of the model of optimal evaluations we find that ratings are more lenient and less accurate when they determine bonus payments and that rewards for accuracy reduce leniency. When supervisors have access to more detailed performance information their ratings vary to a stronger extent with observed performance. In contrast to the model's prediction we do not find that more prosocial supervisors always provide more lenient ratings, but that they invest more time in the rating task and achieve a higher rating accuracy.