On the Accuracy of Influence Functions for Measuring Group Effects