Figure 3:

Evaluation with HCP data with 688 contrasts of sample size 10, comparing accuracy of Negative Binomial (NB) and Delaporte (DEL) distributions for the prediction of the number of contrasts with no significant results (zero foci) based on only significant results (one or more foci). Left panel shows results for voxelwise inference, right for clusterwise inference, both using P_FWE=0.01 to increase frequency of zero foci. For clusterwise datasets, the Negative Binomial confidence intervals always include the observed zero count, while Delaporte ofter underestimates the count. For voxelwise analysis, the Negative Binomial over-estimates the zero frequency substantially, while Delaporte’s intervals include the actual zero frequency in 3 out of 5 splits.