I ran a PLS analysis using two spm-derived betas per subject (which represent the cue and stim portions of a task) and three conditions, using the command line interface. In the results structure that I get back I have 2 LVs, the first of which has a p<.002 in results.perm_result.sprob, yet looking at the voxel saliencies, not a single voxel appears to be anywhere close to passing the bootstrap test. I'm not sure how this could be. To make sure I'm doing this right, here is how I'm getting the z-scores for my saliences:

z=abs(results.u(:,1)./results.boot_result.u_se(:,1));

This is my understanding from reading the 2004 NeuroImage paper, is it correct?

If so, then a max(z) returns 0.16. That means the most salient voxel is nowhere near the standard cutoff of 2.57. In comparing my results to the data shown in the 2004 paper, it appears that both my saliences are somewhat smaller than normal (max(abs(results.u(:,1)=0.0093) and my standard errer seems to be higher (mean(results.boot_result.u_se(:,1)=0.06). At first blush I'd say that I was just looking at a massive ammount of noise, however, doing a standard 2nd level test in SPM using a contrast extremely similar to the LV results in a large set of voxels which pass an FDR-corrected test with p<.01. So I think I must doing something wrong, either in the setting up of the datamat or in the anlysis of the results. Anyone have an guess as to what is going on?