Hi Joe, The problem is this: the input dataset is stored in single precision, not double. Internally PALM converts the test statistic (t in this case) to z-stat. There are many reasons for doing the transformation: to correct across contrasts and design matrices that have different degrees of freedom and to put all spatial statistics into a common scale (and here you indeed use TFCE). This happens regardless of whether the user chooses the option -zstat or not. If the user uses -zstat, the test statistic will be saved as z-stat, otherwise it will be saved as the original statistic, but internally it's always z-stat.
This particular input file is stored using single precision, and this is kept throughout. That is, PALM doesn't internally convert the inputs to double -- what if the user actually wants single precision to, e.g., save memory or allow for a large dataset? To further complicate matters in this case, here we are testing the mean, and some voxels (4853, precisely) have an average signal compared to their variance that is high to the point of going over the limit that would allow conversion to a z-stat without overflowing the single precision limit. Take the first voxel with this issue (voxel index 7): it has a test statistic t = -21.1164818 (all decimal places included). The df of the numerator is 1 (t-test) and of the denominator is 216. Then: >> palm_gtoz(single(-21.1164818), 1, 216) ans = -Inf But if it were double, we would find the correct result, that is not infinite: >> palm_gtoz(double(-21.1164818), 1, 216) ans = -15.536373937167403 So, to fix this, convert the input file to double precision and it should work. Another option is to edit the file palm_miscread.m, and add a new line right after line 304 containing: X.data = double(X.data); Hope this solves the issue! All the best, Anderson From: Joseph Orr <joseph....@tamu.edu> Date: Wednesday, July 25, 2018 at 10:28 To: "Winkler, Anderson (NIH/NIMH) [E]" <anderson.wink...@nih.gov> Cc: HCP Users <hcp-users@humanconnectome.org> Subject: Re: [HCP-Users] issues with palm tfce for HCP task func data Thanks Anderson - I tried seed set to 1 and 'twist' and both resulted in the same error. Joe -- Joseph M. Orr, Ph.D. Assistant Professor Department of Psychological and Brain Sciences Texas A&M Institute for Neuroscience Texas A&M University College Station, TX On Mon, Jul 23, 2018 at 9:55 PM, Winkler, Anderson (NIH/NIMH) [E] <anderson.wink...@nih.gov<mailto:anderson.wink...@nih.gov>> wrote: Hi Joseph, I downloaded your files earlier today but couldn’t yet have a look. It may be that the particular values in the permutation distribution caused the tail fitting to fail (simply an unlucky scenario). If that is the case, try running again with a different seed (there is an option for that, see among the advanced options). The default is 0, so try with a different number (e.g., 1, 2, ...). All the best, Anderson Sent from mobile. Please excuse brevity. ________________________________ From: Joseph Orr <joseph....@tamu.edu<mailto:joseph....@tamu.edu>> Sent: Monday, July 23, 2018 7:59:26 PM To: HCP Users Subject: Re: [HCP-Users] issues with palm tfce for HCP task func data I inserted a breakpoint in palm_competitive at line 104 right before the error (pasted below) occurred. In the first pass. dpv_tstat_fwep, tfce_tstat, and tfce_tstat_uncp are saved (now in the folder linked above), but when i continue, I get the NaN error. The vector dd is length n-permutation - 1, and has a single NaN in row 2499. The workspace at the time of the error is saved as post_tfce_uncorrp_workspace.mat in the linked folder. 100% [Design 1/1, Contrast 1/1, Shuffling 2500/2500, Modality 1/1] Elapsed time with permutations: ~ 4438.79 seconds. Computing p-values. Saving p-values (uncorrected, and corrected within modality and within contrast). Saving file: motor_RH_results_L_test_dpv_tstat_uncp 104 if any(isnan(dd)), Saving file: motor_RH_results_L_test_dpv_tstat_fwep Saving file: motor_RH_results_L_test_tfce_tstat Saving file: motor_RH_results_L_test_tfce_tstat_uncp K>> any(isnan(dd)) ans = logical 1 K>> sum(isnan(dd)) ans = 1 Error using palm_competitive (line 105) Data cannot be sorted. Check for NaNs that might be present, or precision issues that may cause over/underflow. If you are using "-approx tail", consider adding "-nouncorrected". Error in palm_datapval (line 67) [~,cdfG,distp] = palm_competitive(Gvals(:),'descend',true); Error in palm_saveall (line 303) Ptosave = palm_datapval(plm.Gtfce{y}{m}{c},plm.Gtfcemax{y}{m}{c},false); Error in palm_core (line 2489) palm_saveall(plm,opts); Error in palm (line 81) palm_core(varargin{:}); -- Joseph M. Orr, Ph.D. Assistant Professor Department of Psychological and Brain Sciences Texas A&M Institute for Neuroscience Texas A&M University College Station, TX On Mon, Jul 23, 2018 at 3:27 PM, Joseph Orr <joseph....@tamu.edu<mailto:joseph....@tamu.edu>> wrote: I'm having an issue running palm following the CIFTI example (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/PALM/Examples#Example_10:_Using_CIFTI_files<https://urldefense.proofpoint.com/v2/url?u=https-3A__fsl.fmrib.ox.ac.uk_fsl_fslwiki_PALM_Examples-23Example-5F10-3A-5FUsing-5FCIFTI-5Ffiles&d=DwMF-g&c=ODFT-G5SujMiGrKuoJJjVg&r=ZKy1VO33u0kvO-PqY1gpb9Ld-AGhtT8c9PAcpsEyp70&m=RaviqUi2wUI0iRi8KmuXK6qfq6njkV8BCJLflU-UzvM&s=5jdHn9UVkGwcsjlZ9THOxLwYV1FFETS2_XSGDwG-0N4&e=>). I've worked with PALM a lot, but this is the first time using CIFTI files. I am analyzing several functional tasks from 219 participants in the HCP. I'm running a one-sample t-test with two covariates. The subcortical statistics run fine, but I am having issues with the cortical surfaces. I was running 10000 permutations and after completing the permutations the job would stall for 10+ hours. When I tried running less (e.g., 500) permutations I get an error from palm_competitive about NaNs. The dpv fwe-corrected stats save (and look reasonable), so the NaN appears during the tfce stage. For the surface I'm using: HCP_S1200_GroupAvg_v1/S1200.R.midthickness_MSMAll.32k_fs_LR.surf.gii along with the mean surface vertex area of the 219 participants. Inspecting the data in wb_view I don't see anything unusual, and no Inf/NaN are indicated in wb_command -file-infomation. Here are the options I am using (I've tried toggling accel, nouncorrected, and logp in different combinations, but the results are always the same): -i lang_story-math_data_L.func.gi<http://lang_story-math_data_L.func.gi>i -d design_n219_mean.mat -t design_n219_mean.con -o lang_story-math_data_L_cort -T -tfce2D -s L_midthickness.surf.gii L_area.func.gii -logp -n 10000 -ise The relevant files for one example are available here: https://drive.google.com/drive/folders/1uZPQK6Gy2MtTuOjZwin0ZToxBSUN9tct?usp=sharing<https://urldefense.proofpoint.com/v2/url?u=https-3A__drive.google.com_drive_folders_1uZPQK6Gy2MtTuOjZwin0ZToxBSUN9tct-3Fusp-3Dsharing&d=DwMF-g&c=ODFT-G5SujMiGrKuoJJjVg&r=ZKy1VO33u0kvO-PqY1gpb9Ld-AGhtT8c9PAcpsEyp70&m=RaviqUi2wUI0iRi8KmuXK6qfq6njkV8BCJLflU-UzvM&s=q91SdqtP5f9zbWfpbdaTy3-OUhKsnArcuolWz3jUdC0&e=> I've checked everything that I can think of to troubleshoot the issue, but I've hit a wall, so I'd be grateful for any advise/insights. The last thing I'm trying is using the sphere surface, and I'll update if that works. Joe -- Joseph M. Orr, Ph.D. Assistant Professor Department of Psychological and Brain Sciences Texas A&M Institute for Neuroscience Texas A&M University College Station, TX _______________________________________________ HCP-Users mailing list HCP-Users@humanconnectome.org<mailto:HCP-Users@humanconnectome.org> http://lists.humanconnectome.org/mailman/listinfo/hcp-users _______________________________________________ HCP-Users mailing list HCP-Users@humanconnectome.org<mailto:HCP-Users@humanconnectome.org> http://lists.humanconnectome.org/mailman/listinfo/hcp-users _______________________________________________ HCP-Users mailing list HCP-Users@humanconnectome.org http://lists.humanconnectome.org/mailman/listinfo/hcp-users
