All-zero regressor warning

[AndyJHKim]

Hi.

While I am running GLM single, I'm encountering a warning. For one subject only, I repeatedly get the following warning during FITTING TYPE-B MODEL (FITHRF) and CROSS-VALIDATING DIFFERENT NUMBERS OF REGRESSORS.

"Warning: One or more regressors are all zeros; we will estimate a 0 weight for those regressors."

I checked the design matrix carefully. Across the entire experiment (all runs concatenated), every column has at least one event, so there should not be any all-zero regressors. Run-wise, many regressors are zero (because in a lexical decision task experiment each word/pseudoword appears only once), but this is true for all subjects. Yet only this one subject triggers the warning.
(Please see my previous discussion for more experiment design detail, #195 (comment))

If you want, I can share my DESIGNINFO.npy with you via email.

Do you have any idea about this warning?

The below is a part of output.

Thanks in advance!

*** Saving design-related results to /path/to/DESIGNINFO.npy. ***
*** FITTING DIAGNOSTIC RUN-WISE FIR MODEL ***
*** Saving FIR results to /path/to/RUNWISEFIR.npy. ***

*** FITTING TYPE-A MODEL (ONOFF) ***

fitting model...
done.

preparing output...
done.

computing model fits...
done.

computing R^2...
done.

computing SNR...
done.

*** Saving results to /path/to/TYPEA_ONOFF.npy. ***

*** Setting brain R2 threshold to 0.10324589326880167 ***

*** FITTING TYPE-B MODEL (FITHRF) ***

Warning: One or more regressors are all zeros; we will estimate a 0 weight for those regressors.
Warning: One or more regressors are all zeros; we will estimate a 0 weight for those regressors.
Warning: One or more regressors are all zeros; we will estimate a 0 weight for those regressors.
...

*** Saving results to /path/to/TYPEB_FITHRF.npy. ***

*** DETERMINING GLMDENOISE REGRESSORS ***

*** CROSS-VALIDATING DIFFERENT NUMBERS OF REGRESSORS ***

Warning: One or more regressors are all zeros; we will estimate a 0 weight for those regressors.
Warning: One or more regressors are all zeros; we will estimate a 0 weight for those regressors.
Warning: One or more regressors are all zeros; we will estimate a 0 weight for those regressors.
...

*** FITTING TYPE-C MODEL (GLMDENOISE) ***

*** Saving results to /path/to/TYPEC_FITHRF_GLMDENOISE.npy. ***

*** FITTING TYPE-D MODEL (GLMDENOISE_RR) ***

*** Saving results to /path/to/TYPED_FITHRF_GLMDENOISE_RR.npy. ***

*** All model types done ***

*** nothing selected to return ***

[kendrickkay]

It is puzzling. I would need more information to diagnose, like the design matrix.

[AndyJHKim]

I uploaded the design matrix files to my GitHub repository:
https://github.com/AndyJHKim/LDT

For reference:

• sub-001 (and all other subjects) run without any warnings.
• sub-002 triggers the “all zero regressor” warning.

The design matrices are created by the same code for every subject.

If you have a moment to take a look at the design matrix file for sub-002, I would greatly appreciate it.
Thanks a million for your help!

[kendrickkay]

I'm having trouble reading the .npy file into MATLAB (using readNPY). Can you save as a .csv or different format?

…

On Nov 14, 2025, at 9:13 PM, AndyJHKim ***@***.***> wrote: I uploaded the design matrix files to my GitHub repository: https://github.com/AndyJHKim/LDT For reference: sub-001 (and all other subjects) run without any warnings. sub-002 triggers the “all zero regressor” warning. The design matrices are created by the same code for every subject. If you have a moment to take a look at the design matrix file for sub-002, I would greatly appreciate it. Thanks a million for your help! — Reply to this email directly, view it on GitHub <#196 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAU2DEX2X5VIDTCPFUC7Z3L342K6HAVCNFSM6AAAAACMB4K5UGVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTIOJXGQ4DENA>. You are receiving this because you commented.

[AndyJHKim]

I converted the .npy files to .mat format, and the conversion appears to have been successful.
https://github.com/AndyJHKim/LDT/tree/main/DESIGNINFO/mat

[kendrickkay]

I did a spot check. The first run of the subject with an error has an event onset 14 rows before the end of the design matrix. For a TR of 0.25s , this means the onset is 3.5 s before the end.... This may be related to the warning you are getting. Basically, the HRF (if it's a delayed/late one) might still be all 0 for that very last trial. In which case the regressor is all zero. The code should be warning about occurrences like this (see attached). Also, see the attached HRF library figure (where the x-axis is in seconds).

[kendrickkay]

Bottom line is, probably no big deal. Given how close the last trial is the end of the recorded run, there's really nothing you can do to rescue that trial. The code is simply estimating a beta of 0 for that one. Probably no big deal.

[AndyJHKim]

1. Hmm… that is strange, because we manually stopped each run immediately after the final trial. I know “immediately” is subjective, but I don’t think the delay was longer than 8 seconds. But anyhow, the data shows that all other end buffers are indeed not less than 8 seconds. Luckily, I also found that the last trial of the first run is a pseudoword. If it had been a real word, I probably would have discarded the entire subject, because I want to get every beta for each of real word.

2. When I checked the TYPEB_FITHRF, the numbers do not match. By this I mean, the array shape is betasmd: (60, 74, 58, 750) = (x, y, z, trials), and when I counted the number of zero betas, Zero betas: 15,830 / 193,140,000.
   If one whole trial has betas of 0, shouldn’t the total number of zero values be 60 × 74 × 58 x 1 = 257,520? Why is it only 15,830 instead of 257,520?

3. I also noticed that some end buffers are around 7.x seconds. For example, in the first subject (the one without the error), the 4th and 5th runs have end buffers of 7-point-something seconds. I’m assuming that “< 8 seconds” means values like 7.x are still acceptable.

[kendrickkay]

For #3, you'll have to examine the code directly to double check things, as the timing conventions are a bit tricky. As for #2, you'll need to do some focused inspections and troubleshooting to see what's going on. My theory was only a theory/guess. Note that the HRF thing is more complicated than it may seem. FITHRF fits 20 different HRFs; only some HRFs have a long delay, and those HRFs may or may not ultimately be chosen for a given voxel.

[AndyJHKim]

I see. I will double check with the codes and the outputs. Thanks a lot for the detailed explanation and for taking the time to look into my issue!! GLMsingle has been really useful for my project!