Questions regarding cohen's d calculation in PALM

[Yi-Hsuan Lin]

Hello experts,

I'm hoping you can help clarify an output from PALM for a group-level analysis.

I am running a group-level analysis on fMRI task data from the HCP S1200 dataset using PALM (version alpha119 in MATLAB). In the first level, I selected some ROIs from the MMP 1.0 and did the parcellation. At the group level, I entered a behavioral measure of interest into the design matrix to assess whether activity in these regions scaled with behavior. I also included three additional covariates of no interest and accounted for family structure using exchangeability blocks.

I used the -saveglm option and have been reviewing the outputs. For some significant parcels, I've noticed a large discrepancy between my manual calculation and the Cohen's d value that PALM generates directly.

For example, one parcel shows the following results:

• t-statistic = 1.95
• Cohen's d (from PALM's _cohen file) = 0.97

When I manually calculate the effect size from the cope and varcope files (d = cope / sqrt(varcope * N)), the d = 0.15 (with 170 participants).

My questions are: 

1. Is the _cohen output the preferred estimate of effect size in this context?
2. Could this discrepancy be expected due to model complexity (e.g., covariates, relatedness)?
3. Or might there be something I’ve misunderstood in how to analyze, interpret, or compare these values?

Any guidance would be incredibly helpful. Thank you very much for your time. 

Best, 

Yi-Hsuan

[Anderson Winkler]

Hi Yi-Hsuan,

 

The formula you used for manual computation only applies to one-sample t-tests, and even so as long as the regressor is coded as all-ones. If coded, say, as a column full of any other constant value, the result of the test (tstat, pvalue) would be the same, but that formula would also fail. The formula also fails if testing the intercept when there are covariates, as indicated above, if the regressor of interest is continuous.

 

Computing by hand requires rescaling the design (the COPE isn’t scale-free) so such that it matches the formulas Cohen proposed for some particular designs. But also note that Cohen didn’t propose using “d” as a measure of effect size with continuous variables. Although it is possible to find an equivalence (which is what PALM does) that is somewhat a departure from his intention, which was to use correlations as measures of effect size with continuous variables. You can get these with PALM with the option “-rstat”.

 

So, the 0.97 should in principle be correct, but perhaps using correlations will be more helpful or interpretable.

 

All the best,

 

Anderson

 

 

From: Yi-Hsuan Lin <elsie...@gmail.com>
Date: Tuesday, August 5, 2025 at 5:51 AM
To: HCP-Users <hcp-...@humanconnectome.org>
Subject: [hcp-users] Questions regarding cohen's d calculation in PALM

  External Mail

  This email originated outside of The University of Texas Rio Grande Valley.
  Please exercise caution when clicking on links or opening attachments.

--
You received this message because you are subscribed to the Google Groups "HCP-Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hcp-users+...@humanconnectome.org.
To view this discussion visit https://groups.google.com/a/humanconnectome.org/d/msgid/hcp-users/663b7554-67ea-4bef-81bb-7b35b2740941n%40humanconnectome.org.

[Yi-Hsuan Lin]

Hi Anderson, 

Thank you so much for the detailed explanation! 

It's really helpful.

Best, 

Yi-Hsuan

Anderson M. Winkler 在 2025年8月6日 星期三清晨6:16:52 [UTC+8] 的信中寫道：