Can fitts from a condition-level CSPLIN GLM be used for trial-wise MVPA?

AFNI Message Board
1

AFNI version info (afni -ver):Precompiled binary linux_ubuntu_16_64: Feb 24 2026 (Version AFNI_26.0.09 'Pupienus Maximus')

Dear AFNI experts,

I have a question regarding the interpretation and possible use of the fitts output from 3dDeconvolve.

My experiment is a slow event-related design (TR = 0.5 s). Each trial consists of a stimulus presentation followed by an imagery period. I modeled the data using condition-level regressors with CSPLIN basis functions, for example:

-stim_times 1 perception_face.1D 'CSPLIN(0,12,25)'
-stim_times 2 perception_high.1D 'CSPLIN(0,12,25)'
...
-stim_times 8 imagery_nonface.1D 'CSPLIN(0,12,25)'

Thus, each condition has one set of CSPLIN parameters, and the beta estimates are condition-level rather than trial-level.

My question is:

Would it be reasonable to extract the fitted response from fitts at a specific TR/time point after each trial onset and use these trial-wise patterns as input for multivariate pattern analysis?

For example, if I extract the voxel pattern around the expected peak response (e.g., 4–6 s after each trial onset) from the fitts dataset, would this represent a valid trial-wise neural pattern, or would it simply reflect the condition-level model prediction repeated across trials?

My concern is that because the original GLM does not estimate separate beta weights for individual trials, the extracted trial-wise patterns from fitts may not contain trial-specific variability.

Thank you very much for your advice.

2

if I extract the voxel pattern around the expected peak response (e.g., 4–6 s after each trial onset) from the fitts dataset, would this represent a valid trial-wise neural pattern, or would it simply reflect the condition-level model prediction repeated across trials?

My concern is that because the original GLM does not estimate separate beta weights for individual trials, the extracted trial-wise patterns from fitts may not contain trial-specific variability.

Your concern is spot on. In a condition-level model, the implicit assumption is that all trials within a condition share the exact same response (in both shape/profile and magnitude). Consequently, the fitted time series can only reflect that assumption, suppressing any trial-by-trial variability.

Here are a few alternative approaches currently available to estimate trial-level BOLD responses (ordered from simpler/cruder to more refined):

  1. 3dLSS in AFNI: Employs Least Squares Separate (LSS) modeling with a fixed canonical HRF.

  2. GLMsingle: Fits trial-level responses by selecting an optimal HRF from a library of 20 candidate functions.

  3. Profile estimation via regularization: Estimates flexible response shapes using basis functions (e.g., TENT, CSPLIN) combined with regularization. See this discussion paper for details.

Gang Chen