Repository navigation
Volumetric mixed effects #342
Description
Activity
I am also struggeling with making brainstat work for volumetric input. I do not get the error shown above, but multiple problems keep me from getting valid results.
copes = # first level contrast parameter maps firstlvl_data = np.array([cope.get_fdata().flatten(order="F") for cope in copes]) # flattening with Fortran order, as this is specified in the function docs mask_data = img_mni_gm.get_fdata().flatten(order="F").astype(int) term_conf = FixedEffect(df_behav.memory_confidence) term_sub = MixedEffect(df_behav.subject) mixed_model = term_conf + term_sub slm_mixed = SLM( mixed_model, df_behav.memory_confidence, mask=mask_data, # skipping surf argument correction=["fdr", "rft"] ) slm_mixed.fit(firstlvl_data)-
as noted above, the argument terminology is highly confusing for volumetric data, even given the function help. The use of "surf" especially is unclear
-
using rft correction will throw an error that it into works for surface data (so I guess no cluster correction for volumetric data?)
ValueError: Random Field Theory corrections require a surface. -
during FDR correction, a shape mismatch error is thrown even though the input data and mask have the same format
shape of data: (42, 1082035) shape of mask: (1082035,)
--------------------------------------------------------------------------- ValueError Traceback (most recent call last) /home/tumnic/ahechler/ebits_notebooks/bold/2nd_level_nilearn.ipynb Cell 43 in () [4]mixed_model = term_conf + term_sub [6]slm_mixed = SLM( [7]mixed_model, [8]df_behav.memory_confidence, [9]mask=mask_data, [10]correction=["fdr"] ---> [13]slm_mixed.fit(firstlvl_data) File [~/.conda/envs/.../lib/python3.10/site-packages/brainstat/stats/SLM.py:158), in SLM.fit(self, Y) 156 self._unmask() 157 if self.correction is not None: --> 158 self.multiple_comparison_corrections(student_t_test) File [~/.conda/envs/.../lib/python3.10/site-packages/brainstat/stats/SLM.py:234), in SLM.multiple_comparison_corrections(self, student_t_test) 231 def multiple_comparison_corrections(self, student_t_test: bool) -> None: 232 "Performs multiple comparisons corrections. If a (one-sided) student-t 233 test was run, then make it two-tailed if requested." --> 234 P1, Q1 = self._run_multiple_comparisons() 236 if self.two_tailed and student_t_test: 237 self.t = -self.t ... 71 Q = np.ones((self.mask.shape[0])) ---> 72 Q[self.mask] = np.squeeze(Q_tmp[0, :]) 74 return Q ValueError: shape mismatch: value array of shape (126006,) could not be broadcast to indexing result of shape (1082035,)- Even if I leave out the correction, I get an array of t-values that includes two 0 values and every other value is NaN
slm_mixed.t[~np.isnan(slm_mixed.t)]
array([0., 0.])
Brainstat looks like an amazing tool to properly move from bash/FSL to python, especially since nilearn does not offer the mixed model approach. I am of course totally open to brainstat specializing on surface data, but currently it is not clear to me to which extent this (should) work.
-
Hello Experts!
Have there been any updates on using the slm_mixed on volumetric NIFTI data? I understand that both the 'rft' feature as well as all of the plotting seems to rely on having a surface that contains both vertices and faces. The foundational paper does say that volumetric analysis is possible but so far this capability is rather elusive.
For context, I am working with cortical thickness data in the Oasis-TRT template space (from ANTs / Mindboggle).
Best,
AndrewLike the previous posts, I also assume that BrainStat can perform volume analyses, since SurfStat was prepared for the task, and BrainStat is just awesome.
But, also like the others, I cannot get it to work. I cannot figure out how data must be input, since
slm.fitseems to expect surface-based data. I'd be very happy if someone who got it to run on volumes could pitch in.Can VBM data be analyzed in Brainstat, or is it limited to analyzing surface data? Currently, most studies focus on analyzing surface data.
Thanks for the reports, @danjgale @hechlera @ageoly-git @lconcha @Bella0709 — pulling all of these together since you're hitting variations of the same set of issues.
Short version: BrainStat's
SLMclass was written for surface data and its volumetric path is, in practice, only partially functional today. The companion paper describes volumetric capability (inherited from SurfStat), but the Python implementation of the multiple-comparisons machinery, the masking utilities and the plotting helpers all assume a 2Dsubjects × verticesarray on a surface mesh. There is no maintained, end-to-end volumetric tutorial.What works today (with caveats):
- Pass
surf=Noneand feed a 2D(n_subjects, n_voxels)array of voxel values intoSLM.fit. Use Fortran ordering (order="F") when flattening the NIfTIs, as the docstring states. - Coefficient/t/p maps that come out of the linear-model + t-test step are valid; you can
inverse_transformthem back into NIfTI space withNiftiMasker. - FDR correction will run if your mask is consistent — see (b) below.
Why each of you hit an error:
a.
TypeError: can't multiply sequence by non-int of type 'float'(@danjgale): thecontrastargument is being passed as a pandas Series of object/categorical dtype rather than a numeric column. Cast withnp.asarray(design['condition'], dtype=float)before passing it in.
b.shape mismatch: (126006,) could not be broadcast to (1082035,)(@hechlera): the FDR step computes Q over the masked data and then writes back into an array sized bymask.shape[0]. Ifmaskis a flat int array rather than a boolean array — and especially if it has been "applied" before being passed in — the indexing fails. Passmaskas a 1D boolean array of lengthn_voxels(same length as the columns ofY), and letSLMdo the masking.
c.ValueError: Random Field Theory corrections require a surface.(@hechlera): expected. RFT is implemented only for surfaces. For volumes, usecorrection="fdr"only.
d. All-NaNslm.t(@hechlera): symptom of the dtype problem above; once the contrast is numeric and the mask is the right shape, the t-map populates.Volumetric cluster-level inference (RFT/TFCE) is not supported. SurfStat had it for volumes; BrainStat's port does not. Today, the realistic options are:
- Run BrainStat for the GLM and FDR map.
- Use
nilearn.glm.second_levelif you want voxel-level cluster correction at the second level. - Use FSL
randomise/ SPM for permutation-based cluster correction.
Mixed effects with
MixedEffect(subject)does run on volumetric data once (a)–(d) above are sorted, since it's just additional variance components in the design matrix; it doesn't depend on the surface.I'll leave this issue open as a tracking item for a proper volumetric tutorial + a documented
SLMvolumetric mode — the right fix is probably a smallmode="volume"switch that disables RFT and ships with an example notebook. PRs welcome.- Pass
Partially resolved in #381 — every
SLM(...)invocation pattern that was posted in this thread now runs end-to-end on volumetric input.Concretely:
contrast: pandas Series / categorical inputs are coerced to float inSLM.__init__instead of producing the crypticTypeError: can't multiply sequence by non-int of type 'float'(or all-NaN t-values) downstream.mask: integer 0/1 masks (the natural shape ofnib.load(...).get_fdata().astype(int)) are coerced to bool, fixing theshape (126006,) could not be broadcast to indexing result of shape (1082035,)error in_fdr.- The "Random Field Theory corrections require a surface" error now explicitly tells you to use
correction='fdr'for volumes.
The ambitious half of this issue — cluster-level RFT / TFCE on volumes — is still not implemented. SurfStat had it; BrainStat's port doesn't, and a faithful re-port is multi-week work. Leaving this issue open as the tracker for that and labelling
enhancement / help wanted. PRs welcome.- added a commit that references this issue
on May 4, 2026
I am trying to run a mixed effects model using Brainstat with volumetric data (NIFTIs). I was under the impression that Brainstat can do analyses on volumes (as per the companion paper and the documentation landing page). However, I'm not really sure how to go about this, as the tutorials and documentation are surface-centric.
Is the statistics module strictly surface data after all? If not, then what's the way to go about using
SLMwith volumetric data?My code resembles this following basic example, with the
???indicating where I am confused with regards to volumetric input:The last line raises an internal type error if I set
surf=maskinSLM. If I take outsurfandmaskcompletely inSLM, I still get internal type errors. What am I missing in order for it to run on volumetric data, unless I am totally mistaken?Example of the type error I get regardless: