Nilearn: Error in glm.second_level.SecondLevelModel.compute_contrast

Summary of what happened:

I was fitting a simple second-level model and computing contrast to conduct a one sample t test following the tutorials. Instead of directly passing in a list of first level model objects, I passed in a pandas df with 3 columns subject_label , map_name and effects_map_path following the documentation here.

Here’s my code (contrast_df is the pandas dataframe I passed to the model):

design_matrix = pd.DataFrame([1] * len(contrast_df), columns=['intercept']) 
second_level_model = SecondLevelModel().fit(contrast_df, design_matrix=design_matrix)`
z_map = second_level_model.compute_contrast(output_type='z_score')

The fit command runs but the compute_contrast command throws an error:

line 467, in compute_contrast
File "/miniconda3/envs/popcorn/lib/python3.11/site-packages/nilearn/glm/second_level/",line 127, in _check_first_level_contrast
    if isinstance(second_level_input[0], FirstLevelModel):
  File "/miniconda3/envs/popcorn/lib/python3.11/site-packages/pandas/core/", line 3804, in __getitem__
    indexer = self.columns.get_loc(key)
  File "/miniconda3/envs/popcorn/lib/python3.11/site-packages/pandas/core/indexes/", line 3805, in get_loc
    raise KeyError(key) from err
KeyError: 0

Upon looking into the code, the issue seems to be that because my second_level_input is a pandas df, second_level_input[0] tries to index into a column in the df called ‘0’. Obviously that column does not exist. The same line of isinstance(second_level_input[0], FirstLevelModel) also appears in _check_second_level_input which is called in the fit command, and that runs fine for me. In _check_second_level_input the code checks for different cases of second_level_input, so I wonder if doing the same here in _check_first_level_contrast will get rid of the issue?

I am completely new to Nilearn, so any input is appreciated! Thanks.

Hi @jmu,
You need to specify a contrast_def variable that specifies your contrast. Example, if you want to find the difference between columns 1 and columns 2, while controlling for column three, it would be [1, -1, 0].

Perhaps working through tutorials (such as this one Intro to GLM Analysis: a single-session, single-subject fMRI dataset - Nilearn) would help you understand better.


Hi @Steven, thanks so much for your quick reply and for pointing me to the tutorial! In my case I simply wanted to test if column 1 of my design matrix predicts the fmri signal, so if I’m not mistaken, my contrast_def = [1,0,0,…0], where the length of the array = ncols of the design matrix? I previously specified this contrast in my first level model but did not add the argument first_level_contrast in my second_level_model.compute_contrast command. My apologies. But weirdly, now that I added this array as the first_level_contrast, it threw me an error:

File "/miniconda3/envs/popcorn/lib/python3.11/site-packages/nilearn/glm/", line 191, in Tcontrast
    raise ValueError("t contrasts should be length P=%d, "
ValueError: t contrasts should be length P=102, but this is length 16

16 is the number of columns in my first level design matrix and according to the docs of compute_contrast that should be the shape of my first_level_contrast argument? Please let me know what I missed here. Thank you!

Your contrasts for first and second level should not be the same.

First level is typically the subject-level, and includes volume-by-volume covariates such as motion parameters.

Second level is across-subjects analyses, where you might have covariates such as subject age, sex, group, etc.

Without knowing more about your code and data it is hard to say what the correct size of your contrast/design matrices should be.


Thank you for your reply Steven! Sorry to bother you again.

My first-level design matrix has 16 columns, with 1 regressor of interest, 14 nuisance motion regressors and 1 intercept column. To test if the regressor of interest (the first column of the design matrix) correlates with the fmri signal, I used this contrast definition for the first level, which yields an array of (16,) with 1 at the zeroth index and 0 everywhere else:

contrast_matrix = np.eye(first_level_design_matrix.shape[1])[0]

At the second level I have 14 subjects and I’m looking to perform a one-sample t test to test a group effect. Following the tutorial, this is my second level design matrix:

design_matrix = pd.DataFrame([1] * n_subjects, columns=['intercept'])

I did not specify the second_level_contrast argument in SecondLevelModel.compute_contrast, because the docs said

The default (None) is accepted if the design matrix has a single column, in which case the only possible contrast array((1)) is applied

I guess I might have misunderstood the documentation. I thought the second_level_contrast argument in SecondLevelModel.compute_contrast is to specify the second level contrast definition, while the first_level_contrast argument in SecondLevelModel.compute_contrast is identical to the first level contrast definition back when I called compute_contrast on the first level model (ie your screenshot)?

Thank you so much for your patience.