BoundsError when running coeftable on real data

[Alx-mor]

Description

Hello!

While extracting coefficients with coeftable(m) from the model with this formula (fLMM = @formula 0 ~ 1 + condition + (1 | subject)), I receive the following error: BoundsError: attempt to access 5-element Vector{String} at index [3-element BitVector].
This error does not occur with the simulated data from the tutorial, only with real data.

I've consulted the documentation and then verified that my data is:

• An Array Float 64 with shape channel, times, total epochs; namely, 3 channels, 601 times, 158 epochs
• Events contain ~150 entries across 2 subjects, each with 4 condition levels (fully crossed). The only issue here the number of epochs across subjects and conditions is not fully balanced, as I dropped noisy epochs
• In events dataframe, the subject and condition columns are both explicitly converted to categorical
• Model formula is fLMM = @formula 0 ~ 1 + condition + (1 | subject); 1-element Vector{StatsModels.FormulaTerm{StatsModels.InterceptTerm{false}, Tuple{StatsModels.MatrixTerm{Tuple{StatsModels.InterceptTerm{true}, StatsModels.CategoricalTerm{DummyCoding, Matrix{Float64}, 3}}}, RandomEffectsTerm}}}: �[38;2;144;202;249m1�[39m �[38;2;239;83;80m+�[39m condition �[38;2;239;83;80m+�[39m (�[38;2;144;202;249m1�[39m �[38;2;239;83;80m|�[39m subject) (I work in JupyterLab and it seems to be an issue with some julia outputs presentation)
• Type of condition is CategoricalVector{String7, UInt32, String7, CategoricalValue{String7, UInt32}, Union{}} (alias for CategoricalArray{String7, 1, UInt32, String7, CategoricalValue{String7, UInt32}, Union{}})
• Condition is a 4-element Vector{String7}: "a_cr", "b_cr", "b_ncr", "a_ncr"
• The model fits with no issues (m = fit(UnfoldModel, fLMM, events, data, times))

I also ruled out that formulas with only a fixed effect and only a random effect (fLMM = @formula 0 ~ 1 + condition and fLMM = @formula 0 ~ 1 + (1 | subject), respectively) produce no issue with the coeftable.

Also, I were able to extract coefficients using coef(m) function, and it had a size of 3x601x4 and a length of 7212. But while accessing random effects with ranef(m), I once again got BoundsError: attempt to access 5-element Vector{String} at index [3-element BitVector].

I will be grateful for any guidance.

[behinger]

thanks! I successfully reproduced the issue, it appears with any categorical effect with more than two levels.

I'll look into it

[behinger]

I just released a new version, this should be up in 10min or so

[Alx-mor]

Thank you! I updated and checked.

First, the error remained, and I decided to simulate the data to check whether the problem was on my side. I followed your tutorial on multi-subject simulation, with the only difference being my simulated data had 4 conditions, and everything went fine; the coeftable was built successfully. Only then did I notice the biggest difference between my data and the simulated: I have channels, as I reported earlier.

So, the model is built on data with the shape 3 channels, 601 times, 158 epochs. The simulated data only had times and epochs dimensions. When I averaged my data across channels, it finally worked. Maybe that was the case from the very beginning.

I am okay with averaging across channels, however I wonder if it is possible to work with unaveraged data, as was the case with Unfold.jl.

[behinger]

oh dear - I will check it. But the problem existed for sure. Maybe two bugs. Thanks for your patience, it was obviously a mistake to test the models with 2x2 cases only...

[behinger]

Mh. there must be another mistake, With version 0.1.2, I can run data with multiple channels and multi-level effects without error.

e.g.ranef(m)

[Alx-mor]

I restarted the kernel and it worked. Maybe it was a glitch on my side. Thank you very much for testing!

[behinger]

jippi :) thanks for reporting in

[behinger]

@all-contributors please add @Alx-mor for bug

[allcontributors]

@behinger

I've put up a pull request to add @Alx-mor! 🎉