BSplineTerm and continuous terms

[EmmanuelleKris]

Hello
I'm trying to use the data "face_saccades_opendata_fig10.set" in julia.
I epoch data and I define the design:

fstim = @formula(0 ~ 1)
fbutton= @formula(0 ~ 1)
fsacc= @formula(0 ~ 1 + spl(sac_amplitude,10))

bfsacc = firbasis((-0.2, 0.5), sfreq)
bfstim = firbasis((-0.2, 0.9), sfreq)
bfbutton= firbasis((-0.2, 0.9), sfreq)

bfDict = [ "stimonset" => (fstim, bfstim),
"buttonpress" => (fbutton, bfbutton),
"saccade" => (fsacc,bfsacc)]

but when I try to model I get the error
m = Unfold.fit( UnfoldModel, bfDict, eegevent, data_epochs)
ArgumentError: BSplineTerm only works with continuous terms (got sac_amplitude)

And I can't identify the problem. Do you have any suggestions, please?

[behinger]

Thanks for reporting, this helps us improve the usability of unfold.

Can you check typeof(evts.sac_amplitude)? I suspect you get an Any which Unfold/Statsmodels can only interpret as categorical.

evts.sac_amplitude = Float64.(evts.sac_amplitude) would convert it to numeric proper and fix that error, or using the second, more complex converter in the load-eeglab example (untested).

In any case, i will improve the error message to provide a more helpful error message in the future.

[EmmanuelleKris]

Thanks for your answer.

In fact :
typeof(eegevent.sac_amplitude)
Vector{Any} (alias for Array{Any, 1})

so, I convert it.
And now, I have another error:
m = Unfold.fit( UnfoldModel, bfDict, eegevent, data_epochs)
AssertionError: Your events are not of the same size as your last dimension of data
whereas

size(data_epochs)
(40, 1251, 7661)

julia> size(eegevent)
(7661, 14)_

My apologies for the avalanche of no doubt naive questions...

[behinger]

the reason is that you try to fit on epoched data instead of non epoched. Please provide the continuous EEG data

Think of unfold deconvolution as replacing the epoching and the averaging step

PS: This is actually super helpful, it allows me to design much better error messages :)

[EmmanuelleKris]

So, it is working with evts.sac_amplitude = Float64.(evts.sac_amplitude), thanks!

if it's not an existing variable, e.g. sac_amplitude for a stimulus event, do you recommend filling it with NaN, setting it to 0 or leaving it empty in EEGlab structure?

[behinger]

cool! I think if a variable has no value, then it should not be used in the model fit. What I would recommend is not to use the same model-formula for stimulus and fixation events. Most convenient is probably NaN, most correct is missing (in Julia). So I recommend putting it to NaN in matlab, as it will simplify the parsing to Float.