Issues with event types

[EmmanuelleKris]

Hi everyone,

In our event structure (created with Matlab and imported into Julia), there are 6 types of event:

"Saccade"
"Fixation"
"Conflict" (=stimulus)
"Blink"
"SystemDecision"
"Validation"

For each event, the same columns :

For "Fixation" events, there are features of the saccade preceding the fixation (amplitude of the saccade,etc). However, for some fixations, the preceding saccade is not taken into account (typically if the fixation is preceded by a blink), so for this fixation, the ‘saccade amplitude’ is NaN.
For Conflict event, for example, the columns related to saccade and fixation are empty.

So, finally, in "sac_amplitude", across event types, we have three possibilities: value or NaN or [].

And my problems begin when I want to run an EFRP estimation model (i.e. on the “fixation” event) with splines on saccade amplitude, for example.

fstim= @formula(0 ~ 1 )
ffix= @formula(0 ~ 1+spl(sac_amplitude,5))

bfstim = firbasis((-0.2, 0.8), sfreq)
bffix = firbasis((-0.1, 0.8), sfreq)

bfDict = [ "Conflict" => (fstim, bfstim),
"Fixation" => (ffix, bffix)]

m = Unfold.fit(
UnfoldModel,
bfDict,
eegevent,
DataCh,
eventcolumn = "event",
)

• If I let empty arrays for other events and NaN for Fixation , I get an error message when modeling while converting the values to float

  eegevent.sac_amplitude=[float(eegevent.sac_amplitude[i]) for i = 1:size(eegevent.sac_amplitude,1)] (Float64.(eegevent.sac_amplitude) doesn't work because of empty arrays)
  : ERROR: ArgumentError: BSplineTerm only works with continuous terms (got sac_amplitude)
  => As you know, I'm a Julia novice, so I probably don't handle typing conversions very well, so maybe the problem simply is a conversion bug?

• If I fill "sac_amplitude" empty arrays for "Fixations" events by a value (for example averaged amplitude), I get

ArgumentError: quantiles are undefined in presence of NaNs or missing values
whereas sac_amplitude isn't empty or nan for "Fixations" event but it is still empty for other types of events for which this feature doesn't exist (such as "Conflict")

• If I fill all empty and NaN arrays (whatever event type) by a "fake" value (as zero), the modelisation works but this solution bothers me because the value 0, for example, is not so meaningless.

In fact, it seems to me that Unfold's functions initially consider the whole events structure, rather than just feature of interest depending on the event (ie sac_amplitude for Fixations in this example)?

What would you recommend for managing features whose values don't exist for certain events, please?

[behinger]

thanks for your report! I just did some testing to figure out what's going on. Some questions I think I can adress:

1. A continuous effect in evts.sac_amp should be of a Numeric type for modelling as a continuous effect. E.g. if you'd have typeof(evts.sac_amp) = Union{Nothing,Float64} you run into ArgumentError: BSplineTerm only works with continuous terms (got sac_amp_nothing). The reason is that we cannot re-cast a sub-dataframe because we cannot be sure what type exactly you want to have in each sub-dataframe. Thus, the "correct" way to bring forward the data with partially missing columns is to use missing or NaN.

2. If you have missing data in your covariates, then Unfold cannot possibly know what to do. As you indicate, if you replace NaN/Missing with 0., the model runs, but you assume that that "fixation" had an incoming saccade amplitude of 0.

The central question is: What value to put here? This is a problem I don't think there is a right answer to. The general term is "imputation", you could e.g. replace this with an average value of mean(sac_amplitude) - but I think the underlying problem you need to think about is what kind of relationship do you expect? Maybe those fixations simply need to be modelled as their own eventtype, e.g. blink_fixations. I have not looked into them in detail, but I think, especially with eyelink, those "blink-saccades", are an artefact, they often have also unrealistic large velocities, thus I just filter the dataframe to remove them.

What you could think is to simply model blink onset (or offset? or both?) as a dedicated event, that should cover the activity.

I hope my logic is clear!
Best, Bene

edit: I'm working on providing more helpful error messages in #279 but it is not merged yet

[EmmanuelleKris]

Thanks for your answer
I agree with you about blinks and the logic of naming fixation differently.

But this doesn't entirely solve my problem...
Unfold uses evts dataframe which contains all events with their type, their latency and other covariables. Some covariables may be specific to one type of event only, no?

In my example, in evts dataframe, excluding the special case of "blink" event:

for the rows concerning the “conflict” event, sac_amplitude column is empty since it isn't an ocular event...

When I try to model for different types of event (each event, a basis function) with splines on sac_amplitude for only one basis function, I'm stuck.
But maybe I'm missing a key point?

Best,

[behinger]

If I understand correctly, the problem is that your column is currently Union{Float64,Nothing} whereas it should be Union{Float64, missing} (or NaN). Can you try:

evts.sac_amp = something.(evts.sac_amp,missing)

which replace everything that is nothing with missing

[EmmanuelleKris]

More precisely:

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

with

typeof(eegevent.sac_amplitude[9])
Matrix{Union{}} (alias for Array{Union{}, 2})

[behinger]

I see - that is a very strange empty object you have there 🙈 this makes the conversion a little bit more complex:

_sac_amp = replace(x-> isempty(x) ? missing : x ,evts.sac_amp_emptyarray) # replace empty with missing
sac_amp = convert(Vector{Union{Missing,Float64}},_sac_amp) # explicitly cast type away from Any to something StatsModels identifies as continuous

[EmmanuelleKris]

It is working!!! Thank you so much!!! :D

[EmmanuelleKris]

Hello,
I am continuing this discussion because it still concerns the manipulation of events...
Let's imagine that we want to fit a model based on a group of events (for example, the first two fixations in the trials) that are initially named differently (“firstfix,” “secondfix”).
What would be the best way to do this, please:
Should we modify eegevent by giving these two events the same name, or is this an existing feature?

[behinger]

this is not an existing feature, so you would need to rename them indeed

e.g. something like

ix = eegevents.eventtype .∈ Ref(["firstfix","secondfix"])
eegevents.eventtype[ix] .= "firsttwofix"