[Question] How to handle missing covariate values for fitting a model with overlap correction?

[catherinekago]

Description

Hi Unfold team,

I’m new to the Unfold toolbox and looking for guidance on the best conceptual and practical approach in Unfold to handle missing covariate values within the events table for fitting a model with overlap correction .

Context

I model FixationOnset events and include saccade-derived covariates (e.g., Preceding Saccade Amplitude, Preceding Saccade Angle) as spline / circular-spline effects. For some events these covariates are missing because the fixation was preceded by a blink, so there is no valid preceding saccade.

Current formula for FixationOnset

1 + Condition + spl(PrecedingSaccAmp, 5) + circspl(PrecedingSaccAngle, 5, 0.0, 360.0)

Additional event types (each with its own FIR basis)

• StimulusOnset — intercept-only FIR
• Response — intercept-only FIR
• BlinkOnset — intercept-only FIR
• SaccadeOnset — intercept-only FIR

Questions
What is the recommended way to handle these missing saccade covariates (NaN / missing) within a FixationOnset design?

• Drop FixationOnset rows where the covariates are missing?
• Split FixationOnset into AfterSaccade vs AfterBlink (and only include saccade terms for the former)?
• Move the saccade covariates to a SaccadeOnset regressor instead?

Or is there a way for spline terms to tolerate missing such that rows are retained but excluded only for that predictor?

Any advice is highly appreciated! Thanks!

[behinger]

Hey!

Let me see if I get it right. You have the following situation:

- - - S1 S1 S1 F1 - - - - B B B B BF - - - - S2 S2 F2 - - -

Here the --- would be fixation samples, S1/S2 the saccades samples, B the blink samples and BF I call a blink-fixation. Your question now is whether you should model the BF as a F, as something special, or not at all.

I honestly don't know enough about blinks, I would try to find out if there is a blink-off-response. We know the eyes are moving downward inward during a blink (Bell's Phenomenon), and we also know, they need to re-orient after a blink if they were fixating before. But I dont think we know if this elicits a specific brain response.

Unfortunately, there is no "easy" way out because the covariates are not missing at random at all. It is rather the answer is in the assumptions you want to make about what brain responses could exist or not.

Hope that helps already. I will transfer this issue to a discussion question, because I think it is a super valuable question that would get lost in the issues.
Best, Benedikt

[catherinekago]

Hi Benedikt,

Thanks for the quick reply, your summary matches my situation exactly. I see there’s a very similar discussion in #280, so I’ll move the follow-up questions there to keep things in one place.

Cheers,
Kathrin