Baselining eye-tracking pupillometric data: am I doing it right?

17 March 2020 3 5K Report

Hi everybody, I need some help with an analysis of pupillometric data; it’s the first time that I use pupillometry, so I hope I didn’t make too many mistakes or at least that they won’t jeopardize the whole analysis.

I ran a between-subjects experiment in which the participants watched the same visual stimulus in three different conditions; during the stimulus presentation I recorded their eye-tracking data. I'm very interested in pupillometry but here's my problem:

the software I use (iMotions) provides me with the aggregated and auto-scaled data for each of the three conditions: these data are apparently very clean and consistent (there has to be some kind of automated correction of blinks and artifacts).

The software output basically has two columns: timestamp (in milliseconds, identical in the three conditions) and pupil diameter (in cm, strangely enough, but never mind…)

I ran an ANOVA with the condition as the factor and the pupil diameter as the dependent. variable, F(2, 5141) = 119.38, p < .001 ηp2 = .044, (1-β) > .99. Bonferroni corrected post-hocs were all significant p < .001 (see graph 1 in attachment)

I got suspicious: the significance was too high and, above all, the three conditions do not start at the same point on the y axis (ycond1 = 0.47; ycond2 = 0.47; ycond3 = 0.50). I thought that maybe the significant difference could be due to this (let’s say the participants in Condition 3 had larger pupils for some reason); so, to baseline the data, I tried to let them start at the same point.

To do this, I rearranged the columns for them to show me not the pupil diameter but the pupil dilation; I organized them so that the new y value of, let’s say, x = 1 (time frame = 132) was the old y value (pupil diameter) minus the value of y with x = 0) (see attached screenshot) In the example, for condition 1; the new value for timeframe 132 is: 0,48 - 0,47 = 0,01.

I ran the same ANOVA and now the results appear to be more reliable, F(2, 5141) = 42.15, p < .001 ηp2 = .016, (1-β) > .99. There has been a drastic decrease in the F value and partial eta squared. Bonferroni corrected post-hoc analyses revealed that the condition 2 only was significantly different from the other ones (p < .001) (see graph 2).

…and now…question time!

Would you say that this procedure is right? I guess there could be many errors in it, but I’m not an expert and I didn’t manage to take great advantage from reading many papers on this matter.

Would you have trusted the ANOVA results at point 3? Or rather I was right to baseline those data?

To baseline the data, I acted freely and according to nothing but a rule of thumb that came into my mind. Would you suggest other processes?

Alexandre Marois

For your baseline, I would suggest to use a pre-stimulus average (e.g., on all the measures collected 200 ms before the stimulus) rather than subtracting the first value of the pupil at time 0. This value is not necessarily valid.

Concerning the analysis you did, I'm not totally sure to understand what you did. One way you could analyze this is by first averaging the pupil diameter across each trial for all participants to obtain a mean waveform for each subject. Then, on that averaged waveform, you should compute a relative pupillary dilation response to see how the stimuli you presented triggered different pupillary responses or not. To do so, you might need to average the pupil diameter (again, on the mean waveform) around the peak (maybe between 500 and 1000 or 1500 ms post-stimulus). You should then subtract the baseline of the waveform (as explained above) from that average measure of dilation, which will give you a relative pupillary dilation response. For each subject, you will then have a relative response (that you should convert in mm) representative of the type of trial they were exposed to. Then, you can perform a uni ANOVA with the type of trial as the independent variable and the relative pupillary dilation response (computed on the averaged waveform) for all participants as the dependent variable.

Of course, there are a lot of other techniques, but I think this one is quite easy to do and valid.

Alessandro Ansani

Hi Alexandre,

sorry for my late response and thanks very much for your answer. Since I didn't have pre-stimulus values, I finally managed to do it by choosing not the first value, but the average of the first 200 ms of the visual stimulus (in fact, it begins with a fade in from black, so it should be fine). I saw similar procedures in other papers so I'm quite confident now.

Af for your second paragraph, I guess the procedure you describe is absolutely the perfect one; unfortunately I can't do it because due to Covid my lab closed and now I just have aggregated data. It still remains unclear whether and how do these data take into account SD and individual peaks; I'll be talking with some of the software developers asap and I'll keep you posted on this.

Thanks again,

All the best,

Alessandro

How to quantitative analize the data of TNBS assay for gelatin?

SAR X-Band constellation from Capella Space?

Colab - how to extract NDVI time series from Sentinel-2?

The use of Orange, a high level tookit for Machine Learning, data mining and data science for scientific research or just teaching?

How to defend against the reviewer's "not enough trials" objection (non-parametric observed power issues)?

Any difference between "hydroxy naphtol blue", and "hydroxy naphtol blue , disodium salt" for LAMP?

Estimate Enterprise Value using specific accounting (items/lines) variables in regressions?

Effects of grammatical phrasing (2nd person singular vs. 1st person plural) on compliance to suggestions/prohibitions/prescriptions, etc?

How can particle fillers increase the elongation at break of polymers?

How to change features size without padding in CNN?

How to compare two groups with only two measurements?

How to manually calculate p value in ANOVA? what is the equation or formula for it?

If you haven't computer? how you show the photons to pupils?

Can impedance of antenna be less than 50 ohms?

Mixed ANOVA - help?

How to Interpret Partial Eta Squared Numbers For a Repeated Measures Anova?

Can we proceed to perform Two-way Mixed-ANOVA despite normality violations?

Sample size determination for non-parametric tests?

Do you think there is cognitive oversaturation in the current pandemic? What impact can this have? And other manifestations?

Three-way robust ANOVA in R?