Hi Jim,

There's a pretty straightforward way of pre-processing, which entails the following steps:

1) Interpolate blinks from the signal. These are characterised by rapid declines towards 0 at blink onset, and rapid rises from 0 back to a regular value at blink offset. For a blink removal algorithm, see Sebastiaan Mathot's approach (link attached), which is a good way of doing it.

2) Reject artifacts, e.g. by Hampel filtering. The link to a Matlab function is attached.

3) Optionally smooth the data (depending on your parameters, the Hampel filter might actually act as a smoothing function too). A popular approach is to use a moving window, e.g. a Hanning window (see attached Wikipedia link for clarification).

4) Divide the signal (e.g. timepoint 0 ms to timepoint 3000 ms) by the median pupil size during a baseline period (e.g. timepoint -200 ms to timepoint 0 ms). This is an important step, as most trackers tend to work with arbitrary numbers, whereas most papers report changes as a proportional change.

N.B. Please note that some trackers report the pupil AREA, whereas others report the pupil DIAMETER. The proportional increase over time of the AREA will be a different function than the proportional increase over time of the DIAMETER, because the two are not linearly related!

EDIT: Additionally, make sure that no eye movements happened in the intervals from which you collected pupil data. And be sure that your stimuli are equiluminant. This is important, as systematic changes in luminance between conditions will result in a systematic difference in pupil size between conditions because of the pupilary light response. For a recent paper using pupilometry, see our recent publication in Journal of Vision (attached). The design of our experiment is a good example of how to capitalise on the pupilary light response to test spatial attention. It's also a good example of an equiluminant experimental paradigm.

files.figshare.com/1031648/Mathot___A_simple_way_to_reconstruct_pupil_size_during_blinks.pdf

http://uk.mathworks.com/matlabcentral/fileexchange/34795-outlier-detection-and-removal--hampel-

http://en.wikipedia.org/wiki/Window_function

Hi Edwin, thanks very much for your detailed response, what an excellent answer! You mention ensuring there are no eye movements during the period you are measuring the pupil size, why is this (assuming luminance at the eye is constant)? Is it the added noise that the movements might result in?

Thanks again.

Because the eye is a sphere, (especially the horizontal) pupil diameter and pupil surface will be different when looking at the pupil directly, compared to looking at it from an angle (which is what eye trackers do). Therefore, your pupil size readings will be influenced purely by eye position (and thus by eye movements, which are changes in eye position).

Additionally, pupil size is actually influenced by eye movement preparation, especially if the to-be-fixated location is of a different luminance than the currently fixated one, as you can read in attached paper (links to published version and freely accessible pre-print PDF attached).

http://psycnet.apa.org/journals/xhp/41/1/28/

https://peerj.com/preprints/238v1.pdf

Hi Edwin, do you have any references describing the typical error in pupil size produced by changes in eye position? Do you know if anyone have attempted to correct for this?

Hi Thomas,

It's quite a theoretical point, and it should heavily depend on your setup (distance between camera and eye, angle between screen-participant line and camera-participant line, slope of participant's face etc.). You could do a little test by having participants do a smooth pursuit of a moving stimulus, and correlating stimulus eccentricity (as a proxy for the eye's angle in the tracker's image) with pupil size.

Additionally, the point of regard will have an influence on pupil size by moderating the amount of light from the screen that actually reaches the eye (this will be less when looking into the top left corner of the screen than when looking at the centre). So I think it might be difficult to tease apart all contributing factors to the effect of point of regard on pupil size, and it will therefore be very hard to correct for it. I don't know of any post-hoc attempt.

Obviously, you can control for it with a well randomized design (that is: testing different conditions with an equal set of on-screen stimulus locations). Or you can control for it by using a single stimulus location for all conditions.

@Jim

You mention you already have data from a peripheral detection task. Did you test people when they were fixating centrally, by showing them stimuli in their peripheral vision? In this is the case, you needn't worry about potential effects of eye position on pupil size. Any differences you find in pupil size will likely reflect an exogenous shift in attention to the peripheral stimulus (see the paper I attached to my previous answer). The extend to which pupil size is modulated on each trial could serve as a measure for how well a participant attended to the stimulus (but do keep in mind that single-trial pupil data is noisy!).

This is supported by the idea that noradrenaline activity modulates pupil size (see the second paragraph of the Introduction of the attached Frontiers paper), and has been said to modulate sustained attention (see Neuropsychologia paper attached). Therefore, fluctuations in pupil size over time (provided constant luminance) might reflect sustained attention for a task.

Cheers,

Edwin

http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3183372/

http://www.sciencedirect.com/science/article/pii/S0028393208004090

Hi Edwin,

Thanks again for the info. I perhaps naively thought that the eye-tracking software and algorithms may have somehow accounted for changes in measured pupil size caused by changes in eye position, but this may have been wishful thinking! It sounds like the various factors contributing to changes in measured pupil size when the eyes move around would make this difficult to achieve.

In terms of the data we gathered, we actually had two conditions of fixation target - one static and one that was moving randomly. Participants had to keep fixating the target whilst periodically indicating if they saw another object in their periphery. It sounds like it will be safest to just use data from the static fixation target. The pupil size data is a supplementary question to the main focus of the experiment, but it's one that I'm very interested in so would be nice if the data was reliable enough to say something useful!

Thanks,

Jim

Hi Jim,

No worries, glad I could help! The static fixation might indeed be the safest bet. Let me know if you get stuck with the analysis, I'd be happy to help.

Cheers,

Edwin

Hi Edwin,

why would you interpolate blinks from the signal?

Thanks,

Ashwaq

A bit late to the party, but I also asked a similar question recently. I stumbled upon some code on the inter-web. Under "Software for Distribution" there is some matlab code called Pupillary waveform deblinking Matlab code by Greg Siegle at Pitt. Cleans up blinks quite nicely and you can play around with thresholds as an input argument. I found it very useful and super easy to use. Link attached. 

Hope this helps!

Andrew 

http://www.pitt.edu/~gsiegle/

Hi Andrew, thanks for the useful link, I will take a look!

Cheers,

Jim

hello Jim. can you give me your dataset link?

hello Jim, there are, I will send you the target link