Safe and sensible preprocessing and baseline correction of pupil-size data.

Mathôt S, Fabius J, Van Heusden E, Van der Stigchel S. · 2018

Read Safe and sensible preprocessing and baseline correction of pupil-size data. by Mathôt S, Fabius J, Van Heusden E, Van der Stigchel S. (2018) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc5809553.txt.

Introduction

Mathôt and co-authors give safe rules for cleaning pupil-size data. Pupil size tracks arousal and mental effort. Analysts often correct each trace against a baseline stretch. Done badly, that step can fake the effect it seeks.

Blinks and data loss can shrink baseline values to near zero. Dividing by such baselines warps the whole trace. Subtracting the baseline is far safer. The authors add five plain rules, from marking bad data to dropping dud trials. They also warn against effects that appear too fast. Readers get a short, firm guide to honest pupillometry.

Catalog record

CREATORMathôt S, Fabius J, Van Heusden E, Van der Stigchel S.
YEAR2018
COLLECTIONPAPERS
FORMATTXT
FILE SIZE51 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC5809553

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Safe and sensible preprocessing and baseline correction of pupil-size data By Sebastiaan Mathôt, Jasper Fabius, Elle Van Heusden, Stefan Van der Stigchel Published in: Behavior Research Methods DOI: 10.3758/s13428-017-1007-2 ABSTRACT Measurement of pupil size (pupillometry) has recently gained renewed interest from psychologists, but there is little agreement on how pupil-size data is best analyzed. Here we focus on one aspect of pupillometric analyses: baseline correction, i.e., analyzing changes in pupil size relative to a baseline period. Baseline correction is useful in experiments that investigate the effect of some experimental manipulation on pupil size. In such experiments, baseline correction improves statistical power by taking into account random fluctuations in pupil size over time. However, we show that baseline correction can also distort data if unrealistically small pupil sizes are recorded during the baseline period, which can easily occur due to eye blinks, data loss, or other distortions. Divisive baseline correction (corrected pupil size = pupil size/baseline) is affected more strongly by such distortions than subtractive baseline correction (corrected pupil size = pupil size − baseline). We discuss the role of baseline correction as a part of preprocessing of pupillometric data, and make five recommendations: (1) before baseline correction, perform data preprocessing to mark missing and invalid data, but assume that some distortions will remain in the data; (2) use subtractive baseline correction; (3) visually compare your corrected and uncorrected data; (4) be wary of pupil-size effects that emerge faster than the latency of the pupillary response allows (within ±220 ms after the manipulation that induces the effect); and (5) remove trials on which baseline pupil size is unrealistically small (indicative of blinks and other distortions). Pupil size is a continuous signal: a series of values that indicate how pupil size changes over time. In this sense, pupil-size data are similar to physiological measures, such as electroencephalography (EEG), which measures electrical brain activity over time, and it is even more similar to skin conductance, which (like pupil size) fluctuates slowly and correlates with arousal (Bradley, Miccoli, Escrig, & Lang, 2008). Pupil size is different from most behavioral measures, such as response times, that generally provide only a single value for each trial of the experiment. Psychologists are often interested in how pupil size is affected by some experimental manipulation (reviewed in Beatty & Lucero-Wagoner, 2000; Mathôt & Van der Stigchel, 2015). To give a classic example, Kahneman and Beatty (1966) asked participants to remember a varying number (three to seven) of digits. They found that participants’ pupils dilated (i.e., became larger) when the participants remembered seven digits, compared to when they remembered only three; that is, memory load caused the pupil to dilate (become bigger). As was common for pupil-size studies of the time, Kahneman and Beatty (1966) expressed their results in millimeters of pupil diameter; that is, they used absolute pupil-size values. But expressing pupil size in absolute values has a disadvantage: It is affected by slow, random fluctuations of pupil size. These fluctuations are a source of noise that reduce statistical power and make it more difficult to detect the effects of interest (in the case of Kahneman & Beatty, 1966, the effect of memory load). To deal with these fluctuations, researchers often look at the difference in pupil size compared to a baseline period, which is typically the start of the trial. By looking at pupil-size changes, rather than absolute pupil sizes, differences in pupil size that already existed before the trial are taken into account, are no longer a source of noise, and no longer reduce statistical power. This is baseline correction. There are two main ways to apply baseline correction: divisive, in which pupil size is converted to a proportional difference from baseline pupil size (corrected pupil size = pupil size/baseline), and subtractive, in which pupil size is converted to an absolute difference from baseline pupil size (corrected pupil size = pupil size − baseline). There are variations of these approaches, such as using percentage rather than proportion change, or converting absolute differences from baseline pupil size to z-scores; but these are all minor variations of these two general approaches. Here we will therefore focus on the difference between divisive and subtractive baseline correction. There are several reasons why researchers may choose either divisive or subtractive baseline correction. (Although such reasons are rarely provided.) Divisive baseline correction is attractive because it provides an intuitive measure: proportion change. If a paper states that an eye movement caused a 10 % pupillary constriction (Mathôt, Melmi, & Castet, 2015d), you can easily judge the size of this effect: substantial but not enormous. In contrast, if a paper states that a manipulation caused a 0.02-mm diameter change (Bombeke, Duthoo, Mueller, Hopf, & Boehler, 2016), you need a moment to remember (or look up) that human pupils are 2–8 mm in diameter, and that a 0.02-mm effect is therefore tiny. This is, in our view, less intuitive. And if the eye tracker reports pupil size in arbitrary units, typically based on a pixel count of the camera image, then absolute pupil-size differences become even harder to interpret. (In addition, even for eye trackers that report pupil size in seemingly absolute units [mm], measurements may not be entirely invariant to factors such as the distance between the participant and the camera, and may therefore be partly arbitrary as well.) However, despite these disadvantages, subtractive baseline correction may be the natural choice for some researchers because it is the standard approach in EEG research (e.g., Gross et al., 2013; Woodman, 2010). In pupillometry, there is no established standard for applying baseline correction. Based on our experience, most researchers now apply some form of baseline correction (but see, e.g., Gamlin et al., 2007), and variations of subtractive baseline correction (Binda, Pereverzeva, & Murray, 2013; Hupé, Lamirel, & Lorenceau, 2009; Jainta, Vernet, Yang, & Kapoula, 2011; Knapen et al., 2016; Koelewijn, Zekveld, Festen, & Kramer, 2012; e.g. Laeng & Sulutvedt, 2014; Murphy, Moort, & Nieuwenhuis, 2016; Porter, Troscianko, & Gilchrist, 2007; Privitera, Renninger, Carney, Klein, & Aguilar, 2010) seem somewhat more common than variations of divisive baseline correction (Bonmati-Carrion et al., 2016; Herbst, Sander, Milea, Lund-Andersen, & Kawasaki, 2011; Mathôt, van der Linden, Grainger, & Vitu, 2013; H. Wilhelm, Lüdtke, & Wilhelm, 1998). (One paper is listed per research group. This list is anecdotal, and not a comprehensive review.) As far as we know, no-one has systematically studied baseline correction of pupil-size data. However, baseline correction has been studied in the context of EEG/MEG data, as shown by a recent debate about whether or not baseline correction of EEG/MEG data should be abandoned in favor of high-pass filtering (Maess, Schröger, & Widmann, 2016; Tanner, Morgan-Short, & Luck, 2015; Tanner, Norton, Morgan-Short, & Luck, 2016). However, pupil-size data are different from EEG/ MEG data. For example, although in the absence of direct external stimulation the size of the pupil fluctuates in cycles of 1–2 s (Mathôt, Siebold, Donk, & Vitu, 2015a; Reimer et al., 2014), it does not show the slow systematic drift shown by EEG/MEG voltages (Tanner et al., 2016). Our aim is therefore to study baseline correction specifically for pupil-size data. We will use real and simulated data to see how robust subtractive and divisive

Read the complete text