Eyetracking, in which you present still images or movie clips to participants and track where on-screen they are looking, is a powerful and versatile method that is widely used within infant psychology. Most researchers use the built-in software provided by eyetracker manufacturers – which sets out, probably deliberately, to make researchers feel very confident in the quality of infant eyetracking data. When you are viewing a live replay of the gaze data, for example, the use large amounts of smoothing and interpolation to give the impression that the tracking looks very sensible, and real.
You get a very different picture, though, if you look at the raw data that the eyetracker is actually recording. One problem is that the same eyetracker, working in identical conditions, can record very good quality from one child, and very bad quality from another child. This paper identifies two problems in particular with infant eyetracking data quality – one, low precision, and the other, low robustness. We also show that, on average, worse quality data tends to be recorded from younger (relative to older) participants, from fidgetier participants (as opposed to those who sit very still, and later (relative to earlier) in a testing session.
Are Issues with Infant Eyetracking Data Consistent?
This would be OK if it were a random source of error – unpredictable measurement noise that did not systematically influence results. In fact, though, as we also show, data quality shows strong relationships with several different aspects of data quality. See the example below, in which we take a sample of real data obtained during the recording of a face (top left) and analyse it for the proportion of time the participant has spent looking at the eyes vs other areas of the face (top right). We then manipulate the data to simulate the effect of low precision (bottom left) and repeat the same analyses (bottom right). It appears that the bottom sample shows less looking to the eyes area – but in fact the only difference between the two samples is data quality. This is a problem because it would be easy to mistakenly get the impression that one person tends to look more to the eyes than another person – whereas in fact the only difference is that the first person is sitting more still, causing better quality eyetracking. We also run other analyses looking at other aspects of data quality – such as fixation duration (more on fixation time) and various reaction time measures – with similar results.
If you’re interested you can read more about this, including some algorithms that I and my colleage Jukka Leppanen have written to try to minimise these problems, here, here and here. The Matlab algorithms I wrote for that paper to calculate eyetracking data quality can be downloaded here.