Eskenazi, Michael A. "Best practices for cleaning eye movement data in reading research." Behavior Research Methods (2023): 1-11.
There are generally three different data cleaning methods that are applied to fixations in reading studies: fixation merging, temporal cut-offs, and outlier removal. Fixation merging methods were first developed by pioneering researcher Keith Rayner at his UMass eye tracking labs. These methods have been passed down through generations of research training and have made their way into eye-movement processing software as default settings (e.g., Data Viewer from SR Research). There are three stages in the fixation merging process. In the first stage, fixations less than 80 ms are merged with a longer fixation within .50 degrees of visual angle. In the second stage, fixations less than 40 ms are merged with a longer fixation within 1.25 degrees of visual angle. Finally, in the third stage, three fixations of 140 ms or less within a single interest area are merged together. I attempted to determine the original justification for these values by contacting SR Research Support who said that they do not take a position on whether these values and this method should be used but that they were included in the program at the request of an eye tracking researcher. I contacted her to determine the origin and she noted that they came from Keith Rayner’s original work. I next contacted several prominent eye tracking researchers who were trained by or worked closely with Keith Rayner, and though they all kindly responded, none of them could provide any specific justification for why these values were selected. Thus, although these values are commonly used as one of the steps to clean fixations, their origin and justification remain unclear.
The temporal cut-off method is used to eliminate fixations that are not based on lexical processing. It takes a minimum of 80 ms for visual information transduced in the eye to travel through the optic nerve to visual processing areas of the brain responsible for processing words (Rayner et al., 2012). Thus, a fixation that lasts less than 80 ms is unlikely to reflect lexical processing and should not be included in analyses designed to draw conclusions about lexical processing. However, when short fixations are central to the research question, some researchers choose to retain them in the data set (Schotter & Leinenger, 2016). On the upper end, there is less clear evidence about what constitutes a fixation duration that goes beyond typical lexical processing. The average gaze duration is around 250 ms with a distribution that has a long positive skew (Rayner et al., 2012). The data in the long tail on the right end of the distribution has many possible cut-off points, and researchers determine where to draw the line, usually somewhere between 800 and 1200 ms.
Finally, the outlier method is used to remove extreme data points that could bias the results. Outliers can occur for a variety of reasons, including experimenter or participant error, machine error, temporary lapse of attention, or other factors that result in data points that are not based on lexical processing. The typical method for removing outliers is to calculate the participant and condition-specific mean and standard deviation, and then to remove any fixation durations that are 2.5 or 3 standard deviations away from that mean. There are a variety of different methods for removing outliers, and researchers should be careful about which they choose (Berger & Kiefer, 2021). For example, when data have an ex-Gaussian distribution as with reaction time and fixation duration data, the typical outlier method results in trimming more data from the lower end than the higher end of the distribution (Miller, 1991).
Eskenazi, Michael A. "Best practices for cleaning eye movement data in reading research." Behavior Research Methods (2023): 1-11.
There are generally three different data cleaning methods that are applied to fixations in reading studies: fixation merging, temporal cut-offs, and outlier removal. Fixation merging methods were first developed by pioneering researcher Keith Rayner at his UMass eye tracking labs. These methods have been passed down through generations of research training and have made their way into eye-movement processing software as default settings (e.g., Data Viewer from SR Research). There are three stages in the fixation merging process. In the first stage, fixations less than 80 ms are merged with a longer fixation within .50 degrees of visual angle. In the second stage, fixations less than 40 ms are merged with a longer fixation within 1.25 degrees of visual angle. Finally, in the third stage, three fixations of 140 ms or less within a single interest area are merged together. I attempted to determine the original justification for these values by contacting SR Research Support who said that they do not take a position on whether these values and this method should be used but that they were included in the program at the request of an eye tracking researcher. I contacted her to determine the origin and she noted that they came from Keith Rayner’s original work. I next contacted several prominent eye tracking researchers who were trained by or worked closely with Keith Rayner, and though they all kindly responded, none of them could provide any specific justification for why these values were selected. Thus, although these values are commonly used as one of the steps to clean fixations, their origin and justification remain unclear.
The temporal cut-off method is used to eliminate fixations that are not based on lexical processing. It takes a minimum of 80 ms for visual information transduced in the eye to travel through the optic nerve to visual processing areas of the brain responsible for processing words (Rayner et al., 2012). Thus, a fixation that lasts less than 80 ms is unlikely to reflect lexical processing and should not be included in analyses designed to draw conclusions about lexical processing. However, when short fixations are central to the research question, some researchers choose to retain them in the data set (Schotter & Leinenger, 2016). On the upper end, there is less clear evidence about what constitutes a fixation duration that goes beyond typical lexical processing. The average gaze duration is around 250 ms with a distribution that has a long positive skew (Rayner et al., 2012). The data in the long tail on the right end of the distribution has many possible cut-off points, and researchers determine where to draw the line, usually somewhere between 800 and 1200 ms.
Finally, the outlier method is used to remove extreme data points that could bias the results. Outliers can occur for a variety of reasons, including experimenter or participant error, machine error, temporary lapse of attention, or other factors that result in data points that are not based on lexical processing. The typical method for removing outliers is to calculate the participant and condition-specific mean and standard deviation, and then to remove any fixation durations that are 2.5 or 3 standard deviations away from that mean. There are a variety of different methods for removing outliers, and researchers should be careful about which they choose (Berger & Kiefer, 2021). For example, when data have an ex-Gaussian distribution as with reaction time and fixation duration data, the typical outlier method results in trimming more data from the lower end than the higher end of the distribution (Miller, 1991).