Horizontal Bias in Infant Eye Movements during Scene Viewing
Master Thesis of Daan van Renswoude
University of Amsterdam
D.R.vanRenswoude@uva.nl
August 22, 2014
Abstract
Most knowledge about infant development stems from simple looking time measures (Aslin, 2007). Eye-trackers allow for more sophisticated measures of eye movements that make it possible to learn much more about infant development in the near future. This thesis examined eye movements of 43 infants (Mean age = 8.45 months, SD = 3.66, range = 3.48 to 15.47) and 47 adults (Mean age = 21.74 years, SD = 4.54, range = 17.89 to 39.84) when viewing 28 natural scenes. To compare eye movements of infants and adults, fixations and saccades were classified using a new algorithm (Mould et al., 2012). A general effect in scene viewing is that adults make more horizontal than vertical and oblique saccades (Foulsham et al., 2008). Earlier studies used a binning approach to describe saccade directions, but this thesis also modelled the saccade directions with a mixture of Von Mises distributions. The results showed that the binning approach led to spurious results, whereas the Von Mises approach gave a better description of the data. Furthermore, it was shown that infants have a stronger horizontal bias than adults. Finally, the examination of the algorithm showed that thresholds to classify fixations and saccades differ between individuals. This is an important finding, because most algorithms use fixed thresholds that do not capture individual differences.
Introduction
We perceive the world as an environment through which we can smoothly move our eyes. This is amaz-ing given our actual eye movements that are not smooth at all. Instead, our eye movements consist of fixations and saccades. We make approximately three fixations and saccades every second (Rayner, 2009). During fixations our eyes barely move, while the saccades are fast jumps from one fixation point to another. Modeling these eye movements can be very useful in understanding the cognitive develop-ment of young infants. Young infants cannot tell us what they know and how they perceive the world. Being able to model their eye movements, gives an amazing insight in how they perceive the world. This
study examines eye movements of infants observing static natural scenes.
Infant eye tracking studies often use single object stimuli (Bornstein et al., 2011b,a). These single ob-ject stimuli are not a good representation of the real world where objects always occur in context of the environment. Bornstein et al. showed that object-context relations influence eye movements (2011a) and that infants look differently at single object stim-uli than at the same stimstim-uli in context of their envi-ronment (2011b). Therefore, the current study used natural scenes as stimuli. The processing of natu-ral scene stimuli is known as scene viewing; although little is known about scene viewing in infants, scene viewing is studied extensively in adults (Henderson, 2003; Rayner et al., 2009; Tatler, 2007).
−1000 0 1000 2000 1500 500 −500 x y −1000 0 1000 2000 1500 500 −500 x y −1000 0 1000 2000 1500 500 −500 x y
Figure 1: Typical scanning patterns of three young infants, the boxes represent the stimuli, the dots the position of the eyes.
Scene viewing is a very complex axtivity, but nev-ertheless some simple characteristics have been estab-lished in adults. A general effect in scene viewing is that more horizontal than vertical and oblique sac-cades are made (Gilchrist and Harvey, 2006; Tatler and Vincent, 2008). This bias is known as the hori-zontal bias. Foulsham and Kingstone (2012) showed that incorporating the horizontal bias in models that predict eye movements improves these models. Theo-ries explaining the horizontal bias in adults may make certain predictions about the occurence of a horizon-tal bias in infants. Foulsham et al. (2008) describe four theories that may explain the bias: (1) the bias is a laboratory artifact, (2) it is an oculomotor bias, (3) the bias is learned and (4) the bias is caused by image characteristics.
First, Foulsham et al. (2011) showed that labo-ratory artifacts (i.e. monitors are wider than high) could not explain the bias. They, for instance, found that the horizontal bias reduces, but does not disap-pear when stimuli are presented in square and por-trait format. The second explanation, that eye mus-cles make horizontal saccades easier than vertical and oblique ones, was also ruled out for adults. Foulsham and Kingstone (2010) showed that saccades in all di-rections could be made easily. The third theory, that the horizontal bias is learned did not receive much at-tention, but studying infants makes it possible to test this theory. Abed (1991) found cultural differences in
saccade directions; vertical readers make more verti-cal saccades during scene viewing, whereas horizontal readers make more horizontal saccades during scene viewing. This implies the horizontal bias could be acquired while learning to read.
The fourth theory, that image characteristics may explain the horizontal bias, has received most atten-tion. Image characteristics can influence eye move-ments bottom-up or top-down. Top-down process-ing implies that high-level features such as objects (Nuthmann and Henderson, 2010) and the ‘gist’ of a scene (Castelhano and Henderson, 2007; Torralba et al., 2006) influence eye movements. Bottom-up processing implies that low-level features such as, orientation, contrast and color influence eye move-ments (Itti and Koch, 2000; Itti et al., 1998). Althaus and Mareschal (2012) found that top-down process-ing drove eye movements of 12-month-olds, but not of 4-month-olds. The 12-month-olds had a prefer-ence for regions that changed during trials, whereas this preference for changing regions was absent in 4-month-olds. Bottom-up processing also seems to play a minor role in infant eye movements. Amso et al. (2014) showed that bottom-up saliency predicted eye movements well in adults, but not in infants. And Althaus and Mareschal (2012) found that 12-month-olds fixated more in bottom-up salient regions than 4-month-olds. It is thus hypothesized, that if dif-ferential sensitivity to bottom-up and top-down
pro-Figure 2: Five of the natural scene stimuli used in this study.
cessing drives the horizontal bias, that young infants should not show the bias or to a lesser extent than older infants.
To test this hypothesis, analysis methods that are comparable between infants and adults should be used. The literature provides no standard method to study eye movements; there are for instance many different algorithms to classify fixations and saccades. This is problematic, because different algorithms can lead to different results of the same raw data (Shic et al., 2008), that makes comparing studies difficult (Nystr¨om and Holmqvist, 2010). Furthermore, these algorithms often use fixed thresholds to classify fix-ations and saccades. This is problematic because of the idiosyncratic nature of eye movements (Andrews and Coppola, 1999; Castelhano and Henderson, 2008; Poynter et al., 2013). Most algorithms use speed-based thresholds, but there is much variation in sac-cade speeds (Abel et al., 1983), this variation is not captured by fixed thresholds. If there is much vari-ation in adults, it is likely there is also varivari-ation in infants, maybe even to a larger extent, because the visual system of the infants is still developing. It is therefore, especially in this infant study, important to incorporate these individual differences. To this end, the current study uses the non-parametric data-driven algorithm of Mould et al. (2012) that classifies fixations and saccades based on an individually de-rived threshold. This algorithm is explained in the method section.
Another problem with infant eye-tracking data are the many missing values. Figure 1 shows three typical tracking patterns of three infants who saw a stimu-lus (the rectangle box). Looking at Figure 1, it can be seen that some data points exceed the stimulus
boundary. Other sources of missing data are blinks, looking away and eye-tracker failure. To properly study eye movements these missing values need to be dealt with in some suitable way. Unfortunately, the literature provides no systematic way of dealing with missing data. Most studies do not mention what is done with missing values at all, others throw away all trials with missing data and some only mention that trails with too many missing values were not in-cluded. This thesis will handle missing values in a more systematic way, by studying the effects of dif-ferent methods to handle missing data on the speed thresholds.
To summarize, this thesis examines whether in-fants have the horizontal bias. If this bias is driven by bottom-up and top-down processing it is expected that the bias is larger for adults than for infants. Fur-thermore, it is expected that older infants show the bias to a larger extent than younger infants. This thesis also examines two methodological issues: (1) thresholds to classify fixations and saccades are de-rived individually and are expected to differ between persons because of the idiosyncratic nature of eye movements and (2) missing values are handled in a systematic way to explore to what extent methods of handling missing data influence the results.
Method
Participants
Forty-three infants (Mean age = 8.45 months, SD = 3.66, range = 3.48 to 15.47) and 47 adults (Mean age
Figure 3: llustration of the procedure to derive the speed threshold (left panel) and the duration threshold (right panel).
= 21.74 years, SD = 4.54, range = 17.89 to 39.84) participated in a free scene viewing study conducted by Scott Johnson (http://www.babylab.ucla.edu/). The gender ratio in infants (14 males, 29 females) was independent of the gender ratio in adults (19 males, 28 females), χ2 (1) = .31, p = .58. All participants
saw 28 natural scenes for four seconds while the po-sition of one eye was measured with an EyeLink 500 Hz eye-tracker. This resulted in 90 (participants) x 28 (images) is 2520 trials. Each trial consisted of 2000 (500 Hz x 4 s) x and y coordinates indicating the position of the eye at a given time. This data set was used to test all hypotheses.
Classifying Fixations and Saccades
The algorithm of Mould et al. (2012) was used to derive the fixations and saccades from the raw data. This algorithm uses a two step procedure; in the first step a speed threshold is used to separte saccadic from non-saccadic eye movements, in the second step a duration threshold is used to eliminate noise from
the non-saccadic durations.
The first step uses the fact that the speed of eye movements is low during fixations and high during saccades. First, local maxima in speed of eye move-ments are found by selecting the eye movemove-ments that are faster than the previous and following eye move-ments. There are many local speed maxima in fixa-tions, but few local speed maxima in saccades. Sec-ond, these local maxima are used to separate saccadic and non-saccadic movements. This is done by com-paring the distribution of local speed maxima exceed-ing the threshold (gray histogram), with a uniform distribution of local maxima exceeding the thresh-old (dotted line), see the left panel of Figure 3. The difference between these two distributions is given by the gap statistic (red line in the left panel of Figure 3); the optimal speed threshold is where this gap statis-tic reaches a maximum. Eye movements with speeds below the threshold are classified as non-saccadic eye movements and eye movements with speeds above the threshold are classified as saccadic eye movements. After this first step the non-saccadic eye movements
still consist of noise and fixations.
To separate fixations from noise, a duration thresh-old is used. The right panel of Figure 3 clearly shows two distributions; one with many short non-saccadic durations (noise) and one with longer non-saccadic durations approximately normal distributed around 400 ms (fixations). To separate these two distribu-tions the minimum between the two local maxima of the histogram of non-saccadic durations is calculated, see the right panel of Figure 3. This minimum is the optimal threshold (red line in right panel of Figure 3) to separate fixations from noise. All points above this threshold are classified as fixations.
The fixations that are now derived are assumed to have saccades between them, therefore all data points between the fixations are classified as saccades. When there are missing values between two fixations, it cannot be assumed a saccade is made and therefore those data points are classified as missing.
Missing Data
The eye-tracking data of the infants contains many missing values that need to be dealt with in a suitable way. There are two general options to deal with miss-ing data: (1) missmiss-ing values can be interpolated or (2) missing values can be omitted. The most common in-terpolation method is linear inin-terpolation, which was also used in this study. In linear interpolation the sequence of missing values is filled with data forming a straight line from the last recorded data point be-fore the missing sequence to the first recorded data point after the missing sequence. The source of these missing values can be blinks, eye tracker failure, look-ing away by the participant and data points outside the stimuli boundaries. Ideally these different types of missing values are handled in a different way. It may for instance be useful to interpolate blinks to get a clearer picture of the whole scan pattern. On the other hand, interpolating values where someone looked away is not useful at all. This section de-scribes how, looking away, blinks, eye-tracker failure
and data points exceeding stimuli boundaries were classified and dealt with in this thesis.
When dealing with looking away, interpolation is never a good option, because it does not make sense to impute data points when someone did not look. Missing durations longer than 500 ms in infants are unlikely to be blinks (Bacher and Smotherman, 2004), but are more likely to be looking away. There-fore, missing durations longer than 500 ms were clas-sified as looking away and are omitted in the analyses. In the case of blinks, both interpolation and omit-ting missing values can be valid options. The study of Bacher and Smotherman (2004) shows that blinks of 3-month-old infants are likely to fall in the range of 150 to 500 ms. Therefore, missing sequences shorter than 500 ms are classified as blinks. To deal with these blinks, ommision and interpolation were tested against each other.
Pedrotti et al. (2011) found measurements 50 ms preceding and following blinks to be unreliable, be-cause the closing and opening of the eyelid makes it difficult for the eye-tracker to get an accurate mea-sure. Therefore, data points 50 ms before and after blinks were included and interpolated in the analyses to deal with eye-tracker failure.
The last source of missing values are the data points that exceed the stimuli boundaries, the anal-ysis was conducted with these data points omitted and included. Table 1 gives an overview of these six differents methods to handle missing values.
The best way to test which method should be pre-ferred would be to compare the methods with expert judgement. That way the method that comes closest to the judging of the experts can be select as the best method. Unfortunately this was too time consuming for this thesis, therefore another method was used. The speed thresholds were compared between the methods. If different methods lead to approximately the same thresholds and these thresholds correlated highly, this would be an indication that it does not really matter which method is used. It would then
Table 1: Six Methods to Handle Missing Data
No interpolation interpolate missing sequences < 500 ms
interpolate missing sequences < 400 ms ± 50 ms
Omit data points exceeding
stim-uli boundaries 1 2 3
Include data points exceeding
stimuli boundaries 4 5 6
be best to choose a method that results in a low vari-ance between thresholds and makes few assumptions. If different methods yield different results, choosing the best method would be much harder. In that case no interpolation nor exclusion of data points will be used to make as few assumptions as possible.
In addition to comparing the thresholds of differ-ent methods, some more analysis on the missing data were conducted. Adults were expected to have fewer missing data points than infants, this was tested us-ing a t-test. Furthermore, it was tested if age and order of presentation were predictors of missing data points in infants. Age of infants was expected to cor-relate negatively with the number of missing data points, because older infants stay more easily fo-cussed than younger infants. Order was expected to correlate positively with the number of missing data points, because infants are likely to loose their at-tention during the experiment. If these effects were present, age and presentation order would be used as control variables in further analyses.
Thresholds
Speed thresholds of infants and adults were expected to differ between individuals (because of the idiosyn-cratic nature of eye movements (Andrews and Cop-pola, 1999; Castelhano and Henderson, 2008; Poyn-ter et al., 2013)), but not between images. This was tested by comparing two nested multilevel models; one in which intercepts can vary between
individu-als (or images) and one in which all individuindividu-als (or images) were constrained to the same intercept. Be-cause these models were nested, log-likelihood ratio tests could be performed to test these hypotheses. If speed thresholds of infants showed individual differ-ences, it would be tested if these individual differences could be explained by age. This would be done by adding age to the model. The duration thresholds were also expected to differ between individuals, but this could not be tested because there was no varia-tion within these thresholds, since every person only has one threshold. However, it was tested if the du-ration thresholds of infants correlated with age.
Horizontal Bias
To answer the question if and to what extent infants have the horizontal bias, the proportion of horizon-tal, vertical and oblique saccades that infants and adults make was calculated for each trial. The sac-cade direction ranges from zero degrees (straight to the right) to 360 degrees. Saccades made along the 0◦axis (± 30◦) and the 180◦axis (± 30◦) were clas-sified as horizontal saccades, saccades made along the 45◦ axis (± 15◦), the 135◦ axis (± 15◦), the 225◦ axis (± 15◦) and the 315◦ axis (± 15◦) were clas-sified as oblique saccades and saccades made along the 90◦ axis (± 30◦) and the 270◦ axis (± 30◦) were classified as vertical saccades. These propor-tions were chosen so every direction covered the same range (33.3%, i.e. 120◦) of the circle. This was done to be able to make fair comparisons between the three
directions. The mean proportions of the three sac-cade directions collapsed over images of the infants and adults was calculated. It was tested if the ratio between these mean proportions was the same for in-fants and adults. Furthermore, it was tested if the mean proportion of horizontal saccades of the infants correlated with age. To do this properly, a multilevel model was used to incorporate the expected variance within individuals.
Although it is common to bin the saccades in three direction as is described above, this approach may not be optimal. It would probably be better to es-tablish the distribution of saccade directions data-driven instead of classifying them into pre-defined categories (horizontal, vertical and oblique). There-fore, the Von Mises distribution was also used to describe the data. The Von Mises ditribution can be though of as the normal distribution on a circle. Because multiple peaks in the distribution were ex-pected, for instance at zero and 180 degrees repre-senting horizontal saccades, a mixture of Von Mises distributions was fitted on the data of the infants and the adults. The µ (mean) parameters of infants and adults were compared to test for differences in sac-cade directions between infants and adults. The κ (1 / variance) parameters were compared to test sac-cade precision differences between infants and adults and the α (the proportion of each mixed compo-nent) parameters were compared to test if the pro-portion of saccade directions differed between infants and adults.
Results
Algorithm Adjustments
The algorithm of Mould et al. (2012) was used to classify fixations and saccades. Mould et al. used Matlab to conduct their study, in this study R Core Team (2014) was used. In order to obtain the R-code to perform this study the Matlab code provided by Mould et al. was rewritten into R-code. Although
the original algorithm was used as much as possible, some constraints were modified and the algorithm was used in a slightly different way than Mould et al. did, because infant eye-tracking data has some spe-cific characteristics. The main adjustment was that speed thresholds were estimated per trial, whereas Mould et al. used multiple trials of the same person. Although this worked fine for the speed threshold, this approach resulted in too few data points to re-liably estimate the duration threshold. Therefore 28 trials (all images) per person were combined to es-timate the duration threshold per person instead of per trial.
Although the Mould et al. algorithm is aimed to be completely data driven, visual inspection of the classification showed the fit improved after making three small adjustments. First, only trials with at least one second of recorded data were analyzed in order to have enough data points to reliably derive the speed threshold. Second, velocities that exceeded the 1000 deg/ms were not used to derive the speed threshold. Inspection of the data showed that these velocities often occurred before and after missing data points. This is in line with the study of Pedrotti et al. (2011) who also found that eye-trackers produce un-reliable measures before and after blinks. Nystr¨om and Holmqvist (2010) also omit these high velocities in their algorithm, because these velocities are physi-ologically impossible. Third, fixations that were sepa-rated by saccades shorter than 10 ms were combined. This was also in line with Nystr¨om and Holmqvist (2010) who suggested that saccades shorter than 10 ms are physiologically impossible.
Missing Data
To explore the influence of missing data, missing values were handled in six different ways in this study. Data points exceeding stimulus boundaries were omitted and included, missing sequences shorter than 500 ms were and were not linear interpolated and data points around missing sequences were and were not linear interpolated along with the missing
0 5 10 15 20 25
Infants
Thresholds (Deg/s) Images 20 30 40 50 60 70 80 0 10 20 30 40 Thresholds (Deg/s) Individuals 20 30 40 50 60 70 80 0 5 10 15 20 25Adults
Thresholds (Deg/s) Images 20 30 40 0 10 20 30 40 Thresholds (Deg/s) Individuals 20 30 40Figure 4: Variation in speeds thresholds between images (upper panels) and individuals (lower panels) for infants (left panels) and adults (right panels), the colors correspond to the different methods of handling missing data described in Table 1; (1) black, (2) red, (3) green, (4) dark blue, (5) light blue and (6) purple
Table 2: Means and Standard Deviations of the Thresholds of Infants (left) and Adults (right), Grouped over Images for all Six Methods of Handling Missing Data and the Correlations between these Methods
M SD 1 2 3 4 5 M SD 1 2 3 4 5 1 37.73 2.80 21.70 1.53 2 40.02 2.87 0.77 26.00 2.84 0.74 3 40.42 3.12 0.64 0.74 26.05 2.17 0.61 0.74 4 41.40 3.11 0.79 0.76 0.58 22.42 1.71 0.97 0.76 0.57 5 45.09 3.74 0.63 0.72 0.49 0.83 25.35 2.35 0.79 0.86 0.60 0.84 6 45.26 4.89 0.57 0.53 0.50 0.76 0.78 25.04 2.42 0.72 0.65 0.64 0.72 0.66
Table 3: Means and Standard Deviations of the Thresholds of Infants (left) and Adults (Right), Grouped over Individuals for all Six Methods of Handling Missing Data and the Correlations between these Methods
M SD 1 2 3 4 5 M SD 1 2 3 4 5 1 39.44 20.29 21.71 4.52 2 41.23 17.07 0.99 26.03 6.57 0.85 3 42.33 20.53 0.98 0.98 26.10 6.92 0.74 0.83 4 41.77 20.84 0.94 0.96 0.91 22.45 4.82 0.99 0.86 0.75 5 45.45 20.00 0.94 0.96 0.93 0.99 25.38 6.69 0.87 0.94 0.79 0.90 6 45.69 21.26 0.92 0.94 0.93 0.95 0.97 25.08 6.61 0.81 0.90 0.92 0.82 0.89
sequence itself. Combining these three methods re-sulted in six stategies to handle missing data, see Ta-ble 1 in the method section.
Table 2 and 3 show the means, standard deviations and correlations of the speeds thresholds of the data displayed in Figure 4. These speed thresholds may give some insight in which missing data method is the best. Low means and standard deviations were preferred over high means and standard deviations. Standard deviations should be low, because thresh-olds are ideally estimated with as little error as pos-sible. Means should be low, because with little noise in data, the distinction between fixations and sac-cades can be made at lower thresholds. Means and standard deviations also correlated (r = 0.86 (left of Table 2), r = 0.89 (right of Table 2), r = 0.33 (left of Table 3), r = 0.98 (right of Table 3)), this made sense as they both can be thought of as a measure of error. Furthermore, adults had much lower means and standard deviations than infants, this could be because the data of the adults contains less noise than the infant data.
Table 2 and 3 also show that the correlations be-tween the thresholds produced by the six different methods were high, especially between individuals. Because all methods produce similar results, meth-ods with little variance and few assumptions were preferred. The best method by this standard is the method where data points outside the stimuli bound-aries were omitted and nothing was interpolated. It seems reasonable that noise reduces, when data points exceeding stimuli boundaries are omitted. The data used in further analysis is the data without data points exceeding stimuli boundaries and without any interpolated data points.
Missing Data Analysis
There were 14 trials (1%) of the adults and 86 trials (7%) of the infants that had too few data points to an-alyze. Therefore these trials were excluded from the analysis. One infant only had two trials with enough data to estimate the speed thresholds, but two trials are not enough to estimate the duration threshold.
Therefore the data of that infant was excluded for further analysis. Of the remaining 42 infants there were at least 19 trials of each infant and 23 trials of each adult in the analyses. The missingness of com-plete trials did not correlate with age of the infants, r = 0.25, t (40) = 1.6, p = 0.117.
Although this was a very short study, the 28 stimuli were only presented for four seconds, infants always have difficulties keeping their attention on the task. A Welch t-test showed that infants had on average more missing data points over all trials (M = 14970, SD = 13857) than adults (M = 2754, SD = 2769, t(43.93) = 5.61, p < 0.001). The missing data points in infants are probably not missing at random, but are likely to be related to age and presentation order. Because the number of missing data points is likely to differ between individuals a multilevel model was used in which intercepts could differ between indi-viduals. This random intercept model improved the fit of model compared to the single intercept model according to the log-likelihood ratio test, χ2 (1) =
781.052, p < 0.001. To check if slopes should also be estimated per individual, model 1 was fitted with fixed and random slopes. The log-likelihood ratio test showed it was better to estimate the slopes per indi-vidual, χ2(2) = 21.794, p < 0.001.
M issingij= β0j+ β1j× Order + ij (1)
β0j= γ00+ γ01× Age + U0j
β1j= γ10+ γ11× Age + U1j
Model 1 showed that age (γ01= -37.1, SE = 14.43,
t (40) = -2.57, p = 0.014), order (γ10 = 16.21, SE =
3.93, t (1132) = 4.13, p < 0.001) and their interaction (γ11= -1.09, SE = 0.43, t (1132) = -2.56, p = 0.011)
are significant predictors of the number of missing data points. Looking at Figure 5 it becomes clear that trials that were presented later have more miss-ing data points and that younger infants have more missing data than older infants. Figure 5 also sheds some light on the interaction effect; it seems that the relationship between missing data points and age is
stronger for trials presented later during the experi-ment. This can be seen because the slopes in the up-per half of Figure 5 are a bit steeup-per than the slopes in the bottom half.
Obviously it is important to control for these or-der and age effects. To control for the oror-der effect, stimuli were presented in a random order. Therefore there should not be any variation within images. To check for variation in mean number of missing data points within images, models with and without ran-dom intercepts were compared. The log-likelihood ratio test showed that the fit did not improve when intercepts were allowed to vary over images, χ2(1) =
0, p = 0.999. This implies that all images had approx-imately the same number of missing data points and that order did not influence missing values in certain images.
The age effect could influence the data because older infants had more data points. This could eas-ily lead to erroneous conclusions about the horizontal bias, because it could be found that more horizontal saccades were made by older infants than by younger infants, whereas in fact older infants made more sac-cades in every direction. Therefore the proportion of saccades was used to test the horizontal bias hypoth-esis.
Thresholds
Speed thresholds were derived per person per trial and duration thresholds only per person. It was ex-pected that speed thresholds were different for indi-viduals, but not for images. Figure 6 shows plots of the variability in speeds thresholds between images (left panel), between individuals (middle panel) and of variability in duration thresholds between individ-uals (right panel).
Two Welch t-tests showed that infants (M = 38.08, SD = 18.47) had higher speed thresholds than adults (M = 21.71, SD = 4.52, t (45) = -5.6, p < 0.001) and that infants (M = 89.17, SD = 32.72) had higher duration thresholds than adults (M = 60.09, SD =
Age in Months
Number of Missing Data P
oints 0 500 1000 1500 2000 4 6 8 10121416 1 2 4 6 8 10121416 3 4 4 6 8 10121416 5 6 4 6 8 10121416 7 8 9 10 11 12 13 0 500 1000 1500 2000 14 0 500 1000 1500 2000 15 16 17 18 19 20 21 22 4 6 8 10121416 23 24 4 6 8 10121416 25 26 4 6 8 10121416 27 0 500 1000 1500 2000 28
Figure 5: Relationship between age and missing data points for all 28 presentation orders
0 5 10 15 20 25 Images
Speed Thresholds (Deg/s)
20 30 40 50 60 70 Infants Adults 0 10 20 30 40
Speed Thresholds (Deg/s)
Individuals 20 30 40 50 60 70 Infants Adults 0 10 20 30 40 Duration Thresholds (ms) Individuals 40 80 120 160 Infants Adults
Figure 6: Variation in speed thresholds between images (left panel) ,individuals (middle panel) and duration thresholds (right panel) for infants and adults
16.36, t (59) = 5.21, p < 0.001). This is an indica-tion that the adult data contains less noise than the infant data, because a distinction between fixations and saccades could be made at lower thresholds for adults than for infants.
Looking at Figure 6, it seems that the variation in speed thresholds is larger between persons than between images. This can be seen because the confi-dence intervals include the overall mean (dotted ver-tical line) most of the time when means are calculated over images, but not when means are calculated over persons. To test the hypothesis that there is no varia-tion in images, two multilevel models were compared; one in which intercepts could differ between images and one in which only one intercept was estimated. As expected the random intercept model did not im-prove the fit compared to the single intercept model according to the log-likelihood ratio test in infants (χ2 (1) = 0, p = 0.999), but unexpectally it did in adults (χ2 (1) = 10.066, p = 0.002). Although the
improvement was small, as can be seen by the log-likelihood ratio value. To test the hypothesis that there is variation in individuals, two multilevel mod-els were compared; one in which intercepts could dif-fer between individuals and one in which only one intercept was estimated. As expected the random in-tercept model did improve the fit compared to the single intercept model according to the log-likelihood ratio test in infants (χ2 (1) = 890.152, p < 0.001)
and adults (χ2 (1) = 456.082, p < 0.001). Notice that these individual effects are much larger than the effect of image found in adults, as can be seen by the large log-likelihood ratio values.
The variation in both speed and duration thresh-olds of the infants could partially be explained by age. There was a negative correlation of -0.29 between the speed thresholds and age that was marginally signif-icant, t (40) = -1.94, p = 0.06 and there was a sig-nificant negative correlation of -0.34 between the du-ration thresholds and age, t (40) = -2.28, p = 0.028. Older infants had lower speed and duration thresh-olds than younger infants. There was no correlation between the speed and duration thresholds, -0.06, t (40) = -0.4, p = 0.694. The interpretation of the
cor-relation between speed thresholds and age should be done carefully, not only because it is marginally sig-nificant, but more importantly because the variation within individuals is not taken into account. Using a multilevel model to examine the relationship be-tween speed thresholds and age is therefore a better approach. The following multilevel model was fitted to the data of infants:
T hresholdij= β0j+ β1j× Order + ij (2)
β0j= γ00+ γ01× Age + U0j
β1j= γ10+ γ11× Age + U1j
To check if slopes should also be estimated per in-dividual, model 2 was fitted with the fixed slopes and random slopes. The log-likelihood ratio test showed it was better to estimated the slopes per individual, χ2 (2) = 29.222, p < 0.001. The multilevel model
showed that age was a marginally significant predic-tor of speed thresholds, γ01= -1.1, SE = 0.61, t (40)
= -1.81, p = 0.077), but order (γ10 = 0.08, SE =
0.13, t (1072) = 0.56, p = 0.573) and the interac-tion (γ11 = 0.01, SE = 0.01, t (1072) = 0.62, p =
0.536) were not. As expected there were no order effects. Age and speed threshold had a negative re-lationship on almost every image, see Figure 7. That age was a predictor of speed thresholds implies the distinction between fixations and saccades could be made at lower thresholds when infants get older; this is an indication that fixations become more accurate as infants get older.
Horizontal Bias
The Binning Approach
For all saccades, the direction in which the saccade was made was calculated, see Method section. Then the proportion of horizontal, vertical and oblique sac-cades was derived. Figure 8 shows these proportions for infants and adults grouped over images. It is clear that horizontal saccades are made more often
Age Threshold 20 40 60 80 4 6 8 10121416 1 2 4 6 8 10121416 3 4 4 6 8 10121416 5 6 4 6 8 10121416 7 8 9 10 11 12 13 20 40 60 80 14 20 40 60 80 15 16 17 18 19 20 21 22 4 6 8 10121416 23 24 4 6 8 10121416 25 26 4 6 8 10121416 27 20 40 60 80 28
Figure 7: The relationship between age and speed threshold for all 28 images
Age Hor iz ontal 0.0 0.2 0.4 0.6 0.8 1.0 4 6 8 10121416 1 2 4 6 8 10121416 3 4 4 6 8 10121416 5 6 4 6 8 10121416 7 8 9 10 11 12 13 0.0 0.2 0.4 0.6 0.8 1.0 14 0.0 0.2 0.4 0.6 0.8 1.0 15 16 17 18 19 20 21 22 4 6 8 10121416 23 24 4 6 8 10121416 25 26 4 6 8 10121416 27 0.0 0.2 0.4 0.6 0.8 1.0 28
0.1 0.2 0.3 0.4 0.5 0.6 0.7 Propor tion of Saccades
Horizontal Vertical Oblique
Infants Adults
Figure 8: Proportion of saccades in the horizontal, vertical and oblique directions for infants and adults, error bars lie two standard errors around the means
than vertical and oblique saccades; the horizontal bias was found in both infants and adults. Further-more, adults showed this bias to a larger extent than infants. To test if the proportion of horizontal sac-cades varied within images and individuals, fixed- and random intercept models were compared for infants and adults. Log-likelihood ratio tests showed that the random intercept model improved the fit compared to the fixed effects model over images in infants (χ2 (1) = 49.099, p < 0.001), over individuals in infants (χ2
(1) = 15.259, p < 0.001), over images in adults (χ2(1)
= 4.04, p = 0.044) and over individuals in adults (χ2
(1) = 283.255, p < 0.001). These findings indicated that there was variation in the proportion of horizon-tal saccades within individuals and images for both infants and adults. Looking at the log-likelihood ra-tio values, infants seemed to vary more over images, whereas adults had more variation between individ-uals.
To test if age was a predictor of the horizontal bias,
a multilevel model was fitted with age, order and their interaction as predictors of proportion of horizontal saccades and infants as grouping variable:
Horizontalij= β0j+ β1j× Order + ij (3)
β0j= γ00+ γ01× Age + U0j
β1j= γ10+ γ11× Age + U1j
To check if slopes should be estimated per indi-vidual, model 3 was fitted with the fixed slopes and random slopes. The log-likelihood ratio test showed that estimating random slopes was not necessary, χ2
(2) = 2.172, p = 0.338. The multilevel model showed that age is no predictor of the proportion of horizon-tal saccades, γ01 = 0.01, SE = 0.01, t (40) = 0.81, p
= 0.425), but order (γ10= -0.01, SE = 0, t (1036) =
-1.91, p = 0.056) and the interaction between age and order (γ11= 0, SE = 0, t (1036) = 1.74, p = 0.082) are
marginally significant predictors of the proportion of horizontal saccades. Figure 9 shows the relationship between age and proportion of horizontal saccades for all 28 orders.
To check for order effects in adults, a multilevel model was fitted with order as predictor of proportion of horizontal saccades:
Horizontalij = β0j+ β1× Order + ij (4)
β0j= γ00+ U0j
Model 4 was also fitted with fixed and random slopes. The log-likelihood ratio test showed it was not necesarry to estimate slopes per individual, χ2
(2) = 0.193, p = 0.908. Order also was a significant predictor of the proportion of horizontal saccades in adults (β1 = 0.001, SE = 0.001, t (1253) = 2.151, p
= 0.032). But the effect was in a different direction, infants made more horizontal saccades on earlier tri-als, whereas adults made more horizontal saccades
on later trials. However, this finding in adults is con-sistent with the marginally significant interaction be-tween age and order found in infants. As can be seen in Figure 9, older infants do make more horizontal saccades on later trials than younger infants.
Although the method used here to split the direc-tion of saccades into three direcdirec-tions is commonly used (e.g. Foulsham et al. (2008); Foulsham and Kingstone (2010)), this approach may not be opti-mal. This is the case because all saccades are con-strained to three directions, whereas saccades can ac-tually cover the full range of a circle. The Von Mises distribution may be more appropriate to describe this kind of directional data.
The Von Mises Approach
The Von Mises distribution can be thought of as the normal distribution on a circle, which ranges from zero to 360 degrees. The Von Mises distribution has two parameters; µ, the mean which lies somewhere between zero and 360 and κ, which is 1 / variance. The higher the κ value is, the more peaked the distri-bution is. To describe the data in this study one Von Mises distribution was insufficient. There were multi-ple directions in which individuals look and therefore a mixture of Von Mises distributions was used to de-scribe this directional data.
The question remained how many Von Mises dis-tributions should be mixed. Therefore several mix-tures of the Von Mises distribution with an increas-ing number of components were fitted on the data of the infants and adults using the movMF package (Hornik and Gr¨un, 2011) in R (R Core Team, 2014). Table 4 shows that five components were sufficient to fit the data of the infants according to the AIC, BIC and their akaike weights (Wagenmakers and Farrell, 2004). The number of components for the adult data was less clear, the AIC suggested ten components, whereas the BIC suggested six components. To make things even more complicated it would be ideal if in-fants and adults had the same number of components, so they could easily be compared. The five and six
component models were selected, because five of the six components in the adult model were comparable with the five components of the infant model, see Figure 10. The colors in Figure 10 correspond to the component in that model, it can be seen that only the yellow component with the mean around 270◦
was not comparable between infants and adults. A direct comparison between the parameters of the fitted models in infants and adults was not possi-ble, because standard errors were not estimated. To be able to compare the µ, κ and α parameters be-tween infants and adults, bootstrapping was used to create confidence intervals around these parameters. This was done by 100 times simulating data based on the parameters and refitting the simulated data. The bootstrapped means and standard deviations of the parameters are displayed in Figure 11. As Figure 11 shows, the µ parameter was estimated very accu-rate and was almost the same for infants and adults. Both infants and adults had two components look-ing to the right (around 0◦), one component looking up (around 90◦) and two components looking to the left (around 180◦). The κ parameters of adults were higher on all components, except for the vertical up component. This is in line with the idea that adult data contains less noise than the infant data, because there was fewer variance in adult saccades than in in-fant saccades. The α parameter showed differences in component one and three. Adults made more sac-cades up, whereas infants made more sacsac-cades to the left. Interpretation of these α parameters should be done carefully, because the adults had another com-ponent that is not displayed in the plot.
To compare these findings with the findings of the binning approach the alpha values were used to com-pare the proportion of horizontal and vertical sac-cades in infants and adults. Although the binning ap-proach showed that adults made more horizontal sac-cades than infants, the Von Mises approach showed the opposite pattern. Ninety-five percent of infant saccades came from a horizontal component, whereas 73% of adult saccades came form a horizontal com-ponent, see the left panel of Figure 12. Furthermore, none of the components in infants and adults was in
Infants Degrees 0.000 0.002 0.004 0.006 0.008 0.010 0 90 180 270 360 Adults Degrees 0.000 0.002 0.004 0.006 0.008 0.010 0 90 180 270 360
Figure 10: Fit of the mixture of the Von Mises distribution on the infant (left panel) and adult (right panel) data 1 2 3 4 5 Mu components Degree 0 50 100 150 200 Adults infants 1 2 3 4 5 Kappa components Kappa 0 20 40 60 80 100 120 Adults infants 1 2 3 4 5 Alpha components Alpha 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Adults infants
Figure 11: Bootstrapped means and standard deviations of the µ, κ and α parameters of infants (light bars) and adults (dark bars). Error bars lie two standard deviations around the mean
Table 4: AIC, BIC and Corresponding Akaike Weights for Models with 1 to 12 Components Fitted on the Infant and Adult Data
Infants Adults
AIC Weight BIC Weight AIC Weight BIC Weight 1 -35.5 0 -22.32 0 -56.71 0 -41.74 0 2 -636.85 0 -603.9 0 -4026.03 0 -3988.61 0 3 -819.42 0 -766.7 0 -5431.32 0 -5371.44 0 4 -850.56 0 -778.07 0.01 -5480.94 0 -5398.61 0 5 -879.31 0.93 -787.05 0.99 -5566.23 0 -5461.45 0 6 -873.32 0.05 -761.29 0 -5605.94 0 -5478.7 0.96 7 -870.01 0.01 -738.21 0 -5622.26 0.01 -5472.57 0.04 8 -861.33 0 -709.76 0 -5616.51 0 -5444.37 0 9 -870.27 0.01 -698.93 0 -5610.51 0 -5415.92 0 10 -856.89 0 -665.78 0 -5632.83 0.99 -5415.78 0 11 -860.11 0 -649.23 0 -5619.44 0 -5379.95 0 12 -854.37 0 -623.73 0 -5592.58 0 -5330.62 0
the oblique direction.
To gain some insight in this remarkable finding, the binning approach was used on the earlier boot-strapped data. This led to almost exactly the same results as the classifying of the real data. The middle panel of Figure 12 shows the bootstrapped data that follows the same pattern as the real data, displayed in the right panel of Figure 12. The bootstrapped data was simulated under the parameters of the mixture of the Von Mises distribution, therefore the propor-tions of the saccade direcpropor-tions should be similar to those in the left panel in Figure 12, but that was not the case at all. This clearly showed that the bin-ning approach can lead to wrong conclusions, there was for instance no oblique component in both the infant and adult model, yet 26% of the saccades in infants and 21% of the saccades in adults was clas-sified as oblique under the binning approach. Fur-thermore, the middle panel of Figure 12 also shows the switch between the proportion of horizontal sac-cades in infants and adults. Altough the 95% of the saccades of infants were horizontal in the mixture of the Von Mises distribution under which the data is simulated, the binning approach only classified 48% of the saccades as horizontal. This mis-classification also appeared in the adult data; of the 73%
horizon-tal saccades 60% was estimated correctly using the binning approach. Because the mis-classification was less severe in adults it seemed infants made fewer hor-izontal saccades than adults, whereas the Von Mises approach suggested otherwise. That the Von Mises approach gives the best description of the data is ar-gued in the discussion.
Discussion
This thesis examined whether infants have the hori-zontal bias and addressed two methodological issues that were a prerequisite for a reliable analysis of in-fant eye-tracking data: (1) whether thresholds to classify fixations and saccades differed between indi-viduals and (2) if interpolation of missing data points and omitting data points outside the stimuli bound-aries changed these thresholds. The results showed that infants do have the horizontal bias; this could be concluded from both the binning as the Von Mises ap-proach. Although these approaches differed in their conclusions to what extent infants have the bias; the binning approach showed adults have a stronger hor-izontal bias than infants, whereas the Von Mises ap-proach showed that infants have a stronger
horizon-0.0 0.2 0.4 0.6 0.8 1.0
Von Mises Approach (Bootstrapped)
Propor tion of Saccades Horizontal Vertical Infants Adults 0.1 0.2 0.3 0.4 0.5 0.6 0.7
Binning Approach (Bootstrapped)
Propor
tion of Saccades
Horizontal Vertical Oblique
Infants Adults 0.1 0.2 0.3 0.4 0.5 0.6 0.7
Binning Approach (Real Data)
Propor
tion of Saccades
Horizontal Vertical Oblique
Infants Adults
Figure 12: Proportion of saccades in the horizontal, vertical and (oblique) directions based on 100 bootstrapped samples, classified by the Von Mises fit (left panel) and classified by the binning approach (middle panel). Error bars in both plots lie two standard deviations around the means. The right panel displays the real data classified by the binning approach (exactly the same plot as Figure 8)
tal bias than adults. This mismatch is likely to be caused by the binning procedure, because binning continuous data can lead to spurious findings, see di-cussion of this point below. The second main finding of this thesis was that thresholds to classify fixations and saccades differed between individuals. Finally it was shown that interpolating missing data points and omitting data points outside the stimuli boundaries did not result in large threshold differences.
Drawing conclusions about the horizontal bias is difficult with the two approaches giving different re-sults. It is known that binning continuous data into groups can lead to spurious results (MacCallum et al., 2002; Royston et al., 2006). To test if the binning ap-proach could also have led to spurious results in this study, a simulation study was conducted. Data sim-ulated under the parameters of the infant and adult Von Mises mixture distributions was indeed misclas-sified when the binning approach was used. This is an indication that the binning approach can lead to spu-rious effects and the Von Mises approach is a better approach to describe this kind of directional data.
This spurious effect most likely occurred because the variances of the components of the mixture of the
Von Mises distributions are different for infants and adults. In infants the variance of the components was large, whereas in adults the variance of the compo-nents was smaller. Many data points of infants were classified as horizontal under the Von Mises approach because of the wide distributions, whereas relatively few data points were classified as horizontal under the binning apporach because of the relatively small bin widths. In adults, the variance of the compo-nents was more in line with the bin widths, making the misclassification less severe. Because the different approaches led to completely different saccade direc-tions in infants, but to small differences in saccade directions in adults, a spurious effect could occur.
Although the binning approach most likely led to spurious results in this study, the horizontal bias re-ported in the literature is unlikely to be a spurious result. As mentioned in the previous paragraph; the spurious effect probably occurred because the classifi-cation methods had a different effect on infants than on adults. Because earlier research focussed solely on adults, it is unlikely that the horizontal bias is a spurious effect. However, future studies could (and should) benefit from using the Von Mises approach to describe saccade directions. Not only does modelling
the distribution give a better description of the data than binning it, it also reduces the chance of find-ing spurious effects and it increases the power (Fe-dorov et al., 2009; MacCallum et al., 2002; Royston et al., 2006). Although it may take some more time than binning, these are some important advantages of the Von Mises approach. Because the von Mises approach gives a better description of the data than the binning approach, the results of the Von Mises approach will be used to further discus the implica-tions of the findings in this thesis.
That the horizontal bias was found in adults is in line with earlier studies (Foulsham et al., 2008; Gilchrist and Harvey, 2006; Tatler and Vincent, 2008), but that infants also have this bias, and even to larger extent than adults, is a novel finding. In the introduction, four theories explaining the bias were discussed; (1) the bias is a laboratory artefact, (2) it is an oculomotor bias, (3) the bias is learned and (4) the bias is caused by image characteristics. The finding that eye movements of infants are almost ex-clusively in the horizontal direction can shed some light on these theories.
First, this study cannot exclude the fact that the bias is a laboratory artefact, because the experimen-tal setting was not controlled to do so. Except for one image, al images were wider than height which may have caused relatively many horizontal saccades. However, it is unlikely the effects found in this study can be fully attributed to the experimental setting, because Foulsham et al. (2011) showed laboratory artefacts only play a minor role in explaining the hor-izontal bias.
Second, Foulsham et al. (2008) rule out the pos-sibility of an oculomotor bias, because people can easily make saccades in all directions. The finding that infants almost exclusively make horizontal sac-cades may imply oculomotor responses do play a role. It may for instance be the case that the eye mus-cles that control the eye in the horizontal direction develop earlier than the muscles in the vertical di-rection. Furthermore, that saccades in all directions can be easily made, does not necessarily imply
ocu-lomotor responses do not play a role. Ocuocu-lomotor responses may still guide eye movements, when there are no other cues to rely on. That may explain the horizontal bias Foulsham et al. (2008) found in the fractal condition. Participants could not recognize fractals they saw earlier and not correctly state the orientation of the fractal; an indication they had little cues to rely on, however they did show the horizontal bias.
Thrid, the explanation of Abed (1991) that the horizontal bias is acquired when learning to read, be-comes highly unlikely. The results of this thesis would suggest that unlearning is a more likely explanation than learning. Infants made almost all their saccades in the horizontal direction and adults may have un-learned this bias. Cultural differences may still play a role in the unlearning process; this could only be properly tested by studying the horizontal bias in in-fants and adults of different cultural backgrounds.
Fourth, Foulsham et al. (2008) and Foulsham and Kingstone (2010) theorized image characteristics may explain the horizontal bias. Image characteristics are assumed to guide eye movements by top-down and bottom-up features. Since infants made almost all their saccades in the horizontal direction, image char-acteristics can only to a small extent guide eye move-ments of infants. This is the case because infants show the bias on all different images, that makes it unlikely that image characteristics guide their eye movements. If top-down and bottom-up processing do explain the bias it would be expected that in-fants had a smaller horizontal bias than adults, be-cause bottom-up processing guides eye movements in adults, but not infants (Amso et al., 2014). Fur-thermore, Althaus and Mareschal (2012) showed top-down processing does not guide eye movements of younger infants, but does guide eye movements of older infants.
However, this does not necessarily imply that im-age characteristics do not play a role in the horizontal bias at all. Kidd et al. (2012) showed infants have a preference for not too simple nor too complex scenes. The natural scenes used in this study may have been
too complex for the infants to process. It may be the case that, because of this complexity infants went in a sort of default viewing mode in which saccades are primarily horizontal. This could be similar to the fractal condition of Foulsham and Kingstone (2010), where the fractals were too complex for adults to pro-cess and saccades were also primarily horizontal.
Most saccades of adults were also in the horizontal direction, but there were enough differences in direc-tions that may be explained by image characteris-tics. Unfortunately, this thesis cloud not study these differences between images because there were not enough data points available to reliably fit the Von Mises distribution per image. Future studies should use the Von Mises distribution to study the influence of image characteristics on the horizontal bias. The parameters of the Von Mises distributions can then be correlated with image characteristics in order to gain more insight in the mechanisms underlying the horizontal bias.
To sum up the discussion about the horizontal bias; laboratory artefacts, oculomotor responses, learning and image characteristics may all influence the hor-izontal bias. Although, they may do this differently than theorized earlier. The finding that infants make almost all their saccades in the horizontal direction may imply that making saccades in the horizontal di-rection is the default mode of making saccades. The explanations discussed above may not facilitate the bias, they may suppress the bias and also ‘allow’ sac-cades in different directions. Future studies may test this hypothesis, for instance by studying if the bias is still present when people look at a black screen or an image with random noise.
Another major finding of this thesis is that thresh-olds to classify fixations and saccades differed be-tween individuals. Although it is known that peo-ple differ in their saccade speeds (Abel et al., 1983), thresholds are often fixed at the same value for all participants. By incorporating these individual dif-ferences the classification of fixations and saccades could probably be improved. This thesis could not test if the classification improves, because the best
way to do so would be by comparing the different classification methods with expert judgements and that goes beyond the scope of this thesis.
Individual differences were found in both the speed and duration thresholds. Both thresholds correlated with age of infants; older infants had lower speed and duration thresholds than younger infants. Further-more, adults had lower speed and duration thresholds than infants. A possible explanation of these findings may be that the eye movements of infants become more precise as they get older. When the eye is still moving a lot when a fixation is made, it is harder to tell where the fixation stops and the saccade begins, than when eye movements within fixations are lim-ited. Because younger infants may have less control over their eye movements, this may explain the corre-lation between thresholds and age. If precision plays a role could be tested by using a variance measure of fixations and correlated that with the thresholds; it would than be expected that lower speed thresholds are a result of less variance in the fixations, whereas high speeds threholds are the result of more variance in the fixations.
Finally this thesis also studied if omitting data points outside stimuli boundaries and interpolation influenced the thresholds. Thresholds became a bit lower when data points outside the stimuli bound-aries were omitted. This is in line with the idea that the thresholds tell something about the precision of eye movements. After omitting the data points out-side the stimuli boundaries, the eye movements are limited within the stimuli and are thus more precise, that results in lower speed thresholds. Thresholds were largely unaffected by interpolation, a possible explanation is that 500 ms may be a too long inter-val to interpolate. Furthermore, linear interpolation may not be ideal; Wass et al. (2013) showed only short sequences (< 150 ms) should be interpolated and instead of linear interpolation they suggest using the mean of the current fixation location to interpo-late.
Interpolation is mainly used to insert data points when a participant blinked. This is done so all data
points can be classified as either fixations or saccades. In this thesis data points were classified as fixations, saccades and missing. That the used procedure al-lowed data points to be missing is very useful be-cause of two reasons: (1) trials do not have to be excluded from the analysis because of missing data and (2) no assumptions have to be made about what exactly happens during a blink, what classifies as a blink and if blinks should be interpolated. Since trials with missing data points do not have to be excluded there is more data to analyze. That is especially use-ful in infant studies where there are a lot of miss-ing data points. Furthermore, separatmiss-ing blinks from looking away in infants is very difficult, it is therefore useful that this is not necessary in this classification method. Although the classification method seemed to work very well, this was not actually tested. Fu-ture research should test this classification method to validate the Mould et al. (2012) algorithm when used on infant data.
To summarize, two main conclusions can be drawn on the basis of this thesis: (1) infants have a stronger horizontal bias than adults and (2) thresholds to sep-arate fixations from saccades differ between individ-uals. Future studies can incorporate these findings in order to improve models that predict eye movements. Almost all knowledge about infant development is de-rived from simple looking time measures (Aslin, 2012, 2007). Being able to use more sophisticated measures like fixation durartions, saccade directions and scan patterns can tell us much more about infant devel-opment in the near future. This thesis contributed to this future by examining an algorithm that makes these more sophisticated measures possible and by introducing a method to describe saccade directions that outperforms the traditional binning approach.
Acknowledgments
I would like to thank Ingmar Visser and Maartje Ray-makers very much for supervising my thesis project. Your comments on earlier drafts improved this thesis a lot. Furthermore, I would like to thank Scott
John-son and his lab staff for conducting the actual study and providing the dataset.
References
Abed, F. (1991). Cultural influences on visual scan-ning patterns. Journal of Cross-Cultural Psychol-ogy, 22(4):525–534.
Abel, L. A., Troost, B. T., and Dell’Osso, L. F. (1983). The effects of age on normal saccadic char-acteristics and their variability. Vision research, 23(1):33–37.
Althaus, N. and Mareschal, D. (2012). Using saliency maps to separate competing processes in infant visual cognition. Child development, 83(4):1122– 1128.
Amso, D., Haas, S., and Markant, J. (2014). An eye tracking investigation of developmental change in bottom-up attention orienting to faces in cluttered natural scenes. PloS one, 9(1):e85701.
Andrews, T. J. and Coppola, D. M. (1999). Idiosyn-cratic characteristics of saccadic eye movements when viewing different visual environments. Vi-sion research, 39(17):2947–2953.
Aslin, R. N. (2007). What’s in a look? Developmental Science, 10(1):48–53.
Aslin, R. N. (2012). Infant eyes: A window on cog-nitive development. Infancy, 17(1):126–140. Bacher, L. F. and Smotherman, W. P. (2004).
Spon-taneous eye blinking in human infants: a review. Developmental psychobiology, 44(2):95–102. Bornstein, M. H., Mash, C., and Arterberry, M. E.
(2011a). Perception of object–context relations: Eye-movement analyses in infants and adults. De-velopmental psychology, 47(2):364.
Bornstein, M. H., Mash, C., and Arterberry, M. E. (2011b). Young infants’ eye movements over “nat-ural” scenes and “experimental” scenes. Infant Be-havior and Development, 34(1):206–210.
Castelhano, M. S. and Henderson, J. M. (2007). Ini-tial scene representations facilitate eye movement guidance in visual search. Journal of Experimental Psychology: Human Perception and Performance, 33(4):753.
Castelhano, M. S. and Henderson, J. M. (2008). Sta-ble individual differences across images in human saccadic eye movements. Canadian Journal of Ex-perimental Psychology/Revue canadienne de psy-chologie exp´erimentale, 62(1):1.
Fedorov, V., Mannino, F., and Zhang, R. (2009). Consequences of dichotomization. Pharmaceutical Statistics, 8(1):50–61.
Foulsham, T. and Kingstone, A. (2010). Asymme-tries in the direction of saccades during perception of scenes and fractals: Effects of image type and image features. Vision research, 50(8):779–795. Foulsham, T. and Kingstone, A. (2012). Modelling
the influence of central and peripheral information on saccade biases in gaze-contingent scene viewing. Visual Cognition, 20(4-5):546–579.
Foulsham, T., Kingstone, A., and Underwood, G. (2008). Turning the world around: Patterns in sac-cade direction vary with picture orientation. Vision Research, 48(17):1777–1790.
Foulsham, T., Teszka, R., and Kingstone, A. (2011). Saccade control in natural images is shaped by the information visible at fixation: evidence from asymmetric gaze-contingent windows. Attention, Perception, & Psychophysics, 73(1):266–283. Gilchrist, I. D. and Harvey, M. (2006). Evidence for a
systematic component within scan paths in visual search. Visual Cognition, 14(4-8):704–715.
Henderson, J. M. (2003). Human gaze control during real-world scene perception. Trends in cognitive sciences, 7(11):498–504.
Hornik, K. and Gr¨un, B. (2011). movmf: Mix-tures of von mises fisher distributions. R pack-age version 0.0-0, URL http://CRAN. R-project. org/package= movMF.
Itti, L. and Koch, C. (2000). A saliency-based search mechanism for overt and covert shifts of visual at-tention. Vision research, 40(10):1489–1506. Itti, L., Koch, C., Niebur, E., et al. (1998). A model
of saliency-based visual attention for rapid scene analysis. IEEE Transactions on pattern analysis and machine intelligence, 20(11):1254–1259. Kidd, C., Piantadosi, S. T., and Aslin, R. N. (2012).
The goldilocks effect: Human infants allocate at-tention to visual sequences that are neither too sim-ple nor too comsim-plex. PLoS One, 7(5):e36399. MacCallum, R. C., Zhang, S., Preacher, K. J., and
Rucker, D. D. (2002). On the practice of di-chotomization of quantitative variables. Psycho-logical methods, 7(1):19.
Mould, M. S., Foster, D. H., Amano, K., and Oakley, J. P. (2012). A simple nonparametric method for classifying eye fixations. Vision research, 57:18–25. Nuthmann, A. and Henderson, J. M. (2010). Object-based attentional selection in scene viewing. Jour-nal of vision, 10(8).
Nystr¨om, M. and Holmqvist, K. (2010). An adap-tive algorithm for fixation, saccade, and glissade detection in eyetracking data. Behavior research methods, 42(1):188–204.
Pedrotti, M., Lei, S., Dzaack, J., and R¨otting, M. (2011). A data-driven algorithm for offline pupil signal preprocessing and eyeblink detection in low-speed eye-tracking protocols. Behavior research methods, 43(2):372–383.
Poynter, W., Barber, M., Inman, J., and Wiggins, C. (2013). Individuals exhibit idiosyncratic eye-movement behavior profiles across tasks. Vision research, 89:32–38.
R Core Team (2014). R: A Language and Environ-ment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria.
Rayner, K. (2009). Eye movements and attention in reading, scene perception, and visual search.
The quarterly journal of experimental psychology, 62(8):1457–1506.
Rayner, K., Smith, T. J., Malcolm, G. L., and Hen-derson, J. M. (2009). Eye movements and visual encoding during scene perception. Psychological science, 20(1):6–10.
Royston, P., Altman, D. G., and Sauerbrei, W. (2006). Dichotomizing continuous predictors in multiple regression: a bad idea. Statistics in medicine, 25(1):127–141.
Shic, F., Scassellati, B., and Chawarska, K. (2008). The incomplete fixation measure. In Proceedings of the 2008 symposium on Eye tracking research & applications, pages 111–114. ACM.
Tatler, B. W. (2007). The central fixation bias in scene viewing: Selecting an optimal viewing posi-tion independently of motor biases and image fea-ture distributions. Journal of Vision, 7(14):4. Tatler, B. W. and Vincent, B. T. (2008).
System-atic tendencies in scene viewing. Journal of Eye Movement Research, 2(2):1–18.
Torralba, A., Oliva, A., Castelhano, M. S., and Hen-derson, J. M. (2006). Contextual guidance of eye movements and attention in real-world scenes: the role of global features in object search. Psycholog-ical review, 113(4):766.
Wagenmakers, E.-J. and Farrell, S. (2004). Aic model selection using akaike weights. Psychonomic bul-letin & review, 11(1):192–196.
Wass, S., Smith, T. J., and Johnson, M. H. (2013). Parsing eye-tracking data of variable quality to provide accurate fixation duration estimates in infants and adults. Behavior research methods, 45(1):229–250.
