Effects of visual span on reading speed and parafoveal processing in eye movements during sentence reading

Abstract

Abstract

Abstract:

Abstract  The visual span (or “uncrowded window”), which limits the sensory information on each fixation, has been shown to determine reading speed in tasks involving rapid serial visual presentation of single words. The present study investigated whether this is also true for fixation durations during sentence reading when all words are presented at the same time and parafoveal preview of words prior to fixation typically reduces later word-recognition times. If so, a larger visual span may allow more efficient parafoveal processing and thus faster reading. In order to test this hypothesis, visual span profiles (VSPs) were collected from 60 participants and related to data from an eye-tracking reading experiment. The results confirmed a positive relationship between the readers' VSPs and fixation-based reading speed. However, this relationship was not determined by parafoveal processing. There was no evidence that individual differences in VSPs predicted differences in parafoveal preview benefit. Nevertheless, preview benefit correlated with reading speed, suggesting an independent effect on oculomotor control during reading. In summary, the present results indicate a more complex relationship between the visual span, parafoveal processing, and reading speed than initially assumed.

Introduction

Several studies have suggested that the number of identifiable letters without moving the eyes determines the speed of reading (Legge et al., 2007; Pelli et al., 2007; Yu, Cheung, Legge, & Chung, 2007). Gordon E. Legge and colleagues have contributed much to our present understanding of the relationship between the visual span and reading speed. They found, for example, that increasing the visual span through practice resulted in accompanying improvements in reading speed (Chung, Legge, & Cheung, 2004), consistent with the authors' view that the visual span limits the sensory input on each fixation and thereby mediates reading speed. In addition, both visual span and reading speed showed qualitatively similar dependencies on print size and contrast (Legge et al., 2007) and on the spacing between letters (Yu et al., 2007). 

Although Legge, Mansfield, and Chung (2001) took into account the possibility that a combination of visual acuity limitations and letter masking (e.g., crowding) may set the boundaries of the visual span in peripheral vision and thus the limits on reading speed (see also Bouma, 1970), Pelli et al. (2007) argued that crowding and not visual acuity is responsible for the drop-off in reading rate beyond central (foveal) vision. Crowding describes the phenomenon in which objects (e.g., letters) are less accurately identified when surrounded by other objects (see Levi, 2008, for a review). As a matter of fact, the visual span is typically operationalized using a flanked-letter identification task in which subjects are asked to report the letters of a letter trigram presented at different eccentricities around the central fixation position (Bouma, 1973; Legge et al., 2001). In the case of reporting the middle letter, this involves the identification of a letter that is surrounded by distractor letters. Thus, the trigram task can be used to measure the susceptibility to crowding. The visual span based only on identifying the middle letter may be viewed as the uncrowded window around fixation (Pelli et al., 2007), which impairs the speed of reading when it is small. 

RSVP and eye tracking during reading

The main body of research postulating a relationship between the visual span and reading speed comes from studies measuring reading with rapid serial visual presentation (RSVP) methods. Words are presented in isolation one after the other with full experimental control over the presentation time and presentation location of each single word relative to the subject's fixation position. Reading speed is calculated based on the minimum word presentation time at which the subject reads above a given performance threshold (e.g., above 80% words read out loud without error). Moreover, as the words can be presented at different eccentricities relative to the central fixation position, this method also allows measuring reading speed in peripheral vision. However, natural reading differs considerably from the situation in RSVP reading in that both viewing time and viewing location of a word is under full control of the subject. Thus, words are rarely viewed from the same initial fixation position (e.g., readers do not always fixate the word at the word center), and the duration of each fixation varies in order to keep up with changing demands in sentence comprehension. 

Following this definition, natural reading also goes beyond the flash card reading that was used in experiment 2 in Yu et al. (2007). Flash card reading involves the presentation of the whole sentence instead of isolated words and necessitates eye movements. However, although flash card reading abandons experimental control over the subject's fixation location on the words and allows more freedom in planning individual fixation durations, the presentation time of each flash card display remains under experimental control. In the present study, silent reading was investigated online based on monitoring the reader's eye movements while reading single-line sentences for comprehension. Each sentence remained visible until the reader voluntarily gazed into the lower right corner of the monitor to terminate sentence presentation, and reading speed was directly assessed from the fixation durations on the words across all sentences. This method preserved all aspects of the natural gaze selectivity during reading (i.e., selection of fixation locations and fixation durations) while simultaneously permitting the study of the interaction between visual perception, cognitive processing, and oculomotor control at the level of individual reading fixations. 

Perceptual span and fixation-based reading speed

An important theoretical concept in the study of eye movements during reading is the perceptual span. It defines the area from which information is extracted during reading to arrive at the maximum reading speed at which linguistic word-recognition processes and eye movements are optimally coordinated. Research has shown that the perceptual span in reading is quite large, extending from three to four letters to the left up to 14–15 letters to the right of fixation (McConkie & Rayner, 1975; Rayner & Bertera, 1979; Rayner, Well, & Pollatsek, 1980). It is typically assessed with a reading task in which only an experimentally controlled window around the fixation position renders the sentence visible, and outside the window, the sentence is masked (e.g., with Xs). As the window is moving contingent on the gaze position across the sentence, reading speed reaches asymptote when the window size equals the entire visual field from which information is sampled during reading. 

In contrast to the visual span, the perceptual span is based on more coarse information than the letter identity, for example, on information about between-word spaces that are important for oculomotor targeting (see O'Regan, 1990, for a careful distinction between visual and perceptual span). In fact, the (letter) identification span for reading (Underwood & McConkie, 1985) does not exceed seven to eight letters to the right of fixation and is thus more similar in size to the uncrowded window measured by Pelli et al. (2007). However, it is the perceptual span and not the identification span that turned out to be the more prominent concept in explaining the spatiotemporal decisions in eye-movement control during reading. In any case, all these findings indicate the importance of nonfoveal information for reading. Specifically, the fact that reading times become shorter the more information is made available to the right of fixation indicates a relationship between parafoveal processing and fixation-based reading speed. 

The perceptual span and reading speed has mostly been investigated in the context of reading acquisition, comparing the perceptual span sizes of beginning readers at different levels of reading skills with those from skilled readers (Rayner, 1986; for investigations on the identification span, see Häikiö, Bertram, Hyönä, & Niemi, 2009). However, although beginning readers showed somewhat smaller spans, Rayner argued that the observed differences in the perceptual span were not sufficient to explain the slow reading rates of the beginning readers. Recently, Rayner, Slattery, and Bélanger (2010) tested slow (<258 wpm) and fast (>258 wpm) readers with a moving-window task, comparing reading with a one-, two-, and three-word window to the control condition with full sentence view. Slow readers reached asymptote in their reading rate already with a two-word window whereas fast readers could increase their reading speed across all window sizes. Rayner et al. (2010) argued that slow readers had a smaller perceptual span most likely because they spend more time with encoding and processing the fixated word in foveal vision. However, the same would hold if the causality of the argument was reversed: Smaller perceptual spans might have been the cause and not the result of slower reading due to less efficient parafoveal preprocessing with reduced span size. 

Parafoveal processing and preview benefit

Another influential method to study parafoveal processing is to use gaze-contingent display changes of a target word in a sentence-reading task (e.g., Rayner, 1975; for comprehensive reviews on this literature, see Hyönä, 2011; Schotter, Angele, & Rayner, 2012). The basic rationale of such experiments is to control the information about a not-yet-fixated word while it is still in parafoveal vision. Sentences are presented with a target word (i.e., word n + 1) being displayed, for example, as a length-matched random string of letters for all fixations up to the pretarget word n. As the saccade toward word n + 1 crosses the invisible boundary at the end of word n, the nonword preview is replaced with the target word. Fixation durations on the target word are typically shorter when the identical word (i.e., valid preview) was available prior to its fixation compared to when the preview was masked (i.e., invalid preview). Such parafoveal preview benefit suggests the integration of parafoveal information into later foveal word recognition (Inhoff, 1990; Rayner, McConkie, & Ehrlich, 1978). In other words, parafoveal preview benefit increases the efficiency of foveal processing and the speed with which a word can be read when it is finally fixated. Differences in parafoveal preview benefit should thus be a likely factor for individual differences in reading speed. 

There has not been much research on studying individual differences in parafoveal preview benefit—not to mention its relationship to differences in the size of the readers' perceptual span. Two age-comparative studies tested young and older adults in a moving-window experiment (Rayner, Castelhano, & Yang, 2009) and a parafoveal-preview reading experiment (Rayner, Castelhano, & Yang, 2010). Older adults, who are typically the slower readers (Kliegl, Grabner, Rolfs, & Engbert, 2004; Rayner, Reichle, Stroud, Williams, & Pollatsek, 2006), were expected to show smaller perceptual spans in the moving-window task and reduced preview benefit in the parafoveal-preview experiment. Although the results held true for the older adults' perceptual spans being, on average, reduced in their rightward extent (Rayner et al., 2009), the age difference in the preview benefit was less evident and not significant in all fixation measures (Rayner, Castelhano, et al., 2010). Moreover, in a study examining parafoveal preview of even the next word, n + 2, older adults showed a significant benefit of a similar size as that of young readers (Risse & Kliegl, 2011). 

Visual span profiles and individual differences in the boundary paradigm

The evidence summarized above indicates a clear functional dependency between the perceptual span size and the subjects' reading speed at the level of fixation durations. However, the link to parafoveal processing remains not fully understood, and the little evidence there is lacks systematic consistency. In order to investigate the role of parafoveal processing and its relation to perceptual limitations and reading speed, subjects were tested with a flanked-letter identification task, in which they were required to identify only the middle letter of a trigram. In addition, they participated in a parafoveal-preview reading experiment in which their eye movements were recorded during single sentence reading. The data of the latter experiment were initially published with a different focus in Risse and Kliegl (2014). 

The decision to use a perceptual task that measures the visual span rather than the perceptual span was motivated by the fact that the eccentricities that were of interest in the parafoveal-preview experiment were rather small (i.e., the distance from the gaze position in the preboundary word n to the target word n + 1 after the boundary). Thus, individual differences in the visual span should already show noticeable consequences on how much can be parafoveally processed in the present reading experiment. Moreover, the flanked-letter identification task used for measuring the visual span provides important information beyond the size of the span. One major finding with this task is that the identification accuracy across eccentricities can be described as a split Gaussian curve (Legge et al., 2001). The amplitude of such visual span profiles (VSPs) accounts for the reader's visual acuity in foveal vision, and the often-larger standard deviation for the right than for the left side of the curve characterizes the typical right visual field advantage for word-like stimuli (Ducrot & Grainger, 2007; Pollatsek, Bolotzky, Well, & Rayner, 1981). Thus, VSPs provide indicators for both foveal and parafoveal processing (i.e., at the level of letter-identification performance) and offer more information than is typically obtained from the perceptual span assessed with the moving-window task. 

VSPs were fitted for each subject, and parameters were derived that uniquely described the subjects' visual spans. Please note that the VSPs are exclusively based on the identification performance of the middle letter only. The VSP parameters were then used to predict the participant's individual reading speed and target word fixation durations in the eye-tracking reading experiment. The predictions were straightforward: If reading speed is determined by the amount of visual input that can be acquired during each fixation as has been demonstrated using methods other than eye tracking (e.g., Legge et al., 2001; Pelli et al., 2007), participants with a larger visual span should be the faster readers in that they show shorter average fixation durations. In addition, if parafoveal processing of upcoming words (i.e., to the right of fixation) mediates reading speed in more natural reading situations that allow completely unconstrained movements of the eyes, the rightward extent of the visual span should be a better predictor of the reader's speed than the leftward extent. 

Having said that, it may be the case that individual differences in the VSP amplitudes are predominant, and foveal processing efficiency at least partially compensates for a lack of parafoveal efficiency. Therefore, the relationship between VSP parameters and the preview benefit effect size—a more direct indicator of the amount of parafoveal processing during reading—was additionally investigated. Looking at the difference between target word fixation durations in the case of valid preview and target word fixation durations in the case if invalid preview (i.e., the preview benefit) should control for differences in foveal target word viewing times. As such, readers with greater rightward visual span parameters were expected to show a larger preview benefit. 

Method

Subjects

Sixty students from a local high school and the University of Potsdam were participating both in the trigram letter-identification experiment and in the parafoveal-preview reading experiment. The trigram experiment lasted about 20 min whereas the boundary experiment was about 1 hr. Order of experiments was counterbalanced across subjects. Two subjects were excluded from analyses: one who did not complete the trigram experiment and one who performed very badly so that no VSP could be computed. Another subject was excluded during later steps in the analysis, also due to bad performance in the trigram task (see below). The remaining 57 subjects (11 male and 46 female) were, on average, 21 years old (16–33 years; SD = 3.4). 

Apparatus and procedure

Figure 1, left panel, describes the flanked-letter identification task adapted from Legge et al. (2001). From a viewing distance of 60 cm, subjects were instructed to identify the middle letter of a random-letter trigram that could appear at seven different locations on the computer monitor (Iiyama Vision Master Pro 514, 22-in., 41-cm physical monitor width, 1024 × 768 pixels resolution). Trigrams were presented centered on the horizontal meridian with the target letter centered on 0°, 2°, 4°, or 6° of visual angle to the right or to the left of the vertical meridian. Exposure time of the trigrams was fixed to 100 ms at each eccentricity. 

Figure 1

 

Illustration of the two experiments performed by each participant. The left panel shows a schematic description of the task adapted from Legge et al. (2001) to derive individual VSPs. The right panel gives an example for the two preview conditions of the target word n + 1 (in orange) in the eye-tracking experiment during natural reading. Asterisks represent the position of the eyes relative to the invisible boundary (dashed line). More details are provided in the text.

Trigrams were displayed in black Courier bold font with a size of 14 points on a white background. The black font had an average luminance of 2.69 cd/m², and the white background had an average luminance of 204 cd/m². Courier is a monospace font type with a fixed x height, i.e., 11 pixels for each character for the given font size and monitor resolution. The space between letters accounted to 0.04°, and each letter subtended 0.42° of visual angle. Flanker and target letters were randomly selected from the set of 26 letters of the alphabet. Letters could hence repeat within a trigram and between trigrams of successive trials. Trigrams appeared 60 times at each eccentricity with the locations presented in random order. 

Each trial started with a display of 31 hash signs (#) covering the entire area at which letter stimuli could appear. The initial fixation stimulus was centered on the screen midpoint with a capitalized “I” above and below the middle hash sign. Subjects were instructed to fixate the position between the two “I”s throughout the entire experiment even after the fixation stimulus disappeared. After 500 ms, a trigram appeared for 100 ms at one of the seven locations (−6°, −4°, −2°, 0°, 2°, 4°, 6°) on the horizontal meridian followed by a blank screen until the participant pressed one of the 26 letter keys on the keyboard to respond. The key press released the display of the initial fixation stimulus and the start of the next trial. Every 42 trials, subjects were encouraged to make a short break. Pressing the space bar, they started the next block of trials. 

Figure 1, right panel, illustrates the parafoveal-preview reading task (see also Risse & Kliegl, 2014). In this experiment, subjects were instructed to read simple single-line sentences for comprehension while their eye movements were tracked with two head-mounted infrared video cameras interfaced with an EyeLink II system (eye monitoring: binocular; sampling rate: 500 Hz; spatial resolution: 0.01°; SR Research, Osgoode, Ontario, Canada). Sentences were displayed on an Iiyama Vision Master 514 Pro monitor (screen size: 22 in.; screen resolution: 1024 × 768 pixels; monitor refresh rate: 150 Hz). Participants were seated 60 cm in front of the monitor with their heads positioned in a chin rest to reduce head movements. Sentences were presented in black letters on a white background using 14-pt Courier regular as a monospace font (horizontal width per letter: 8 pixels or 0.29° of visual angle, default letter-to-letter spacing). 

Subjects read six practice and 120 test sentences for comprehension with the preview manipulation counterbalanced across subjects. After a fixation control close to the left monitor border, the sentences were displayed on the horizontal midline of the monitor such that the center of the first word replaced the stimulus of the fixation control. Subjects ended each trial by gazing into the lower right corner of the monitor. Each sentence contained a target word with an invisible boundary set at the beginning of the blank space to the left of the target word. As long as both online measured eye positions were to the left of that boundary, a preview for the target word was presented at the parafoveal position. Once the first eye exceeded the boundary, the preview was replaced with the target word. The parafoveal preview was either (a) identical to the target word (i.e., valid preview) or (b) a different word than the target word (i.e., invalid preview). Thus, in the valid preview condition, the target word was replaced by itself when the boundary was crossed. Completion of display changes needed, on average, 8.3 ms after the first eye crossed the invisible boundary (ranging from 5 to 11.7 ms). Sentences were removed post hoc in which the display change occurred after fixation onset on the target word.¹ Subjects were calibrated with a nine-point grid at the beginning of the experiment and afterward every 15 sentences or if there was a repeated error in the fixation control at the beginning of each sentence presentation. 

Data analysis

Data were analyzed using R, a free software for statistical computing (version 3.0.1; R Core Team, 2013). Subjects' VSPs were estimated based on a nonlinear least squares method provided in the “stats” package. More precisely, identification accuracies for the seven trigram locations were fit for each individual subject with an asymmetric or split Gaussian function as suggested by Legge et al. (2001). Identification probabilities P(x) at different eccentricities x were given by  

 

with A, σ_L, and σ_R denoting the amplitude, the standard deviation for the left, and the standard deviation for the right Gaussian curves, respectively. 

Spearman rank correlations were then employed to test the dependency between the VSP parameters, the reading rate, and the parafoveal preview benefit. Spearman correlation is less prone to the influence of outlier participants compared to Pearson correlation. Moreover, it is the more general approach as it is a nonparametric measure that makes no assumptions on the probability distribution of the variables. The conclusions from the analysis were not changed when using Pearson correlation. Importantly, the correlation analysis looked at relationships at a rather global level of reading speed based on subject means. 

Linear mixed-effects models (LMMs) followed up the correlation analyses and were estimated using the lmer function from the “lme4” package (Bates, Maechler, & Bolker, 2013). For each individual sentence in the eye-tracking reading experiment, single fixation durations (SFDs) on the target word n + 1 were predicted as a linear function of the preview condition of word n + 1 before its fixation (pvn1: valid preview vs. invalid preview) and the target word's processing difficulty indicated by its word frequency (fqn1: high frequency vs. low frequency). In addition, the subject-related measures of global reading rate, the foveal VSP area (aF), the right VSP area (aR), and the left VSP area (aL) were submitted as covariates to the model (see below for more details). Because of correlations between the between-subject covariates (e.g., r = .44 for the correlation between aF and aR), collinearity was removed from the VSP measures by regressing them on the set of highly correlated predictors, and the residuals were used as the covariates of interest. The VSP area measures were selected for the reported LMM because they showed the strongest influence in the correlation analyses. However, the results were similar when using any of the other VSP parameters. 

All fixed-effects variables were centered on zero, and the model intercept reflected the average target-word SFD (grand mean) across all subjects and conditions. The deviation of each subject and the deviation of each item (target word) from the grand mean were estimated as crossed random factors. Thus, other between-reader influences, such as differences in linguistic knowledge or reading strategy, were captured in the random intercept for the readers and yielded an estimation of the fixed effects adjusted for such differences. As such, the LMM analysis looked at a local level of reading speed reflected in the temporal decision of how long to fixate the target word. Moreover, in contrast to the correlational approach based on subject means, the LMM recognized the full repeated-measures design based on 3,490 observations and tested all hypotheses of interest in one single analysis. Besides the resulting increase in statistical power, this analysis allowed a direct test of the theoretically interesting third-order interactions between VSP measures, the effect of parafoveal preview, and differences in the global reading rate of each subject. Including all interactions up to the third order, this model contained 64 parameters (fixed effects) and two variance estimates (random effects) resulting in 67° of freedom, including residuals. 

Results

Reader variables

Visual span parameters

Several parameters were derived from the readers' VSPs as indicators of different properties of their visual spans. The VSPs are shown in Figure 2. The average VSP across all subjects is illustrated in red, the highlighted area below the curve corresponds to the region of foveal vision (the central 2° of vision; e.g., Hirsch & Curcio, 1989). The subjects' amplitudes A were constrained to meaningful probability values (0–1) and ranged from .28 to 1.00 (M = .88, SD = .12), indicating the subjects' peak performance at the fixation position. The standard deviations σ_L for the left and σ_R for the right side of the VSP curves were left unconstrained and ranged from 1.71 to 4.94 (M = 2.99, SD = .62) and from 2.08 to 4.73 (M = 3.12, SD = .58), respectively. 

Figure 2

 

Visual span profiles for each subject. Identification accuracies for middle letters in letter trigrams presented at different positions relative to fixation position at 0°. In red, the average split Gaussian curve across subjects. The colored area depicts the region of foveal vision. The example sentence illustrates the mapping of the visual span task to the reading task.

One shortcoming of using the standard deviations in the VSPs as an indicator for parafoveal processing is that a large standard deviation does not necessarily correspond with a high letter-identification performance in parafoveal vision. Given a subject with low vision, his/her poor performance at every eccentricity will translate into a small amplitude A but simultaneously into large standard deviations σ_L and σ_R. Therefore, Yu et al. (2007; see also Legge et al., 2001) quantified the size of the visual span in terms of the information transmitted from the area under the VSP. For the present purpose, the area under the VSPs was additionally split into three regions, indicating foveal vision and parafoveal vision to the left and to the right of fixation. More precisely, the areas were calculated from the individual VSPs as the definite integrals across the intervals (−6°, 6°) for the entire VSP area but also across the intervals (−6°, −1°), (−1°, 1°), and (1°, 6°) for the parafoveal left, the foveal, and the parafoveal right VSP areas, respectively. This allowed testing the selective influence of the different regions of the visual field on fixation-based reading speed and the preview benefit. To further investigate the asymmetry of the visual span, the ratio between the foveal and parafoveal VSP areas was calculated. 

Global reading rate

The global reading rate of each subject was computed directly from the eye-movement data in the reading experiment based on the sentences with valid preview. In these sentences, there was no display change (i.e., the preview was identical to the target and thus replaced by itself) and reading was natural. For each sentence, the average reading time per word was computed as the ratio of the sum of all fixation durations in the sentence (except the initial word) and the number of words minus one in the sentence. After averaging the word reading times per sentence across subjects, reading times (in milliseconds) were converted into reading rate measured as wpm. 

Parafoveal processing parameter

As an indicator for parafoveal processing, the amount of preview benefit was computed for each subject as the difference in mean fixation durations on word n + 1 in the case of invalid preview compared to valid preview of word n + 1 before its fixation. Thus, preview benefit should translate into a positive difference. Typically, more useful preview is gained from easy words that are frequently used in a given language rather than from more difficult and less frequent words (e.g., Inhoff & Rayner, 1986; Kennison & Clifton, 1995). Therefore, preview benefit was computed for high-frequency words n + 1 and for low-frequency words n + 1 separately. Results are reported for SFDs only. This implies all cases in which a single fixation on word n + 1 was sufficient for word recognition and integration processes to not program a refixation of that word in the first pass of reading the sentence. Results on other fixation-duration measures were similar and confirmed the present conclusions. Figure 3a shows the preview benefit for each subject ordered by magnitude. 

Figure 3

 

Correlation results for visual-span parameters, reading rate, and preview benefit. (a) Subjects sorted by preview-benefit effect size measured in the eye-tracking experiment for high-frequency (HF) and low-frequency (LF) target words n + 1. (b) Correlation coefficients for reading rate and preview benefit of HF and LF targets with the letter-identification performance at different eccentricities. Correlations outside the highlighted areas are significant at the 5% alpha level (uncorrected for multiple tests). None of the correlations remain significant after Bonferroni correction. (c) Table of Spearman ρs for correlations of VSP parameters with reading rate and preview benefit. Significant results are highlighted in bold (** p < 0.01, * p < 0.05). FPR: foveal-parafoveal ratio.

Correlation between visual span, reading speed, and preview benefit

If the visual span determined reading speed also at the level of fixation durations, we should observe substantial correlations between the visual span parameters and the subjects' global reading rate. The strongest correlation was observed for the VSP amplitude and reading rate (see correlation table; Figure 3c), indicating that subjects with higher letter-identification accuracy at the central fixation position were the faster readers. The standard deviations of the subjects' VSPs correlated only very weakly with the reading rate. However, looking at the measures of the VSP area, the entire VSP area showed very similar results than obtained for the VSP amplitude. In addition, there were moderate correlations for the foveal and the right parafoveal VSP region with only the latter reaching statistical significance. The ratio between foveal and parafoveal areas was not significantly correlated with reading speed. 

Figure 3b shows the correlations of fixation-based reading rate with the letter-identification performance for different locations in the visual span. The illustration confirms that the subject's average reading speed in the eye-tracking reading experiment was most strongly related to the reader's letter-identification performance in foveal vision. The correlations rapidly decreased outside the narrow foveal VSP region, which comprised identification accuracies at locations ranging from −1° to 1° around the central fixation position. Moreover, the correlation strength dropped off more quickly in the left visual field than in the right visual field. 

As a next step, the relationship between VSP parameters and parafoveal processing was examined. If the size of the visual span, particularly into reading direction, is conditional on parafoveal processing during reading, a strong correlation between parameters characterizing the right side of the VSP and the subjects' preview benefit should be expected. However, there were no such correlations between any of the VSP measurements and the size of the preview benefit. Yet, in the case of high-frequency words, the n + 1 preview benefit correlated with reading speed. Fast readers with a high reading rate showed a smaller preview benefit from high-frequency targets than slow readers having a low reading rate. This is the first evidence that parafoveal processing is related to reading speed, but the relationship was opposite than expected. Given that the visual span determines reading speed because a larger span size reduces the time needed to read words that were previously presented in parafoveal vision (Legge et al., 2007; Pelli et al., 2007; Rayner, Slattery, et al., 2010) fast readers should show greater benefit from parafoveal preview than slower readers. However, the present finding suggests less preview benefit for fast readers with a high reading rate. 

Analysis of SFDs with LMMs

Individual differences in reading speed were additionally tested in an LMM predicting fixation durations on the target word n + 1 in the eye-tracking reading experiment. Table 1 summarizes the model coefficients for the intercept, the six main effects, and the significant interactions. The interactions that were not significant are not reported in the table but were kept in the model. Reading rate was significant, confirming that individual differences in the subjects' global reading rate (averaged across all sentences) translated into the local control of fixation durations on the target word. Fast readers yielded shorter SFDs on word n + 1 than slow readers. In contrast, there was no evidence for a main effect of the VSP areas (aF, aL, aR). Nevertheless, target-word reading times were significantly shorter when valid preview was available of word n + 1 prior to fixation compared to when preview was denied, indicating the expected main effect of preview benefit (pvn1). The frequency main effect (fqn1) was not significant. 

Table 1

 

Summary of fixed-effects estimates (b: regression slopes), their standard errors (SE), and t values (b/SE). Notes: For the random effects, variances (var) and standard deviations (SD) are provided. Significance (p) of parameters was derived from nested model comparisons with likelihood ratio tests. Note that only significant interactions are reported although the subordinate interactions were kept in the model.

				p(>χ2)
Intercept (Intc.)	5.4	.02	301.2	<0.001***
Preview (pvn1)	.04	.01	4.25	<0.001***
fqn1:pvn1	−.14	.02	−7.32	<0.001***
Reading rate (rr)	−.001	<.01	−6.69	<0.001***
rr:aR:pvn1	−.01	<.01	−2.91	0.004***
rr:fqn1:pvn1	.001	<.01	3.08	0.002***
VSP area foveal (aF)	−.01	.43	−.03	0.975***
aF:aR:pvn1	13.1	6.2	2.12	0.034***
VSP area right (aR)	−.65	.41	−1.56	0.123***
VSP area left (aL)	−.67	.41	−1.65	0.104***
Frequency (fqn1)	.01	.01	1.29	0.196***
				p(>χ2)
Items:
Intc.	.008	.09		<0.001***
Subjects:
Intc.	.011	.10		<0.001***

Because VSP measures, reading rate, and preview benefit showed only moderate correlations and not a full deterministic relationship in the correlation analysis above, it made sense to also look at their interactions in the LMM. The three significant third-order interactions are illustrated in Figure 4. Most noteworthy, readers with a small right VSP area showed reliable preview benefit irrespective of whether they were slower or faster readers (i.e., a constant difference between the solid and the dashed yellow lines in Figure 4a). In contrast, readers with a large right VSP area showed almost no preview benefit when they were fast readers but large preview benefit when they were slow readers. In fact, Figure 4a suggests that slow readers with a large right VSP area slowed their fixation durations particularly in response to an invalid preview of word n + 1. 

Figure 4

 

Higher-order interactions in the LMM. (a) Target-word fixation durations after valid and invalid preview as a function of the subjects' global reading rate and right VSP area. For illustration purposes, the continuous variable of the right VSP area (aR) was dichotomized by median split. (b) Fixation durations on high-frequency (HF) and low-frequency (LF) target words n + 1 after valid and invalid preview as a function of each subject's global reading rate. (c) Fixation durations for preview validity as a function of the foveal and right parafoveal VSP area. The continuous variable aR was dichotomized by median split.

Figure 4b shows the preview benefit (i.e., shorter SFDs for valid preview) only for high-frequency targets and slower readers. Faster readers' fixation durations were generally shorter and did not differ between cases of valid and invalid preview, irrespective of the target word's frequency. This interaction nicely replicated the finding from the correlation analysis reported above. In addition to the correlation analysis, the LMM revealed that the larger preview benefit for slow readers was indeed ascribed to a reduction in fixation durations in case of valid preview. Finally, Figure 4c shows that readers with a small foveal but large parafoveal VSP area were less likely to modulate their target-word fixation durations with respect to the preview validity. 

Discussion

Limiting the sensory input on each fixation, the visual span (or “uncrowded window”) is assumed to determine the speed of reading, mostly investigated with tasks that involve RSVP of single words (Legge et al., 2007; Pelli et al., 2007). The present study tested whether the same relationship was true for a more natural reading situation in which all words of a sentence were presented at the same time and the reader controlled when and where a word was fixated. In contrast to Yu et al. (2007; experiment 2), the reader's eye movements were recorded and reading speed was mainly determined by the duration of individual fixations on the words. Moreover, eye-tracking reading studies have shown that parafoveal preview of a word prior to its fixation reduces later word-recognition times. The present study further tested whether the amount of such preview benefit is directly linked to the size of the subject's visual span and thus mediates individual differences in reading speed. 

Visual-span differences in fixation-based reading speed

Participants showed substantial individual variability in their VSPs based on identification of the middle letter in a trigram task adapted from Legge et al. (2001) and also in their overall reading rate measured in an eye-tracking sentence-reading task. The results from a correlation analysis on reader averages and from a LMM on single target-word fixation durations converged to the following picture: VSP parameters were predictive for individual differences in the subjects' reading rates also at the level of fixation durations in the natural reading situation, supporting the conclusion from Legge et al. (2007) and Pelli et al. (2007) that reading speed is controlled by the number of identifiable letters on each fixation. This is also in line with findings from reading experiments varying the window-size of visible text gaze-contingent on the reader's fixation position. Such experiments suggested that slower readers have a smaller perceptual span than faster readers (Rayner, Slattery, et al., 2010). 

However, the correlations in the present study were relatively small. This may reflect the fact that, in unconstrained reading situations, readers do not proceed at their speed limit but attempt to find an optimal balance between visual input and sentence comprehension. In other words, reading for comprehension and not for speed may increase the relative importance of nonvisual, potentially linguistic, information in controlling the durations of reading fixations. Nevertheless, the present results confirm a positive relationship between the visual span and reading speed, and this relationship was obtained even for unconstrained reading and at the level of the temporal decisions of when to move the eyes further in the text. 

Moreover, dividing the visual span into three regions, the present results mainly indicated a relationship between reading speed and the readers' foveal identification performance. This seems not surprising given that the fovea yields the highest visual acuity and the reader could freely choose to fixate the words and to bring them into foveal vision for a fast inspection. There was also some indication that the right side of the VSP correlated with the subjects' reading rate. However, this was only the case when looking at the (partial) area below the VSPs but not so for the estimated standard deviation of the subjects' right VSP curves. If there was a strong functional relationship between the rightward extension of the visual span and reading rate, one would have expected a reliable correlation in both VSP measures. This apparent contradiction may be related to the fact that the VSP area depends on both the size of the standard deviation and the amplitude of the VSP curve whereas the standard deviation is, in principle, independent of the estimated amplitude. In other words, the selective finding in the right VSP area may emphasize the particular importance of the foveal letter-identification accuracy on the timing of eye movements during reading. 

Visual-span differences in parafoveal processing during reading

The reading task involved an additional gaze-contingent manipulation of the parafoveal preview of a target word (Rayner, 1975). As expected, local fixation durations on the target word were shorter when valid (informative) rather than invalid (uninformative) preview of that word was obtained in parafoveal vision before its fixation, and the size of this preview benefit varied across readers. Taking the amount of preview benefit as an indicator for the readers' ability to effectively use parafoveal information during reading, a correlation between the preview benefit and particularly the right part of the visual span (e.g., the right VSP area) was predicted. However, this was not observed in the present study. A recent study also found no relationship between the size of the preview benefit and the subject's threshold for identifying the middle letter in a letter trigram (Frömer et al., 2014). However, larger identification thresholds for single letters corresponded with a larger preview benefit at least in some fixation duration measures. Although the task used to measure preview benefit was word list reading and not sentence reading, the results support the present notion that crowding of inner letters within words does not impair parafoveal processing during reading much. 

Nevertheless, preview benefit correlated with the overall reading rate during sentence reading. Surprisingly, larger preview benefit was associated with lower reading rates suggesting that slow readers benefitted more strongly from valid parafoveal information during reading than fast readers. In addition, slow readers with a larger right VSP area seemed to suffer disproportionately from invalid parafoveal preview. One explanation for this surprising result may be that slower readers use parafoveal information more efficiently because they make longer fixation durations that simultaneously increase the time the upcoming words can be preprocessed in parafoveal vision. Such an interpretation seems particularly plausible if it is assumed that during reading, attention is gradually distributed across multiple words, and word recognition proceeds in parallel (e.g., Engbert, Nuthmann, Richter, & Kliegl, 2005; Reilly & Radach, 2006). 

Alternatively, the absence of preview benefit in fast readers may be the result of physiological time constraints in the oculomotor system. Given that saccade programming requires about 150–170 ms (e.g., Abrams & Jonides, 1988; Rayner, Slowiaczek, Clifton, & Bertera, 1983), fast readers may have no opportunity to further reduce their fixation durations on the target word. Assuming that fast readers are the more skilled readers, their (visual) word recognition processes may be highly automatized (Stanovich, 1980) and thus rapid enough to make up for the lack of useful preview simply during the foveal fixation of the target word. At the same time, they would have no additional advantage from valid preview because their fixation durations are already at the lower limit of the saccade latencies. Although this interpretation would not allow any conclusions about the actual amount of preview the fast readers gained in the present experiment, the results clearly indicate that, when reading sentences for comprehension, fixation-based reading speed can benefit from but is not fully determined by parafoveal processing. 

Conclusion

Taken together, the nature of individual differences in the visual span and reading rate appeared more complex than expected in eye-tracking reading experiments. Readers' fixation durations positively correlated with their letter-identification performance in the visual field, particularly in foveal vision. A clear modulating role of parafoveal preview benefit on reading speed, however, could not be established. Some of the findings in the present study were even opposite to expectations from assumptions on the role of the perceptual span during reading. Whether this suggests qualitative differences between the visual span and the perceptual span in the context of eye-movement control in natural reading situations needs to be addressed in future experiments. However, the present study offers a promising approach to combine methods to investigate the visual span (or “uncrowded window”) with eye-tracking research, allowing inferences on the basis of variables with a high time resolution, such as individual fixation durations. For this purpose, LMMs provide an optimal multivariate statistical method to cope with experimental variables and between-subject covariates simultaneously and to reveal additional information about individual differences at the level of fixation durations rather than subject means. 

Acknowledgments

This research was supported by Grants KL 955/6–1 from the Deutsche Forschungsgemeinschaft as part of the Research Group 868 “Computational Modeling of Behavioral, Cognitive, and Neural Dynamics.” Data and R scripts are available at Potsdam Mind Research Repository (http://read.psych.uni-potsdam.de/pmr2) and the Mind Research Repository (http://openscience.uni-leipzig.de/index.php/mr2). 

Commercial relationships: none. 

Corresponding author: Sarah Risse. 

Email: [e-mail]. 

Address: Department of Psychology, University of Potsdam, Potsdam, Germany. 

Footnotes

¹  The influence of noticing display changes on the preview benefit was carefully analyzed in Risse and Kliegl (2014) in which the full details of the boundary experiment are provided. The analyses revealed no evidence for a significant modulation of eye-movement parameters with respect to the subject's self-reported sensitivity toward target word replacements. A similar display change analysis was reported in Hohenstein and Kliegl (2014).