Beneficial influence of in-context predictability when young adults read with a simulated central scotoma

Abstract

Conflicting results have been reported regarding the effect of word predictability when reading with eccentric vision. The present study aims to shed light on these discrepancies by investigating how in-context word predictability influences reading performance with a simulated scotoma, while considering the visual and lexical features of words. Thirty-five healthy young people read aloud sentences presented using the self-paced reading paradigm. A group of 22 participants practiced reading with a 10° diameter, gaze-contingent simulated central scotoma, with the other group serving as controls. Each participant underwent two in-lab sessions, reading 304 sentences (2–4 hours, depending on their group). Reading time, fixation number, and duration were analyzed for each target word using mixed-effect models. When reading with a simulated scotoma, in-context predictability shows a significant effect on performance, with a 35% decrease in reading time for highly predictable words compared with unpredictable ones (2.5 seconds vs. 1.6 seconds). This effect is modulated by practice, with the decrease dropping to 22% (1.3 seconds vs. 1.0 seconds) after only few hours of scotoma exposure. This effect seems to be driven by the total number of fixations required to identify words and is absent in the control group. These results support the hypothesis that reading with eccentric vision, which limits visual access to text, results in a stronger in-context predictability advantage. Moreover, this effect has a greater impact early in eccentric reading practice. This suggests greater reliance on linguistic inferences to compensate for impaired visual input, compared with central reading, at least until functional adaptation occurs.

Introduction

Maculopathies—such as age-related macular degeneration (AMD), Stargardt's disease, diabetic retinopathy, or even some forms of severe myopia—result in central visual field loss (CFL). In industrialized countries, AMD alone affects approximately one in eight individuals over the age of 60 years. Hence, 200 million people worldwide have to cope with CFL because of AMD (Vyawahare & Shinde, 2022), and this rate is increasing constantly. CFL is characterized by the presence of a blind spot located in the center of an individual's visual field and called a scotoma. Because reading requires high visual acuity and steady eye fixation, both of which are provided by central vision, CFL impairs one's capacity to read fluently. As a result, reading difficulty is the most common clinical complaint reported by CFL patients, who have to use eccentric vision (Brown et al., 2014). 

To restore functional reading with CFL, great efforts have been invested to understand what factors could be manipulated to improve reading performance in those visually impaired readers (Legge, Mansfield, & Chung, 2001; Chung, 2020). To date, many visual aspects have been investigated, including font type (Tarita-Nistor, Lam, Brent, Steinbach, & González, 2013; Bernard, Scherlen, & Castet, 2016; Chung & Bernard, 2018; Xiong, Lorsung, Mansfield, Bigelow, & Legge, 2018, Beier, Oderkerk, Bay, & Larsen, 2021), interline spacing (Chung, 2004; Bernard, Scherlen, & Castet, 2007; Calabrèse et al., 2010), text orientation (Subramanian, Yu, Wagoner, & Legge, 2010; Calabrèse, Liu, & Legge, 2017), and text scrolling (Bowers, Woods, & Peli, 2004; Walker, Bryan, Harvey, Riazi, & Anderson, 2016). However, none of these visual factor manipulations seem to yield clinically relevant reading speed improvement (Chung, 2020). 

Recently, a different lead has been investigated, trying to examine the content of a text itself, rather than its sole visual aspects. Indeed, in the presence of a scotoma, bottom-up visual input is less reliable, possibly forcing patients to rely much more on top-down linguistic inference than normally sighted readers (Calabrèse, Bernard, Faure, Hoffart, & Castet, 2016). According to that hypothesis, reading performance in CFL readers should be much more influenced by the linguistic properties of a text than it has been reported before with normally sighted individuals, for which word frequency (the number of occurrences of a word in a given language) and word orthographic neighborhood (the span of words that orthographically differ by only a single letter; e.g. fan, van, man, and pan) are the most influential factors. 

Indeed, word frequency was found to have a much stronger facilitator effect on reading speed in French CFL readers compared with normally sighted readers, even when taking into account visual factors such as word length (Stolowy et al., 2019). This first result confirms the specific influence of lexical factors on reading speed with CFL compared with normal vision, supporting the need for further investigations. Second, word neighborhood size, estimated with Coltheart's N (i.e., the number of words that can be obtained by changing one letter while preserving the identity and positions of the other letters) (Coltheart, Davelaar, Jonasson, & Besner, 1977) was found to have the opposite effect in CFL readers compared with normally sighted readers (Sauvan et al., 2020). For French readers with normal vision, word neighborhood size shows a facilitator effect on reading speed; the more neighbors, the faster a word is identified. For French CFL readers, word neighborhood size showed an opposite inhibitory effect during sentence reading. The authors concluded that, in the presence of a pathological central scotoma, a large neighborhood size tends to decrease reading speed because of the increased confusion yield by the numerous potential word candidates. Thanks to an indirect measure of in-context word predictability, the authors were also able to report that the inhibitory effect of neighborhood size was weaker for words that were highly predictable from the sentence context. The authors hypothesized that, because neighbors rarely share similar meaning, confusion between neighbors would be reduced thanks to contextual information. This result highlights the crucial influence of in-context predictions when reading meaningful text. 

In normal vision, effects of in-context (or contextual) word predictability on reading performance have been documented for many years. In 1995, Bullimore and Bailey (1995) showed that skilled readers read text faster than random strings of words. Mr. Chips., an ideal-observer model of reading, exhibited a 30% increase in reading speed when reading with context cues, confirming a facilitating effect of word predictability in typical readers (Legge et al., 2002). Syntactic and semantic relationships between words make continuous text more predictable than random words. The effect of in-context predictability on normal readers has also been studied from an eye-movement perspective. When a word can be predicted from context, readers often skip over it or spend relatively little time fixating on it (Ehrlich & Rayner, 1981). Computational models of eye movements in reading such as E-Z Reader (Reichle, Pollatsek, Fisher, & Rayner, 1998) and SWIFT (Engbert, Nuthmann, Richter, & Kliegl, 2005) integrate predictability as one of the two main linguistic variables that influence eye movement decisions (when and where to move the eyes), with the other one being word frequency. 

In the case of CFL, which requires reading with eccentric vision, in-context word predictability has also been studied, yielding discrepant results and several potential hypotheses. One of them suggests that decoding visually degraded input is costly for individuals with CFL, who may then have fewer cognitive resources available for contextual analysis compared with normal readers. This hypothesis is supported by results from Sass, Legge, and Lee (2006), showing that removal of sentence context through word scrambling had a greater negative impact on normal compared with low vision. However, these results were obtained from a heterogeneous group of low vision individuals, including both CFL and other visual deficits. A second hypothesis suggests that in-context gain is equivalent for CFL and control readers. This hypothesis is supported by empirical data showing a strong increase in reading speed with context, but no significant difference between normal vision and CFL reading (Fine & Peli, 1996) or between central and peripheral reading with normal vision (Fine, Rubin, Hazel, & Petre, 1999). Other authors have also shown a 15% to 30% reading-speed advantage for sentences over random words in a heterogenous group of readers with low vision, which is roughly consistent with findings from controls (Legge, Ross, Luebker, & LaMay, 1989; Legge et al., 2002). A third hypothesis suggests that, if individuals with CFL can decode text well enough, they may rely more on linguistic inferences. In doing so, they would rely more on predictability compared with normal readers and experience a stronger context advantage. This hypothesis is supported by (Bullimore & Bailey, 1995), who reported higher reading speed for meaningful text compared with random words for both control and AMD readers, the advantage being significantly greater for AMD. To date, the influence of in-context predictability on reading performance in the presence of a central scotoma is not identified clearly. Importantly, it is impossible to objectively quantify the amount of adaptation to central field loss in patients tested at one time point during their pathology. Therefore, none of the studies described elsewhere in this article involving pathological CFL took into account the interindividual differences in adapting to the use of eccentric vision, possibly explaining results discrepancies. Given the recent investigations of psycholinguistic word features when reading with CFL and their obvious relationship with predictability, it seems timely to reevaluate the in-context word predictability effects when reading with eccentric vision. 

Because central scotomas force individuals to use their eccentric vision, the normal periphery has been used as a model to study visual behavior with CFL. Recent evidence has demonstrated that simulating CFL can effectively recreate the challenges faced by actual patients with CFL in a variety of daily living tasks (Tjan, Kwon, & Nandy, 2011; Kwon et al., 2012; Kwon, Nandy, & Tjan, 2013; Walsh & Liu, 2014; Maiello, Kwon, & Bex, 2018; Almutleb & Hassan, 2020; Maniglia, Jogin, Visscher, & Seitz, 2020; Maniglia, Visscher, & Seitz, 2023; Biles, Maniglia, Yadav, Vice, & Visscher, 2023), including reading (Barraza-Bernal, Rifai, & Wahl, 2017a, Barraza-Bernal, Rifai, & Wahl, 2017b; Yu & Kwon, 2023). Indeed, scotoma simulation provides a useful proxy for examining eye movements and visual performance alterations induced by CFL, forcing normally sighted individuals to use their eccentric vision (Macnamara, Chen, Schinazi, Saredakis, & Loetscher, 2021). In this context, reading with a gaze-contingent central scotoma is equivalent to reading with eccentric vision (Aguilar & Castet, 2011; Harvey & Walker, 2014; Akthar, Harvey, Subramanian, & Walker, 2021)—the process of relying on the peripheral visual field around the fovea and macula. Yet, the assumption that the use of peripheral vision may be comparable between patients with CFL and healthy individuals is still debated (Chung & Legge, 2025). 

One of the most notable adaptations to eccentric vision is to redirect the eyes such that targets fall onto a peripheral location of the retina outside the scotoma, referred as the preferred retinal locus (PRL) (Timberlake et al., 1986; Tarita-Nistor et al., 2023). Despite the development of one or more PRL(s), the patients’ functional vision is impaired and eye movement patterns are affected, exhibiting: fixation instability, longer saccade latency, and smaller saccade amplitude (Seiple, Szlyk, McMahon, Pulido, & Fishman, 2005; Calabrèse et al., 2011, Calabrèse et al., 2014; Kumar & Chung, 2014; Calabrèse et al., 2016). In the context of simulated scotoma, eccentric viewing adaptation strategies have also been reported in healthy individuals, allowing for a stable PRL to emerge within a few hours of exposure to the scotoma (Kwon et al., 2013; Liu & Kwon, 2016; Chen et al., 2019; Yu & Kwon, 2023). Recent findings suggest that the development of such a PRL, in response to an artificial scotoma, may represent a strategy, rather than a genuine visuomotor adaptation. In other words, normally sighted subjects forced to read with an imposed artificial scotoma would learn the necessary fovea offset (in amplitude and direction) required to deal with the occlusion of their visual field, rather than really adopting a non-foveal location as the PRL, through re-referencing (Ağaoğlu, Fung, & Chung, 2022). Still, using a simulated scotoma with naive healthy subjects presents the advantage to control for the amount of time spent learning to read with eccentric vision and estimate how it may affect performance. 

Therefore, the goal of the present work was to investigate the effect of in-context word predictability when reading with eccentric vison, using a simulated central scotoma, while taking into account word lexical features. A multifactorial approach was chosen, accounting for the fact that reading limitations with eccentric vision are likely to result from multiple contributing factors. We hypothesize that eccentric vision yields a stronger in-context predictability advantage, compared with central vision, because readers must rely more on linguistic inferences to compensate for the degraded visual input, at least until they adapt their behavior to the presence of the simulated scotoma. To investigate this hypothesis, we examine the influence of predictability on reading accuracy and reading time. Given that changes in reading time can be attributed to either a change in fixation number, a change in fixation duration, or both, we also analyzed these two measures to identify the underlying factors of performance changes. Each of these analyses takes into account the lexical and visual word features that have been shown to have a significant influence on reading speed in individuals with acquired CFL (Stolowy et al., 2019; Sauvan et al., 2020). Importantly, we wanted to test whether predictability effects were influenced by the individuals’ fluency and their adaptation to the presence of the scotoma. Indeed, for our hypothesis to be validated, we must show that the amplitude of the in-context predictability effect decreases as readers adjust to the presence of the scotoma and need to rely less on inference. Therefore, we tested normally sighted participants to whom we simulated a gaze-contingent central scotoma over several practice sessions, to record performance over the adaptation process since the first exposure to the simulated scotoma. 

Methods

Participants

Forty-four healthy participants (27 females) were recruited between March 2022 and June 2023. Among all recruited participants, four were excluded from the study owing to poor recording of their eye tracking data. The mean age of the remaining 40 individuals who participated was 23.5 ± 2.8 years old at the time of the experiment. All participants had normal visual acuity (20/20 or better) with their prescribed correction and were tested for ocular dominance. All of them had French as their first language and no diagnosis of either speech or neurological pathology. Written informed consent was obtained from all participants before starting the experiment. The research was approved by the Human Protection Committee of the French National Center for Scientific Research (n*2023-A00152-43). 

Procedure

All tests were done monocularly with the dominant eye while the contralateral eye was patched. Near visual acuity was measured monocularly with the ETDRS acuity chart on an iPad using the FLEX visual acuity app (Konan's Chart2020). Maximum reading speed (MRS) and critical print size (CPS) were estimated monocularly using the MNREAD iPad app (Calabrèse et al., 2018; Mansfield, Ahn, Legge, & Luebker, 1993). Visual acuity, MRS and CPS were compared with baseline measures for age-matched controls to include only participants’ whose performance fall within normal range. During the main experiment, participants were asked to read aloud short sentences displayed on a screen in front of them (Figure 1). At the beginning of the experiment, each participant was assigned to one of two groups: the scotoma group (n = 27), who had to read the sentences with a gaze-contingent artificial scotoma, or the control group (n = 13), who read the sentences without any specific manipulation of the display. This number was chosen optimally after conducting a preliminary study (n = 15; not reported in the present work) to estimate how many sentence trials were required before participants started to experience stable reading time when reading with a 10° simulated scotoma under the self-paced reading paradigm. Reading all 19 blocks took on average 2 hours for the controls and 4 hours with the simulated scotoma and was divided across two in-lab sessions, separated by less than 2 weeks (mean, 3.9 ± 2.7 days). 

Figure 1.

 

Reading material. The French sentence corpus developed for Albrengues, Lavigne, Aguilar, Castet, and Vitu (2019) was used in this experiment. (a) Within a pair, sentences were composed of identical word sequences differing by only one word (the prime word), which was either semantically related to the target word (predictable condition) or semantically unrelated to the target word (unpredictable condition). (b) From the 304 pairs of sentences, we created two different sets containing one or the other sentence from each pair, with an equal amount of predictable and unpredictable sentences. Each subject was assigned to one sentence set.

Reading material

We selected 304 pairs of French sentences from a larger corpus developed by Albrengues et al. (2019). The sentences we used contained 31 to 69 characters (mean, 50.6 ± 7.3) and 6 to 14 words (mean, 9.0 ± 1.4). Each sentence contained both a prime and a target word, with the prime word always appearing in second or third position and the target word appearing on average 3.8 words later, but never at the last position. The two sentences of a given pair were identical except for the prime word, which was either semantically related or unrelated to the target word (Figure 1). Hence, each target word was presented in a context in which it was highly predictable and a context in which it was rather unpredictable. A precise measure of predictability, ranging from 0 to 1, was estimated for each target word within a given sentence in a previous study using a cumulative cloze task (Albrengues et al., 2019). Thus, each target word had been tagged with two predictability scores. In the unpredictable condition, predictability score of the target word ranged from 0.00 to 0.04 (mean, 0.005 ± 0.013), meaning that the word was correctly predicted by 0% to 4% of the participants tested. In the predictable condition, the predictability score ranged from 0.22 to 1.00 (mean, 0.67 ± 0.23), meaning that the word was correctly predicted by 22% to 100% of the participants tested. Target words ranged from 2 to 13 letters long (mean, 6.06 ± 1.96), with a frequency ranging from 0 to 1,031 occurrences/million (mean, 66.93 ± 124.5) and a number of orthographic neighbors ranging from 1 to 25 (mean, 3.83 ± 4.85). 

Experimental design and stimuli

Each participant read only one sentence per pair, with an equal number of predictable and unpredictable sentences. Sentences were organized into 19 blocks of 16 sentences each (Figure 2). Within a block, sentences were presented in a random order. Sentences were presented on a 21-inch CRT monitor with a refresh rate of 120 Hz using the PsychoPy library (Peirce, 2007, Peirce, 2009). At a viewing distance of 40 cm, the display area of the monitor was 56° × 42° (1,152 × 864 pixels). Characters were black (0 cd/m²) on a gray background (32 cd/m²), displayed in Courier font. Based on previous measurements of CPS in peripheral vision (Chung, Mansfield, & Legge, 1998; Chung, Legge, & Cheung, 2004), we chose an optimal letter size of 1.3°. This print size was adopted because it allowed to fit long sentences on the screen while subtending approximately twice the CPS for normal vision at 5° in eccentricity and once the CPS at 10°. Sentences were left justified, vertically centered on the screen and displayed over two to four lines depending on their length. Participants were instructed to read aloud as quickly and accurately as possible once the sentence appeared. Sentences were presented using the self-paced reading paradigm (Just, Carpenter, & Woolley, 1982), in which all words are present on the screen at the onset of the trial but are replaced by strings of xs (Figure 2). To read, participants have to reveal each word successively with a keyboard press. Within a sentence, only one word is available at a given time, and previously revealed words are masked again as the participant progressed. This paradigm was chosen because it requires to progress through lines of text, yielding eye movements quite similar to natural reading, while enabling the experimenter to identify which word is read at all times, representing a crucial asset when analyzing eye movement behavior in eccentric reading. Although only one word is visible at any given time (preventing readers from making use of full word preview from parafoveal processing), the overall structure of the sentence is preserved, including spaces between words. This feature is extremely important because it allows readers to demarcate words, estimating the length of upcoming words and therefore making predictions about them. Overall, the self-paced reading paradigm induces an average 44% increase in reading time compared with natural sentence reading, both with central and eccentric vision (unpublished data collected in a group of 57 young healthy participants). At the end of each trial, once participants finished reading the sentence out loud, the experimenter ended the trial with a button press. For each target word, reading time (in seconds) was recorded automatically as the duration of its unmasked period. Accuracy was also recorded for each target word. 

Figure 2.

 

Experimental design and self-paced reading paradigm. (a) Sentences were presented within experimental blocks of 16 trials where 8 predictable and 8 unpredictable sentences were randomly displayed. Each participant read 9 blocks during the first testing session and 10 blocks during the second session. (b) Sentences were presented with the self-paced reading paradigm, where the whole sentence is displayed at the onset of a trial, but words are masked by strings of xs. Participants had to click on a keyboard in front of them to unmask each word sequentially, until they read the last word. Once a word was unmasked, the previous one was again replaced by xs, so that only one word was readable at any given time. Participants in the control group read with no manipulation of the visual display, while participants in the simulated scotoma group read with a gaze-contingent 10° diameter circle, textured with random white noise.

Eye movement recording and scotoma simulator

Gaze position of all participants was recorded monocularly using an EyeLink 1000 eye-tracker (SR Research Ltd., Kanat, Ontario, Canada) in head-fixed mode using the tower mount configuration. Eye position was estimated 500 times per second, as the center of the best-fit ellipse around the pupil. A 9-point calibration was performed at the beginning of each experimental block, followed by a 9-point validation. Calibration and/or validation were repeated until the validation error was less than 1.5° on average and less than 2.0° for the worst point. For participants recruited in the scotoma group, eye position data were used in real time, following recommendations from Aguilar and Castet (2011), to display a gaze-contingent mask on the screen that was always centered on the gaze. This central simulated scotoma was a 10° diameter circle, textured with random white noise, forcing participants to read with eccentric vision. Ocular data were extracted and processed offline in R (R Core Team, 2024) using the FDBeye and eyelinker packages. For each participant and each target word, we extracted the total number of fixations and their duration (in milliseconds). Fixations shorter than 80 ms or longer than 1,000 ms were not taken into account. Before data recording, all participants were presented with a few practice sentences displayed with the self-paced reading paradigm without scotoma, to ensure they were familiar with the protocol and understood how to perform the task. Participants in the scotoma group did not have any prior exposure to the simulated scotoma. 

While preprocessing eye position data, we found that in many instances (18% of the total number of fixations) participants had clicked to display the upcoming word while maintaining a steady fixation. In such cases, one single fixation would occur during the presentation of two consecutive words: during the last milliseconds of word n and the first ones of word n + 1. These in-between words fixations were split to assign each word, n and n + 1, with the most accurate fixation number and duration, taking into account the actual time each word was on (see Figure 3 for a detailed example). 

Figure 3.

 

In-between word fixations processing. In this example, fixation 3 is an in-between fixation, that starts during the word n “store” and ends during the word n + 1 “sells.” The initial in-between fixation (fixation 3) lasts 250 ms. However, 40% of this fixation occurs during the display of the word n and 60% during the display of the word n + 1. During processing (right), its duration is split at the onset of the word n + 1, which is set at 300 ms. The split leads to a fixation of 100 ms for n and 150 ms for n + 1. The count of fixation is one for all regular fixations and corresponds with a percentage for in-between-word fixations; here, 0.4 and 0.6 correspond with the 40% and 60% of the fixations’ repartition.

Adaptation to the simulated scotoma

Overall, participants in the scotoma group read between 2.5 and 4.4 hours with a simulated scotoma (mean, 3.3 ± 0.5). This duration was the time necessary to go through the 304 trials and was dependent on each reader's fluency. To verify that this amount of exposure was long enough for participants to adapt to the relatively large 10° scotoma and experience stable performance, individual learning curves were plotted and analyzed using segmented regression (also called broken-line or piecewise regression) (Seber & Wild, 1989). Such regression may provide a slightly less optimal fit than an exponential model. However, its most attractive feature is the estimation of a breakpoint (also called knot or joint point, which represents the critical value where a change in slope occurs), along with a direct reading of the slope of the surrounding phases. This feature makes it very well-suited for the analysis of performance curves (Cudeck & Harring, 2010). For each participant, a curve of average reading time as a function of block number was plotted (Figure 4). For each curve, we first determined a statistical plateau for which all reading speeds measured outside the plateau was at least 1.65 standard deviations slower than the average speed on the plateau (representing a confidence level of 95% with a one-tail distribution). The last point on the plateau was considered a breakpoint and used as a starting point for fitting a piecewise linear regression model in R using the Segmented package. Each individual curve was segmented by fitting a bilinear model. The slope of each segment and the block number at breakpoint were extracted. Learning curves of representative subjects are presented in Figure 4. Overall, 81% of our participants (22/27) experienced more or less steep learning curves followed by a statistically significant plateau of performance across five blocks or more (flat segment which slope was comprised between −0.05 and 0.05). It is likely that these participants have been able to develop an efficient attentional strategy, rather than experiencing a genuine visuomotor adaptation with the development of a PRL-like behavior. Through only a couple hours of practice, they may have learned the necessary fovea offset required to deal with the occlusion of their visual field, being able to improve their fixation stability and overall performance, but without the occurrence of an oculomotor re-referencing. Figure 4 presents data from four representative participants (Figure 4, A–D, top row). The other five participants never experienced a statistically significant plateau of performance at the end of the experiment (Figure 4, E–I, bottom row), suggesting that they may have required more time to adapt efficiently to the presence of the simulated scotoma. To interpret performance changes through a somewhat complete learning course (from first exposure to proficiency), these five participants were excluded from the final analysis. 

Figure 4.

 

Reading time changes throughout the experiment, for participants in the simulated scotoma group. Subplots show mean word reading time as a function of block number for four representative participants included in the analysis (top row) and five participants excluded from the analysis (bottom row). For all nine participants, the mean reading time values (pink and blue circles) and their standard deviation (gray error bars) are superimposed by a bilinear fit (black solid line), which breakpoint is marked by a yellow circle. In green, participants A to D (top row) experienced a similar learning pattern: a more or less steep learning curve (in pink) followed by a statistically significant plateau of performance across five blocks or more (in blue). Plateau was defined as a segment which slope was comprised between −0.05 and 0.05. In red, participants E to I (bottom row), who did not show a significant plateau of performance toward the end of the experiment, were excluded from the final analyses. Average data from the control group (small black circles) are displayed in each subplot for comparison.

Statistical analysis

Statistical analyses were carried out using R (R Core Team, 2024) with the following additional packages: dplyr, tidyr, stringr, lme4, and ggplot2. A total of four analyses were performed to inspect the effect of predictability on: reading accuracy (analysis 1), reading time (analysis 2), fixation number (analysis 3), and fixation duration (analysis 4) of the target word only. Analysis 1 was performed with a generalized linear model. The other three analyses were performed with linear mixed-effects models. For each analysis, several models were first constructed with the target word accuracy (model 1), reading time (model 2), fixation number (model 3), or fixation duration (model 4) as the dependent variable. Word predictability, length, frequency, and neighborhood size were entered as fixed effects, along with the block number and the presence/absence of scotoma. All mixed-effects models included a random intercept for participants, assuming a different baseline performance level for each individual, as well as random intercept for each target word. To satisfy the assumptions of parametric statistical tests (Howell, 2009), variables were inspected and transformed as follows. Dependent variables (reading time, fixation number and fixation duration) were transformed into natural logarithm (ln) units. Independent variables (word length, frequency, neighborhood size and length) were centered around their mean. For each analysis, the Akaike information criterion was used to assess an optimal random-effects structure (Zuur, Ieno, & Elphick, 2010). The significance of the fixed effects in each model was then estimated using t values. Results were considered significant for t-values greater than 2 or less than −2, corresponding with a 5% significance level in a two-tailed test (Gelman & Hill, 2006; Baayen, Davidson, & Bates, 2008). In the Results, fixed-effects estimates are reported along with their t values (Bates, Mächler, Bolker, & Walker, 2015). 

Results

Participants

Table 1 shows demographic information and reading performance measures for the 35 participants included in the analyses. 

Table 1.

 

Characteristics of the 35 participants included in the analysis. Values are mean ± standard deviation.

 

Accuracy (Analysis 1)

Accuracy was excellent across all trials, participants, and conditions (scotoma versus control): 99.9% accuracy in the control group (SD = 0.7) and 99.6% accuracy in the artificial scotoma group (SD = 1.6). Model 1 revealed no significant effect of predictability on accuracy and no significant difference in accuracy between the two conditions (scotoma vs. control). 

Reading time (Analysis 2)

Results from model 2 are plotted in Figure 5 and summarized in Table 2 (full table results are provided in the Supplementary Material). For the control group and during the first block of the experiment, participants’ average reading time was 0.60 seconds for unpredictable words and 0.59 seconds for highly predictable words (Figure 5, dashed orange line). This difference was not significant (t = −0.34). As the number of blocks increased, reading time for unpredictable words decreased significantly from 0.60 seconds to 0.34 seconds from first to last block, showing a significant effect of practice in the control group (t = −16.85). Still for the control group, there was no significant interaction between predictability and block number (t = −0.30), indicating that the same effect of practice was observed for predictable words, with a significant decrease from 0.59 to 0.33 seconds from first to last block. Hence, the predictability effect for that group remained null throughout the experiment, and reading time was improved only through practice, regardless of the level of predictability. 

Figure 5.

 

Reading time as a function of in-context word predictability, grouped by condition and block number. Data points show median word reading time for each value of predictability, broken down by condition (scotoma in filled circles vs. control in open triangles) and block number (orange for the first block, red for the middle block, and purple for the last block). Lines represent slopes estimated by model 2, along with their 95% confidence intervals (solid lines for the scotoma group; dashed lines for the control group).

Table 2.

 

Fixed effect estimates from models 2, 3, and 4. All models share the same fixed and random structure, except for the dependent variable, which are respectively word reading time (model 2), fixation number (model 3) and fixation duration (model 4). Models’ estimates are reported in the table, along with their t-values. ^* p < 0.05. ^** p < 0.01. ^*** p < 0.001.

 

In the presence of the simulated scotoma, reading time of unpredictable words increased significantly by a factor of 4.18 compared with the control group (2.51 against 0.60 seconds) during the first block. We also found a significant effect of predictability in the scotoma group (t = −10.72): during the first block, reading time decreased from 2.51 seconds to 1.62 seconds between low and high predictability (Figure 5, solid orange line). Similar to the control group, we found a significant effect of practice when learning to read with the simulated scotoma (t = −2.72): when predictability was at its lowest, reading time decreased from 2.51 seconds in the first block to 1.29 seconds in the last block. On top of that practice effect, we found a significant interaction between predictability, block number, and condition, meaning that the effect of predictability changed over the course of the experiment, but only in the presence of a scotoma: as block number (and therefore practice) increased, the amplitude of the predictability effect decreased significantly (t = 3.10) (Figure 5). Overall, word length and word frequency had a significant effect on reading time in the control group (t = 2.71 and t = −2.47, respectively), which was not significantly different in the scotoma group (t = −1.77 and t = −1.52, respectively). 

Oculomotor pattern changes

As shown in analysis 2, adaptation to the presence of the scotoma significantly decreases reading time over the course of the experiment. This performance improvement is likely to reflect changes in oculomotor behavior, because a decrease in reading time can be attributed to either a decrease in fixation number, a decrease in fixation duration, or a mixture of both. Visual inspection of the eye traces (Figure 6) suggests that reading with a simulated scotoma initially leads to an increased number of fixations and more erratic eye movement patterns. By the final block of the experiment, fixation patterns tend to resemble those of a control participant. We analyzed fixation number (analysis 3) and fixation duration (analysis 4) to identify the underlying factors of these performance changes. 

Figure 6.

 

Eye movement traces from two representative participants: one reading with a simulated scotoma (top row) and a control participant (bottom row). Each panel represents a sentence, displayed across two or three lines, color coded by line: blue (first), green (second), and orange (third). Left panels correspond with the first experimental block and right panels with the final block. Each dot corresponds with a fixation, with horizontal positions normalized between 0 (left) and 1 (right) for each sentence. The y axis (fixation index) indicates the chronological order of fixations.

Fixation number (Analysis 3)

Results from model 3 are plotted in Figure 7 and summarized in Table 2 (full table results are provided in the Supplementary Material). For the control group and during the initial block of the experiment, participants’ average number of fixations per target word was 1.55 for unpredictable words and 1.46 for predictable words (Figure 7, dashed orange line). This difference was not significant (t = −1.19), which indicates that, at the beginning of the experiment, the number of fixations were similar for control participants, regardless of the level of predictability. The number of fixations required to read low-predictability words decreased significantly from 1.55 fixations per word to 1.08 from first to last block. There was no significant interaction between block number and predictability for the control group (t = 0.65), which means that the predictability effect for the control group remained null throughout the experiment and that fixation number was only improved through practice, regardless of the level of predictability. 

Figure 7.

 

Fixation number as a function of in-context word predictability, grouped by condition and block number. Data points show the median fixation number of target words, for each value of predictability, broken down by condition (scotoma in filled circles vs. control in open triangles) and block number (orange for the first block, red for the middle block and purple for the last block). Lines represent slopes estimated by model 3, along with their 95% confidence intervals (solid lines for the scotoma group; dashed lines for the control group).

Overall, there was a significant effect of the presence of the artificial scotoma on the number of fixations (t = 13.49). At the lowest predictability and during the first block of the experiment, the average number of fixations increased from 1.55 in the control condition to 7.24 in the scotoma condition. Contrary to the control group, there was a significant effect of predictability on the average number of fixations in the scotoma group (t = −8.08). Hence, across the first block, whereas low-predictability words required 7.24 fixations, high-predictability words only yielded an average of 4.29 fixations. As for the control group, there was a significant effect of block number when reading with a scotoma (t = −5.58). When predictability was low, the number of fixations decreased from 7.24 in the first block to 3.71 in the last block. Moreover, there was a significant interaction between predictability and block number in the scotoma condition: the higher the block number, the smaller the amplitude of the predictability effect (t = 3.47). 

Finally, among the word characteristics, only the word target length showed a significant effect on fixation number in both groups: the longer the word the greater the number of fixations. Each word length increase of 1 character leads to a significant increase in reading time by a factor of 1.05 in the control group and of 1.02 in the scotoma group. 

Fixation duration (Analysis 4)

Results from model 4 are plotted in Figure 8 and summarized in Table 2 (full table results are provided in the Supplementary Material). We found no significant effect of the presence of the simulated scotoma on the average fixation duration (Table 2) (t = 0.65). During the first block of the experiment, no significant effect of predictability was observed on fixation duration for either the control (t = −0.41) or the scotoma group (t = 0.86). For both groups, the average fixation duration did not significantly decrease from block to block (Figure 8). 

Figure 8.

 

Fixation duration as a function of in-context word predictability, grouped by condition and block number. Data points show the median fixation duration (in milliseconds) of target words, for each value of predictability, broken down by condition (scotoma in filled circles vs. control in open triangles) and block number (orange for the first block, red for the middle block and purple for the last block). Lines represent slopes estimated by model 4, along with their 95% confidence intervals (solid lines for the scotoma group; dashed lines for the control group).

Discussion

The goal of the present work was to investigate the hypothesis that, compared with central vision, eccentric vision yields a stronger in-context predictability advantage because readers rely more on linguistic inferences to compensate for the degraded visual input, at least until they adapt their behavior to the presence of the simulated scotoma. To do so, we forced young normally sighted participants to read with eccentric vision by imposing a gaze-contingent simulated scotoma. We estimated how in-context word predictability affects reading time and accuracy, both with and without a simulated scotoma, and how this effect changes over time and practice while taking into account visual and lexical features of the words. Underlying changes in eye movement patterns were also investigated. 

Our first main result is that in-context predictability shows a strong significant effect on reading time in the presence of a simulated scotoma. Highly predictable words are read more than 1.5 times faster compared with non-predictable ones, representing a 35% increase in reading speed, solely explained by the access and use of contextual information. This first result goes against the hypothesis according which decoding eccentric (and therefore degraded) visual input requires so much cognitive resources that readers do not have enough available left to process and take full advantage of contextual cues (Sass et al., 2006). In contrast, we found no significant effect of in-context predictability for control readers, for whom average reading time remained constant, regardless of the predictability of a word within a given sentence. These results are in favor of the hypothesis suggesting that individuals reading with eccentric vision can use linguistic inference efficiently, experiencing a stronger context advantage compared with normal readers. By extension, one may argue that this reasoning could also apply to individuals with CFL, forced to use eccentric vision to read. This rationale would be in line with the results reported in actual patients by Bullimore and Bailey (1995), who found that contextual cues increase reading speed more dramatically in AMD than in normally sighted readers. However, our results differ on several major points, likely owing to methodological differences. 

First, we found a smaller in-context advantage in the presence of a simulated scotoma, with a 1.5 decrease factor of reading time, where Bullimore and Bailey (1995) reported a 2.9 increase factor of reading speed for in-context words. The nature of the reading material and our distinct ways to estimate the influence of context may explain this difference. In the present study, target words were presented within a grammatically correct sentence and given a precise measure of predictability, ranging from 0 to 1, previously measured with a cumulative cloze task in a large cohort of control readers (Albrengues et al., 2019). Therefore, each word was attributed an empirical measure of predictability, estimated from the semantical and syntactical context of the sentence it was presented in. In contrast, Bullimore and Bailey (1995) compared meaningful texts against series of random words, estimating the overall influence of contextual cues on reading speed, with no specific control over the level of word predictability within texts. Therefore, the present study allows to estimate the influence of in-context predictability on a more precise scale. Additionally, these authors emphasized that average word length varied across their experimental conditions, possibly explaining part of the effect they reported, whereas word characteristics such as length, frequency and even orthographic neighborhood size were controlled for in the present study. Finally, while Bullimore and Bailey (1995) reported results from individuals with pathological CFL, the present work focuses on eccentric reading from scotoma simulation. As already mentioned, differences in the use of peripheral vision between patients with CFL and healthy individuals are subject to debate (Chung & Legge, 2025) and are very likely to explain part of the discrepancy in the predictability advantage amplitude reported between pathological and simulated conditions of eccentric reading. 

The second major difference between our results and those from Bullimore and Bailey relates to the in-context predictability effect reported for the normal vision group. In the present work, control readers do not benefit from highly predictable words. Yet, several studies have demonstrated that predictability facilitates reading in typical readers (Bullimore & Bailey, 1995; Legge et al., 2002). Such discrepancy may be due to the use of different reading paradigms. Indeed, the sequential reading imposed by the self-paced reading paradigm in the present work, may not allow to capture the sensitive predictability effect (in the range of milliseconds) experienced by control readers. We must also consider that different context effects may exist for different text displays. Indeed, for dynamic forms of text presentation, such as rapid serial visual presentation reading and drifting text, stronger context effects have been reported in normal readers (Fine & Peli, 1996). 

Our second main result is that the effect of in-context predictability is modulated by the amount of practice with eccentric reading. In the scotoma group, the amplitude of the predictability effect decreased significantly as exposure time to the scotoma increased. At first exposure to the simulated scotoma, average reading time ranged from 2.51 seconds for low predictability words to 1.62 second for highly predictable words, representing a 35% improvement in reading speed, solely accounted for by predictability influence. At the end of the experiment, this improvement was reduced to 22%, with reading time ranging from 1.29 to 1.00 second. This progressive decrease in amplitude implies that readers rely less and less on linguistic inference, which probably results from their adaptation to eccentric reading and the potential development of compensatory strategies. One could argue that this modulation of the predictability effect through practice might explain the discrepant literature results on in-context gain reported in low-vision individuals (Legge et al., 1989; Fine & Peli, 1996; Fine et al., 1999), for whom it was impossible to control for the amount of practice with eccentric vision. 

In the literature, overall context effects are reported to be smaller in good than in poor readers (Briggs, Austin, & Underwood, 1984). Such a distinction between good and poor readers is used mainly for children learning to read and relates to their developing cognitive and linguistic skills. Still, it pertains to the achievement of proficient reading ability, which is impaired when forced to read with eccentric vision. If we consider our participants forced to read with a simulated scotoma as being poor readers at the beginning of the experiment and good (or at least better) readers toward the end, our results align with the literature. Even more interesting, one could argue that, if participants had been given more exposure time to the scotoma, the amplitude of the effect would have kept decreasing, potentially getting closer to the null effect observed in the control group. This point, which remains to be tested, is in favor of enforcing visual rehabilitation and daily practice of reading for patients with acquired CFL, to reduce the negative influence of low-predictability contexts. 

Characterizing specific eye movement changes in response to eccentric reading adaptation was not the focus of the present work. However, identifying underlying components of performance changes as a function of predictability and adaptation required to describe fixation characteristics. It appears that changes in reading time yield by variations of in-context word predictability could be attributed to changes in the total number of fixations, whereas fixation duration remained constant. Indeed, analyses of reading time (Table 2; Figure 5) and fixation number (Table 2; Figure 7) revealed a similar effect of in-context predictability (both in terms of amplitude and significance level), while no effect was found on fixation duration (Table 2; Figure 8). When reading with eccentric vision, highly predictable words required significantly less fixations than non-predictable words (from 35% less at the beginning of the experiment, to 22% less at the end). This result is in line with findings from numerous studies on young healthy readers, which show that words that are shorter, higher in lexical frequency, or more predictable yield shorter reading times and have lower fixation probabilities (Ehrlich & Rayner, 1981; Kliegl, Grabner, Rolfs, 2004; Rayner, Reichle, Stroud, Williams, & Pollatsek, 2006; Choi, Lowder, Ferreira, Swaab, & Henderson, 2017). Hence, most studies on reading use word skipping as a proxy to estimate the influence of in-context predictability, starting from the postulate that, if the reader can make efficient use of predictability, highly predictable words are more likely to be skipped. Because the self-paced reading imposes to fixate on each word, skipping rates could not be investigated in the present study, preventing us from comparing our results to current literature results on predictability using word skipping. 

Because reading in eccentric vision is very costly, participants could have adopted a risky reading strategy, similar to that observed in healthy older adults (Choi et al., 2017; Paterson et al., 2020), relying more on contextual cues, skipping or partially processing some words, thus making fewer individual word fixations. However, this was not observed in our group of eccentric vision readers. Instead, they consistently prioritized accuracy over speed, carefully reading each word before moving to the next one. The adaptive strategies selected by young, healthy participants for peripheral reading contrast with those typically seen in healthy older adults, who often adopt a risky reading approach to compensate for slower processing. Because of its word-sequential reading nature, the self-paced reading paradigm may have also muted any tendency toward a risky strategy by requiring participants to process each word individually without previewing surrounding words, reinforcing a word-by-word reading approach. 

It is worth mentioning that, aside from the question of predictability, reading with a simulated scotoma showed a strong significant impact on the number of fixations but did not alter significantly the overall fixation durations. These results are consistent with the ones reported by (Scherlen, Bernard, Calabrese, & Castet, 2008), who observed that the number of saccades almost doubled when reading short sentences under simulated central scotomas, while fixation duration was not significantly increased (from 216 ms to 241 ms). Yet, our results differ in amplitude, from the report of a recent investigation of altered eye movements during reading with simulated scotomas (Yu & Kwon, 2023). In this work, the authors showed that when normally sighted individuals had to read a text naturally, compared with reading with a 10° diameter simulated central scotoma and no prior training, their number of fixations increased significantly by 89% and their fixation duration increased significantly by 13% (from 217 ms to 245 ms). In the present study, we report a slightly smaller change in fixation duration, representing a non-significant 10% increase (from 206 ms for the controls to 226 ms with the scotoma, during the first block and for high predictability words). The average fixation number, however, was increased more dramatically in the present work, with a 195% increase when reading with a simulated scotoma. This large difference translated into reading performance, with a reading speed increase factor of 2 between the two conditions in (Yu & Kwon, 2023), against a reading time decrease factor of 3 in our study. These differences could be attributed to the different reading tasks used, involving different processes: sentence versus text reading on one hand, mixed with natural versus self-paced reading on the other one. 

Finally, the present work adds to the numerous studies using gaze-contingent scotoma simulation to investigate eccentric reading. Although PRL establishment can take up to 6 months from the onset of a pathological bilateral central scotoma (Crossland, Culham, Kabanarou, & Rubin, 2005), PRL-like behavior has been reported to develop rapidly (within the range of hours) in normally sighted subjects viewing with an artificial central scotoma (Kwon et al., 2013; Walsh & Liu, 2014; Liu & Kwon, 2016). Here, participants in the scotoma group read between 2.5 and 4.4 hours with a simulated scotoma (mean, 3.3 ± 0.5). Eighty-one percent of them showed learning curves with stable performance for the last five blocks or more (Figure 4), suggesting that this amount of time was sufficient for them to adapt and learn an efficient strategy to deal with the occlusion of their central field, without any evidence that this constitutes a PRL-like behavior. Yet, examining how participants adjust their eye movement patterns to the presence of a simulated scotoma to achieve functional reading should be investigated further to understand how it relates to the longer time course of PRL(s) development in people with naturally occurring central scotomas (Legge & Chung, 2016; Ağaoğlu et al., 2022). 

Conclusions

The results presented here seem to confirm that eccentric reading, which limits optimal visual access to text, yields a stronger in-context predictability advantage compared with central vision, because readers rely more on linguistic inferences to compensate for their weaker visual input. This effect appears to be driven solely by the total number of fixations required to identify words. Moreover, this effect mainly impacts eccentric readers in the early stages of their practice. When interpreting the present results, one must remain aware that the normal periphery in young adults is not a perfect model for reading in the presence of central vision loss (Chung & Legge, 2025) and should be very cautious when making predictions about actual patients with CFL. Still, in the hypothesis that these results could apply to pathological CFL, individuals who have stopped reading or have not developed effective peripheral viewing strategies may be most influenced by context effects. Thus, the need to enforce patients’ visual rehabilitation and daily practice of reading in the early stages of their binocular impairment to promote adaptation and reduce the negative influence of low-predictability contexts (on top of the solely positive effect of practice with eccentric reading). Last, one must consider that the present results may not apply throughout the life span. Indeed, in normal vision, the influence of age on the effect of word predictability has been investigated, with conflicting results. Some studies report no influence of age (Rayner et al., 2006), whereas others suggest a greater facilitation of predictability for younger readers (Kliegl et al., 2004) or, conversely, a greater facilitation effect for older readers (Steen-Baker et al., 2017). Given that most forms of CFL occur later in life, the results reported in the present study remain to be replicated with an older population before making any strong prediction about CFL readers. 

Acknowledgments

The authors thank Françoise Vitu for her help in gathering the linguistic materials for this experiment. 

Supported by the Institut Carnot-Cognition. 

Commercial relationship: none. 

Corresponding author: Aurélie Calabrèse. 

Email: [email protected]. 

Address: CRPN CNRS UMR 7077, Aix-Marseille Univ., Marseille 13284, France.