Faces and words are both associated and dissociated as evidenced by visual problems in dyslexia

Faces and words are traditionally assumed to be independently processed. Dyslexia is also traditionally thought to be a non-visual deficit. Counter to both ideas, face perception deficits in dyslexia have been reported. Others report no such deficits. We sought to resolve this discrepancy. 60 adults participated in the study (24 dyslexic, 36 typical readers). Feature-based processing and configural or global form processing of faces was measured with a face matching task. Opposite laterality effects in these tasks, dependent on left–right orientation of faces, supported that they tapped into separable visual mechanisms. Dyslexic readers tended to be poorer than typical readers at feature-based face matching while no differences were found for global form face matching. We conclude that word and face perception are associated when the latter requires the processing of visual features of a face, while processing the global form of faces apparently shares minimal—if any—resources with visual word processing. The current results indicate that visual word and face processing are both associated and dissociated—but this depends on what visual mechanisms are task-relevant. We suggest that reading deficits could stem from multiple factors, and that one such factor is a problem with feature-based processing of visual objects.

lists here instead of combining them, reading speed and accuracy values for a missing list were dropped instead of imputed from the reading speed and accuracy of the two other lists, unlike in the main manuscript.

S2. Laterality Effects
All other things equal, a face will be processed more efficiently if it appears in the left visual field (projecting to the right hemisphere) compared to the right visual field (projecting to the left hemisphere; for discussion on contralateral biases in high-level visual cortex, see e.g. Hemond, Kanwisher, & De Beeck, 2007;Kay, Weiner, & Grill-Spector, 2015;Pitcher et al., 2019). People are also generally likely to report that a chimeric face made of two left halves, rather than right halves, of an original face image resembles the original, presumably again because the left half of the original face image tends to fall within the left visual field (for a review on lateralization of face processing, see Sigurdardottir & Jozranjbar, 2019). However, such effects appear to be dependent on the type of processing involved. A right visual field (left hemisphere) advantage in a same-different task for faces has been reported when faces can only be distinguished based on a single feature (e.g. different eyes, same nose and mouth; Fairweather et al., 1982;Hillger & Koenig, 1991) or possibly one or two features (Sergent, 1982), and has been attributed to the requirements for parsing and analyzing the local elements of a face (Hillger & Koenig, 1991).
In the current study, faces were always shown simultaneously in three locations (left, center, right) and their orientation could be toward the left, right, or straight ahead. Average performance in feature-based face matching was noticeably better on right-facing than left-facing trials. The opposite was true for global form face matching, where people tended to perform better on left-facing as compared to rightfacing trials. There are at least two possible drivers of these effects (assuming that they replicate). It is possible that the main diagnostic "face" part of a left-facing central sample stimulus tends to fall within the left visual field, and that this in turn triggers preferential right hemisphere processing, while the opposite could be true for a right-facing sample. Another possibility is that faces oriented to the left engage attentional deployment toward the left visual field, while a rightward orientation of faces directs attention to the right visual field, as gaze evokes fast reflexive attentional shifts (Driver et al., 1999;Friesen & Kingstone, 1998; see also Sigurdardottir, Michalak, & Sheinberg, 2014). Lateralized attentional deployment could suffice to engage partially lateralized feature-based vs. configural processing mechanisms as neurons high up in the ventral visual stream mainly process objects at the center of attention (Brooks, Sigurdardottir, & Sheinberg, 2014).
Laterality effects in face matching were found through exploratory data analysis and should be interpreted with caution, including the following exploratory analyses. between reading speed (list 1: IS-FORM common word forms; list 2: IS-FORM uncommon word forms; list 3: IS-PSEUDO pseudowords) and laterality effects (percent correct for leftfacing stimuli minus percent correct for right-facing stimuli) for feature-based face matching (featural laterality) and global form face matching (global laterality). Blue correlation coefficients are significant, and red are non-significant according to traditional criteria, but readers should keep in mind that this analysis is exploratory, and that the dyslexic group has fewer participants.
As mentioned in the main manuscript, deficient feature-based processing in dyslexic readers appeared to be independent from any possible group differences in feature-based face processing laterality effects. Feature-based laterality effect might not be specifically related to dyslexia but instead to reading performance and experience. To explore this possibility, we correlated the feature-based laterality effects with reading speed and accuracy of the three reading lists (IS-FORM common word forms, IS-FORM uncommon word forms, IS-PSEUDO pseudowords) separately for the two groups of typical and dyslexic readers. Laterality indices below zero are consistent with left hemisphere lateralization and those above zero are consistent with right hemisphere lateralization.
There was no detectable association between reading accuracy and featurebased laterality effects in either group (all absolute rs < 0.22). This could be due to truncated range of reading accuracy, particularly in the typical reader group as most typical readers read quite accurately while there was more variability in reading accuracy in the dyslexic group, see supplementary figure s2. In the dyslexic group, correlations between reading speed of the three reading lists and feature-based laterality effects were small and negative (supplementary figure s3, left panel).
However, there was a positive correlation between reading speed and feature-based laterality effects in the typical reader group (supplementary figure s3, right panel). We discuss this association in the Discussion chapter of the main manuscript.
It is also interesting to look at possible group differences in correlations with global form laterality effects. While correlations with global laterality indices are small and negative in the typical reader group, they are somewhat larger and positive in the dyslexic reader group. So, unlike that for typical readers, faster reading of dyslexic readers is possibly associated with increased right hemisphere lateralization of global form face processing. This is very speculative, but it could be in accordance with the possibility that in order for dyslexic readers to gain more reading competence, they need to compensate for a feature-based processing deficit by implementing reading strategies that tap into other types of processing such as global or holistic processing. We note that this is reminiscent of the results of Brady et al. (2020), who report that greater holistic processing predicts better reading of dyslexic readers but worse reading for typical readers.
We finally want to mention that the reading speed of all three lists is highly correlated in the dyslexic group while individual differences in the typical reader group seem to separate into the reading of familiar and unfamiliar material, as evidenced by the lower correlation between reading speed for common word forms and pseudowords (supplementary figure s3). This fits with the possibility that while typical readers rely on different strategies for the reading of familiar and unfamiliar material, dyslexic reader might rely on a similar mechanism in both cases. Skilled reading has been suggested to involve flexibly applying several reading subskills, such as logographic, alphabetic, and orthographic, depending on the requirements of the reading task (Lachmann, 2018). The inability to do so may impair reading. Table s1 shows the results for simple group comparisons of performance in face matching with and without participants who reported a previous ADHD diagnosis.

S3. ADHD Measures
Excluding these participants slightly increased the group difference for feature-based face matching, but otherwise had minimal effect. The ADHD-I and ADHD-II questionnaires were not significantly correlated with performance in face matching (table s2). ADHD-I and ADHD-II scores were on the other hand significantly correlated with several reading measures (table s3). There was, in other words, a significant proportion of the variance in reading problems that could be statistically accounted for by symptoms of ADHD. This highlights that it might not be sufficient, as done in some studies, to simply exclude participants with a previous ADHD diagnosis, as ADHD symptoms are still associated with reading problems even if participants with such a diagnosis are excluded, so a modulatory role of ADHD symptoms in such studies cannot be ruled out.  Table s3. Correlation (Pearson's r) between reading measures and Behavioral Evaluation Questionnaire for Adults I and II, intended to assess symptoms of ADHD. Significant correlation coefficients are underlined in bold. Group is coded as 0: typical reader and 1: dyslexic reader. In our original hierarchical regression analyses, we factored out variance related to ADHD by partialling out ADHD-I, ADHD-II, and ADHD diagnosis, allowing ADHD symptoms to account for all variance shared with reading problems and leaving only the variance in reading problems not accounted for by symptoms of ADHD. With the exception that only participants without a previous ADHD diagnosis are included, Table s4 is directly comparable with Table 1 in the main text. Table s4. Summary of regression models with group membership ARHQ, reading speed and reading accuracy as dependent variables, and measures of ADHD and face processing as independent variables. Unstandardized regression coefficients (b) and p-values are bold for significant independent predictors (blue when b is positive and red when b is negative in online color version). Only participants without a previous ADHD diagnosis (N = 58) are included.

S4. Regression Models Accounting for Visual Search
Our data indicate that reading problems are associated with greater symptoms of ADHD. Furthermore, as detailed in the "sister" manuscript of the current primary manuscript, reading problems are also associated with slower visual search, especially conjunction search (Sigurdardottir, Omarsdottir & Valgeirsdottir, 2021).
General attentional problems as indicated by symptoms of ADHD and specific problems with visual attention as captured by difficulties with visual search, particularly conjunction search, have previously been documented in the literature. Is Table s6. Summary of regression models with group membership (logistic regression, (10) = 40.631, p < 0.001, R 2 Nagelkerke = 0.665), ARHQ (linear regression, F(10,49) = 4.212, p < 0.001, R 2 = 0.462, R 2 adjusted = 0.353), reading speed (linear regression, F(10,49) = 2.431, p = 0.019, R 2 = 0.332, R 2 adjusted = 0.195) and reading accuracy (linear regression, F(10,49) = 2.062, p = 0.046, R 2 = 0.296, R 2 adjusted = 0.153) as dependent variables, and measures of ADHD, face matching, and visual search as independent variables. Unstandardized regression coefficients (b) and p-values are bold for significant independent predictors (blue when b is positive and red when b is negative). Regression coefficients for search slopes and intercepts are in seconds. It should be noted that the number of independent predictor variables is quite high (10) compared to the number of participants (60). While this could possibly pose problems for logistic regression, two subjects per variable appears to suffice to accurately estimate regression coefficients and their confidence intervals in linear regression (Austin & Steyerberg, 2015).