Cognitive and Language Abilities Associated With Reading in Intellectual Disability: A Systematic Review and Meta-Analysis

We searched four databases, PubMed, PsycInfo, Web of Science, and ERIC (Educational Resources Information Center). The search strategy included a combination of keywords related to reading (reading, literacy, decoding, word recognition), ID (ID, mental retardation, mental deficiency, intellectual developmental disorder, developmental disability), and relation (relation, relationship, prediction, correlation, regression, association). A full description of the keywords for each database is in the online supplements. No filters were used during the search. We performed several searches to ensure that articles published or made available since the first search were found and included in the final meta-analysis. The first search was conducted in August 2017, the second search in March 2021, and the third search in August 2022. The first search yielded 1,677 articles (1,333 after removing duplicates). For the subsequent searches, the same search strategy was adopted. One database had changed the interface of its advanced search; consequently, a minor change was made to the search strategy (see online supplements). The second and third searches yielded 3,040 and 3,457 articles, respectively (1,202 and 387 articles, respectively, after removing duplicates within the searches and duplicates of those found in the previous searches).

In addition to the systematic search, other methods were used to find additional records. The first author sent inquiries for file drawer data to all 23 authors of already included articles who had valid email contact information in the publications, and 10 replied. Furthermore, the first author performed the following additional search methods: reference lists in all the included articles were scanned for suitable articles, articles citing the included articles were identified via Google Scholar and screened (completed in May 2021), and a request for file drawer data was made in a poster presentation at the 2019 Society for Scientific Studies of Reading conference.

For the abstract and full-text screening, the inclusion criteria were the presence of: (a) measures of decoding and/or reading comprehension; (b) correlational data; (c) a sample with a mean IQ at or below 70, and a maximum (i.e., range) IQ of 85 (any standardized assessment of cognitive ability was approved. Whenever multiple measurements were reported, the order of preference was non-verbal, full scale, and verbal); (d) a minimum sample size of 10 participants; and (e) participants with nonspecific ID (unknown etiology), Down syndrome, WS, or mixed etiology (i.e., participants with various aetiologies within the same sample). The most recent and current definitions of ID give much more emphasis to adaptive abilities (APA, 2022); however, the most usual way to operationalize ID in research has been through a measure of IQ or related cognitive ability, as a result, our inclusion criteria reflected this approach. The meta-analysis included articles in dissertations, and all articles had to be written in English. Articles were excluded if the focus was another syndrome, another disability, if the participants also had autism, or if the article was a review. Data collection and reporting were guided by the directions of Preferred Reporting Items for Systematic Reviews and Meta-Analysis (PRISMA, Page et al., 2021). For a more detailed description of identified records, of included and excluded records, and the reasons for exclusion, see Figure 1. The PRISMA checklist and an abstract that adheres to the PRISMA guidelines are in the online supplements.

Two of the authors screened the abstracts and full-text articles independently. After initial abstract screening, the following number of articles remained for full-text screening after each search: first search (k = 166), second search (k = 199), and third search (k = 34). Inter-rater reliability figures were calculated for the abstract screening in the second and third search using Cohen’s $\kappa$, and the agreement ranged between 0.56 and 0.70 which is considered as fair to good agreement according to Fleiss et al. (2003). The percentage agreement ranged between 91% and 96%. The two authors marked each study as “include,” “exclude” or “unsure,” and only the studies marked with “exclude” from both authors were excluded before the next step. Studies marked as “unsure” and any studies where there was a difference between the two authors were discussed and resolved. The screening process was conducted in Zotero (first search) and Excel (second and third search). Through the systematic searches and the screening processes, we identified 24 articles that met our inclusion criteria. In addition, the inquiries for file drawer data and forward and backward citation searches made by the first author yielded one additional article plus one additional data set, which was later published. Hence, a total of 26 articles were identified and included in the meta-analysis. In addition, the study by Levy (2011) involves two independent samples, one with DS and one with nonspecific ID. These samples are reported separately in the primary study and thus treated as two separate studies in this meta-analysis, meaning that the final set consists of 27 studies.

Two of the authors coded the descriptive data and the correlations. Agreement between coders was 93% for the descriptive data and 96% for the correlations. All disagreements were discussed and resolved. The articles in this meta-analysis were mainly cross-sectional studies. There also were some relevant longitudinal studies (k = 3) and for these, the pre-test data was used. Our meta-analysis targeted variables that have been found to correlate with reading comprehension and decoding in research on individuals with ID and with typical development. Hence, we coded all pairwise correlations between the following variables: reading comprehension, decoding, phonological awareness, listening comprehension, vocabulary, phonological STM, visual STM, executive-loaded working memory (ELWM), RAN, IQ, chronological age, letter-sound knowledge, and grammatical comprehension. The mean age reported in the primary studies was coded in months, and when age was reported in years in the primary study it was converted to months by the coding author.

The quality of the included articles was assessed with the AXIS tool (Downes et al., 2016), which was created by an expert panel for the purpose of critically appraising cross-sectional studies. In addition, the tool also provides a user-friendly guidance document. Note that the AXIS tool was used to critically appraise the pre-test data from three longitudinal studies that we coded and analyzed; it was not used to assess the methodological quality of the longitudinal designs as only the pre-test data was of interest for our meta-analysis. The AXIS tool consists of 20 items divided into sections that correspond with the outline of a scientific article (one item in the method section is “Was the sample size justified?”). Each item was scored with yes, no, or unsure. The unsure option was used when there was not enough information provided in the primary article to make a decision. The quality assessment did not aim at excluding studies from the meta-analysis, but rather at providing information that could aid in rating the certainty of evidence and the interpretation of the results. Two of the authors conducted the quality assessment independently. Assessments were compared, and any differences were discussed and resolved.

The number of relevant variables varied across the primary studies. Hence, some of the coded correlations were reported in many of the primary studies, whereas some correlations only occurred in one or two primary studies. To minimize the risk of bias where the result is driven by a few primary studies, we decided to only analyze correlations between variables that were reported in five or more studies. As a result, our meta-analysis included correlations between decoding and these variables: phonological awareness, listening comprehension, vocabulary, phonological STM, visual STM, ELWM, RAN, IQ, and chronological age and between reading comprehension and these variables: decoding, listening comprehension, vocabulary, and IQ.

All analyses were made using R (R Core Team, 2017), and the following R packages: metafor (Viechtbauer, 2010), tidyverse (Wickham, 2017), readxl (Wickham & Bryan, 2019), knitr (Xie, 2015), dplyr (Wickham et al., 2019), and robvis (McGuinness, 2019). The manuscript was formatted using papaja (Aust & Barth, 2017), and citr (Aust, 2016). The effect size of interest in this meta-analysis was the correlation coefficient r. Effect sizes were calculated using the escalc() function in the metafor package based on the studies’ reported correlation coefficients and sample sizes. As recommended in the metafor package, the correlations were transformed to z-scores using Fisher’s r-to-z transformation to reduce bias, and these scores were then used in the analyses. In a meta-analysis, a fixed-effects model or random-effects model could be used. A fixed-effects model assumes that there is one underlying effect that is the same for all studies. A random-effects model assumes that the effect could be different in each study, depending on, for example, participant characteristics (different mean IQ in different studies) or design of the studies. For the current meta-analysis, it was reasonable to assume that the underlying effect could be different across studies and, therefore, we used a random-effects model for the analyses. In some cases, primary studies used multiple measures for the same variable (e.g., subtests of phonological awareness such as blending and elision). These effect size estimates could not be assumed to be independent and to deal with the dependencies we used a robust variance estimation (Pustejovsky & Tipton, 2022) for the affected correlations. The random-effects model was fitted on the z-transformed data, and the effect sizes were then back-transformed to correlations for easier interpretation of results.

To test the degree of heterogeneity between studies we used three different heterogeneity measures. Cochran’s Q-test was conducted to examine whether heterogeneity was different from zero. As this test is sensitive to the number of studies included, the I²-statistic was used to determine the proportion of observed variance reflecting true variation between effect sizes (Higgins et al., 2003). We used the categorization proposed by Higgins et al. (2003), namely that I² values of 25%, 50%, and 75% are regarded as low, moderate, and high, respectively. Finally, the amount of between-study heterogeneity (${\tau }^2$) was estimated using the restricted maximum-likelihood estimator (Viechtbauer, 2005). We used the same rule as implemented by Hjetland et al. (2020), namely that a ${\tau }^2$ larger than 0.1 indicates a large variation between studies. To assess whether study quality could explain heterogeneity in the results, we calculated an overall study quality indicator for each study. We did this by assigning different scores to the three risk of bias levels, where low risk of bias gave 2 points, some concerns (or not reported) gave 1 point, and high risk of bias gave 0 points. Study quality was used as a moderator in the analyses that included at least 10 primary studies, because performing moderator analyses on smaller samples can introduce other problems (Higgins et al., 2019). Publication bias was tested using two methods, namely funnel plots and the Rank Correlation test examining Kendall’s $\tau$. Publication bias is indicated through asymmetrical funnel plots and a significant Rank Correlation test. We used Cook’s distances to identify studies that were likely to be influential in the context of the model (Viechtbauer & Cheung, 2010). Studies with Cook’s distance larger than the median plus six times the interquartile range of Cook’s distances were considered to be influential.

The Grading of Recommendations Assessment, Development, and Evaluation (GRADE) tool (Balshem et al., 2011) is a valuable assessment of the certainty of intervention effects; because our meta-analysis concerns correlations, we used a modified version of GRADE to obtain measures of certainty (see Yousefifard & Shafiee, 2023 for a discussion on this topic). This involved four different indicators: (a) an overall assessment of risk of bias, (b) inconsistency of results (a confidence interval [CI] larger than 0.3), (c) imprecision of results (an I² larger than 50%), and (d) a significant publication bias test.

Reading comprehension correlated significantly with all four variables, with strong correlations for decoding (r = .63, p = <.001) and vocabulary (r = .51, p = <.001), and moderate correlations for listening comprehension (r = .43, p = <.001) and IQ (r = .32, p = .001).

The heterogeneity measures found in Table 1 indicate a high proportion of true heterogeneity between studies for decoding and vocabulary. The proportion of true heterogeneity between studies for listening comprehension was low, and for IQ it was moderate. A moderator analysis was performed on the association between reading comprehension and decoding (k = 10) using study quality as a moderator, and it was not significant (p = .437). The average effect sizes for each variable associated with reading comprehension are visualized in Figure 2. Forest plots visualizing all effect sizes from the primary studies can be found in Figures S1–S4 in the Supplements.

The rank correlation test was nonsignificant for all variables associated with reading comprehension (see Table 1), indicating that no publication bias was present. However, a visual inspection of the funnel plots revealed a slightly asymmetrical distribution for both listening comprehension and vocabulary. The funnel plots can be found in Figure S14 in the Supplements. For results from sensitivity analyses, please see Supplements.

Decoding correlated significantly with all the chosen variables, except for listening comprehension (r = .12, p = .119). Decoding strongly correlated with phonological awareness (r = .52, p = < .001). Furthermore, moderate correlations were found with phonological STM (r = .46, p = < .001), vocabulary (r = .34, p = < .001), RAN (r = -.44, p = <.001), IQ (r = .30, p = <.001), and ELWM (r = .32, p = .014). Finally, correlations in the lower range were found with chronological age (r = .26, p = .035) and visual STM (r = .24, p = .007).

The heterogeneity measures in Table 1 indicate low true heterogeneity between studies for phonological STM, listening comprehension, and visual STM. A moderate proportion of true heterogeneity between studies was found for phonological awareness, RAN, IQ, chronological age, vocabulary, and ELWM. A moderator analysis was performed on the association between decoding and the following variables: phonological awareness (k = 10), IQ (k = 18), and vocabulary (k = 12) using study quality as a moderator. Study quality was a significant moderator for the association with phonological awareness accounting for 47.35% of the heterogeneity (p = .010), but not for the association with IQ (p = .560), or vocabulary (p = .522).

The average effect sizes for each variable associated with decoding are visualized in Figure 3. Forest plots visualizing all effect sizes from the primary studies can be found in Figures S5–S13 in the Supplements. Note that the most common way of measuring RAN is for the participant to name items as fast as possible. Consequently, the lower the score the better the performance, and a negative correlation is expected with literacy measures. RAN was reported in six primary studies, and three of them originally reported negative correlations. Two of the studies used a “per minute” score, and hence the correlations were converted to negative scores. One study reported a positive correlation, despite using the common test procedure, and this correlation was not converted.

The Rank Correlation test was non-significant for all variables associated with decoding (see Table 1), indicating that no publication bias was present. A visual inspection of funnel plots revealed quite symmetrical distributions. The funnel plots can be found in Figures S15–S16 in the Supplements. The funnel plot of RAN in Supplemental Figure S16 shows a wide distribution along the x-axis, which is more indicative of heterogeneity between studies than publication bias. For results from sensitivity analyses, please see Supplements.

The primary studies all showed similar patterns in the quality assessment. Studies generally described clear aims (k = 20), and used a design that was appropriate for their aims (k = 20). All studies except one failed to justify their sample size (k = 25) and the vast majority did not report about the use of a sampling frame (k = 25), meaning that they also failed to provide a sample that can be regarded as representative for the population (k = 26). None of the studies provided information about non-responders (k = 26), therefore a decision about non-response bias was not possible to make. The majority of the studies used appropriate tests for measuring the outcome variables (k = 22), but not all studies used standardized tests (k = 11). In general, studies described their methods in a way that could enable replication (k = 24) and the basic data was adequately described (k = 24). The discussions and conclusions were most often justified by the results (k = 22), but many studies failed to discuss limitations of the study (k = 10). Results regarding certainty of evidence, and a risk of bias plot (see Figure S17) are in the Supplements.

This meta-analysis showed that reading comprehension was strongly associated with decoding and moderately associated with listening comprehension. This is similar to research findings about typically developing children (Lervåg et al., 2018), and children with reading comprehension difficulties (Catts et al., 2006), and these associations support the SVR (Gough & Tunmer, 1986). However, other variables also were associated with reading comprehension, something that is not predicted by the SVR. For example, vocabulary was also shown to correlate strongly with reading comprehension, a relationship that has been identified in research on typically developing children as well. Ouellette (2006) found that measures of receptive vocabulary breadth and depth of vocabulary knowledge accounted for 28.5% of the variance in reading comprehension in a sample of typically developing students in Grade 4, even when decoding was taken into account. In another study, Ouellette and Beers (2010) argued for a not-so-SVR when their results showed that vocabulary accounted for 15.3% of the variance in reading comprehension in typically developing students in Grade 6, even when decoding and listening comprehension were also entered into the analysis.

These findings about the relevance of vocabulary to reading comprehension support the LQH (Perfetti, 2007), where both vocabulary size and quality of the reader’s lexical representations are assumed to affect reading comprehension directly. The strong associations between reading comprehension and decoding, listening comprehension, and vocabulary in the current analysis indicate that combining the theoretical frameworks might be a successful way of explaining reading comprehension in individuals with ID. Other support for a combined model comes from Verhoeven and van Leeuwe (2008) who found in a longitudinal study on typically developing children, that a combined model of decoding, listening comprehension, and vocabulary predicted reading comprehension. In another study, focusing on children with ID enrolled in special education classes, reading comprehension was predicted by decoding and a composite variable of vocabulary and syntactic skills (Verhoeven & Vermeer, 2006).

Furthermore, the current study found a moderate correlation between reading comprehension and IQ. This is consistent with findings from Verhoeven and Vermeer (2006), where reading comprehension was predicted by non-verbal IQ, decoding and language skills, in a sample of 10- and 12-year-old children with ID. Studies on typically developing children have also found significant correlations between IQ and reading comprehension both in early and later grades (Hulslander et al., 2010; Scarborough, 1998). In a longitudinal study by Hulslander et al. (2010), full scale IQ was found to be the only significant longitudinal predictor of later reading comprehension over and above decoding abilities and initial measures of reading comprehension. In contrast, two other longitudinal studies did not find any predictive contribution of IQ when earlier reading skills were controlled for (Cunningham & Stanovich, 1997; Scarborough, 1998).

Our meta-analysis revealed many similarities between the pattern of correlations between reading comprehension and other cognitive variables in individuals with ID and individuals with typical development. In addition, although the pattern of correlations in individuals with ID was found to be consistent with the SVR, the presence of other significant correlations supported the LQH. This suggests that a combined theoretical framework is needed to explain reading comprehension in both typical development and the development of individuals with ID.

Our meta-analysis suggests that the ability to process speech sounds (phonological awareness, RAN, and phonological STM) is related to decoding measures. There was a strong and significant correlation between decoding and phonological awareness and a moderate and significant correlation with RAN. Similar findings have been reported in a large body of research on typically developing children (Melby-Lervåg et al., 2012; Moll et al., 2014; Schatschneider et al., 2004), and also correspond with previous research on individuals with ID (Barker et al., 2014; Kennedy & Flynn, 2003; Saunders & DeFulio, 2007). A moderate and significant association was found between phonological STM and decoding, which has been reported in studies on individuals with ID (Channell et al., 2013; Conners et al., 2001; Henry & Winfield, 2010).

Decoding also showed a significant and moderate correlation with vocabulary. The association between decoding and vocabulary have been shown in several studies focusing on individuals with DS (Boudreau, 2002; Cardoso-Martins et al., 2009; Naess et al., 2011). Notably, the meta-analytic review by Naess et al. (2011) found that vocabulary, not phonological awareness, predicted non-word decoding. It could be the case that the moderate association between vocabulary and decoding found in the present study is driven by the primary studies focusing on individuals with DS. There are no clear answers as to why vocabulary seems to be important for decoding in the DS group. Boudreau (2002) suggested that language skills could be the foundation for building literacy skills, or that it might be the case that both language and literacy draw upon a third independent cognitive process.

Furthermore, this meta-analysis showed that decoding correlated moderately with IQ. The relationship between decoding and IQ has been debated in the literature. For children with typical development, it appears that IQ is not related to decoding ability (Gustafson & Samuelsson, 1999; Stanovich, 2005). However, for the population with ID—this is still an open question. van Tilborg et al. (2014) found that non-verbal intelligence was the only significant predictor of word decoding in a group with non-specific ID, explaining a total of 33% of the variance, even though letter knowledge and phonological awareness were used as predictors in the regression. In contrast, a study by Conners et al. (2001) found that two groups with mixed etiology ID divided by level of decoding ability did not differ on an IQ measure. Instead, the better decoders in the study scored significantly higher on measures of language ability and phonemic awareness, which is more in line with research on typical readers and individuals with dyslexia. Boudreau (2002) also found only a weak relationship between non-verbal mental age and decoding (both decoding of words and nonsense words) in a group with DS. The moderate association found in the current study indicates that IQ could play an important part in the development of decoding ability for individuals with ID, although this possibility needs to be evaluated in investigations of the shared variance between IQ and other predictor variables.

Decoding also correlated moderately with ELWM. The role of working memory for reading has mainly been emphasized in relation to reading comprehension, but there are studies on typically developing children suggesting that working memory plays an important role in decoding as well (Christopher et al., 2012). It would be reasonable to assume that working memory has a larger impact on decoding during the early stages of development when decoding is expected to be an effortful task. Many individuals with ID, both children and adults, decode words with an alphabetic strategy (Ratz & Lenhard, 2013), which could explain the association between decoding and ELWM found in this meta-analysis. However, it also is the case that in typical readers there is a relationship between working memory and decoding even at higher grades where an alphabetic strategy would not be expected (Peng et al., 2017). Further research is needed to understand the reasons for this relationship in individuals with ID.

As with reading comprehension, in individuals with ID the pattern of correlations between decoding and other variables was similar to the pattern reported in relation to those with typical development. An interesting exception to this was the correlation between decoding and IQ. In individuals with ID, unlike those with typical development, this correlation was found to be significant; and this could be a useful topic for future research.

The tests of heterogeneity indicated significant between-study variability for many variables. A standard procedure is to explore if between-study heterogeneity can be explained by moderators. However, due to the small number of primary studies in most analyses, moderator analyses were only feasible for four associations. These analyses identified study quality as a significant moderator for only one association, between decoding and phonological awareness.

The quality assessment showed that all primary studies had a potential risk of bias, often related to participant recruitment and sample size justification. These are common challenges in disability research due to smaller population sizes. About half of the studies also had a high risk of bias because they used potentially unreliable, study-specific measures.

A more general limitation of the meta-analysis is the possibility of measurement error. Because individuals with ID have intellectual difficulties, a test of visual STM could instead be an assessment of the degree to which the participant with ID understood the instruction. This phenomenon is discussed by van Wingerden et al. (2018), who suggested that the cognitive demands of phonological awareness tasks could lead them to capture higher-order skills like working memory. This potential limitation is a concern with participants who have cognitive difficulties, although validated standardized tests can help to address the issue.

Except for two associations, certainty of evidence was rated as low or very low, mostly due to the risk of bias, inconsistency, and imprecision of results. As previously discussed, the frequent use of unreliable measures likely contributed to these limitations. In addition, researcher decisions regarding study inclusion may have influenced heterogeneity. Our decision to include studies across different aetiologies enabled a meta-analysis by increasing the number of primary studies but may have introduced a larger amount of heterogeneity.

A further consideration is the influence of orthographic transparency, as primary studies were conducted in languages with varying levels of transparency. Readers of transparent orthographies tend to read faster and more accurately than those of opaque orthographies (Patel et al., 2004). Moreover, the predictors of decoding appear to differ depending on orthographic transparency, with RAN consistently predicting reading fluency, whereas the relationship between decoding and phonological awareness is complex and interactive (Landerl et al., 2019). Although no clear patterns related to orthographic transparency were observed in the forest plots, this factor warrants further investigation in future research. One reviewed study compared German- and French-speaking children with ID, finding that spoken language did not predict progress in word and non-word reading (Sermier Dessemontet & de Chambrier, 2015). Because the orthographic transparency of German and French are not fundamentally different, a comparison between languages with greater orthographic differences would provide a more stringent evaluation of these effects. For example, a comparison between French- and English-speaking individuals might have produced other results as onset entropy (an index of orthographic transparency) is almost twice as high for English compared with French (Ziegler et al., 2010).

The limitations of our systematic review and meta-analysis reflect many of the more general limitations of research concerned with reading, and especially research concerned with ID. Future studies should address these issues while ensuring a high level of transparency when reporting their findings. Adopting open science practices to enhance reproducibility and transparency would be highly beneficial.

This meta-analysis focussed on the predictor variables of reading comprehension and decoding in individuals with ID. When considering our findings, it is important to bear in mind the limitations we have identified. The analysis has confirmed that many variables identified in studies on typically developing children and other groups are related to decoding and reading comprehension in individuals with ID. An exception to this general pattern was that IQ correlated moderately with decoding in individuals with ID, whereas studies on individuals with typical development indicate that there is no significant relationship between these two variables (Gustafson & Samuelsson, 1999; Stanovich, 2005). Furthermore, this meta-analysis indicates support for a combination of the SVR and the LQH as reading comprehension is significantly related to decoding, listening comprehension, and vocabulary. All this taken together confirms that theoretical frameworks describing the relationships between components of reading in typically developing children and other groups could reasonably be applied to individuals with ID.

Consequently, our results imply that abilities associated with reading comprehension and decoding in individuals with ID are similar to those of typically developing children and struggling readers. This indicates that support and interventions developed for typically developing children and struggling readers might be cautiously applied to students with ID, with the addition of feasible adaptations. This is in line with findings from intervention research that reading interventions could be very effective for students with ID if the delivery and context, not the content, are adapted to student needs (see, for example, Allor et al., 2014).

Please find the following supplementary material below: additional figures, tables, and results. Please find the following supplementary material on OSF https://doi.org/10.17605/OSF.IO/U7P8G: data files, analysis script, PRISMA checklist, PRISMA abstract, and a detailed search strategy.