Introduction

Rapid automatized naming (RAN) has been considered an important predictor of reading abilities in children since Denckla and Rudel (1974, 1976). Recently, five meta-analyses of research on the RAN-reading relationship have been published (Araújo et al., 2015; Chen et al., 2021; McWeeny et al., 2022; Song et al., 2016; Swanson et al., 2003), reporting broadly consistent results (see Table 1). However, they all focused specifically on monolingual children, while much less is known on the RAN-reading relationship in bilingual children. A number of studies in bilingual reading development investigating the RAN-reading relationship concurrently and over time reported inconsistent results, finding moderate (e.g., Chung et al., 2021; Vataja et al., 2021) to weak RAN-reading correlations (e.g., Fleury & Avila, 2015). Likewise, empirical reports on cross-linguistic transfer (i.e., RAN measured in L1 as a predictor of reading success in L2) vary in findings, with some studies identifying it (McBride-Chang & Suk-Han Ho, 2005), yet others failing to do so (Li et al., 2011). With large numbers of second language speakers in early primary classrooms all over the world, it is essential to understand the nature of the RAN-reading relationship in bilingual children. A comprehensive meta-analysis of RAN and reading in bilingual children is in place to provide more insight into the nature of both within language and cross-language RAN-reading relationships. Since this is the first meta-analysis of such kind, our main aim is to examine the strength of concurrent as well as longitudinal RAN-reading relationships in bilingual participants, both within one language (L1 RAN–L1 reading and L2 RAN–L2 reading) and between languages (L1 RAN–L2 reading and L2 RAN–L1 reading).

Rapid Automatized Naming (RAN) is the speed of naming highly familiar stimuli, such as colour patches, line drawings of common objects, letters or digits. RAN assessment usually follows the setup adopted in the seminal work of Denckla and Rudel (1974, 1976): 5 different stimuli, each repeated 10 times in a pseudorandom order, are presented on an A4 sheet; the outcome is the time taken to name all 50 stimuli. However, alternative arrangements (e.g., the number of stimuli named in 30 s) are also used.

RAN has attracted a great deal of research and clinical interest, as this seemingly simple task correlates with ecologically valid outcomes, such as ADHD diagnoses (Ryan et al., 2017), arithmetic performance (Hornung et al., 2017) and, most notably, literacy skills across languages in monolingual children (e.g.,Caravolas et al., 2013; Georgiou et al., 2008; Landerl et al., 2019). Numerous studies have found slow RAN performance to be a core deficit in developmental dyslexia (Fernandes et al., 2017; Kirby et al., 2008; Wolf et al., 2002).

Detecting reading impairments in bilingual children is challenging due to the necessity of disentangling between problems caused by insufficient language proficiency and core reading impairments. Consequently, the validity of established predictors of reading impairment ought to be re-examined in the bilingual context. This applies to RAN too: are RAN measures a valid predictor for reading impairment in bilinguals? While the current study does not investigate reading impairment, it can address this question indirectly, by examining whether the pattern of RAN-reading correlations in bilingual learners is similar to that observed in monolinguals. If it is similar (in terms of strength of correlations and their moderators)—especially in longitudinal studies—then the validity of RAN measures in the bilingual context is probably secure. The results of this meta-analysis are thus of practical importance for anyone concerned with the detection of reading impairments in bilingual children.

RAN-reading relationships: theoretical accounts

While there is broad agreement that differences in RAN performance predict differences in reading skills (both concurrently and longitudinally), the debate continues as to the mechanisms underlying their relationship. That debate is certainly warranted given that RAN tends to be a non-redundant predictor: it often accounts for a significant (if small) proportion of variance in reading even when other obvious predictors (such as prior reading skills or phonological awareness) are controlled for (for a review, see Landerl et al., 2022; for specific examples, see e.g. Caravolas et al., 2019; Kirby et al., 2008; Landerl et al., 2019).

Proponents of the phonological processing account argue that RAN tasks measure the rate of access and retrieval of phonological codes from long-term memory (Torgesen et al., 1994; Wagner et al., 1994, 1997). From that perspective, RAN performance and reading performance share variance insofar as they both rely on efficient phonological access and phonological retrieval processes (which can ultimately be explained by the quality of underlying phonological representations).

Wolf and colleagues (2000) proposed the orthographic processing account, which emphasises the importance of precise temporal coordination (i.e., timing) of several cognitive subprocesses (visual, phonological, semantic, and articulatory). That domain general timing mechanism is crucial for fast naming (of alpha- and nonalphanumeric stimuli) as well as the efficient mapping of letter strings onto phonology.

The speed of processing account (Kail & Hall, 1994, 1999) suggests that skilled performance in reading as well as rapid naming rely on the rapid execution of underlying processes.

The working memory account (Amtmann et al., 2007) postulates that the RAN-reading relationship reflects a shared requirement for the maintenance of a set of names in working memory (specifically, its phonological loop component) that allows to integrate phonological and orthographic representations of words in a timely fashion.

The attentional processes account of the RAN-reading relationship (Semrud-Clikeman et al., 2000; Shao et al., 2013) emphasises response competition inherent in rapid naming, and so the role of effective response inhibition processes in fast production of correct responses.

The cascaded processing hypothesis introduced by Protopapas et al. (2013) postulates that the mechanisms of the RAN-reading relationship change over the course of development, but in advanced readers RAN is related uniquely to reading fluency (of connected text or of lists of words). This is because RAN and fluent reading both rely on cascaded processing, namely one where “Multiple items are processed simultaneously, so that one item may be processed phonologically while the preceding one is articulated and the next one is recognized visually” (Protopapas et al., 2013, p. 924).

The theoretical accounts listed here need not be mutually exclusive, as RAN tasks, by their very nature, must involve orchestration of several cognitive sub-processes: attentional (sustained attention, response inhibition), perceptual (feature analysis, pattern recognition), lexical (phonological, semantic) and motor (articulation), and the failure of any of these processes may compromise both fluent naming and fluent reading (McWeeny et al., 2022; Wolf et al., 2000). Indeed, the relative importance of each of these processes for naming as well as reading may vary as a function of reading proficiency (e.g., Georgiou et al., 2014).

RAN-reading relationship: review of evidence

Over the last two decades, several meta-analyses have explored RAN-reading relationships across different orthographies. Monolingual RAN studies identify it as a unique and unidirectional predictor of reading across languages with minor differences between orthographies (for a review, see Landerl et al., 2022). Four out of five meta-analyses included cross-sectional data (Araújo et al., 2015; Chen et al., 2021; Song et al., 2016; Swanson et al., 2003), whereas the one by McWeeny et al. (2022) focused on the longitudinal data only. Table 1 compares the scope and summarises the main findings of those studies.

Table 1 The RAN-reading relationship in monolingual participants: A summary of the main findings
Full size table

Previous meta-analyses (Table 1) have identified a number of potential moderators of the RAN-reading relationship. RAN tasks (e.g., alphanumeric vs. non-alphanumeric), reading domain (e.g., word reading vs. pseudoword decoding vs. reading comprehension), reading outcome measure (accuracy vs. fluency), writing system (alphabetic vs. non-alphabetic), orthographic consistency of the alphabetic writing (e.g., opaque vs. transparent), grade level, and reading level (e.g., average/skilled vs. at risk vs. dyslexic readers) have been shown as a source of variation across studies in monolingual children.

Given the above-mentioned moderators have been well-explored in previous meta-analyses, in the current study we have decided to check whether the patterns observed in research with monolingual children hold for bilingual children as well.

RAN and bilingualism

Research on L2 reading suggests considerable overlap between reading processes in L1 and L2 (e.g., August & Shanahan, 2006; Verhoeven et al., 2019). Theoretical accounts of this cross-linguistic relationship have traditionally relied on the notion of transfer (see Chung et al., 2019 for a review). Historically, transfer was a term in the contrastive analysis hypothesis as formulated by Lado (1957) and referred specifically to the transfer of linguistic forms from L1 to L2. An entirely different understanding of transfer was provided by Cummins (1979) in his common underlying proficiency (CUP) framework and, specifically, the interdependence hypothesis, which assumes the transfer of skills and conceptual knowledge between languages that takes place in the common underlying proficiency. Further research and theorising focused on establishing what is transferred in the common underlying proficiency, drawing on both above-mentioned approaches to varying degrees. For example, Koda (2008), in the transfer facilitation model, states that L1 metalinguistic awareness is transferred in learning to read in an L2 and this transfer is modulated by the typological distance between languages. Cummins’ model was developed as the common underlying cognitive processes model by Geva and Ryan (1993) who, in addition to L2 proficiency, identify general cognitive processes, such as working memory and executive function, as significant predictors of L2 reading that do not merely transfer between languages, but underlie both L1 and L2 reading. The interactive transfer framework (Chung et al., 2018, 2019) acknowledges that both language-specific (e.g., orthographic processing, L1–L2 distance, L2 complexity) and domain-general factors interact, i.e., jointly affect bilingual reading outcomes. Since RAN has components that can be considered both domain-general, like general processing speed, cross-modal temporal integration and the capacity for cascaded processing, and language-specific, like retrieval of language-specific word forms that depends on the degree of automaticity that is, in turn, related to proficiency in that language, we consider the interactive transfer model to be the most appropriate theoretical framework to analyse the RAN-reading relationship in bilingual children.

In agreement with the predictions of the interactive transfer model, there is evidence that, due to its domain-general component, RAN measured in L1 might correlate with literacy skills in L1 and a typologically distant L2, regardless of the differences in orthography and writing systems (Shum et al., 2016). However, there is also evidence that bilingualism per se might affect basic processing mechanisms involved in RAN tasks. Especially relevant for serial naming, as in RAN tasks, is the finding that bilinguals take longer to name pictures than monolinguals and this difference does not disappear with repetition (Ivanova & Costa, 2008). Bilingual lexical processing can be further affected by differences in executive function (Hartanto et al., 2019) and general processing speed (Hilchey & Klein, 2011), at least in some bilingual contexts, which might allow the bilinguals to compensate for slower lexical access. All this means that bilingual lexical processing is essentially different from monolingual lexical processing and the extent to which these differences affect the RAN-reading relationship is unknown, which means that the RAN-reading relationship in bilinguals merits systematic investigation.

Moreover, bilingualism is not a categorical variable (Luk & Bialystok, 2013), which means that bilingual outcomes are sensitive to a plethora of contextual factors which, unfortunately, are often underreported in empirical studies. Both RAN and reading outcomes might be modified by factors related to bilingualism such as language proficiency- and exposure-related factors (e.g., the amount and quality of input in either language, daily rates of L1/L2 use, length of L2 exposure). In the present meta-analysis, we define bilingualism broadly as an ability to use two languages, irrespectively of the level of proficiency in either language, which is incomparable across studies. We also decided to include the type of bilingualism as a moderator in our analysis in order to examine the role of the most basic exposure-related factors such as the order of acquisition (simultaneous vs. sequential) and country of residence (L1- or L2-speaking). Simultaneous bilinguals acquire both languages from birth. Sequential bilingualism refers to cases when the second language is acquired after the first language has been established. While sequential and simultaneous bilinguals acquire the L2 in naturalistic settings, L2 learners are explicitly taught the language in the classroom, while residing in the L1-speaking country.

Overall, in addition to RAN stimuli type (alphanumeric vs non-alphanumeric), reading outcomes (word reading, pseudoword decoding, text reading, and reading comprehension), writing systems (alphabetic vs other), the type of bilingualism is a potential moderator in the RAN-reading relationships in bilingual children we aim to explore.

The current study

The purpose of this study is to assess the strength of concurrent and longitudinal RAN-reading relationship in bilingual learners, both within one language (L1 RAN–L1 reading, and L2 RAN–L2 reading) and between languages (L1 RAN–L2 reading, and L2 RAN–L1 reading). Moreover, the study aims to explore the most likely moderators of that relationship. Our research questions are:

  1. (1)

    What is the strength of the relationship between RAN and reading in bilingual children concurrently and longitudinally?

  2. (2)

    Does that strength differ when measured within language (L1–L1or L2–L2) as opposed to between languages (L1–L2 or L2–L1)?

  3. (3)

    Does that strength vary as a function of (a) RAN task (alphanumeric vs non-alphanumeric); (b) reading domain (word reading; pseudoword decoding; reading comprehension; text reading); (c) Method of reading assessment (fluency vs accuracy); (d) reading group (typical, at risk, dyslexic); (e) reading level; (f) type of bilingualism; (g) writing system (alphabetic vs other); and (h) orthographic consistency of the alphabetic writing?

  4. (4)

    Is the strength of the RAN-reading relationship similar or different for monolingual and bilingual children?

  5. (5)

    Are the moderators of RAN-reading relationships any different in monolinguals than in bilinguals?

Method, search strategies and inclusion criteria

To identify the studies for inclusion in this meta-analysis, PsycINFO, Scopus, Web of Science databases were searched for publications from January 1970 up till September 2021. The procedure, inclusion and exclusion criteria are presented in Fig. 1. The following search terms were used: (RAN OR “rapid naming” OR “rapid automatized naming” OR “rapid serial naming” OR “naming speed”) AND (“second language” OR “bilingual’’) AND (reading). This search yielded a number of 334 articles, book chapters and dissertations. We also checked the reference lists of previous meta-analyses of bilingual research (Gunnerud et al., 2020; Melby-Lervåg & Lervåg, 2014) to identify studies to be included in the meta-analysis. This allowed us to add 5 additional references. We limited our search to published studies only. The first author assessed each title from the search list, obviously irrelevant articles were excluded. After removing duplicates, a list of 205 articles was compiled for the next stage of abstracts evaluation. Two screeners reviewed abstracts independently, 86 studies were agreed to be included for the full-article revision stage (see Fig. 1). At this level the inter-rater agreement was 92%, the consensus was reached by discussions. Finally, after the full articles screening procedure only 38 met the inclusion criteria and were examined in this meta-analysis (see Online Resource 1).

Fig. 1
figure 1

Flow diagram for searching the studies in the meta-analysis (adapted from https://prisma-statement.org)

Full size image

The following inclusion criteria were used:

  1. (1)

    Study reported original empirical data;

  2. (2)

    Study reported zero-order correlations between RAN and reading outcomes;

  3. (3)

    Sample size was reported;

  4. (4)

    Mean sample age below the age of 18;

  5. (5)

    Study reported data on bilingual groups only or compared bilingual and monolingual samples;

  6. (6)

    Standard RAN tasks were administered;

  7. (7)

    Study was published in English.

We operationalized some broad constructs as follows. To be considered a bilingual study, it had to include one of the following groups: (1) equal or simultaneous bilinguals—children acquiring two languages simultaneously; (2) sequential bilinguals, emergent bilinguals—children using L1 (one language) at home and L2 (another language) at school; (3) ESL learners, or second language learners—children learning a second language at school while resident in a country where their native language is spoken. The following tasks were considered a measure of reading: speed or accuracy of reading lists of words, pseudowords or connected text, or responding to comprehension checking questions (multiple choice, true–false, etc.). To be considered a RAN measure, tasks had to include rapid naming of a set of colours, digits, letters, or objects. Discrete rapid naming tasks were not regarded as a measure of RAN in this study. After applying these criteria, we ended up with 38 studies, which include 313 effect sizes and 5312 participants from 47 independent samples.

Coding procedure and variables

In our meta-analysis, Pearson’s correlation coefficient (r) was used as the effect size unit. Only zero-order correlations between RAN and reading outcomes were extracted and coded for the analyses. If subtest scores of RAN tasks were reported along with a composite score of RAN tasks, only individual scores were extracted and coded.

Depending on the scoring type of RAN tasks, the Pearson correlation coefficients reported in studies are positive or negative. When a study measured time of RAN tasks, the correlations were negative (e.g., less time spent on RAN tasks, higher reading outcome scores). Whereas, if a rate (a number of correctly named items per fixed time) was a RAN task measure, reported correlations were positive (e.g., more items are named on RAN tasks, better reading outcomes). At the first stage of coding, correlations were extracted as negative or positive as they were initially reported in the studies. Since most studies reported correlations between RAN time and reading scores, correlations concerning rate were multiplied by − 1 to make the direction of all correlations consistent.

All selected studies were coded according to the following categories: (a) language pair, (b) RAN task, (c) reading domain, (d) method of reading assessment (accuracy, fluency), (e) sample features (reading level, reading group), (f) type of bilingualism, (g) writing system, and (h) orthographic consistency of the alphabetic writing. Studies that provided longitudinal data were also coded in accordance with the time interval between RAN and reading outcomes measures.

Language pair

All studies were assigned to four categories according to the language, in which RAN tasks and reading outcomes were measured. When both RAN and reading outcomes were measured in L1, we assigned studies to a within language (L1) category. Respectively, when both RAN and reading outcomes were measured in L2, studies were assigned to a group of within language (L2). If RAN was measured in L1 and reading outcomes in L2, studies were labelled as a cross language (L1–L2) group. Whereas, when RAN was measured in L2 and reading outcomes in L1, we assigned studies to a cross language (L2–L1) category.

RAN task

RAN tasks were coded into four groups: RAN letters, RAN numbers, RAN objects, and RAN colors. Some studies also reported composite scores of RAN (e.g., objects and colors). These data were excluded from the moderator analyses but included in the overall mean effect size analysis.

Reading domain

In our meta-analysis, reading measures were divided into four domains: word reading, pseudoword decoding, text reading, and reading comprehension. Word reading category included measures of single real word reading tasks. Pseudoword decoding group comprises measures of pseudoword reading. Text reading category covers all tasks that measure sentence or passage reading accuracy (e.g., number of correctly read words in the sentence or text). Reading comprehension includes all the measures of passage, sentence or text reading tasks where the outcome were responses to comprehension checking questions (multiple-choice, close-ended or open-ended). When a study reported correlations for a reading task that did not fall into any of those categories, they were excluded from the moderator analysis (e.g., chain word or loanword reading tasks).

Method of reading assessment

We assigned each reading outcome into one of two categories: accuracy (e.g., number of correctly read words/pseudowords) or fluency (e.g., correctly read words/pseudowords per unit of time).

Reading level

This variable was coded in four groups: pre-readers (kindergarten), novice readers (1–2 grade), intermediate readers (3–4 grade), proficient readers (5 grade and above). When a study provided participants’ mean age, we assigned it to a respective group based on the provided information.

Reading group

Three categories were created to code studies according to the reading status of the sample: (a) typical readers—sample was described as not having reading difficulties, (b) at risk group—participants were from families with dyslexia history, (c) dyslexic—samples were identified as dyslexic/poor readers.

Type of bilingualism

Different terms are often used in bilingualism research to refer to different types of bilingualism. To make comparison across studies possible, we created three coding categories: simultaneous bilinguals, sequential bilinguals, and L2 learners. The category of simultaneous bilinguals includes participants who acquired two languages since birth. Sequential bilinguals’ category includes participants who acquired L1 at home while their L2 was the language of instruction at school or it was the dominant language of their community. The category of L2 learners includes those children, for whom L2 was only taught as a part of the school curriculum, while L1 was the language of instruction (as well as the home language and the dominant language of the community).

Writing systems

Two categories were used to code studies regarding their writing systems. Studies were categorised as “alphabetic” (e.g., English, French) or “other” (e.g., Chinese (logographic), Japanese (syllabic), depending on the language writing system, in which reading outcomes were measured.

Orthographic consistency

Three categories were used to code studies according to their writing systems: opaque (e.g., English), intermediate (e.g., Dutch), and shallow (e.g., Greek) (Seymour et al., 2003). Each study was assigned to a specific category based on the orthography of reading measures.

Time interval

Sixteen out of 38 selected studies reported not only concurrent but also longitudinal data on RAN-reading relationships, the latter including 109 effect sizes. However, methodological differences between studies such as frequency of testing sessions, different time-intervals between testing sessions made the classification of longitudinal data quite perplexing. Therefore, we extracted only correlations between RAN measured in the kindergarten and at the beginning of the first grade (pre-reading stage) and reading domains measured at the end of the first and second grades. Within that range two categories were created: “kindergarten—grade 1” and “kindergarten—grade 2”. The first group includes all correlations between RAN measured at the pre-reading stage (kindergarten, beginning of first grade) and reading domains measured at the end of the first year. The second group includes correlations between RAN also measured at kindergarten and the beginning of the first grade, and reading domains measured in the second grade.

Statistical methods

To address each research question, three separate datasets were created (1) bilingual concurrent dataset includes correlations of RAN and reading domains measured at the same time point (e.g., RAN and reading outcome measured in the first grade); (2) bilingual longitudinal dataset contains effect sizes of RAN and reading domains measured at different time points (e.g., RAN measured in kindergarten and reading outcomes in the first grade or second grade); (3) paired concurrent dataset includes data for monolinguals and bilinguals compared directly within a study.

Reported effects were transformed from Pearson correlations to Fisher’s Z-scores to normalise their distribution for analyses. Later, they were transformed back to Pearson correlations to facilitate the interpretation of the results.

To accommodate multiple effect sizes per study, we used correlated effects models using robust variance estimation (RVE) with the R package robumeta (Fisher et al., 2017; R Core Team, 2022). Using the correlated effects method, we assumed a general between-outcomes correlation of ρ = 0.80. A random-effects model was used for the overall model (i.e., it included all effect sizes). To assess the presence of effect size heterogeneity we computed Cochran’s Q and \(I_{RVE}^{2}\). The estimate of \(\tau_{RVE}^{2}\) describes the variance of the true effect sizes, i.e., it represents true heterogeneity of effect sizes between studies, and not spurious heterogeneity due to sampling error. Related to this, \(I_{RVE}^{2}\) represents the percentage of variation between studies that is due to true heterogeneity rather than the sampling error. Cochran’s Q-statistic (which follows Chi-square distribution) is the difference between the observed effect sizes and the fixed-effect model estimate of the effect size.

Moderators were tested in separate meta-regression models. In addition to an overall model, we examined whether effect sizes differed based on combinations of RAN-reading domains, language pairs, RAN tasks, measures of reading assessment, writing systems, reading levels, reading groups, type of bilingualism, time intervals. For these categorical moderators, we employed ANOVA-analogue models. To compare effect sizes for different groups Wald test was applied, which uses the clubSandwich package in R (Pustejovsky, 2022).

We included p-values and 95% confidence intervals for each point estimate. To detect potential publication bias we produced funnel plots and tested their asymmetry using the Trim-and-Fill analysis and the Vevea and Hedges weight-function model (Vevea & Hedges, 1995), which uses the weightr package (Coburn & Vevea, 2017) in R.

Results

The final sample for concurrent RAN-reading analyses in bilingual children was drawn from 27 independent samples with 182 effect sizes. An intercept-only model was calculated to assess our first research question about an overall strength of the relationship between RAN and reading in bilingual children. This model yielded an average effect size of Z = − 0.42 [95% CI (−.47, −.65), p < .001]. This corresponds to r = −.39 [95% CI (−.44, −.35), see Table 2]. This may be interpreted as a weak to moderate effect (Schober et al., 2018). In addition, the degree of the between-study variability was found to be quite high \(\tau_{RVE}^{2}\) = 0.028, Q(181) = 851.2, p < .001, \(I_{RVE}^{2}\) = 85%.

Table 2 Concurrent moderator analyses for RAN-reading relationship in bilingual children
Full size table
Table 3 Longitudinal moderator analyses for RAN-reading relationship in bilingual children
Full size table

The longitudinal dataset for bilingual children included 16 independent studies with 109 effect sizes. An overall magnitude of RAN-reading relationship across two-year period was not significantly different from the concurrent effect size and produced the result of Z = − 0.40 [95% CI (−.46, −.34), p < .001], which corresponds to r = −.38 [95% CI (−.42, −.32), see Table 3]. The degree of the between-study variability was lower than in the concurrent analysis but still relatively high \(\tau_{RVE}^{2}\) = 0.01, Q(108) = 272.86, p < .001, \(I_{RVE}^{2}\) = 60%.

To answer the fourth and fifth research questions concerning the magnitude of RAN-reading relationship for monolinguals and bilinguals, a dataset from 3 studies with 50 concurrent effect sizes for bilinguals and 22 effect sizes for monolinguals was analysed, which directly compared monolingual and bilingual groups. An average effect size for bilinguals was found to be Z = − 0.49, p = .067 [95% CI (− 1.07, 0.09)], equivalent to r = −.45 [95% CI (−.79, .09)] and Z = − 0.53, p < .05 [95% CI (−.80, −.25)] with r = − .48 [95% CI (−.66, −.25)] for the monolingual group. There was a considerable between-study variability for both groups: \(\tau_{RVE}^{2}\) = 0.025, Q(49) = 161.9, p < .001, \(I_{RVE}^{2}\) = 69% for bilinguals and \(\tau_{RVE}^{2}\) = 0.018, Q(21) = 34.2, p < .001, \(I_{RVE}^{2}\) = 54% for monolinguals, which together with a small number of available effect sizes severely limited the opportunities for meaningful analysis. Therefore, we relegated the analyses of these results to Online Resource 2.

Moderator analyses

Considering the presence of such variability, moderator analyses for concurrent (see Table 2) and longitudinal (see Table 3) datasets in bilinguals were conducted to identify the sources of that variability. Wald-tests within moderators’ results for the bilingual populations are presented in Table 4.

Table 4 Wald-test within moderators’ analysis
Full size table

Language pair

In the concurrent analyses RAN correlations within languages were somewhat stronger than between languages, though the difference was not statistically significant. Longitudinal analyses showed quite similar (and so also not statistically different) correlations within languages and between languages.

RAN task

Our meta-analysis broadly replicated previous findings indicating that alphanumeric RAN (letters and digits) tends to predict reading better than non-alphanumeric RAN (colors and objects). This was the case for both concurrent and longitudinal analyses. However, most differences between alphanumeric and non-alphanumeric conditions failed to reach statistical significance. Only differences between RAN digits and RAN colours F (1, 6.2) = 15.4, p = .007 and RAN digits and RAN objects F (1, 11) = 4.84, p = .05 reached statistical significance, and for longitudinal data only.

Measure of reading assessment

The analyses of concurrent data revealed that RAN correlated stronger with fluency than with accuracy F (1, 9.9) = 21.5, p = .001. In the longitudinal dataset the correlation was also stronger for fluency than accuracy, yet the difference was small and not statistically significant.

Reading domain

In the concurrent analysis RAN demonstrated significantly stronger associations with text reading tasks F (1, 2.7) = 43.5, p = .010, pseudoword decoding F (1, 7.5) = 12.7, p = .010 and word reading tasks F (1, 5.9) = 9.3, p = .023, compared to reading comprehension tasks, which showed the weakest association. Moreover, the association was stronger for text reading than word reading tasks: F (1, 2.0) = 28.4, p = .032. In the longitudinal analysis, RAN showed the strongest correlation with text reading tasks, followed by reading comprehension tasks, then word reading and pseudoword decoding tasks. The only significant difference was, again, between text reading and word reading tasks: F (1, 10) = 34.8, p < .001

Type of bilingualism

Concurrent RAN-reading relationships appeared to be the strongest for simultaneous bilinguals, followed by L2 learners, and sequential bilinguals. No significant differences were found between the three groups. Longitudinal analysis showed the same rank order of correlation strength (highest for simultaneous bilinguals, weakest for sequential ones). The Wald test revealed significantly stronger RAN-reading correlations for simultaneous over sequential bilinguals F (1, 8.1) = 28.6, p = .001 and over L2 learners F (1, 3.8) = 17.2, p = .016.

Reading group

Concurrent analyses revealed the strongest RAN-reading relationship for the dyslexic group, followed by typical, and at risk groups. Two of these contrasts (dyslexic vs. at risk and dyslexic vs. typical) approached significance (p < .10). In the longitudinal analysis, RAN correlated stronger with reading in an at risk group than in a typical group. Longitudinal data for a dyslexic group was not available in the dataset. The Wald test showed a significant difference between at risk and typical groups F (1, 13) = 7.49, p = .018.

Reading level

Concurrent RAN-reading correlations for intermediate and novice readers were stronger than for pre-readers and proficient readers. The difference was significant between intermediate and pre-readers group only F (1, 11) = 5.48, p = .004. In the longitudinal analysis, only two groups (pre-readers and novice readers) were included, showing similar strength of the RAN-reading relationship.

Writing system

RAN-reading correlations in alphabetic and other languages groups were similar, both concurrently and longitudinally.

Time interval

RAN-reading correlations were somewhat stronger for the shorter time interval (kindergarten-grade 1) than the longer one (kindergarten-grade 2). The difference was not significant, however.

In relation to orthographic depth, our bilingual dataset contained studies with data predominantly in opaque languages (e.g. English) with a few studies reporting data in intermediate orthographies (e.g., Dutch, Norwegian). Considering the dominance of opaque orthography (English) in our database, we were not able to include orthographic depth in the moderator analyses.

Publication bias

To assess whether our results were driven by publication bias, we ran the Trim-and-Fill analysis and Vevea and Hedges’ likelihood ratio test. The results revealed that our meta-analyses for RAN-reading relationships were at minimal risk of publication bias. The funnel plot for the bilingual concurrent dataset (see Fig. 2) showed that no studies were missing to reach the funnel plot symmetry. The Vevea and Hedges’ likelihood ratio test results chi2(6) = 6.78, p = .342 were not statistically significant, indicating low risk of publication bias. The funnel plot for the longitudinal data (see Fig. 3) showed that only 9 imputations were needed to reach the funnel plot symmetry. Having added them, the overall effect size stayed effectively the same [changed from Z = − 0.3968 (r = −.38) to Z = − 0.4184 (r = −.40)]. The Vevea and Hedges’ likelihood ratio test results chi2(5) = 2.13, p = .831 were not statistically significant either, indicating low risk of publication bias.

Fig. 2
figure 2

Funnel plot for RAN-reading concurrent relationship. Black circles indicate observed effect sizes

Full size image
Fig. 3
figure 3

Funnel plot for RAN-reading longitudinal relationship. Black circles indicate observed effect sizes, white circles indicate imputed effect sizes

Full size image

Discussion

The present study is the first meta-analysis of the RAN-reading relationship in bilingual children, thus filling in the important gap in previous research. It adds to previous research by reporting both concurrent and longitudinal RAN-reading relationships.

Our first research question was to determine the strength of concurrent and longitudinal RAN-reading relationship in bilingual children. Our findings revealed the average correlation between RAN and reading to be r = −.39, which is comparable to those reported in previous meta-analyses in monolingual children (r = .43: Araújo et al., 2015; r = .44: Chen et al., 2021; see Table 1). The magnitude of longitudinal RAN-reading association was r = − .38, that is in line with previous longitudinal meta-analyses results in English-speaking monolinguals (r = − .38: McWeeny et al., 2022) in the English language. Our findings indicate that RAN tasks can predict reading ability in bilingual learners just as effectively as in monolinguals.

Our second research question was to compare the strength of RAN-reading relationships measured within language and between languages. To our knowledge, this study is the first systematic review to explore this. We found that the RAN-reading correlations were comparable whether measured within or between languages. The same was the case whether those relationships were analysed concurrently and longitudinally. Thus, RAN tasks have similar potential to predict individual differences in reading not only within language but also across languages. Within the interactive transfer model (Chung et al., 2019), we can explain the RAN-reading relationship across languages, irrespective of their typological difference, by interaction of both domain-general and language-specific components of RAN.

Considering between-study variability, our third research question aimed to analyse moderators to gain some insights into RAN and reading relationships in bilingual children. Our analysis broadly replicated previous meta-analyses (Araújo et al., 2015; Chen et al., 2021) insofar as alphanumeric RAN tasks showed stronger correlations with reading than non-alphanumeric RAN tasks. However, in our analysis most of those differences were fairly small and fell short of statistical significance. In addition, RAN correlated significantly more strongly with fluency than with accuracy measures, though in the concurrent analysis only. This finding is in line with previous reports (Araújo et al., 2015; Song et al., 2016—though note McWeeney et al. (2022), where no significant difference was detected) and support the claim that the RAN-reading correlation can be partly explained by the shared processing speed requirement (Araújo et al., 2015). We also examined whether reading domain (the type of reading tasks) had a moderator effect on the RAN-reading relationship. Here, the findings are harder to interpret as they were not fully consistent across concurrent and longitudinal analyses. In the concurrent analysis, text reading emerged as the strongest correlate of RAN, while reading comprehension clearly the weakest. This pattern broadly replicated Araújo et al. (2015) analyses (where the differences were smaller though). In the longitudinal analysis it was, once more, the text reading that emerged as the strongest correlate, while pseudoword reading was the weakest. The same moderator analysis allowed us to compare word and pseudoword reading—a contrast analysed, with mixed results, in several previous meta-analyses (Araújo et. al., 2015; McWeeney et al., 2022: significant difference in favour of real words; Chen et al, 2021: no difference whatsoever). Our results showed no clear outcome: in the concurrent analysis, correlations were stronger for pseudoword decoding than real word reading, while reverse was the case in the longitudinal analysis, with neither difference being statistically significant. Thus, the moderating role of the reading domain still requires further clarification.

A major contribution of the current study is the analysis of the type of bilingualism as a moderator of the RAN-reading relationship, as no previous meta-analytic review has explored the role of bilingualism-related factors. Our analyses produced some (albeit rather weak) evidence that the type of bilingualism may be a moderator of the RAN-reading relationship. While RAN and reading correlated significantly in all three bilingual groups, the correlations were the strongest for the simultaneous bilinguals, and weakest for the sequential bilinguals, with L2 learners falling in between. This pattern was observed in concurrent and longitudinal analyses alike. However, none of those differences were significant in the concurrent analyses. Although the longitudinal analysis revealed a significant difference between the simultaneous group and two other groups (L2 learners and sequential), we have to treat this result with caution as only two effect sizes from one study were available for the simultaneous group.

We also found (both concurrently and longitudinally) the mean effect size for the dyslexic group to be larger relative to the typical readers and at risk groups. Yet, the differences between groups were not significant concurrently. Only longitudinal comparison between at risk and typical groups reached significance (p = .018). However, an unproportionally smaller number of effect sizes for at risk and dyslexic groups drawn from a few studies compared to typical readers does not allow us to make strong conclusions as to the moderator effect of bilinguals’ reading proficiency.

Exploration of the reading level moderator revealed no significant differences between groups, except for intermediate and pre-readers groups. Also, we did not find any difference between the parameters of the writing system, and time interval (for longitudinal data).

Our fourth and fifth research questions were to compare the strength of RAN-reading correlations for monolingual and bilingual participants directly, i.e., holding the methods of assessment constant. Unfortunately, a small number of such direct comparison studies did not allow moderator analyses. Though we computed an average effect size for the monolingual group (r = −.48, p = .014, n = 22) and for the bilingual group (r = −.45, p = .067, n = 50) (the latter effect failing to reach significance), we could not answer our research questions. One recommendation for future research, which follows from our analysis, is simply that more studies comparing mono- and bilingual participants are needed.

By and large, not only does RAN predict reading equally well both within and across languages in bilingual children, but also the effects of moderators studied in previous meta-analyses in monolingual children have been found to be similar in bilingual children. Thus, our findings regarding the moderators’ effects are broadly consistent with the findings from previous meta-analyses (e.g., Araújo et al, 2015). These findings provide support for the validity of RAN measures in the assessment of reading difficulties in the bilingual context.

Limitations

Some limitations of this study should be noted. First, we limited our meta-analysis only to published studies, which could potentially reflect the publication bias. Yet, statistical methods aimed at identifying one revealed that our study was at a low level of the publication bias risk. Second, the description of bilingualism (e.g., languages spoken at home, age of second language acquisition, etc.) varied considerably across studies. The information on language status and proficiency of the participants was often incomplete and presented inconsistently, making classification (into bilingualism types) difficult. While language status could be deduced from the study description, insufficient information about the participants’ language proficiency did not allow any meaningful comparisons across studies. Moreover, the majority of the studies in our dataset dealt with either sequential bilinguals or L2 learners, with simultaneous bilinguals being grossly underrepresented. A low number of the reported effect sizes might skew the interpretation of the results, thus pointing to a need for further research into the effect of the different bilingual contexts on the RAN-reading relationship. A small number of samples and reported effect sizes also influenced the analyses of the categories of reading level, writing system, language pairs, and time interval.

Another issue of this meta-analysis is methodological differences between studies. Some studies focusing specifically on RAN reported a larger number of effect sizes with detailed description of the methodology, RAN tasks and reading outcomes, whereas others reported RAN effect sizes only as a part of an assessment battery providing more general information on how RAN was measured. Therefore, while some of our results can be considered preliminary, they clearly identify areas where there is a need for further research.

Conclusion

This first meta-analysis of the RAN-reading relationship in bilingual children builds on the body of research that already exists in monolingual children. Our findings demonstrate that, in bilingual children, significant RAN-reading relationships (of weak to moderate strength) exist, both concurrently and longitudinally. The strength of those relationships is comparable to those observed in monolingual participants. Not only was the strength of the RAN-reading relationship robust within one language, but it also proved to be a strong predictor between languages, demonstrating a cross-language effect. We suggest this can be accounted for if we assume that RAN performance relies on some universal, domain-general processes, which also underlie reading. Our analysis also suggests that the type of bilingualism may be a significant moderator of the RAN-reading relationship (at least longitudinally), though here the relevant data are insufficient to reach a firm conclusion. We also identified other moderators the role of which remains unclear, suggesting areas for further research.