Grammatical knowledge, in general, refers to knowing the rule of grammar application in reading text construction of verb form, word order, and sentence structure (Shiotsu, 2010; Purpura, 2013). Under the framework of the cohesive tie, past studies have confirmed grammatical knowledge as a prerequisite of efficient reading for reading both L1 and L2 texts (Grabe, 2009; Jeon and Yamashita, 2014). This is because grammatical knowledge determines the sentence structure and further impacts reading comprehension through the word function sequence (e.g., Gahl and Garnsey, 2004; Vinyals et al., 2015), which means that grammatical knowledge may assist individual word comprehension or semantic chunk comprehension when readers use the rule of sentence structure for decoding sentence meanings (Tunmer, 1989; Rego and Bryant, 1993); that is, the better the proficiency in grammatical knowledge application, the better the text reading comprehension performance. Readers could take the grammatical rule to assist text comprehension through the indirect contribution of text semantic meaning inference (Rego and Bryant, 1993).

A consensus has been reached on the fact that fully developed grammatical knowledge could allow readers to monitor syntactic information through text comprehension progress. The relationship between grammatical knowledge and reading comprehension is associated with the reader’s ability to generate coherence in the text and monitor meaning acquisition during the reading process (Fender, 2001; Grabe and Stoller, 2012). For example, the grammatical clues offered coherence hints or information to aid readers in constructing text and understanding discourses (Zwaan and Rapp, 2006; Aryadoust and Baghaei, 2016), which indicates that the grammatical knowledge could be regarded as an ability to process syntactic structures and complicated sentences in text comprehension (Zwaan and Rapp, 2006; Aryadoust and Baghaei, 2016).

Previous studies, however, reported various results regarding the role of grammatical knowledge in text reading comprehension, from a strong correlation to a weak correlation. Some studies reported a strong correlation between grammatical knowledge and reading comprehension (Bentin et al., 1990; Muter et al., 2004; Shiotsu and Weir, 2007; Kim and Cho, 2015). Conversely, other empirical surveys and experimental research reported a weak association due to the development effect of lexical inference (Kern, 1989; Kleeck, 2008; Zhang, 2012), knowledge base (Schatschneider et al., 2004; Silva and Cain, 2015; Brimo et al., 2017), and phonology knowledge (Demont and Gombert, 1996; Blackmore and Pratt, 1997; Farnia and Geva, 2013). Therefore, the correlation between grammatical knowledge and text reading comprehension was unclear.

The current study selected grade group and language type as potential moderators for the following reasons.

In the past two decades, a few studies (e.g., Jeon and Yamashita, 2014) reviewed the contribution of grammatical knowledge to reading comprehension. For example, Rodd et al. (2015) used a meta-analytic approach to report the effects of localizing semantic grammatical knowledge and written comprehension on brain function. Jeon and Yamashita (2014) reported the correlation between grammatical knowledge and general L2 acquisition. However, past meta-analysis studies failed to solve the significant heterogeneity problem through the moderator analysis, which means that the theoretical contribution to the correlation between reading comprehension and grammatical knowledge requires further investigation. In addition, previous studies only selected a small number of empirical studies for correlation calculation, meaning that the selected studies were not representative. The current picture of the interaction effect with reading stage development and learning cognitive condition statement thus remains unclear.

This article extends the current literature concerning the influence of the developmental relations of grammatical knowledge on reading comprehension, investigating the overall correlation between grammatical knowledge and reading comprehension from the recent research published in the past 20 years. This study further examines the interaction effects of the reading stage and learning cognitive condition on the correlation between reading comprehension and grammatical knowledge.

To avoid the comprehension system mistake (García and Cain, 2014), this study only selected materials written in Chinese and English. The Chinese materials were selected from the CNKI database, which included almost all possible academic resources written in Chinese, and the English materials were searched from Google Scholar, PsycINFO, ERIC, and Pre-Quest. Relevant studies were identified through two groups of keywords; the first group of words was related to grammatical knowledge (grammar*, grammatical*, syntax*, syntactic*, oral cloze test*, sentence completion*, error recognition*, implicit knowledge*, and explicit knowledge*) and the second group of words was related to reading comprehension (sentence comprehension*, paragraph comprehension*, passage comprehension*, text comprehension*, reading comprehension*, reading ability*, comprehension ability*, reading acquisition*, reading performance*, comprehension ability*, and comprehension performance*). This study attempted to select all possible materials with publication dates from 1 January 1998 to 1 December 2022, including published or unpublished research articles, dissertations, and conference articles. Thus, 1,104 articles were selected from the database.

Materials eligible for inclusion in this meta-analysis were as follows: (a) empirical studies with a minimum of 30 participants, (b) the selected studies should have correlation scores on the correlation between grammatical knowledge and reading comprehension, (c) the participants should be students from grade 1 of primary school to undergraduate students, (d) students were not diagnosed with deaf or blind, (e) studies investigated the concurrent correlation between grammatical knowledge and reading comprehension, (f) studies should provide enough indicators which could be transformed into Fisher’s z-transformation, and (g) grammatical knowledge and reading comprehension should come from the same scripts (e.g., L1 grammatical knowledge and L1 reading comprehension, L2 grammatical knowledge and L2 reading comprehension). If the grammatical knowledge and reading comprehension came from different scripts (e.g., L1 grammatical knowledge and L2 reading comprehension), the related study was removed. Thus, 128 articles fulfilled the inclusion criteria.

Studies were coded according to the characteristics of the participants and the measurements by two independent coders. Coding was based on (a) study number, (b) first author’s name, (c) publication year, (d) material type (journal article or dissertation), (e) sampling area, (f) sample size, (g) grade group (L refers to grades 1–3 of primary school, H refers to grades 4–6 of primary school, S refers to secondary school, U refers to university), (h) target language scripts, and (i) correlation effect size. Any unclear information was emailed to the article’s author for clarification. Because this study did not investigate the transfer effect in grammatical knowledge and attempt to examine the hypothesis from the learning cognitive condition theory, the two coders were asked to remove those articles in which the grammatical knowledge and reading comprehension did not come from the same scripts, for example, the correlation between L1 grammatical knowledge and L2 reading comprehension. If one article provided more than one correlation that came from different grade groups, this study treated them as an independent study (Mol and Bus, 2011). If one article provided more than one available correlation indicator in the same grade group, first, this study preferred to select the correlation indicator from the standardized measurements than from the research-developed test. After this step, if the article still had more than one available indicator, Hedges et al. (2010) cluster regression was applied to calculate the final Fisher’s z-transformation for this study, ensuring that each study only provided one correlation indicator for further meta-analysis (Mol and Bus, 2011). The internal reliability for two independent coders was 0.96 in the first round of coding. The difference came from sampling area coding; however, this problem was solved by using the country’s name as the sampling area. Thus, 65 articles remained after the coding. Detailed information can be found in Table 1. All materials selected for the meta-analysis have been listed in references with “*.”

This study removed two articles conducted by Zhang (2011, 2012), and one study conducted by Brimo et al. (2017) due to the effect size, which was over 3.5 standard deviation (García and Cain, 2014). As a result, 62 articles with 86 studies were included in the final analysis. The researchers input the correlation indicator through comprehension meta-analysis 3.0 and transformed it into Fisher’s z for further analysis. This study selected Fisher’s z-transformation because z is approximately constant, and z has asymmetrical distribution (Borenstein et al., 2010). The values of Fisher’s z were 0.10 (r = 0.10), 0.31 (r = 0.30), and 0.50 (r = 0.50), denoting small effect size, moderate effect size, and large effect size, respectively (Borenstein et al., 2010).

This study used the random-effects model to report the effect size (Borenstein et al., 2010) and also reported a 95% confidence interval (CI). The effect size was interpreted as significant if the CI did not include zero. Next, the Q value was reported to examine the heterogeneity within materials. If the Q value reached a significant level (p < 0.05), then a meta-regression analysis was carried out to further analyze the effects of selected moderators (Hedges and Pigott, 2004; Borenstein et al., 2010). To compare the effect sizes across factors, this study applied the following equation to examine the difference of Teta (Durlak, 2009; Lenz, 2013): Teta = Diff / SE, Diff = z ₁ – z ₂, SE = sqrt (Variance z ₁ + Variance z ₂).

If |Teta| ≥ 2.58, the factor difference was interpreted as significant (p < 0.01). To examine publication bias, this study examined Rosenthal’s fail-safe number, a funnel plot through the trim-and-fill approach, the rank correlation test, and Egger’s regression test (Borenstein et al., 2010).

A total of 62 articles with 88 studies (N = 14,852) were included in this meta-analysis. In total, 16 studies were master’s dissertations and doctoral theses, and the other 72 studies were published in peer-reviewed journals. Specifically, 20 studies (n = 3,324) investigated the correlation between grammatical knowledge and reading comprehension in group L, 16 studies (n = 2,749) reported the correlation in group H, 29 studies (n = 5,652) investigated the correlation in secondary school students (group S), and 21 studies (n = 2,747) examined the correlation in higher education students (group U). For language type, 41 studies (n = 6,505) investigated the correlation between grammatical knowledge and reading comprehension in L1 scripts, and 47 studies (n = 8,347) focused on the relationship between L2 grammatical knowledge and L2 reading comprehension.

Table 2 provides the results of the correlation effect size between grammatical knowledge and reading comprehension. The overall effect size was large (Fisher’s z = 0.54); however, the heterogeneity analysis showed that the materials’ heterogeneity was significant (Q = 271.48, p < 0.001). Moderator analysis through meta-regression showed that grade group explained 59% variance of materials’ heterogeneity and language type explained 13% variance of materials’ heterogeneity. Publication bias examination showed the effect size-distributed symmetry (see Figure 1): Rosenthal’s fail-safe number was 8,824, the tau value in the rank correlation test was insignificant (tau = 0.04, p > 0.05), and the intercept value in Egger’s regression test was not significant (intercept = 0.26, p > 0.05). The publication bias examination showed that the current study did not have significant publication bias.

Because the majority of materials’ heterogeneity resulted from the grade group, this study further investigated the correlation effect size between grammatical knowledge and reading comprehension in four grade groups. Regarding group L, the results showed that the correlation effect size was moderate (Fisher’s z = 0.36). The heterogeneity analysis showed that the materials’ heterogeneity was not significant (Q = 23.29, p > 0.05). Publication bias examination showed the effect size-distributed symmetry (see Figure 2): Rosenthal’s fail-safe number was 1,935, the tau value in rank correlation test was insignificant (tau = 0.06, p > 0.05), and the intercept value in Egger’s regression test was not significant (intercept = 0.41, p > 0.05). the results showed that the current study did not have significant publication bias.

Regarding group H, the results showed the correlation effect size was nearly large (Fisher’s z = 0.49), and heterogeneity analysis showed that materials’ heterogeneity was not significant (Q = 11.52, p > 0.05). Publication bias examination showed the effect size-distributed symmetry (see Figure 3): Rosenthal’s fail-safe number was 2,255, the tau value in rank correlation test was insignificant (tau = 0.03, p > 0.05), and the intercept value in Egger’s regression test was not significant (intercept = 0.37, p > 0.05). The results showed that the current study did not have significant publication bias.

Regarding group S, the results showed that the correlation effect size was large (Fisher’s z = 0.61), and heterogeneity analysis showed that materials’ heterogeneity was not significant (Q = 9.81, p > 0.05). Publication bias examination showed the effect size-distributed symmetry (see Figure 4): Rosenthal’s fail-safe number was 2,555, the tau value in rank correlation test was insignificant (tau = −0.09, p > 0.05), and the intercept value in Egger’s regression test was not significant (intercept = −0.16, p > 0.05). The results showed that the current study did not have significant publication bias.

Regarding group U, the results showed the correlation effect size was large (Fisher’s z = 0.69), and heterogeneity analysis showed that materials’ heterogeneity was not significant (Q = 10.77, p > 0.05). Publication bias examination showed the effect size-distributed symmetry (see Figure 5): Rosenthal’s fail-safe number was 5,188, the tau value in rank correlation test was insignificant (tau = 0.14, p > 0.05), and the intercept value in Egger’s regression test was not significant (intercept = 0.36, p > 0.05). The results showed that the current study did not have significant publication bias.

Effect size comparison examination showed that the difference between each of the two groups’ effect size was significant (Teta _{H & L} = 4.60, p < 0.01; Teta _{H &S} = 4.90, p < 0.01; Teta _{S & U} = 3.27, p < 0.01).

A large effect size was found in the overall correlation between grammatical knowledge and reading comprehension. Jeon and Yamashita (2014) reported that the correlation between grammatical knowledge and reading comprehension in L2 scripts was extremely high (r = 0.85) in groups S and U, which was very different from the current results. Jeon and Yamashita (2014) had some obvious limitations. First, they only included 18 articles in their meta-analysis, and the results showed that the Q value from heterogeneity analysis was significant. Second, Jeon and Yamashita (2014) did not remove over 3.5 standard deviation effect sizes as the outlier, which showed that the correlation was unclear and did not generate convincing conclusions due to the significant heterogeneity problem. Through the literature search, the current results were consistent with most empirical studies, which showed that the average correlation between grammatical knowledge and reading comprehension was nearly large (Mecartty, 2000; Gersten et al., 2001; Perfetti et al., 2005), indicating that the grammatical knowledge significantly determined the comprehension process. Moreover, this result informed how the cohesive tie function of grammatical knowledge might have great determining power on reassembling decoded words into phrases and clauses, allowing readers to be more efficient in detecting and correcting reading errors to enhance their comprehension process indirectly (e.g., Slobin, 1966; Bowey, 2005; Kempen et al., 2012).

The current results showed that the higher grade group had a significantly larger effect size than the lower grade group. This result was consistent with the majority of past longitudinal studies (e.g., Oakhill et al., 2003; Farnia and Geva, 2013). The reasons for the consistency could be that the requirements, the complicated structure, and the organization level of text comprehension were higher in the higher grade group, resulting in the fact that the requirement of the cohesive tie function was higher in the text coherence clues process (Cain, 2003; Westerveld et al., 2008; Nuske and Bavin, 2015) and organization structure cognition (Oakhill and Cain, 2000; Vaughan and Dillon, 2006; Kendeou et al., 2007). This result informed that the function of the cohesive tie might continue developing after the age of 16, which was different from the majority of comprehension factors, such as decoding ability, vocabulary knowledge, and metalinguistic knowledge (Mol and Bus, 2011; García and Cain, 2014; Trapman et al., 2014). The current result also echoed reading stage theory regarding grammatical knowledge function on text reading comprehension, where the more complicated-level text comprehension needed a higher application of grammatical knowledge for coherence inference.

In each grade group, the language type was not a significant moderator on the correlation between grammatical knowledge and reading comprehension. This result echoed the previous study, which showed that linguistic skills could work similarly across languages’ scripts (Cummins, 1986; Gottardo and Mueller, 2009; Perfetti and Stafura, 2014). In other words, grammatical knowledge could have a transfer effect across all languages, suggesting that grammatical knowledge had an independent cognition function on text comprehension. Readers received text information and delivered target comprehension information into the situation model-building process. The grammatical knowledge worked together with other linguistic skills on text meaning coherence, which indicates that this process encompassed the integrated process and the synthesizing process on word reading and sentence meaning judgment in situation model construction. According to the current results, grammatical knowledge may have an independent effect on reading comprehension across various scripts’ text reading, which means that, for bilingual learners’ reading comprehension, a transfer or compensation effect would help readers to apply either L1 or L2 grammatical knowledge to other language scripts’ comprehension (Sparks et al., 2008). Therefore, readers only need to fully develop grammatical knowledge in one language script, and then, the grammatical knowledge can work in other scripts’ text comprehension processes.

The current study has several limitations. First, this study selected materials that were only written in Chinese or English. Materials written in other languages were not considered. Second, the selected materials only reported the correlation between grammatical knowledge and reading comprehension in Chinese reading, Dutch reading, English reading, and French reading. Next, although past studies mentioned that grammatical knowledge should involve morphological knowledge and verb forms, due to inconclusive findings, this study excluded these two variables. In future studies, if the researcher has an agreed definition in the grammatical knowledge category, it should be further explored. Finally, meta-regression only presented how the extent of the heterogeneity of the materials could be explained by the moderator, while the interaction effect mechanism of each moderator was not identified. For example, this study reported that the grade group and language types were two significant moderators on the overall correlation but did not investigate the internal working interaction effect between these two moderators.