Mediation in the Relation of Orthographic Processing on the Lexical and Sublexical Level with Reading and Spelling Skills. A Large Cross-Sectional Study in Elementary School Children in Germany

ABSTRACT

Purpose

Two types of orthographic knowledge were examined: i) knowledge of permissible letter combinations (general orthographic knowledge) and ii) knowledge of whole words (word-specific orthographic knowledge), to gain further insights into the relationship of general and word-specific orthographic knowledge with literacy skills.

Method

Mediation models were estimated using a sample of 2,636 3rd and 4th grade children (50.6% female, ethnic backgrounds were not surveyed) to examine whether general orthographic knowledge predicts reading fluency and spelling performance and whether this path is mediated by word-specific orthographic knowledge. Using confirmatory factor analyses (CFA), we examined whether the responses to correctly spelled words/legal letter patterns and pseudohomophones/illegal letter patterns in word-specific and general orthographic knowledge tasks were generated by separate latent constructs.

Results

Our results confirm that word-specific orthographic knowledge is a facilitatory mediator in the relationship between general orthographic knowledge and literacy skills. The relationship of general and word-specific orthographic knowledge with literacy skills varies by item type, with a tendency toward higher correlations between literacy skills with pseudohomophones (vs words) and illegal (vs legal) pseudowords, which is in line with the two-factor solution of the CFA.

Conclusion

From a theoretical perspective, we conclude that general orthographic knowledge facilitates orthographic learning of words. From a methodological perspective, we encourage future researchers to distinguish between targets and foils in orthographic decision tasks.

Introduction

Becoming a skilled reader and speller requires knowledge of one’s orthography on different levels, including the knowledge of letters, letter patterns, and whole words. Relying on such knowledge is often collectively referred to as orthographic processing. Over the last decades, orthographic processing has become a widely used term with numerous definitions (Apel et al., 2019; Conrad & Deacon, 2023). While some definitions focus on the stored mental representations of a word (Ehri, 1980; Olson et al., 1994), other definitions focus on the ability to recognize frequent letter patterns to understand and apply print conventions of a writing system (Conrad et al., 2013; Deacon, 2012; Georgiou et al., 2008). There is evidence that children are sensitive to basic conventions of the writing system of their language even before they start learning to read (e.g., “bdc” looks more like a word than “bbb;” Ferreiro et al., 1996; Pollo et al., 2009). Furthermore, children are sensitive to permissible letter patterns within their language (e.g., doubled consonant letters at the beginning of a word are not permissible or very rare in most orthographies) at the beginning of reading and spelling acquisition (Apel et al., 2012; Cassar & Treiman, 1997; Pacton et al., 2001; Rothe et al., 2014; Zhang & Treiman, 2021). This sensitivity increases during reading and spelling acquisition (Deacon et al., 2012, 2019; Ise et al., 2014; O’Brien, 2014), and higher sensitivity to permissible letter patterns is related to higher reading and spelling skills (Bakos et al., 2018; Cassar & Treiman, 1997; Deacon et al., 2019; Rothe et al., 2015; Zarić et al., 2021).

Sensitivity to permissible letter patterns is often referred to as orthographic processing at the sublexical level, or general orthographic knowledge. It is commonly measured by pseudoword forced-choice or decision tasks. In the pseudoword-forced choice task participants are asked to indicate which of two alternative pseudowords looks more like a real word when one of the pseudowords violates orthographic regularities (e.g., “baff” – “bbaf;” Cassar & Treiman, 1997). In a decision task, participants indicate whether an individually presented pseudoword could be a real word (e.g., “ferrab”) or not (e.g., “fferab;” Rothe et al., 2015). Both tasks have been used in the literature (e.g., Rothe et al., 2015): the main difference is whether participants see one item at a time and make a “yes”/“no” decision, or whether they simultaneously see both a legal and illegal variant of a pseudoword and indicate which of them is more word-like. From a methodological perspective, the former task helps avoid ceiling effects in children and adults (Rothe et al., 2015; Schmalz et al., 2019). Overall, a point of criticism of many implementations of the general orthographic knowledge tasks is that it is often not specified which regularities are being manipulated (Apel et al., 2019): for example, violations can be of positional constraints (e.g., words in English rarely start with a double consonant; exception: “llama”) or relate to the legality of letter clusters (e.g., words in English often contain the double consonant “ll,” but rarely contain the double consonant “jj”).

In addition to general orthographic knowledge, there is evidence that children store orthographic representations of whole words during reading and spelling acquisition (for an overview, see Share, 1995). This orthographic processing on the lexical level, or word-specific orthographic knowledge, is commonly measured by orthographic choice or decision tasks, in which participants distinguish between correctly spelled word from a pseudohomophonic foil (e.g., “rane” – “rain;” Olson et al., 1994), or indicate whether a single word or a pseudohomophone is spelled correctly (Bergmann & Wimmer, 2008). Word-specific orthographic knowledge – a mental representation that is easy to access – is required for automatized reading, as it allows the reader to directly access the meaning and pronunciation of a word, rather than requiring letter-by-letter decoding (Castles & Nation, 2010). Word-specific orthographic knowledge in the form of a precise, high-quality lexical representation is also essential for spelling, as it contains detailed knowledge of the letters and their order (Banfi et al., 2021; Mehlhase et al., 2019; Moll & Landerl, 2009).

Both types of orthographic knowledge have been studied in relation to reading acquisition. However, reported correlations between general orthographic knowledge (sublexical level) with reading and spelling skills are mostly smaller (r = .45–.69) compared to word-specific orthographic knowledge (lexical level) (r = .63–.90; Arab-Moghaddam & Senechal, 2001; Bekebrede et al., 2009; Cunningham & Stanovich, 1993; Deacon, 2012; Georgiou et al., 2008; MacKay et al., 2022; Roman et al., 2009; Rothe et al., 2014). There is also evidence for an orthographic processing deficit in children with dyslexia, when compared to chronological- or reading-age matched controls, on the lexical level (e.g., Bergmann & Wimmer, 2008; Georgiou et al., 2012) and on the sublexical level (e.g., Breznitz & Meyler, 2003; Rothe et al., 2015). In line with findings of stronger relations for word-specific than general orthographic knowledge to reading and spelling skills, a meta-analysis by Georgiou et al. (2021) revealed stronger deficits for word-specific than general orthographic knowledge in dyslexic children.

There are several potential reasons for the lower correlations between general orthographic knowledge and reading ability compared to word-specific orthographic knowledge with reading and spelling. First, word-specific orthographic knowledge is a measure to examine the mental representation of words and is designed to measure a level of orthographic knowledge that is directly relevant for word recognition. High-quality of these representations is required for automatized reading and spelling of familiar words. In contrast, general orthographic knowledge might act on reading and spelling via mediating factors (e.g., word-specific orthographic knowledge), rather than directly (e.g., Henbest & Apel, 2018; Schmalz et al., 2021). Second, the size of the correlation is, on average, limited by the reliability of a task (e.g., Vul et al., 2009). If the reliability of a typical general orthographic knowledge task is lower than that of a typical word-specific orthographic knowledge task, we would expect lower correlations of general orthographic knowledge with reading ability. This explanation would predict a difference in the size of the correlation between word-specific and general orthographic knowledge tasks, in the absence of a theoretically relevant reason for this difference.

While it is well established that there is some form of relationship between both types of orthographic knowledge and reading and spelling ability, the causal chain underlying this relationship is less clear. In particular, it is unclear how general orthographic knowledge relates to reading and spelling, and to word-specific orthographic knowledge. As outlined above, theoretical and experimental work is relatively clear on the importance of word-specific orthographic knowledge. For general orthographic knowledge, while a correlation with reading ability is well established, it is not clear whether this relationship is causal in the direction of superior general orthographic knowledge leading to better reading ability (Deacon et al., 2012). However, as general orthographic knowledge is evident in kindergarten even before the onset of reading instruction, it is likely that general orthographic knowledge has a causal role in later reading and spelling acquisition. Furthermore, numerous plausible causal theories propose that general orthographic knowledge acts as a cause for reading-related processes (e.g., Conrad & Deacon, 2023; Ehri, 2015; Treiman & Kessler, 2022).

In understanding the proposed causal relationship between general orthographic knowledge and reading ability, word-specific orthographic knowledge may act as a mediator (Henbest & Apel, 2018): general knowledge about legal letter patterns may constrain the number of possibilities in which a particular word may be spelled and may thus facilitate the formation of word-specific orthographic representations. These, in turn, help children to read fluently and spell accurately. General orthographic knowledge may also have a direct effect: for reading fluency, it may constrain participants’ expectations of upcoming letters (Chetail, 2015, but see; Schmalz & Mulatti, 2017). In the case of spelling, a direct causal link between general orthographic knowledge and spelling ability may exist, because knowledge of legal letter patterns should constrain the plausible spelling of a word (e.g., a child who has no word-specific orthographic knowledge for the word “quick” may use their general orthographic knowledge to decide that it is unlikely to have the spelling kwick, even though this would be a phonologically plausible spelling).

In implementing an orthographic knowledge task, researchers need to make a decision about whether to use a choice (pick between two different spellings) or a decision (yes/no) -task. The latter provides a possibility for a more fine-grained analysis, because separate accuracy rates for targets (“yes”-responses for words or legal pseudoword) and nontargets (“no”-responses for pseudohomophones or illegal pseudowords) can only be calculated for decision tasks. For choice task, the two items are provided simultaneously in an alternative-forced-choice format, so there is only one single response for both the pseudoword and its corresponding base word. Previous studies have found, both at the lexical and sublexical level, that dyslexic children have more difficulty in correctly identifying nontargets (orthographic misspellings/illegal letter patterns) as “no”-responses, than identifying targets as “yes”-responses (correctly spelled words/legal letter patterns; lexical level: Bergmann & Wimmer, 2008; sublexical level; Rothe et al., 2015). Note that this is an empirical phenomenon which becomes relevant when designing or analyzing studies. The theoretical explanation for this finding is less clear. In fact, a single explanation may not be able to account for the phenomenon both in the word-specific and in the general orthographic knowledge task. For word-specific orthographic knowledge, a mental orthographic representation of a word allows a child to correctly indicate a pseudohomophone as not correctly spelled. Conversely, an erroneous “yes”-response to a pseudohomophone (e.g., “brane”) might be explained by an imprecise or absent orthographic representation, and on participants’ relying on sublexical decoding (decoding by grapheme-phoneme conversion). Decoding a pseudohomophone activates a phonological word form and matches an entry in the phonological and semantic lexica (as “brane,” pronounced/bræɪn/, sounds like the word “brain”). If no orthographic representation for the phonological entry exists, the participant will rely on the match with the phonological and semantic lexica as a kind of back-up strategy (e.g. Siegler & Shrager, 1984). Even if at later stages of reading acquisition, children are likely to have some degree of word-specific orthographic knowledge, their orthographic representations may be imprecise. Thus, older children with a deficiency in their lexical processing ability may provide an erroneous “yes”-response to a pseudohomophone if the orthographic representation is imprecise (Gangl et al., 2018). For words, any combination of sublexical decoding and lexical knowledge would lead to a correct “yes”- decision, even if there is no or only an imprecise representation of the word, because that the participant cannot detect the spelling error.

When it comes to general orthographic knowledge, an erroneous “yes”-response to an item containing an illegal cluster means that the graphotactic constraint has not been internalized by the participant and the participant fails to detect the illegal letter cluster. A correct “no”-response to an illegal letter cluster may be the result of either rule-like or statistical knowledge about a specific graphotactic constraint, or of an unsuccessful search process of the letter cluster in the mental lexicon, while a correct “yes”-response might be the result of either rule-like or statistical knowledge, or of a successful search process. However, a correct “yes”-response may also be given when the participant does not detect illegal letter clusters because graphotactic constraints have not been internalized.

Briefly, words and legal letter clusters can be answered correctly even if orthographic representations (words) or knowledge about specific graphotactic constraints (legal pseudowords) are imprecise, while precise orthographic representations or knowledge about specific graphotactic constraints are needed to correctly answer pseudohomophones and illegal pseudowords in order to detect spelling errors or illegal letter clusters. As a consequence, pseudohomophones and illegal letter clusters should be more strongly related to reading and spelling skills in dyslexic as well as in typical developing children, as these are more direct measures of orthographic representations and internalized graphotactic constraint.

To summarize, open questions remain about the psychometric properties of tasks measuring orthographic knowledge, especially when it comes to general orthographic knowledge. In addition, little is known about the relationship of general orthographic knowledge, reading and spelling ability, and word-specific orthographic knowledge. While there are several plausible causal pathways, it is not clear whether a causal relationship even exists. The main aim of the current study is to gain further insights into the cross-sectional relationship of orthographic processing on the lexical and sublexical level with reading and spelling skills. We analyzed a large-scale dataset with over 2000 children to address these questions. In a confirmatory factor analysis, we investigated whether responses to correctly spelled words/legal letter patterns and pseudohomophones/illegal letter patterns, both in the word-specific and general orthographic knowledge tasks, reflect different constructs, and whether this two-factor solution, distinguishing between items with a “yes” versus a “no”-response, gives a good fit. Next, we compare response times of correctly identified legal letter clusters and correctly identified illegal letter clusters to examine whether identifying an illegal letter cluster is the result of either the knowledge about a specific graphotactic constraint (would result in same or shorter response times as for correctly identified legal letter clusters) or an unsuccessful search process of the letter cluster in the mental lexicon (would result in longer response times as for correctly identified legal letter clusters). Furthermore, we hypothesized that (1) lower correlation for general orthographic knowledge with reading and spelling skills compared to word-specific orthographic knowledge can be confirmed; that (2) the correlation differs as a function of item type, with stronger correlations of literacy with accuracy to pseudohomophones and illegal item accuracy for word-specific and general orthographic knowledge, respectively, compared to correlations of literacy with accuracy to real words and legal items; and (3) that the relationship between general orthographic knowledge and literacy (reading and spelling skills) is mediated by word-specific orthographic knowledge.

Materials and methods

Families were initially invited to participate in a study on comorbidities between specific learning disorders and psychopathology in elementary school children (Visser et al., 2020).

Recruitment

A total of 52,734 families with children in 3rd and/or 4th grade from the two German federal states Hesse (through the Hessian Ministry of Culture; n = 25.000) and Bavaria (addresses provided by local registration offices; n = 27,734) were invited by letter to participate in the study. Families were invited to use a web-based application to assess children’s academic skills and psychopathological profile over five sessions of 30 to 45 minutes. Children processed all tests and questionnaires individually from home on a tablet or smartphone within a timeframe of eight weeks. For all participating children, the legal guardians gave their informed consent and the children gave their assent via the web-based application prior to the assessment and inclusion in the study. The ethics committees of the University Hospital of the Ludwig Maximilian University Munich (project ID: 438–16; date of approval: 25 August 2016) and of the DIPF | Leibniz Institute for Research and Information in Education, Frankfurt am Main (project ID: FoeDises; date of approval: 2 April 2017) reviewed and approved the study in accordance with the Declaration of Helsinki.

Participants

A total of 4,542 families started the application (response rate 8.6%). We excluded 1,817 cases of children who did not completed all relevant digital tests (up to session five) or if the result on any task was flagged as implausible (e.g., very fast response times, suggesting random selection of response options). For details about the plausibility checks please see Visser et al. (2020). We excluded further 26 cases because of an IQ ≤ 70 and 63 cases as the parents answered to an open question that the child had hearing or visual problems, neurological diseases, or chromosomal defects. In total, we excluded 1,906 cases, resulting in the final study sample of 2,636 children. The final sample is approximately equally distributed in terms of gender and grade level (grade 3: boys = 610, girls = 632; grade 4: boys = 692, girls = 702). For 70.68% of participating children, parents indicated that their child is monolingual German, while 23.25% were bilingual and for 6.07% this information was missing or ambiguous. With 1,458 mothers (55.3%) that have completed the highest German school degree, the sample is not representative in terms of socioeconomic status.

Collection of data

For this study, a software company transformed all standardized psychometric tests and questionnaires from their paper-pencil versions to an online tablet/smartphone version. Tests and questionnaires were grouped into sessions that had to be worked on five separate days. Each session lasted 30 to 45 minutes. For more details on the development of the software see Visser et al. (2020).

Materials

Reading fluency

Reading fluency achievement was assessed using a digitized version of the standardized “Wuerzburger Silent Reading Test – Revised” (WLLP-R; Schneider et al., 2011). In a study by Rothe et al. (2022) the relationship between the digitalized and the paper-pencil test was high (r = .80, n = 233) and was similar to the test-retest reliability of the original paper-pencil test (WLLP-R; Schneider et al., 2011, r = .80–.82). Children were presented with a series of written words and asked to select the corresponding image among four options within five minutes. The digital version uses the same items as the paper-pencil test. Scoring is based on the number of correctly selected images. Raw values (max. score 180) were used for statistical analyses and robustness analyses using grade-specific t-values instead of raw scores were performed. Third and fourth grade children responded to the same items.

Spelling

Spelling performance was assessed using a digitized version of the standardized “Weingarten spelling test for basic vocabulary” (Birkel, 2007a, 2007b). The internal consistency of the digitized version is good for the present sample (WRT3+ Cronbach’s α = .93; WRT4+ Cronbach’s α = .94). In a study by Rothe et al. (2022) the relationship between the digitalized and the paper-pencil test was high (WRT 3+: r = .74, n = 68, WRT 4+: r = .86, n = 165) but was slightly lower than the test-retest reliability of the original paper-pencil test (Birkel, 2007a, 2007b; 3^rd grade: r = .93, 4^th grade: r = .94). After a sentence frame with a target word was read aloud to the children, they were asked to insert the target word by using the virtual keyboard presented on the screen of their device without a time limit (third grade: 55 words, fourth grade: 60 words). Scoring was based on the number of correctly written words. In order to analyze the spelling performance of the third and fourth grade children (who wrote different items and a different number of items) together, grade-specific norms of the manual from the paper-pencil test were used for statistical analyses.

Nonverbal intelligence

Children’s nonverbal cognitive skills were assessed using a digitized version of three of the four subtests (series, classifications, matrices) of the “Culture Fair Intelligence Test” (CFT 20-R; Weiß & Weiß, 2008, test reliability r = .92). The internal consistency (McDonald’s omega = .75 for the total scale) and construct validity (correlation between the subtests and the overall score, series: r = .75, classification: r = .82, matrices: r = .82) of the digitized version is good (Visser et al., 2022). The fourth subtest could not be adapted to an online version. Therefore, we could not use the IQ norms of the manual, so we have developed grade-specific IQ-norms based on the complete sample that used the web-based application. For more details on the development of the norms, see Visser et al. (2020).

Phonological awareness

Phonological awareness was assessed by a phoneme deletion task. Children hear a pseudoword and were asked which word results if they delete a specified phoneme from the pseudoword (e.g., “Knater” without /n/, results “Kater,” English: tomcat). In all pseudowords a consonant within a consonant cluster must be deleted. For each item three response options (target and two distractors) were presented as an illustration. One distractor word was derived from the target word by exchanging the first grapheme (e.g., target word: “Kater,” distractor: “Vater,” English: father). The other distractor word was derived from the pseudoword by exchanging the deleting phoneme by one other phoneme (e.g., pseudoword: “Knater,” distractor: “Krater,” English: sinkhole). Children indicate their answer by tapping the illustration of the word. The test consists of 21 items. Children receive practice items before starting the test. The internal consistency of the task is acceptable (Cronbach’s α = .79). An example of the test is shown in Figure 1.

Figure 1. Illustration of the phoneme deletion task for the pseudoword knater without/n/.

Task format of orthographic knowledge

Orthographic knowledge is commonly measured by choice or decision tasks. A frequent way to construct orthographic choice tasks (word-specific) is to replace one grapheme of an orthographically correct spelled word by a homophonous grapheme. In a choice task, the correct and incorrect response then vary by one grapheme only: for example, if a German speaker is asked whether the word /bro:t/ (meaning “bread”) is spelled as Brot or Broht, their attention will be focused on the vowel spelling. This might trigger the use of explicitly learned orthographic rules, which are much practiced in German primary schools. Conversely, presenting either the word or pseudohomophone (decision task) does not draw attention to a specific grapheme and therefore it is a more direct way to measure the retrieval of mental representations. The same applies for general orthographic knowledge. This assumption is supported by ceiling effects observed in choice tasks for both general and word-specific orthographic knowledge already in 8–10-year-old German primary school students, while decision tasks seem to be more robust against ceiling effects (Rothe et al., 2015) even in adolescents (Bergmann & Wimmer, 2008). Therefore, when examining the retrieval of mental representations in a typically developed sample, it is appropriate to use a decision task rather than a choice task.

Word-specific orthographic knowledge (WOK): Orthographic processing on the lexical level was operationalized by measuring word-specific orthographic knowledge with an orthographic word decision task. Children were told that they will see single words. The task of the child was to indicate by pressing a green tick or red cross whether the visually presented stimulus is spelled orthographically correct or not. There were two types of experimental stimuli: words and pseudohomophones, both sound like real words. Words were spelled orthographically correctly, whereas pseudohomophones were misspelled. Some of the pseudohomophones contained a rarer spelling, such as “ai” for/aɪ̯/, which is more frequently spelt “ei” in German (e.g., word: Leiter, pseudohomophone Laiter, English translation: ladder). Other pseudohomophones contained a more frequent spelling, than the base word, such as “eu” for /ɔʏ/ which, depending on the root word, can also be spelled “äu” (e.g., word: Bäume, pseudohomophone: Beume, root word: Baum, English translation: trees/tree). Thus, on average, the typicality of the spelling was matched between words and pseudowords: The words were all high frequency words based on the childLex corpus (Schroeder et al., 2015); as such, we expect participants to have strong representations of the spoken forms of the words. ChildLex comprises a total of approx. 9.9 million words distributed over approx. 180,000 different words (types). According to their frequency in the corpus, the words are distributed over approx. 140,000 frequency ranks. The words in the word-specific orthographic knowledge task are selected to be among the 5% most frequent words in the corpus according to their frequency rank. Pseudohomophones were derived from these high-frequency words by exchanging one phonologically identical grapheme. Words and pseudohomophones were matched on the number of letters and on bigram- and trigram-frequencies according to childLex. This item selection was done to measure word-specific orthographic knowledge instead of a general orthographic knowledge about positional constraints and legality of letter clusters. There were 30 stimuli of each, words and pseudohomophones. The total of 60 stimuli was divided into two blocks (15 words and 15 pseudohomophones per block). Each stimulus was presented once as a single item on the screen. Stimuli were presented in a pseudorandomized order to control that no more than four consecutive stimuli of the same type were presented. Children received practice items before starting the test. The outcome variable is the number of correct responses. The internal consistency of the task is good (Cronbach’s α = .83). All items of the word-specific orthographic knowledge task and their bi- and trigram frequencies are presented in Table S1 in the online supplement material.

General orthographic knowledge (GOK): Orthographic processing on the sublexical level was operationalized by measuring general orthographic knowledge with an orthographic pseudoword decision task. Children were told that they will see single pseudowords. The task of the child was to indicate by pressing a green tick or a red cross whether the visually presented stimulus could be a real German word or not. Here, children saw a single item on the screen at a time and made “yes”/“no” decisions about the wordlikeness. This version of the task is preferable for this age group to avoid ceiling effects (Rothe et al., 2015), and it allows us to separately examine the responses for legal and illegal pseudowords. There were two types of experimental stimuli: legal pseudowords and illegal pseudowords. Both were orthographically and phonologically unfamiliar. While legal pseudowords contained only letter-strings that follow the German orthography, illegal pseudowords contained one grapheme which violated German orthography (e.g., legal pseudoword: fruhl, illegal pseudoword: fruul – while the grapheme “uh” occurs in the middle of German monosyllabic words such as “Stuhl” [chair], this is not true to the homophonous two-letter grapheme “uu”). Twenty-four out of the thirty illegal pseudowords contain double consonants or vowels that do not occur in German orthography such as “üü,” “ää,” “zz” “uu,” “kk” and “ii.” The remaining 6 items contains a “ß” in an illegal cluster-position (i.e., after a consonant). Legal pseudowords and illegal pseudowords were only matched on the number of letters as they cannot be matched by bi- and trigram frequency due to the illegal letter sequences, which have a frequency of 0. There were 30 stimuli of each, legal pseudowords and illegal pseudowords. The 60 stimuli were divided into two blocks (15 legal pseudowords and 15 illegal pseudowords per block). Each item was presented once, as a single stimulus in the middle of the screen. Stimuli were presented pseudorandomized to control that no more than four consecutive stimuli of the same type were presented. Children received practice items before starting the test. The outcome variables are the number of correct responses and the response times. The internal consistency of the task is good (Cronbach’s α = .84). All items of the general orthographic knowledge task are presented in Table S2 in the online supplement material.

Statistical analysis

First, we computed descriptive statistics (M and SD) to explore the distribution of age, reading fluency, spelling performance, IQ, phonological awareness, word-specific orthographic knowledge (accuracy: total, words, and pseudohomophones), and general orthographic knowledge (accuracy: total, legal and illegal pseudowords) for grades 3 and 4 and the total sample.

We estimated confirmatory factor analyses (CFA) to examine whether the responses to correctly spelled words/legal letter patterns and pseudohomophones/illegal letter patterns in word-specific and general orthographic knowledge tasks were generated by separate latent constructs. This analysis serves to identify if, based on the response patterns across items and participants, the individual items reflect the same theoretical construct (see Spencer et al., 2015, for a similar approach). For this, two-dimensional measurement models were specified for both tasks: words and pseudohomophones for word-specific orthographic knowledge and legal and illegal pseudowords for general orthographic knowledge. We employed the following cutoff values for evaluating model fit indices: Comparative Fit Index (CFI) ≥ .95, Root Mean Square Error of Approximation (RMSEA) ≤ .06, and Standardized Root Mean Square Residual (SRMR) ≤ .08 (Hu & Bentler, 1999). However, since employing rigid rules of thumb when evaluating model fit in the SEM framework can lead to type I errors (Marsh et al., 2004), these cutoff points were interpreted flexibly. We placed less emphasis on evaluating the p-value of the chi-square (χ2) test due to its high sensitivity to sample size. Considering the dichotomous nature of the data, we used tetrachoric correlation matrices (Drasgow, 1986) and the diagonally weighted least squares (DWLS) estimator in its robust weighted least squares mean and variance adjusted (WLSMV) variant (DiStefano & Morgan, 2014). Finally, we report information about the standardized factor loadings, factor correlations (φ), and McDonald’s (1999) omega (ω) reliability coefficients. With a paired t-test we compare the response times of correctly identified legal letter clusters with correctly identified illegal letter clusters.

Next, we estimated Pearson correlations to examine associations between constructs. Additionally, we tested whether the relationships of general orthographic knowledge (accuracy) with reading and spelling were significantly different from those between word-specific orthographic knowledge (accuracy) with reading and spelling (hypothesis 1). We also examined if these relationships varied as a function of item type (hypothesis 2). We calculated the difference between the correlation coefficients we sought to compare in 1,000 bootstrapped samples to test both hypotheses. This allowed us to generate an empirical distribution of these differences values and calculate a 95% confidence interval for each.

Finally, we estimated mediation models. In these, we examined whether accuracy in general orthographic knowledge predicts reading fluency and spelling performance and whether this path is mediated by accuracy in word-specific orthographic knowledge (hypothesis 3). In case of higher correlations for pseudohomophones in word-specific orthographic knowledge and illegal pseudowords in general orthographic knowledge, only these items are used in the mediation analyses. We used a bias-corrected bootstrapped confidence interval with 1,000 samples for this analysis.

All analyses were conducted using the R language (version “4.1.2;” R Core Team, 2021). We used the lavaan package (version “0.6.9;” Rosseel, 2012) to estimate the mediation and CFA models, and semTools (version “0.5.5;” Jorgensen et al., 2021) to estimate McDonald’s ω. Supplementary material regarding the statistical analysis can be retrieved from the OSF page (https://osf.io/t9uan/?view_only=669bfa92bff44a62ae6c485ec44f0c3f).

Results

Descriptive analysis

Descriptive statistics (M and SD) for grades 3 and 4 and the total sample are presented in Table 1.

Table 1. Descriptive statistics (M and SD) for grades 3 and 4 and the total sample.

	3rd grade (N = 1,242)	4^th grade (N = 1,394)	Total (N = 2,636)
Age in months	110.6 (4.3)	121.9 (4.5)	116.6 (7.2)
Reading fluency^a	93.4 (16.2)	104.0 (15.2)	99.0 (16.5)
Spelling^b	53.7 (10.2)	52.4 (9.9)	53.0 (10.0)
IQ^c	102.0 (14.0)	101.7 (13.9)	101.8 (14.0)
Phonological awareness^a (max 21)	15.6 (3.9)	16.4 (3.7)	16.0 (3.8)
WOK^a (max 60)	53.3 (5.6)	55.5 (4.2)	54.5 (5.0)
Words (max 30)	27.9 (2.2)	28.5 (1.6)	28.2 (1.9)
Pseudohomophones (max 30)	25.4 (4.1)	27.0 (3.2)	26.3 (3.7)
GOK^a (max 60)	46.4 (7.5)	48.0 (7.1)	47.2 (7.3)
Legal pseudowords (max 30)	23.5 (5.7)	24.4 (5.7)	24.0 (5.7)
Illegal pseudowords (max 30)	22.9 (4.1)	23.6 (3.8)	23.3 (3.9)

Note. ^aRaw score, ^bt-score, ^cIQ-score, WOK = word-specific orthographic knowledge.

GOK = general orthographic knowledge.

Confirmatory factor analysis

First, we tested a CFA model using all the general orthographic knowledge items, distinguishing between legal and illegal pseudowords. This model showed a poor fit to the data, χ2 (1,709, N = 2,636) = 9,249.45, p < .001, CFI = .862, RMSEA = .041, 90% CI [.040, .042], SRMR = .096. Additionally, we found negative factor loadings for all the items that included the uniquely German letter Eszett (“ß”), all belonging to the illegal pseudowords factor. For this reason, we decided to rerun this analysis, excluding such items (n = 6). Here, the p-value of the chi-square statistic was statistically significant, χ2 (1,376, N = 2,636) = 3,635.49, p < .001. However, CFI, RMSEA, and SRMR values showed an acceptable model fit, CFI = .951, RMSEA = .025, 90% CI [.024, .026], SRMR = .067. The significant p-value of the chi-square statistic could suggest the presence of model misspecifications (Ropovik, 2015). However, because the chi-square statistic is highly sensitive to sample size and the rest of the fit indices were within the acceptable range, we opted not to modify the model based on this result (e.g., eliminating additional items or correlating errors). All factor loadings of the model were positive and statistically significant at the .001 level. The correlation between the two factors was medium-sized (φ = .56). The McDonald’s ω coefficient for illegal pseudowords was .77 and .89 for legal pseudowords. Finally, based on the results of this last CFA model, we decided to exclude all the items that included an Eszett (“ß”) from the general orthographic knowledge task in the subsequent analyses (i.e., correlations and mediation models).

Then, a two-dimensional CFA was also estimated with the word-specific orthographic knowledge items, establishing a distinction between words and pseudohomophones. Once again, the chi-square’s p-value was statistically significant, χ2 (1709, N = 2636) = 2265.589, p < .001. As in the previous case, this is likely due to the high sample size used to estimate the model. Nevertheless, a good fit was obtained for the following indices: CFI = .954, RMSEA = .011, 90% CI [.010, .012], and SRMR = .075. All factor loadings were positive and statistically significant at the .001 level, and the correlation between the two factors was high (φ =.81). The McDonald’s ω reliability coefficient was .82 for pseudohomophones and .59 for words.

Response times of general orthographic knowledge

A paired-samples t-test revealed significant shorter response times for the correctly identified illegal letter clusters (M = 1841.29 ms, SD = 1305.45) than for the correctly identified legal letter clusters (M = 2829.05 ms, SD = 25717.75, t (2635) = 1.97, p = .049, d = .04).

Correlations

Pearson’s correlation coefficients between study variables are shown in Table 2. Word-specific orthographic knowledge correlated moderately or highly with reading (r = .43), and spelling (r = .55). In line with hypothesis 1, general orthographic knowledge showed lower correlations, compared to word-specific orthographic knowledge, with reading (r = .09) and spelling (r = .14). The differences in the correlations between word-specific versus general orthographic knowledge were significant at the .05 level both for reading and spelling (see Table 3).

Table 2. Correlation coefficients between study variables with confidence intervals.

	1	2	3	4	5	6	6a		6b	7	7a
1. Age in months
2. Reading fluency	.26***
	[.22, .29]
3. Spelling	−.09***	.42***
	[−.13, −.06]	[.38, .45]
4. IQ	−.03	.14***	.29***
	[−.07, .00]	[.11, .18]	[.25, .32]
5. Phonological awareness	.05*	.29***	.43***	.24***
	[.01, .08]	[.26, .33]	[.39, .46]	[.21, .28]
6. WOK	.17***	.43***	.55***	.17***	.35***
	[.13, .20]	[.40, .46]	[.52, .57]	[.14, .21]	[.31, .38]
6a. Words	.11***	.29***	.36***	.12***	.25***	.77***
	[.07, .14]	[.26, .33]	[.32, .39]	[.08, .16]	[.21, .28]	[.76, .79]
6b. Pseudohomophones	.17***	.43***	.55***	.17***	.34***	.95***	.52***
	[.13, .21]	[.40, .46]	[.52, .58]	[.13, .21]	[.30, .37]	[.94, .95]	[.50, .55]
7. GOK	.08***	.09***	.14***	.19***	.19***	.17***	.14***		.15***
	[.05, .12]	[.05, .13]	[.10, .17]	[.16, .23]	[.15, .23]	[.13, .20]	[.11, .18]		[.11, .18]
7a. Legal pseudowords	.06**	.06**	.09***	.17***	.16***	.10***	.10***		.08***	.91***
	[.03, .10]	[.03, .10]	[.05, .13]	[.13, .21]	[.12, .20]	[.06, .14]	[.06, .14]		[.04, .12]	[.91, .92]
7b. Illegal pseudowords	.08***	.09***	.16***	.15***	.17***	.21***	.16***		.20***	.74***	.40***
	[.04, .12]	[.05, .13]	[.12, .20]	[.12, .19]	[.13, .21]	[.17, .24]	[.12, .19]		[.16, .23]	[.72, .76]	[.37, .43]

Note. Values in square brackets indicate the 95% confidence interval for each correlation, WOK = word-specific orthographic knowledge.

GOK = general orthographic knowledge.

*p < .05, **p < .01, ***p < .001.

Table 3. Differences between correlation coefficients between GOK and WOK with reading and spelling skills.

Correlation 1	Correlation 2	Difference between correlations
WOK ~ reading, r = .43***	GOK ~ reading, r = .09***	*
WOK (words) ~ reading, r = .29***	WOK (PsH) ~ reading, r = .43***	*
GOK (legal PsW) ~ reading, r = .06**	GOK (illegal PsW) ~ reading, r = .09***
WOK ~ spelling, r = .55***	GOK ~ spelling, r = .14***	*
WOK (words) ~ spelling, r = .36***	WOK (PsH) ~ spelling, r = .55***	*
GOK (legal PsW) ~ spelling, r = .09***	GOK (illegal PsW) ~ spelling, r = .16***	*

Note. WOK = word-specific orthographic knowledge, GOK = general orthographic knowledge, PsH = pseudohomophones, PsW = pseudowords.

*p < .05, **p < .01, ***p < .001.

To test hypothesis 2, we calculated the correlations between word-specific and general orthographic knowledge and the literacy outcomes separately for the two types for items. For the word-specific orthographic knowledge task, the relationship between rates of correctly identified pseudohomophones with reading (r = .43) and spelling (r = .55) skills were significantly stronger than between rates of correctly identified words with reading (r = .29) and spelling (r = .36) skills. For the general orthographic knowledge task, the relationship between response accuracy to illegal pseudowords (r = .16) and spelling is stronger than the relationship between response accuracy to legal pseudowords (r = .09) and spelling. This difference between correlations was not significant when comparing the correlation between reading with response accuracy to illegal pseudohomophones (r = .09) and reading with response accuracy to legal pseudohomophones (r = .06).

To test the robustness of correlations with reading we repeated the analyses with t-values instead of raw-scores. The pattern of results was very similar, with the exception of the correlations between reading fluency and the accuracy to legal pseudowords in the general orthographic knowledge task, which became non-significant.

Given the good fit of a two-dimensional model of both word-specific and general orthographic knowledge, and the (mostly) stronger correlations with literacy skills for the pseudohomophones and illegal items, we included only these items for all analyses below.

Mediation analysis

The effect of illegal general orthographic knowledge on spelling was highly and partially mediated via pseudohomophone word-specific orthographic knowledge (see Figure 2). Each increase in accuracy by one item for illegal general orthographic knowledge was associated with an accuracy increase of a = 0.21 (SE = 0.02) pseudohomophone word-specific orthographic knowledge-items. Adjusting for illegal general orthographic knowledge, each increase in accuracy by one item for pseudohomophone word-specific orthographic knowledge was associated with a b = 1.45 (SE = 0.04) increase in spelling (t-score). Increases in illegal general orthographic knowledge were indirectly related to increases in spelling through increases in pseudohomophone word-specific orthographic knowledge. Specifically, for every a = 0.21 unit increase in the association between illegal general orthographic knowledge and pseudohomophone word-specific orthographic knowledge, there was an ab = 0.31 (SE = 0.03) increase in spelling units. Notably, the bias-corrected bootstrapped confidence interval with 1,000 samples was above zero, 95% CI [0.24, 0.37]. Last, illegal general orthographic knowledge was also associated with spelling independently of its association with pseudohomophone word-specific orthographic knowledge, c′ = 0.16 (SE = 0.05).

Figure 2. Mediation effect of pseudohomophones (illegal word-specific orthographic knowledge) in the relationship between illegal pseudowords (illegal general orthographic knowledge) and reading or spelling.

The effect of illegal general orthographic knowledge on reading was highly and fully mediated via pseudohomophone word-specific orthographic knowledge (see Figure 2). Every 1-unit increase in illegal general orthographic knowledge was associated with an increase of a = 0.21 (SE = 0.02) pseudohomophone word-specific orthographic knowledge units. Adjusting for illegal general orthographic knowledge, every 1-unit increase in pseudohomophone word-specific orthographic knowledge was associated with a b = 1.90 (SE = 0.08) increase in reading units. Increases in illegal general orthographic knowledge were indirectly associated with increases in reading through increases in pseudohomophone word-specific orthographic knowledge. Specifically, for every a = 0.21 unit increase in the association between illegal general orthographic knowledge and pseudohomophone word-specific orthographic knowledge, there was an ab = 0.40. (SE = 0.04) increase in reading units. The bias-corrected bootstrapped confidence interval with 1,000 samples was above zero, 95% CI [0.32, 0.48]. Last, illegal general orthographic knowledge was not associated with reading performance independently of its association with pseudohomophone word-specific orthographic knowledge, c′ = 0.03 (SE = 0.09).

The results from the mediation analysis provide support for hypothesis 3, showing that the relationship between general orthographic knowledge (i.e., illegal pseudowords) and reading and spelling skills importantly depends on word-specific orthographic knowledge (i.e., pseudohomophones). Interestingly, however, word-specific orthographic knowledge partially mediated the effect of general orthographic knowledge on spelling and fully mediated the effect of general orthographic knowledge on reading.

To test the robustness of the reading prediction, we repeated the mediation analysis with t-values instead of raw-scores. The pattern of results was similar.

Discussion

The present study investigated the relationship between orthographic processing at the lexical and sublexical level and reading and spelling skills in 3rd and 4th grade children. First, we examined the dimensionality of two tasks measuring orthographic knowledge, one at the lexical or word-specific level (word-specific orthographic knowledge) and the other at the sublexical or general level (general orthographic knowledge). The confirmatory factor analyses suggest that both tasks have a two-dimensional structure, consisting of legal and illegal items. For the general orthographic knowledge task, we compared the by-participant average response times for correct trials with “yes”-responses (legal pseudowords) versus “no”-responses (illegal pseudowords). Contrary to our expectations, we found shorter reaction times for the latter. This suggests that illegal letter clusters were identified based on knowledge about specific graphotactic constraints and not via an unsuccessful search process of the letter clusters stored in the mental lexicon.

We then proposed that the relationships between general orthographic knowledge with reading and spelling would be lower than between word-specific orthographic knowledge and both literacy skills (hypothesis 1) and that the relationship would differ as a function of item type, with stronger relationships of literacy (reading and spelling) with accuracy on pseudohomophones (word-specific orthographic knowledge) and illegal pseudowords (general orthographic knowledge) (hypothesis 2). The differences in the correlation coefficients obtained are in line with both hypotheses, with the exception of the relationship between item type and reading for the general orthographic knowledge task.

Finally, we proposed that the relationship between general orthographic knowledge and literacy (reading and spelling skills) is mediated by word-specific orthographic knowledge (hypothesis 3). For these mediation analyses, we use the pseudohomophone word-specific orthographic knowledge and illegal general orthographic knowledge items. Word-specific orthographic knowledge significantly mediated the relationship between general orthographic knowledge and reading and spelling. However, the mediation was total for the case of reading and partial for the case of spelling (see Figure 2).

Previous studies have found stronger correlations with literacy for “no”-responses in the orthographic knowledge tasks (Bergmann & Wimmer, 2008; Rothe et al., 2015). For word-specific orthographic knowledge, this means the ability to accurately identify pseudohomophones as misspelled words, and for general orthographic knowledge, to identify pseudowords with illegal clusters as being not-word-like. Here, we replicated this effect, and followed up by performing a CFA to test if “yes” and “no”-responses indeed reflect different constructs. For both word-specific orthographic knowledge and general orthographic knowledge, we found a good fit of the two-factor model, mapping onto the “yes” versus “no”-responses. Furthermore, accuracy for “no”-responses was more strongly correlated with literacy (see Table 2). This suggests that, indeed, accuracy to the different types of items may reflect different cognitive processes or types of knowledge. From this empirical finding, we can draw a clear methodological implication: future studies using orthographic decision tasks may increase the effect size by analyzing only one type of item: pseudohomophones for the word-specific orthographic knowledge task, and illegal pseudowords for the general orthographic knowledge task. Further work is required to fully identify the theoretical implications of this finding.

For general orthographic knowledge, it is noteworthy that the correlation between accuracy for legal and illegal pseudowords was small (r =.1; compared to r = .5 for the correlation between the accuracy for words versus pseudohomophones in the word-specific orthographic knowledge task). This further suggests a theoretical distinction between “yes” and “no”-responses. In the word-specific orthographic knowledge task, both types of items (correctly spelled words and pseudohomophones) require word-specific knowledge. However, different types of knowledge may be tapped by legal versus illegal letter strings. We initially hypothesized that the decision to reject a pseudoword as illegal may be based on a search process, where the letter clusters are compared against clusters stored in a mental lexicon as a result of the exposure to words (Ehri, 2015; Taft, 1993). For an item with an illegal cluster, the cognitive system would compare the illegal cluster against all clusters represented in the mental lexicon and make a “no”-decision when no match is found. For a word with legal clusters only, the search would only continue until a match has been found, which should take less time than going through an exhaustive list of mental representations of legal letter clusters (or waiting until a threshold time for reaching the decision has elapsed). This hypothesis was not confirmed, as reaction times for illegal words were shorter, not longer, than reaction times for legal words. As one possibility, we propose as a preliminary working model that participants store a set of rules about possible constraints. Making decisions about the legality of an item may then still reflect a search process, but rather than matching the visual input against a list of permissible letter clusters, the cognitive system matches the input against a set of rules. In this case, a correct “no”-response for an item with an illegal cluster would reflect a match between a constraint and its features; for an item with only legal clusters, the “yes”-response would be made once no match is found in the exhaustive list of rules, or when a certain amount of time has elapsed, and no match has been found. Here, we make the assumption that knowledge about graphotactic constraints is stored as a set of rules. This working hypothesis is contentious, as it relates to a decade-long debate about rules versus statistics (e.g., Marcus et al., 1995), but it is not the main issue at stake in the current article. Thus, while the data suggests a distinction in the way that items with “yes” versus “no”-responses are processed, future research is required to identify the mechanisms.

We further replicated previous findings that word-specific orthographic knowledge has a stronger relationship with literacy than general orthographic knowledge (Georgiou et al., 2008). By analyzing the psychometric properties of the word-specific and general orthographic knowledge tasks, we found that these two tasks have similar reliabilities: for our analyses, we divided the task into two factors and performed the analyses for the factors measuring accuracy on pseudohomophones and illegal pseudowords, respectively. For these two factors, the reliability was comparable (ω ≈ .8). Thus, we have excluded one potential explanation for this finding: that the general orthographic knowledge task is more prone to noise and therefore shows lower correlations with literacy tasks (Hedge et al., 2018). The explanation for this finding is therefore more likely theoretical than methodological. We propose that the relationship between word-specific orthographic knowledge and literacy is more direct than the relationship between general orthographic knowledge and literacy. Specifically, lexical orthographic representations are directly required both for instant lexical access, enabling fluent reading, and for producing a correct spelling. The link between general orthographic knowledge and literacy is, however, less clear, with different potential causal pathways. From this perspective, the lower correlation between general orthographic knowledge and literacy than word-specific orthographic knowledge and literacy is in line with the Noisy Chain Hypothesis (Schmalz et al., 2021): when there is a causal chain between a cognitive skill and literacy, each additional mediator dilutes the correlation between this skill and literacy, because at each step, other (measured and unmeasured) predictors are likely to influence the mediator.

We found that word-specific orthographic knowledge acted as a mediator for the relationship between general orthographic knowledge and literacy. For the mediation model predicting reading fluency, the mediation was full, as general orthographic knowledge no longer significantly predicted reading fluency after word-specific orthographic knowledge was taken into account. There are several theories that would explain how general orthographic knowledge may affect word-specific orthographic knowledge. In the self-teaching hypothesis (Share, 1999), decoding ability is the strongest predictor of orthographic learning, which is the learning process that leads to good word-specific orthographic knowledge. General orthographic knowledge may act as an additional factor affecting orthographic learning, or it may facilitate decoding ability. Supporting the notion that general orthographic knowledge may independently facilitate orthographic learning, Henbest and Apel (2018) found that graphotactic regularities affect orthographic learning in a fast mapping task, a task where the participants (beginning readers aged 5–6 years old) did not have the time to decode the pseudoword that the participants had to learn. Alternatively or additionally, general orthographic knowledge may facilitate decoding by allowing children to identify common letter clusters and learning that these specific clusters relate to certain sounds (e.g., the cluster -ight to/ɑɪt/, as in “light” or “sight”).

In contrast to the contentious role of graphotactic probabilities in adults (Schmalz & Mulatti, 2017; but see also Chetail, 2015), the results of the present study suggest that general orthographic knowledge helps children to build word-specific orthographic knowledge by facilitating the buildup of orthographic lexical representations. The full mediation in the reading model suggests that general orthographic knowledge may not facilitate reading fluency in any other way. Reading dysfluency might stem from a deficit in efficiently accessing mental representations. For spelling, the quality of the lexical representation might be more important than the cognitive system’s efficiency in accessing it (Banfi et al., 2021; Mehlhase et al., 2019). This opens the question how general orthographic knowledge additionally facilitates orthographic learning (i.e., the buildup of high-quality representations for word-specific orthographic knowledge).

In the mediation analyses, we assumed that a causal relationship would go from general to word-specific orthographic knowledge and to literacy skills. Theoretically, it is possible that there is, additionally or alternatively, a causal relationship between word-specific and general orthographic knowledge, where superior word-specific orthographic knowledge leads to higher general orthographic knowledge. To support the causal link general orthographic knowledge ➔ word-specific orthographic knowledge, previous studies suggest that some degree of general orthographic knowledge develops before the onset of literacy instruction, when we would not expect word-specific orthographic knowledge to be present above chance level (Treiman et al., 2022; Zhang & Treiman, 2021). Furthermore, general orthographic knowledge relies on shallower processing compared to word-specific orthographic knowledge, which is associated with both semantic and phonological activation in addition to a visuo-orthographic analysis. At the same time, it is also clear that general orthographic knowledge cannot develop without exposure to print. As such, it cannot be excluded that some degree of word-specific orthographic knowledge is needed for general orthographic knowledge.

Finally, it is worth considering to what extent our findings are transferable to English, given that most research findings on reading are based on English-speaking participants (Blasi et al., 2022; Share, 2008, 2021). Compared to English, German is a shallow orthography for the reading direction (e.g., Frith et al., 1998; Moll & Landerl, 2009; Schmalz et al., 2015). For the current study, this has affected the choice of our literacy variables: Reading accuracy reaches ceiling soon after the onset of reading instruction; as such, reading fluency measures are generally used with German-speaking children (e.g., Landerl & Wimmer, 2008; Seymour et al., 2003). With a focus on reading fluency as opposed to accuracy, the speed of access or processing plays an important role in addition to the buildup of high-quality orthographic representations (Mehlhase et al., 2019). Thus, if the study was replicated in English using a reading accuracy measure instead of fluency, we might find a higher overlap in the patterns of results between reading and spelling.

In a rare longitudinal study on word-specific and general orthographic knowledge, Deacon et al. (2012) examined the relationship of word-specific and general orthographic knowledge with reading in English. They found that reading predicting orthographic knowledge but not the other way around. However, Deacon et al. measured accuracy (which quickly reaches ceiling in German-speaking primary school children) and emphasized that it is important to examine this direction of the relationship for reading fluency as well. The results of the present study support our assumption that general orthographic knowledge facilitates the formation of word-specific orthographic knowledge, which in turn facilitates reading fluency and spelling. In a next step, longitudinal studies are needed to investigate the causal chain underlying the relationship between the different types of reading (accuracy and fluency) and orthographic knowledge (general and word-specific).

German and English are similar in terms of syllabic complexity (Seymour et al., 2003). Thus, we would not predict that general orthographic knowledge plays a different role across these two orthographies. This may be different for orthographies with different syllabic structures: For example, the simpler syllabic structure of Italian poses stricter restrictions on permissible letter patterns, while a wider range of consonant strings is permitted in Slavic languages such as Czech (Caravolas & Bruck, 1993). At this stage, however, there are too few studies to speculate how such cross-linguistic differences would affect graphotactic sensitivity, its relationship to orthographic learning, or literacy.

Conclusion

Here, we examined the role of word-specific orthographic knowledge, general orthographic knowledge, and their relationship to reading and spelling. First, we found that “yes” and “no”-responses on the orthographic decision task measuring word-specific and general orthographic knowledge reflect different constructs, with accuracy for “no”-responses (pseudohomophones and pseudowords with illegal letter clusters, respectively) showing higher correlations with literacy skills. Second, we found that word-specific orthographic knowledge is a stronger predictor of literacy than general orthographic knowledge and interpret this in line with the Noisy Chain Hypothesis: the link between word-specific orthographic knowledge and literacy is more direct, while the link between general orthographic knowledge and literacy is likely to involve intervening mediators. Third, we identified word-specific orthographic knowledge itself as a mediator of the relationship between general orthographic knowledge and literacy. Thus, general orthographic knowledge is likely to affect literacy skills by helping participants build up word-specific orthographic knowledge. In addition, general orthographic knowledge is likely to directly affect spelling by constraining the number of plausible spellings.

Acknowledgments

We thank the children and their families for participating in the current study and Meister Cody GmbH for their role in developing the online assessment. In addition, we thank Marcus Hasselhorn, Linda Visser, Katharina Grunwald, Janosch Linkersdoerfer (DIPF); Julia Kalmar, Ruth Görgen and Stefan Haberstroh (LMU) for their support in data collection and data preparation.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Supplementary material

Supplemental data for this article can be accessed online at https://doi.org/10.1080/10888438.2024.2352402.

Additional information

Funding

This study was funded by a grant from the German Federal Ministry of Education and Research (BMBF) under the registration’s numbers 01GJ1601A & 01GJ1601B.