Inhibitory control predicts growth in irregular word reading: Evidence from a large-scale longitudinal study.

Inhibitory Control in Irregular Word Reading

When the DRC model ( Coltheart et al., 2001 ; Figure 1 ) reads irregular words (e.g., “break”), the decoded output (/bri:k/) and the lexical output (/breɪk/) conflict with each other at certain phoneme slots (e.g., the vowel, /i:/ or /eɪ/). To resolve this conflict, a large value was set for the Phoneme to phoneme inhibition parameter at the phoneme recognition node. This reduces the activation level at the phoneme slots where the decoded and lexical output disagree, and allows more processing cycles to determine the phoneme (e.g., /i:/ or /eɪ/; will settle on the lexical phoneme /eɪ/ due to its stronger activation than the decoded one). Therefore, the DRC model proposes the following cognitive processes might be employed when we read irregular words: inhibiting from reading aloud straight away, allowing time to settle on an output, and finally reading aloud the lexical output.

Analysis of error patterns in human readers revealed that the majority of errors in reading irregular words were the regularized decoded output (e.g., reading “break” as /bri:k/; Treiman et al., 1995 , Experiments 3 and 4, in which readers read known words). This indicates that readers have the predisposition to read aloud the decoded output, even when the lexical output is also available. These data therefore also suggest that to correctly read irregular words, human readers need to inhibit the predisposition to read aloud the decoded output without considering the lexical output. Any effect of inhibitory control on irregular word reading should only be observable once readers have learned GPCs in addition to some irregular words, as only then will the predisposition to read aloud the decoded output be established and the lexical output be available.

Inhibitory Control in the Development of Irregular Word Reading

When the ST-DRC model ( Pritchard et al., 2018 ; Figure 2 ) self-teaches irregular words (e.g., “chef”), the decoded output (/tʃef/, “ch” being pronounced as in “church”) conflicts with the plausible lexical candidate (/ʃef/, “ch” being pronounced as “sh” as in “shake”) at certain phoneme slots (/tʃ/ or /ʃ/). To resolve this conflict, a large value was set for the PhonemePhonlexInhibition parameter at the phoneme recognition node. This parameter reduces activation at the phoneme slots where output disagrees (e.g., /tʃ/ or /ʃ/), and prevents the model from using the decoded /tʃef/ to develop the lexical route (i.e., connecting the /tʃef/ to the orthographic representation CHEF in the orthographic lexicon, which then connects to semantics and letters). Instead, the inhibition parameter allows more processing cycles to recognize the correct phonological lexical output /ʃef/, connect it to the orthographic lexical representation, and develop the full lexical route. Therefore, the ST-DRC model suggests that human readers might need to inhibit themselves from using the decoded output to develop the lexical route, and instead allow time to recognize the correct phonological lexical output and use that to develop the correct lexical route. This is supported by empirical data. Analysis of children’s self-teaching outcomes showed that 55% of irregular words were pronounced using the decoded output ( Murray et al., 2022 ), indicating children’s predisposition to overuse the decoded output when self-teaching irregular words. To correctly self-teach irregular words, children might need to inhibit this predisposition.

The ST-DRC model assumes that, after the phonological lexical output is recognized (e.g., /ʃef/), it can be smoothly connected to the correct orthographic representation (CHEF), which will be stored in the orthographic lexicon. However, empirical studies show that developing the correct orthographic representation and connections is not easy for irregular words (e.g., Wang et al., 2011 , 2012 ). Wang et al. (2011 ; Experiment 2) examined 7- to 9-year-old children’s self-teaching of novel written irregular words. First, children were familiarized with oral vocabulary knowledge (semantics and phonological lexicon) of the words. They were then instructed to use this knowledge to assist reading aloud of these words, either in contextually rich stories (context condition) or lists (no-context condition). Their orthographic representations of these words were tested immediately and after a 10-day delay. Results showed that, across conditions (context, no-context; immediate, delay) and orthographic tests (spelling, orthographic choice, orthographic decision), around 50% of words were self-taught with a regularized orthographic representation (e.g., SHEF). This reflects children’s robust predisposition to overgeneralize GPCs when developing orthographic representations for irregular words. Therefore, this predisposition might need to be inhibited before a stable orthographic representation (e.g., CHEF) can be established in the orthographic lexicon and the lexical route can be correctly developed.

The Current Study

The evidence so far suggests that inhibitory control might be important for irregular word reading and its development. However, this has not been tested. The current study aimed to fill in this gap. We hypothesized that (a) inhibitory control would correlate with concurrent irregular word reading; (b) inhibitory control would predict concurrent irregular word reading after controlling for decoding and vocabulary, but only at later years when the sublexical and lexical routes have developed to a certain extent; and (c) inhibitory control would predict growth in irregular word reading, after controlling for decoding and vocabulary. We also planned to test whether and to what extent decoding and vocabulary predict growth in irregular word reading. To test these hypotheses, a secondary data analysis was conducted on data collected at three time points in the Surrey Communication and Language in Education Study (SCALES), a longitudinal U.K. population-based study that tracked children’s language, academic, and cognitive development ( Norbury, 2022 ).

Decoding

Decoding was measured by accuracy score on the CC2 nonwords subtest, which was administered using the same procedure as for the irregular word subtest. Cronbach’s α was reported to be .94 ( Moore et al., 2012 ).

Vocabulary

Vocabulary was measured by the Receptive One-Word Picture Vocabulary Test (ROWPVT; Martin & Brownell, 2010 ) and the Expressive One-Word Picture Vocabulary Test (EOWPVT; Martin & Brownell, 2011 ). In ROWPVT, participants heard a word and were instructed to select the corresponding picture from four choices. In EOWPVT, they were instructed to name objects, actions, or concepts illustrated in pictures. According to the manual, Cronbach’s α and test–retest reliability coefficients for both measures are above .90.

Inhibitory Control

Inhibitory control was measured by a Go/No-Go task ( Gooch et al., 2016 ). Participants were instructed to press a response key as quickly as possible when the Go stimulus (bug) appeared but to inhibit their response when the No-Go stimulus (ladybird) appeared. Each stimulus was preceded by a fixation cross and a varied lag (300, 600, or 900 ms). Participants completed eight practice trials followed by 80 randomized test trials (20 No-Go and 60 Go trials). Prior to the Go/No-Go test phase, participants completed 33 Go trials to establish the prepotent response, which were not included in the current analysis.

We planned to use six measures from the task to generate a latent variable of inhibitory control. Following Brocki and Bohlin (2004) , these included (a) commission errors (responding to No-Go stimuli); (b) impulsivity errors (key pressing before the onset of stimulus); (c) mean reaction time (RT) in correct Go trials; and (d) omission errors (missing Go targets). Two additional measures were also used since they were proposed to be supported by the inhibitory control brain network ( Bellgrove et al., 2004 ; Garavan et al., 2002 ): (e) the intraindividual variability (IIV) in RT ( Dykiert et al., 2012 ), which was calculated as the ratio of the standard deviation of the individual’s RT to the mean of the individual’s RT in Go trials, and (f) the post error slowing (PES; Dutilh et al., 2012 ), calculated by averaging the RT difference between post error trials and pre error trials.

Using 5,000 random splits, the Spearman–Brown corrected reliability estimates ( Parsons, 2021 ; Pronk et al., 2021 ) in Years 1, 3, and 6, respectively, were .71, .66, .72 for commission errors, .93, .89, .93 for impulsivity errors, .83, .74, .81 for omission errors, .86, .88, .90 for mean RT in correct trials, and .75, .77, .73 for IIV. The measurement of PES was based on consecutive trials with a certain pattern (correct Go trial, incorrect No-Go trial, correct Go trial). Therefore, trials were not split to estimate reliability for PES.

Transparency and Openness

This study’s analysis plan and hypotheses were preregistered ( https://osf.io/vah82 ). All data are available at the UK Data Service and can be accessed at https://doi.org/10.5255/UKDA-SN-8968-1 . Analysis code has been made available on the Open Science Framework ( https://osf.io/qkvyf/ ). Regarding the research materials, CC2 tests are openly available ( Castles et al., 2009 ; https://www.motif.org.au/tests ). ROWPVT and EOWPVT are only available from the publisher ( Martin & Brownell, 2010 , 2011 ). For access to the Go/No-Go task, please contact Gooch et al. (2016) .

Results

Descriptive statistics for observed measures in Years 1, 3, and 6 are shown in Table 1 . Means increased over time for irregular word reading, decoding, and vocabulary measures, and decreased for inhibitory control measures (lower scores indicate better inhibitory control ability; cf., a lack of consensus for PES; Gupta et al., 2009 ; Wiersema et al., 2007 ). Developmental trajectories of irregular word reading are presented in Figure 3 . Trajectories of other measures are available in Figures S1–S9 in the online supplemental materials .

Note . The trajectory in bold is based on means at each time point.

Structural equation modeling was conducted with Lavaan (Version 0.6-9; Rosseel, 2012 ) in R (Version 4.1.1). Robust full information maximum-likelihood estimation was used to account for missing data and deviations from normality (Shapiro–Wilk tests, p s < .01, in all measures across time points). Participants with complete data have better vocabulary knowledge than those with missing data. The two groups do not differ in terms of irregular word reading, decoding, or Go/No-Go measures ( Table S1 in the online supplemental materials ). The SCALES oversampled children with lower language abilities, which might have counteracted missing data from children with poor vocabulary knowledge.

A latent variable of vocabulary was constructed with two indicators, accuracy scores of both ROWPVT and EOWPVT. To construct the latent variable of inhibitory control, the covariance matrix between Go/No-Go measures was explored for each time point ( Table S2 in the online supplemental materials ). Commission error rate, IIV, and impulsivity error rate were intercorrelated across time points with coefficients between .43 and .62, indicating that their shared variance was explained by a common factor. Scatter plots ( Figures S10–S18 in the online supplemental materials ) showed that these intercorrelations were true effects and were not driven by outliers. No other cluster of measures was intercorrelated with effect sizes greater than .30 at any time point. Therefore, the commission error rate, IIV, and impulsivity error rate were used to construct the latent variable of inhibitory control. The standardized factor loadings are reported in Table 2 .

Factorial invariance tests ( Widaman et al., 2010 ) showed that the two latent variables changed over time in structure, as the model fit was significantly better when each parameter (indicator, factor loading, intercept) of the latent variables was freely estimated than when constrained to be equal across time points ( p s < .05; lack of strong factorial invariance). This suggests that neither of the latent variables is strictly comparable across time points. Therefore, relations (correlations and regressions) that involve the latent variables are also not strictly comparable across time points and caution should be taken when interpreting any such comparisons. However, it should be noted that the main aim of the current study was to test whether there are effects of inhibitory control on irregular word reading and its growth at each time point, not to compare effect sizes over time. Therefore, the lack of strong factorial invariance did not affect our focal analysis.