The Effect of Facilitative Versus Inhibitory Word Training Corpora on Word Reading Accuracy Growth in Children With Dyslexia

Abstract

We modeled word reading growth in typically developing (n = 118) and children with dyslexia (n = 20), Grades 2–5, across multiple exposures to 30 words. We explored the facilitative versus inhibitory effects of exposures to differential mixes of words that support high- versus low-frequency vowel pronunciations. One training corpus contained a ratio of 80%–20% high- to low-frequency pronunciations (e.g., for ea; 80% ea pronounced as /i/ as in bead and 20% ea pronounced /ε/ as in dead), whereas the other consisted of a ratio of 20%–80%. We also modeled accuracy at the final exposure for a subset of 12 shared words across conditions using item-level crossed-random effects models with reading skill (i.e., typically developing vs. dyslexic), condition, word frequency, and vowel pronunciation (i.e., high- vs. low-frequency vowel pronunciation) as predictors in the model. We were particularly interested in the interaction between condition and vowel pronunciation across reading groups. Results suggest typically developing children were influenced by the interaction between condition and vowel pronunciation, suggesting both facilitation and inhibition, whereas children with dyslexia were influenced by condition and vowel pronunciation without an interaction. Results are interpreted within the overfitting model of dyslexia.

Keywords

word reading accuracy dyslexia word training

Writing systems, in general, provide readers with maximal phonological and semantic information using minimal orthographic units for processing (Frost, 2012; Seidenberg, 2011), allowing the orthography to concisely represent the language’s phonological space and the way it represents meaning. Thus, what constitutes an efficient orthography in one language may be quite different than what constitutes an efficient system in another. Consider, for example, the different manners in which two alphabetic writing systems capture the phonological properties of the words they represent. In Spanish, there is a nearly one-to-one mapping between letters and phonemes, whereas in English, phonemes can be represented by either a single letter (e.g., p in pan) or letter cluster (e.g., ph in graph), and many graphemes, particularly vowels, can be pronounced in more than one way (cf. pint vs. hint, bead vs. head). This ambiguity in the English orthographic–phonological mapping system (referred to as quasi-regularity) poses significant challenges to beginning readers of English. For instance, Seymour et al. (2003) reported that children who were acquiring reading in orthographically consistent languages (Greek, Finnish, German, Italian, Spanish) were close to ceiling in familiar high-frequency word reading by the middle of first grade. In contrast, English-speaking children performed extremely poorly (34% correct). Danish (71% correct), Portuguese (73% correct), and French (79% correct) children showed somewhat reduced levels of reading accuracy, which is in line with the reduced consistency of these languages.

In English, much of the ambiguity associated with the pronunciation of vowels can be resolved by considering the context in which they occur (see Venezky, 1999). For instance, ea is pronounced as /i/ in beat, /ɛ/ in head, and /eɪ/ in steak. Because /i/ is the most frequent of these pronunciations, a decoding system that operates on each grapheme independently would misread head and steak. However, if the following consonant (i.e., the coda) is taken into account, then /i/ is the most frequent pronunciation in –eat, but /ɛ/ is the most frequent pronunciation in –ead. In a corpus analysis, Kessler and Treiman (2001) found that the consistency of vowel pronunciations increases significantly when the syllable coda is considered. Thus, a decoding process based on multigrapheme units could successfully decode both beat and head. (Note that steak would still be misread. In some cases, the only context that reliably indicates the correct pronunciation is the whole word.)

There is substantial evidence that both children and adult readers make use of knowledge of regularities involving units larger than individual graphemes and phonemes. For example, Treiman et al. (2003) and Treiman et al. (2006) observed that how readers pronounce a nonword containing an ambiguous vowel (e.g., ea) depends on the context in which it occurs. Thus, whereas cheam is almost always read as rhyming with beam, chead is sometimes read as rhyming with bead and sometimes as rhyming with head, suggesting that the decoding process is sensitive to the context in which a grapheme appears. Treiman et al.’s studies demonstrate that this sensitivity to grapheme context develops early (i.e., first grade), continues through elementary school (i.e., fifth grade), and is most pronounced in adults. Overall, the results of Treiman et al. (2006) suggest that children become sensitive to the statistical regularities representing context-dependent orthographic to phonological relationships that exist in the English orthography.

Steacy et al. (2019) recently explored how various factors predict vowel pronunciation of nonwords with and without context-dependent vowel pronunciations in developing readers. Specifically, they extended the Treiman et al. (2006) study by considering a more varied set of nonwords, partitioning item variance across items and participants, and including a diverse set of child predictors (e.g., phonemic awareness, rapid automatized naming, set for variability, visual statistical learning, vocabulary, and reading skill) and a nonword predictor (a continuous measure of the support for the alternative pronunciation of each item based on a type ratio between words containing the rime with the conditional vowel pronunciation to the total occurrences of the rime). Steacy et al. were particularly interested in trying to understand why some developing readers may be more willing to consider context-dependent orthographic to phonological relations when reading nonwords and why certain rime patterns may have a stronger influence on supporting context-dependent vowel pronunciations than others. Results suggested that general reading skill at the child level and rime support at the nonword level uniquely facilitated context-dependent vowel pronunciations in developing readers. Results were interpreted as supporting a developmental model of word reading, in which children are more likely to use an alternative vowel pronunciation in a nonword as they become more proficient readers and as the occurrence of the alternative vowel pronunciation is increasingly supported by the corpus of words. As would be expected in such a model of word reading development, child and corpus attributes work to “tune” variant vowel pronunciations across individual children and words.

This concept of “vowel tuning” suggests that children’s assignment of a pronunciation in words with variant vowels arises from complex interactions between experiential and child-specific factors related to word reading development (see Whitehurst & Lonigan, 1998). The results of Steacy et al. (2019), Treiman et al. (2003), and Treiman et al. (2006) suggest that as a child learning to read English encounters words such as pin and mint with the repeated co-occurrence of the grapheme i and the phoneme /ɪ/, the organization of the reading system becomes attuned to this regularity. Importantly, the quasi-regularity of a writing system such as English creates a challenge for the learning mechanism, and indeed, even in skilled adult readers, seeing the word hint interferes with the subsequent reading of pint, while facilitating the reading of mint (Bowers et al., 2002). Thus, the relative strength of associations between –int and /ɪnt/ (as in mint), –int and /aɪnt/ (as in pint), and –ind and /aɪnd/ (as in find) will affect the assignment of vowel pronunciation (i.e., tuning) in words such as stint and blind in developing readers.

In addition, how a child processes words may have important consequences on the tuning of vowels across words with variant vowel pronunciations. For instance, Ehri (2005, 2014) has argued that skilled readers successfully map all graphemes in a word onto all phonemes in the pronunciation, creating a consolidated representation. Whereas children with dyslexia lack sufficient “graphophonic” knowledge to fully analyze matches between orthographic and phonological units to store complete word-specific representations (see Ehri & Saltmarsh, 1995). Deficits in phonemic awareness skill likely limit the growth of important subword orthographic-to-phonological connections in children with dyslexia (see Ziegler & Goswami, 2005). As a result, dyslexics exhibit a general tendency to process only partial information about words and further to rely on other sources of information that are considerably less efficient, to facilitate word recognition (see Harm & Seidenberg, 2004; Stanovich, 1980, 1984; Stanovich et al., 1984). This is consistent with the view that children with dyslexia may be over reliant on a global processing strategy that affords insufficient attention to individual letters or groupings of letters (Byrne, 1992; Compton et al., 2014; Ehri & Saltmarsh, 1995; Frith, 1985; Siegel et al., 1995) and the corresponding phonological representations. The result is that children with dyslexia tend to add word-specific representations to the lexicon without associated growth in subword orthographic to phonological connections. A lack of sufficient attention to orthographic to phonological connections when reading words may limit important subword feedback of vowel pronunciations in words with variant vowels and limit the natural tuning of vowels in words.

Little is currently known about how children come to assign vowel pronunciations in words with variant vowels and further how this is affected by reading experience. To begin to examine how reading experience and reading skill affect word reading accuracy in words with variant vowel pronunciations, we modeled word reading growth using “localized” reading exposure to a corpus of words with certain vowel pronunciations that should facilitate or inhibit the reading of words with variant vowels. (By localized exposure, we mean short-term experience reading a set of training words that supports a particular pronunciation of a vowel). Typically developing and children with dyslexia were provided with multiple exposures to 30 words, each repeated 4 times per day across 3 consecutive days, resulting in 360 total exposures with 12 exposures per unique word (referred to as the short-term exposure to phonological regularities [STEPR] learning paradigm, described below). We were interested in exploring the facilitative versus inhibitory effects of exposures to differential mixes of words (i.e., training corpora) that support high- versus low-frequency grapheme-phoneme correspondences (GPCs) of the vowel pronunciation in words with variant vowels. One training corpus contained a ratio of 80%–20% high- to low-frequency vowel GPC pronunciations (e.g., for ea; 80% ea pronounced as /i/ as in bead and 20% ea pronounced /ε/ as in dead), whereas the other consisted of a ratio of 20%–80% (e.g., for ea; 20% ea pronounced as /ε/ as in bead; 80% ea pronounced /i/ as in dead).

To accomplish this, we first contrasted growth of reading accuracy of words over the 12 exposures to examine general growth in accuracy across exposures in typically developing and dyslexic children. We then looked exclusively at word accuracy at the final exposure across a subset of words shared across the 80%–20% and 20%–80% conditions. Using item-level analyses, we were interested in how accuracy on the final exposure to a word with a variable vowel differed as a function of child reading skill (i.e., typically developing vs. dyslexic), condition (i.e., corpora supporting high- vs. low-frequency vowel GPC pronunciation), and vowel GPC (i.e., high- vs. low-frequency vowel GPC pronunciation). Specifically, we focused on the interaction between condition and vowel GPC across reading groups. Our hypothesis was that typically developing children would demonstrate greater sensitivity to localized experiences with words by exhibiting an interaction between condition and vowel GPC variables leading to both facilitation and inhibition of word reading (i.e., facilitation of word reading accuracy when condition and vowel GPC matched and inhibition when they did not), whereas dyslexic children would show decreased sensitivity to localized experiences with words due to a lack of subword orthographic to phonological feedback.

Method

Participants

Participants were 138 children in Grades 2–5 from private and public schools in the United States. Demographic data for the sample are presented in Table 1. Sample raw and either scaled scores ( $\underline{X}$ = 10, SD = 3) or standard scores ( $\underline{X}$ = 100; SD = 15) for assessments of phonological awareness, rapid automatized naming, vocabulary, set for variability, and general word reading and decoding skill are provided in Table 2 to describe the sample. Data are provided for the entire sample, children who were considered typically developing, and children with dyslexia. In this study, we oversampled children who were struggling to learn to read words as represented by the depressed age-adjusted scaled and standardized scores representing phonemic awareness, rapid naming, and speeded word reading and decoding. However, untimed word and nonword reading accuracy were in the normal range for the sample. Children were identified as dyslexic in our sample if they scored below the 25th percentile on both the standardized measures of untimed word and nonword reading skill. The typically developing (n = 118) and children with dyslexia (n = 20) differed significantly on measures of phonological awareness (d = 1.04), rapid automatized naming (d = 1.01), word identification (d = 1.23), word attack (d = 1.69), sight word efficiency (d = 1.40), and phonemic decoding efficiency (d = 1.38) favoring the typically developing children. No difference between the two groups on vocabulary performance was detected.

Table 1.

Demographic Statistics.

Variable	Full sample N = 138
Variable	n	%	M	SD
Age (years)			9.37	1.12
Gender
Female	70	50.72
Male	68	49.28
Race
African American	70	50.72
Hispanic	13	8.42
Caucasian	55	39.86

Table 2.

Descriptive Statistics.

Variable	Total sample (N = 138)				TD group (n = 118)				Dyslexic group (n = 20)
Variable	M	SD	Minimum	Maximum	M	SD	Minimum	Maximum	M	SD	Minimum	Maximum
Phonological awareness
Raw score	21.68	6.40	7	33	22.60	6.17	22.60	6.17	16.64	5.32	7	27
Scaled score	8.26	2.80	3	15	8.74	2.69	4	15	5.59	1.68	3	9
Rapid automatized naming
Raw score	20.71	5.62	12	48	20.40	5.57	12	48	22.36	5.72	14	36
Scaled score	8.82	2.28	3	14	9.09	2.16	4	14	7.32	2.38	3	12
Set for variability
	25.73	10.83	8	52	26.86	10.97	8	52	19.10	7.09	9	31
Word identification
Raw score	45.19	8.51	24	67	46.63	7.92	25	67	37.18	7.27	24	48
Scaled score	97.50	14.83	9	129	100.46	13.59	9	129	80.91	9.75	56	90
Word attack
Raw score	17.56	6.29	4	30	18.84	5.80	4	30	10.50	3.8	5	17
Scaled score	98.70	10.31	71	125	101.20	8.93	80	125	84.86	5.1	71	90
Sight word efficiency
Raw score	53.92	14.95	9	82	56.78	13.31	11	82	37.91	13.66	9	62
Standard score	89.82	16.02	55	123	93.20	14.6	56	123	70.95	9.07	55	91
Phonemic decoding efficiency
Raw score	22.28	10.97	0	55	24.16	10.46	0	55	11.77	7.23	0	27
Standard score	86.58	14.57	55	126	89.54	13.28	55	126	70.00	9.57	55	89
WASI vocabulary
Raw score	23.20	5.39	8	36	23.28	5.26	8	36	22.73	6.18	14	36
Scaled score	9.40	2.82	3	17	9.58	2.77	3	17	8.41	2.99	4	16

Note. WASI = Wechsler Abbreviated Scale of Intelligence; TD = typically developing.

Procedure

Test examiners were research assistants who had been trained on tests until procedures were implemented with a minimum of 80% fidelity. All tests were given individually, audio recorded for reliability/fidelity purposes, and scored by the original examiner. Children received small school-related prizes for participating in each testing session. All tests were double-scored and double-entered; discrepancies were resolved by a third examiner. Average fidelity of test administration procedures (based on a random selection of 20% of the taped assessment sessions) exceeded 94% for all tests.

Child Measures

Condition

The STEPR learning paradigm comprised 3 days of exposure in which children were presented with multiple exposures to 30 words, each randomized and repeated 4 times per day, resulting in 360 total exposures with 12 exposures per unique word. Children were exposed to words with the following variant vowel pronunciations: ea (e.g., bead vs. dead), oo (e.g., hoot vs. foot), in (e.g., hint vs. pint), ow (e.g., town vs. grown), o (e.g., cove vs. love), and u (e.g., rut vs. put). The training corpus of words contained different mixtures of orthographic to phonological vowel regularities (representing the high- and low-frequency vowel GPCs) with corrective feedback. One training corpus contained a ratio of 80%–20% high- to low-frequency pronunciations (e.g., for ea; 80% ea pronounced as /i/ as in bead and 20% ea pronounced /ε/ as in dead), whereas the other consisted of a ratio of 20%–80% (e.g., for ea; 20% ea pronounced as /ε/ as in bead; 80% ea pronounced /i/ as in dead). An example of how STEPR training words vary by condition is as follows for the oo grapheme: A child in the strong 80%–20% condition of oo as /u/ would see the words hoot, boot, shoot, and spook (representing /u/) and foot (representing /ʊ/), whereas in the strong 80%–20% condition of oo as /ʊ/ the child would see foot, book, look, and shook (representing /ʊ/) and spook (representing /u/) repeated 4 times each day over the 3 days (12 exposures/word). Children’s response (i.e., correct or incorrect) was recorded for each exposure to a word yielding 12 responses per word. The list of words used in each condition is provided in the appendix.

Untimed word identification

Word identification was measured with the Word Identification subtest from the Woodcock Reading Mastery Tests–Revised/Normative Update (Woodcock, 1998). For this task, children were asked to read words aloud one at a time. The test was not timed, but children were encouraged to move to the next item after a 5-s silence. Correct pronunciations were counted as correct, and the total score was the sum of correct items. Basal and ceiling rules were applied. The examiner’s manual reports split half reliability exceeding .85 for the grade range of our sample (Woodcock, 1998).

Timed word identification

The Sight Word Reading Efficiency subtest (SWE) of the Test of Word Reading Efficiency�2nd Edition (TOWRE-2; Torgeson et al., 2012) is a norm-referenced measure of sight word reading accuracy and fluency. The children were presented with a list of 104 words ordered in ascending difficulty and asked to read aloud as many as possible in 45 s. The authors report an alternate forms reliability of .91.

Untimed nonword reading skill

Nonword decoding was measured using the Woodcock Word Attack subtest from the Woodcock Reading Mastery Tests–Revised/Normative Update (Woodcock, 1998). For this task, children were asked to read isolated pseudowords aloud (e.g., ree, ip, and weaf) one at a time. The test was not timed, but children were encouraged to move to the next item after a 5-s silence. Correct pronunciations were counted as correct, and the total score was the sum of correct items. Basal and ceiling rules were applied. This measure was administered at the end of Grades 1–4. The examiner’s manual reports split half reliability exceeding .85 for the grade range of our sample (Woodcock, 1998).

Timed nonword reading skill

The Phonemic Decoding Efficiency subtest (PDE) of the TOWRE-2 is a norm-referenced measure of decoding accuracy and fluency. The children were presented with a list of 63 nonsense words ordered in ascending difficulty and asked to read (decode) aloud as many as possible in 45 s.

Phonemic awareness

The phonemic awareness task was the Elision task from the Comprehensive Test of Phonological Processing (CTOPP; Wagner et al., 2013). Children were asked to delete phonological units from words. The authors report test–retest reliability of .93 (Wagner et al., 2013).

Rapid automatized naming

To test for rapid automatized naming, we used the letter naming task from the CTOPP (Wagner et al., 2013). For this task, children were asked to name a series of letters as fast as they could without making mistakes. The total score was the number of seconds children took to name all of the letters. Test–retest reliability is .72 for children of ages 8–17 years according to the test manual (Wagner et al., 2013).

Vocabulary

The vocabulary subtest from the Wechsler Abbreviated Scale of Intelligence (WASI-II; Wechsler, 2011) was used to measure expressive vocabulary. The test requires children to identify pictures and define words. Interrater reliability for elementary age children ranges from .92 to .94 (McCrimmon & Smith, 2013).

Set for variability

Based on the work of Tunmer and Chapman (1998, 2012), SfV was assessed by participants’ ability to determine the correct pronunciation from spoken words that were “mispronounced” based on common decoding rules, as they might be if they were regularized or partially decoded (e.g., /brikfəst/ for /brɛkfəst/). Kearns et al. (2016) report a coefficient alpha of .82 for ages 7–11 years.

Reader group

Children were coded as dyslexic (1) or typically developing (0) based on the criteria provided above (see “Participants” section).

Word Measures

Frequency

The metric used for word frequency was the standard frequency index (SFI) from the Educator’s Word Frequency Guide (Zeno et al., 1995). SFI represents a logarithmic transformation of the frequency of word type per million tokens within a corpus of more than 60,000 samples of texts from various sources. These sources range from textbooks to popular literature. The words used in this study had a mean SFI of 54.45 (SD = 10.28; range: 24.2–71.0).

Vowel GPC

Words were coded for whether they contained the high- or low-frequency vowel GPC pronunciation (high-frequency GPC = 1).

Analyses

Growth models

To contrast the growth of word reading accuracy in the words over the 12 exposures in typically developing and dyslexic children, we conducted latent growth modeling using Mplus software (Muthén & Muthén, 2017). Because words varied across the two conditions (see the appendix), accuracy across the words was aggregated to the level of the group (i.e., typical and dyslexic) across each set of exposures. All models were centered at the final exposure. A linear growth model provided inadequate fit of the growth data. To address this lack of fit, we included a quadratic term, which resulted in a better fitting model. We then used reader group to predict growth parameters of the model. A detailed description of the model building process is outlined below.

Crossed-random effects models

To examine individual differences in word reading accuracy at the final exposure, we employed item-level crossed-random effects models to explore child- and word-level predictors for words that overlapped across the two conditions (N_words = 12). These cross-classification multilevel models were used to predict children’s reading of words coded as a dichotomous response (correct or incorrect) using child-level (condition, reader group, phonological awareness, rapid automatized naming, set for variability, vocabulary) and word-level (vowel GPC and word frequency) predictors. We conducted these analyses using Laplace approximation available through the lmer function (Bates & Machler, 2009) from the lme4 library in R (R Development Team, 2012). Random intercepts were included for child and word. Fixed effects were included for all child and word predictors and interactions. We estimated the variability explained by calculating the reduction in child and word variance from the base model using the formula $(r_{010 (B a s e m o d e l)} - r_{010 (M o d e l n)}) / r_{010 (B a s e m o d e l)}$ , where n represents the model to which the base model was compared (Bryk & Raudenbush, 1992). Due to the small sample size, we used an adjusted p value of .10 as a threshold of statistical significance.

Results

Growth Models of Experimental Words

Unconditional model

The data used for modeling growth were aggregated up to the word level for typically developing and dyslexic children separately by calculating the proportion of the sample who read each word correctly at each exposure time point. Thus, the base data for these analyses are the proportion of the sample reading the words correctly at each time point rather than random effects for students and words. Data consisted of accuracy scores for each of 48 unique words presented across two groups (typically developing vs. dyslexic). Therefore, we modeled the growth for individual words rather than individual children, and in this case, time was nested within word. The 12 observed points were coded -11, -10, -9, -8, -7, -6, -5, -4, -3, -2, -1, 0 from the first observation. In this coding scheme, the intercept term represents expected performance at the final data point (exposure 12). We first fit a linear model to the data with only slope and intercept terms included in the model. This approach resulted in a model with unreasonable fit, χ²(73) = 286.25, CFI = .91, TLI = .91, RMSEA = .16 (95% CI = [.14, .18]). We then included a quadratic term in the model, χ²(69) = 199.33, CFI = .94, TLI = .94, RMSEA = .13 (95% CI = [.11, .15]), that resulted in a significantly better fit, Δχ²(4) = 86.92, p < .001; however, improvements were made to the quadratic model through the addition of three error covariances (i.e., times 3 and 4; times 5 and 9; times 5 and 10), χ²(66) = 145.14, CFI = .97, TLI = .97, RMSEA = .10 (95% CI = [.08, .12]). These error covariances were also added to the original linear model that produced reasonable fit, χ²(70) = 218.19, CFI = .93, TLI = .94, RMSEA = .13 (95% CI = [.11, .15]); however, the revised quadratic model fit significantly better than the revised linear model, Δχ²(4) = 73.05, p < .001. Estimated means and variances for this quadratic model are presented in Table 3 and indicated that the estimated mean at the last time point was 85.30. Although the linear portion of the trajectory was not significant (−0.29, p = .173), the quadratic effect was (−0.11, p < .011), and the negative direction of the estimate suggested that students’ word reading growth decelerated over the measurement occasions. The variances implied that students significantly differed from the sample mean estimates of intercept and slopes. The growth parameters were highly correlated (correlation between intercept and linear slope is .38; correlation between quadratic and intercept is .59; correlation between quadratic and linear slope is .81).

Table 3.

Fixed and Random Effect Estimates and Variance Explained for Unconditional and Conditional Word Growth Models Across Exposures.

Unconditional model
Fixed effect	Coefficient	SE	p
Intercept	85.302	1.319	.000
Slope	−0.294	0.216	.173
Quadratic	−0.111	0.018	.000
Random effect	Variance component	SE	p
Intercept	200.027	26.907	.000
Slope	3.027	0.810	.000
Quadratic	0.015	0.006	.007
Conditional model (dyslexic)
Fixed effect	Coefficient	SE	p
Intercept	−18.913	1.989	.000
Slope	−.672	0.419	.109
Quadratic	−.054	0.035	.125
Random effect	Variance component	SE	p
Intercept	109.90	15.53	.000
Slope	2.76	0.785	.000
Quadratic	0.013	0.005	.015

To understand the relation between dyslexic status and intercept, slope, and curvature, we included dyslexic as a predictor in the model (i.e., conditional model). We found that being classified as dyslexic resulted in significantly lower estimated word reading scores by 18.91 points (p < .001). That is, across words, the estimated percentage of children in the dyslexia group who read the words correctly at final intercept was approximately 19% lower than in the typically developing group. There was not a statistically significant difference between typically developing and dyslexic groups in the linear (p = .109) or quadratic (p = .125) estimates of growth; however, due to the sample size of dyslexic students being quite small, we are statistically underpowered to reflect small or moderately sized differences. Model implied growth curves for the typically developing and dyslexic groups are provided in Figure 1.

Figure 1.

Growth trajectories for typically developing and children with dyslexia across 12 exposures.

Item-Level Analyses

To explore the effects of child-level (i.e., reader group, condition, phonological awareness, rapid letter naming, set for variability, and vocabulary) and word-level (i.e., vowel GPC and word frequency) characteristics on the probability of reading words correctly at exposure 12 (i.e., final exposure), we conducted a series of item-level analyses. These analyses were based on the 12 words that were overlapping between the two conditions (i.e., bead, dead, spook, foot, hint, pint, wow, grown, stone, love, rut, put). Correlations among child predictors are provided in Table 4. All child and word predictors were grand mean centered before being entered into the cross-classified models. The intraclass correlations for child and word were .35 and .13, respectively, with the remaining portion of the variance due to the logit scale. A series of models were fit to the data: (a) an unconditional model with only a random intercept for child and word, (b) a model with only reading group as a predictor, (c) a model with only condition as a predictor, (d) a main effects model with all child and word covariates, and (e) an interaction model. Results of these models are provided in Table 5. Models 2 and 3 indicated that there was no main effect for condition, but there was a main effect for reader group. Typically developing children had a probability of .97 of reading the words correctly at exposure 12, whereas children with dyslexia had a probability of .84 of reading the words correctly. In the main effects model, significant predictors included reader group (i.e., typically developing vs. dyslexic; γ = −1.169, z = 3,252), set for variability (γ = 0.069, z = 3.378), and word frequency (γ = 0.076, z = 2.950). The intercept for this model indicates that a typically developing child at the mean on all child measures reading a word of mean frequency would have a probability of .96 of reading the word correctly. This model further indicates that typically developing outperformed dyslexic children, children with higher set for variability outperformed those with lower, and high-frequency words were read correctly more often than low-frequency words. More specifically, with all other child and word measures at the mean, children 1 SD above the mean on set for variability had a probability of .98 of reading the word correctly, whereas students 1 SD below the mean had a probability of .91 of reading the words correctly; words 1 SD above the mean frequency had a probability of .98 of being read correctly, whereas words 1 SD below the mean frequency had a probability of .91 of reading the word correctly. The main effects model accounted for 56% of child variance and 52% of word variance.

Table 4.

Zero-Order Child-Level Correlations.

Variables	1	2	3	4
1. Phonemic awareness	—
2. Rapid letter naming	−.15	—
3. WASI vocabulary	.33	−.02	—
4. Set for variability	.66	−.20	.40	—

Note. WASI = Wechsler Abbreviated Scale of Intelligence.

Table 5.

Fixed Effects and Variance Estimates Predicting Logit Word Reading Accuracy at Exposure 12.

Fixed effects parameter	Unconditional model				Condition model				Group model				Main effects model				Interaction model
Fixed effects parameter	Est.	SE	z	p	Est.	SE	z	p	Est.	SE	z	p	Est.	SE	z	p	Est.	SE	z	p
Intercept (γ₀₀₀)	3.314	.344	9.62	<.001	3.183	.381	8.354	<.001	3.522	.350	10.064	<.001	3.178	.409	7.776	<.001	3.728	.483	7.724	<.001
Child covariates
γ₀₀₁ Condition	—	—	—	—	0.261	.346	0.754	.451	—	—	—	—	0.151	.284	0.534	.593	−0.754	.431	−1.751	.080
γ₀₀₂ Reader group	—	—	—	—					−1.832	.392	4.673	<.001	−1.169	.360	−3.252	.001	−1.914	.591	3.236	.001
γ₀₀₃ Phonological awareness	—	—	—	—	—	—	—	—	—	—	—	—	0.021	.031	0.665	.506	0.024	.032	0.753	.452
γ₀₀₄ Rapid letter naming	—	—	—	—	—	—	—	—	—	—	—	—	−0.024	.023	−1.006	.314	−0.024	.024	−1.003	.316
γ₀₀₅ Set for variability	—	—	—	—	—	—	—	—	—	—	—	—	0.069	.020	3.378	.001	0.068	.021	3.335	.001
γ₀₀₆ WASI vocabulary	—	—	—	—	—	—	—	—	—	—	—	—	0.044	.029	1.536	.125	0.044	.029	1.515	.130
Word covariates
γ₀₀₇ Vowel GPC	—	—	—	—	—	—	—	—	—	—	—	—	0.463	.531	0.872	.383	−0.308	.619	−.497	.619
γ₀₀₈ Frequency	—	—	—	—	—	—	—	—	—	—	—	—	0.076	.026	2.950	.003	0.077	.026	2.945	.003
Interactions
γ₀₀₉ Condition × Reader Group	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	1.402	.856	1.637	.102
γ₀₁₀ Reader Group × Vowel GPC	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	0.876	.599	1.463	.144
γ₀₁₁ Condition × Vowel GPC	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	1.420	.486	2.922	.003
γ₀₁₂ Condition × Reader Group × Vowel GPC	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	—	−1.614	.891	1.812	.070

Note. p < .10 for all variables in bold. WASI = Wechsler Abbreviated Scale of Intelligence; GPC = grapheme-phoneme correspondence.

To explore the interaction between child reading skill, condition, and vowel GPC, we added exploratory interaction terms to our item-level analyses. The three-way interaction among reading skill, condition, and dominant vowel GPC (p = .07) was explored via a simple slopes analysis. This interaction is illustrated in Figure 2, which demonstrates the relationship between condition and vowel GPC for children in the typically developing and dyslexic group, respectively. In the typically developing group, the training appears to be facilitative for the high-frequency vowel GPC words. That is, when the training favors the high-frequency vowel pronunciation and the target word contains a high-frequency vowel pronunciation, a higher percentage of children read the words correctly. Conversely, when the training favors the high-frequency vowel pronunciation and there is a mismatch between that training and the target word, a lower percentage of children read the words correctly. This demonstrates some interference from the training favoring the alternate pronunciation. For the low-frequency condition, there was limited effect of match versus mismatch words, suggesting little facilitation or inhibition in this training condition. For the students with dyslexia, the probability of reading a word correctly was higher if the children were in the condition favoring high-frequency GPC pronunciations and for words with high-frequency GPCs. Results in the dyslexic children suggests simple main effects of condition and vowel GPC with little indication of an interaction suggesting either facilitation or inhibition.

Figure 2.

Accuracy at final exposure as a function of condition (favoring high- vs. low-frequency vowel GPC pronunciation) and vowel GPC (high- vs. low-frequency vowel GPC pronunciation) for the typically developing sample and the sample with dyslexia.

Discussion

The quasi-regular nature of the English writing system requires developing readers to form context-dependent connections, which represent the probabilistic co-occurrences and constraints that exist between subword orthographic and phonological forms, often requiring connections at larger orthographic grain size units such as rimes or syllables (Treiman et al., 1995). These context-dependent connections between orthography and phonology seem to form as a result of interactions between experiential and child-specific factors related to word reading development (see Steacy et al., 2019), allowing a grapheme such as ea to be pronounced as /i/ in beat, /ɛ/ in head, /ɛə/ in pear, and /eɪ/ in steak. Little is currently known about how children form these context-dependent connections in words with variant vowel pronunciations as a function of localized exposure (a process we have referred to here as vowel tuning) and whether the process varies across typically developing and children with dyslexia.

Our goal in this study was to begin to explore how reading experiences and reading skill affect word reading accuracy in words with variant vowel pronunciations by examining possible facilitative versus inhibitory effects resulting from differential training corpora that support high- versus low-frequency vowel GPC pronunciations in words. The STEPR paradigm was designed to examine the effects of reader group (i.e., typically developing vs. dyslexic), condition (i.e., corpora supporting high- vs. low-frequency vowel pronunciation), and vowel GPC (i.e., high- vs. low-frequency vowel GPC pronunciation) on word reading accuracy after 12 exposures to words across 3 days. Results of individual growth modeling across words suggest that there was a significant difference in word reading accuracy between the typically developing and dyslexia groups at final intercept. This finding indicates that, overall, a lower percentage of children in the dyslexia group read words correctly at the end of 12 exposures equally distributed across 3 days compared to the typically developing group.

We were particularly interested in the interaction between condition and vowel GPC in the typically developing and children with dyslexia. Our item-level analyses at the final exposure suggest both facilitative and inhibitory effects of training but only in the high-frequency vowel pronunciation condition and only for the typically developing children. Specifically, in the high-frequency vowel pronunciation condition, words with the high-frequency pronunciations were facilitated (e.g., bead), whereas words with the low-frequency pronunciation were inhibited (e.g., dead). Results for this condition suggest that the localized corpus had an effect on how individual words were read even after 12 exposures spread evenly over 3 days. We suspect that the differential condition effect was influenced by experience with the larger English corpus that by definition supports the localized high-frequency pronunciation while inhibiting the localized low-frequency pronunciation. We speculate that the 12 exposures in the low-frequency vowel pronunciation were insufficient to counteract the effects of experience with the larger English corpus. As we consider implications for instruction, we are struck by the power of experience with the general English corpus to favor the higher frequency vowel pronunciation and the potential need for extended exposures to the low-frequency vowel pronunciation examples.

Interestingly, the facilitation and interference effects observed in the typically developing sample were not present in the dyslexic sample. In the children with dyslexia, we did not observe a difference between words that matched the training corpus and words that did not but instead just a main effect favoring high- versus low-frequency vowel pronunciation conditions and high-frequency vowel GPC pronunciation. The finding suggests that children with dyslexia, unlike their typically developing peers, fail to benefit from corpus feedback, supporting the hypothesis that children with dyslexia may be processing only partial information from words by placing insufficient attention on individual letters or groupings of letters (Byrne, 1992; Ehri & Saltmarsh, 1995; Frith, 1985; Siegel et al., 1995) and the corresponding phonological representations. From a grain size perspective (see Ziegler & Goswami, 2005), attention to the courser whole-word phonological grain size prevents the development of finer orthographic to phonological connections that support the development and expansion of important context-dependent subword orthographic to phonological connections (Harm et al., 2003) that are likely needed to drive localized corpus effects that tune vowel pronunciations.

Results are consistent with connectionist models that have explored the effects of moderate to severe phonological impairments on word and nonword reading skills in children with dyslexia (for details, see Harm et al., 2003; Harm & Seidenberg, 1999, 2004). In this work, mild phonological impairment only affects nonword reading, whereas more severe phonological impairment affects word reading as well. In the impaired phonological system, more work has to be done by the hidden units because the phonological system is less able to repair degraded or incomplete phonological representations that result from the decoding process. This increased workload on the hidden units causes them to be more likely to “memorize” word forms and form item-specific representations. Harm and Seidenberg (1999) hypothesize that the effect of requiring the hidden units to perform a more exact computation results in overfitting of the training data, which interferes with generalization. As such, the impaired network fails to find the sublexical structure shared by similar words, representing them more like unanalyzed, individual wholes with less overlapping structure. The impaired model cannot take advantage of the similarity between words when reading the nonword, even though it can correctly pronounce real words. The model provides a computational account of why poor phonological representations limit the sublexical feedback that one would expect from the localized training corpus. Consistent with previous findings (e.g., Ehri & Saltmarsh, 1995), our results suggest that dyslexic children need extended support to develop the sublexical orthographic to phonological connections needed to benefit from sublexical feedback across word exposures.

Limitations and Future Research

There were limitations to the study outlined above, and conclusions should be drawn with these limitations in mind. First, because we aggregated up to the level of the word for the growth models, we had limited power in this study. Future work would benefit from more words in the training corpus. Furthermore, there were large variance estimates in the high-frequency training condition that made it difficult to draw specific conclusions about the effects of the training on certain kinds of words. Because of the limited number of words, one word could greatly affect the variance.

Practical Implications

The results from this study suggest that students with dyslexia have difficulty attending to sublexical features of words and have particular difficulty applying lower frequency GPCs. This behavior is certainly consistent with the view that children with dyslexia may be over reliant on a global processing strategy that results in insufficient attention to individual letters or groupings of letters (Ehri & Saltmarsh, 1995; Frith, 1985; Siegel et al., 1995) and the corresponding phonological representations. For instance, Ehri and Saltmarsh (1995) have hypothesized that children with dyslexia lack sufficient “graphophonic” knowledge to fully analyze matches between orthographic and phonological units to store high-quality representations that include complete sublexical associations between orthography and phonology. Unfortunately, the design of the study does not provide insights into what can be done to promote the use of alternate (i.e., low-frequency GPC) vowel pronunciations in students with dyslexia. For example, it is not clear whether both pronunciations of a GPC should be taught early in instruction or whether focus should first be placed on strengthening the high-frequency GPC pronunciation followed by the associated low-frequency GPC pronunciation. Previous work suggests that students with dyslexia may need to be encouraged to engage with and analyze words at the sublexical level to promote this flexibility (e.g., Ehri & Saltmarsh, 1995). Programs such as PHAST (see Lovett et al., 2000) that explicitly teach this flexibility seem to be particularly effective at developing flexible decoders who have been explicitly taught to try various pronunciations of vowels, a strategy known as “vowel alert,” until an unknown word is correctly decoded (Steacy et al., 2016). More work is needed to understand the optimal instructional paradigms to promote sublexical feedback and flexibility in students with dyslexia.

Footnotes

Appendix

List of Words Used in Each Condition.

High frequency GPC80%/20%		Low frequency GPC20%/80%
/i/	/e/	/i/	/e/
bead	dead	bead	dead
heat			head
seat			sweat
knead			bread
/u/	/ʊ/	/u/	/ʊ/
hoot	foot	spook	foot
boot			book
shoot			look
spook			shook
/ɪ/	/aɪ/	/ɪ/	/aɪ/
hint	pint	hint	find
mint			kind
print			pint
stint			blind
/aʊ/	/oʊ/	/aʊ/	/oʊ/
town	grown	wow	flow
down			grown
wow			snow
brown			mow
/oʊv/	/ʌv/	/oʊv/	/ʌv/
cove	love	stone	done
drove			none
stone			love
wove			shove
/ʌ/	/ʊ/	/ʌ/	/ʊ/
rut	put	rut	put
hut			bull
nut			pull
dull			full

Note. GPC = grapheme-phoneme correspondences.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was supported in part by Grant P20HD091013 from the National Institute of Child Health and Human Development. Statements do not reflect the position or policy of these agencies, and no official endorsement by them should be inferred.

ORCID iD

Yaacov Petscher

References

Bates

Machler

(2009). Package “lme4”: Linear mixed-effects models using S4 classes (Version 0.999375-32). http://cran.r-project.org/web/packages/lme4/lme4.pdf

Bowers

J. S.

Damian

M. F.

Havelka

(2002). Can distributed orthographic knowledge support word-specific long-term priming? Apparently so. Journal of Memory and Language, 46(1), 24–38. https://doi.org/10.1006/jmla.2001.2802

Bryk

A. S.

Raudenbush

S. W.

(1992). Hierarchical linear models for social and behavioural research: Applications and data analysis methods. SAGE.

Byrne

(1992). Studies in the acquisition procedure for reading: Rationale, hypotheses, and data. In Gough

P. B.

Ehri

L. C.

Treiman

(Eds.), Reading acquisition (pp. 145–174). Lawrence Erlbaum. https://login.proxy.lib.fsu.edu/login?url=https://search.proquest.com/docview/618212390?accountid=4840

Compton

D. L.

Miller

A. C.

Elleman

A. M.

Steacy

L. M.

(2014). Have we forsaken reading theory in the name of “quick fix” interventions for children with reading disability? Scientific Studies of Reading, 18, 55–73. https://doi.org/10.1080/10888438.2013.836200

Ehri

L. C.

(2005). Learning to read words: Theory, findings, and issues. Scientific Studies of Reading, 9, 167–188. https://doi.org/10.1207/s1532799xssr0902_4

Ehri

L. C.

(2014). Orthographic mapping in the acquisition of sight word reading, spelling memory, and vocabulary learning. Scientific Studies of Reading, 18(1), 5–21. https://doi.org/10.1080/10888438.2013.819356

Ehri

L. C.

Saltmarsh

(1995). Beginning readers outperform older disabled readers in learning to read words by sight. Reading and Writing, 7, 295–326. https://doi.org/10.1007/BF03162082

Frith

(1985). Beneath the surface of developmental dyslexia. Surface Dyslexia, 32, 301–330.

10.

Frost

(2012). Towards a universal model of reading. Behavioral and Brain Sciences, 35, 263–279. https://doi.org/10.1017/S0140525X11001841

11.

Harm

M. W.

McCandliss

B. D.

Seidenberg

M. S.

(2003). Modeling the successes and failures of interventions for disabled readers. Scientific Studies of Reading, 7, 155–182. https://doi.org/10.1207/S1532799XSSR0702_3

12.

Harm

M. W.

Seidenberg

M. S.

(1999). Phonology, reading acquisition, and dyslexia: Insights from connectionist models. Psychological Review, 106, 491–528. https://doi.org/10.1037/0033-295X.106.3.491

13.

Harm

M. W.

Seidenberg

M. S.

(2004). Computing the meanings of words in reading: Cooperative division of labor between visual and phonological processes. Psychological Review, 111(3), 662–720. https://doi.org/10.1037/0033-295X.111.3.662

14.

Kearns

D. M.

Rogers

H. J.

Koriakin

Al Ghanem

(2016). Semantic and phonological ability to adjust recoding: A unique correlate of word reading skill? Scientific Studies of Reading, 20(6), 455–470.

15.

Kessler

Treiman

(2001). Relationship between sounds and letters in English monosyllables. Journal of Memory and Language, 44(4), 592–617.

16.

Lovett

M. W.

Lacerenza

Borden

(2000). Putting struggling readers on the PHAST track: A program to integrate phonological and strategy-based remedial reading instruction and maximize outcomes. Journal of Learning Disabilities, 33, 458–476. https://doi.org/10.1006/jmla.2000.2745

17.

McCrimmon

A. W.

Smith

A. D.

(2013). Review of Wechsler Abbreviated Scale of Intelligence, second edition (WASI-II). Journal of Psychoeducational Assessment, 31(3), 337–341. https://doi.org/10.1177/0734282912467756

18.

Muthén

L. K.

Muthén

B. O.

(2017). Mplus user’s guide (7th ed.).

19.

R Development Core Team. (2012). R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing. Vienna, Austria. ISBN 3-900051-07-0. http://www.R-project.org/

20.

Seidenberg

M. S.

(2011). Reading in different writing systems: One architecture, multiple solutions. In McCardle

Miller

Lee

Tzeng

(Eds.), Dyslexia across languages: Orthography and the brain-gene-behavior link (pp. 151–174). Paul H. Brookes. https://login.proxy.lib.fsu.edu/login?url=https://search.proquest.com/docview/1314705555?accountid=4840

21.

Seymour

P. H. K.

Aro

Erskine

J. M.

(2003). Foundation literacy acquisition in European orthographies. British Journal of Psychology, 94(Pt. 2), 143–174. https://doi.org/10.1348/000712603321661859

22.

Siegel

L. S.

Geva

(1995). Evidence for superior orthographic skills in dyslexics. Psychological Science, 6(4), 250–254. https://doi.org/10.1111/j.1467-9280.1995.tb00601.x

23.

Stanovich

K. E.

(1980). Toward an interactive-compensatory model of individual differences in the development of reading fluency. Reading Research Quarterly, 16, 32–71.

24.

Stanovich

K. E.

(1984). The interactive-compensatory model of reading: A confluence of developmental, experimental, and educational psychology. Remedial and Special Education, 5(3), 11–19. https://doi.org/10.1177/074193258400500306

25.

Stanovich

K. E.

Cunningham

A. E.

Feeman

D. J.

(1984). Relation between early reading acquisition and word decoding with and without context: A longitudinal study of first-grade children. Journal of Educational Psychology, 76, 668–677. https://doi.org/10.1037/0022-0663.76.4.668

26.

Steacy

L. M.

Compton

D. L.

Petscher

Elliott

J. D.

Smith

Rueckl

Sawi

Frost

Pugh

(2019). Development and prediction of context-dependent vowel pronunciation in elementary readers. Scientific Studies of Reading, 23, 49–63. https://doi.org/10.1080/10888438.2018.1466303

27.

Steacy

L. M.

Elleman

A. M.

Lovett

M. W.

Compton

D. L.

(2016). Exploring differential effects across two decoding treatments on item-level transfer in children with significant word reading difficulties. Scientific Studies of Reading, 20(4), 283–295.

28.

Torgesen

J. K.

Wagner

Rashotte

(2012). Test of word reading efficiency 2. Pro-Ed.

29.

Treiman

Kessler

Bick

(2003). Influence of consonantal context on the pronunciation of vowels: A comparison of human readers and computational models. Cognition, 88(1), 49–78. https://doi.org/10.1016/S0010-0277(03)00003-9

30.

Treiman

Kessler

Zevin

J. D.

Bick

Davis

(2006). Influence of consonantal context on the reading of vowels: Evidence from children. Journal of Experimental Child Psychology, 93(1), 1–24. https://doi.org/10.1016/j.jecp.2005.06.008

31.

Treiman

Mullennix

Bijeljac-Babic

Richmond-Welty

E. D.

(1995). The special role of rimes in the description, use, and acquisition of English orthography. Journal of Experimental Psychology: General, 124(2), 107.

32.

Tunmer

W. E.

Chapman

J. W.

(1998). Language prediction skill, phonological recoding ability and beginning reading. In Hulme

Joshi

R. M.

(Eds.), Reading and spelling: Development anddisorder (pp. 33–67). Erlbaum.

33.

Tunmer

W. E.

Chapman

J. W.

(2012). Does set for variability mediate the influence of vocabulary knowledge on the development of word recognition skills?. Scientific Studies of Reading, 16(2), 122–140.

34.

Venezky

R. L.

(1999). The American way of spelling: The structure and origins of American English Orthography. Guilford Press.

35.

Wagner

R. K.

Torgesen

J. K.

Rashotte

Pearson

N. A.

(2013). Comprehensive test of phonological processing (2nd ed.). Pro-Ed.

36.

Wechsler

(2011). Wechsler Abbreviated Scale of Intelligence–second edition (WASI-II). NCS Pearson.

37.

Whitehurst

G. J.

Lonigan

C. J.

(1998). Child development and emergent literacy. Child Development, 69(3), 848–872. https://doi.org/10.2307/1132208

38.

Woodcock

R. W.

(1998). Woodcock Reading Mastery Tests–Revised/normative update: Examiner’s manual. American Guidance Service.

39.

Zeno

S. M.

Ivens

S. H.

Millard

R. T.

Duvvuri

(1995). The educator’s word frequency guide (CD-ROM). Touchstone Applied Science Associates.

40.

Ziegler

J. C.

Goswami

(2005). Reading acquisition, developmental dyslexia, and skilled reading across languages: A psycholinguistic grain size theory. Psychological Bulletin, 131, 3–29. https://doi.org/10.1037/0033-2909.131.1.3