Modeling Complex Word Reading: Examining Influences at the Level of the Word and Child on Mono- and Polymorphemic Word Reading

Abstract

Purpose:

The probability of a child reading a word correctly is influenced by both child skills and properties of the word. The purpose of this study was to investigate child-level skills (set for variability and vocabulary), word-level properties (concreteness), word structure (mono- vs. polymorphemic), and interactions between these properties and word structure within a comprehensive item-level model of complex word reading. This study is unique in that it purposely sampled both mono- and polymorphemic polysyllabic words.

Method:

A sample of African American (n = 69) and Hispanic (n =6) students in grades 2–5 (n =75) read a set of mono- and polymorphemic polysyllabic words (J=54). Item-level responses were modeled using cross-classified generalized random-effects models allowing variance to be partitioned between child and word while controlling for other important child factors and word features.

Results:

Set for variability and the interaction between concreteness and word structure (i.e., mono- vs. polymorphemic) were significant predictors. Higher probabilities of reading poly- over monomorphemic words were identified at lower levels of concreteness with the opposite at higher levels of concreteness.

Conclusions:

Results indicate important predictors at both the child- and word-level and support the importance of morphological structure for reading abstract polysyllabic words.

Decades of research supports the essential role of automatic word recognition in skilled reading (see Ashby & Rayner, 2006). Despite the fact that most words in English have more than one syllable (Baayen, Piepenbrock, & van Rijn, 1993), the vast majority of the literature on word reading development has focused on the study of monosyllabic word recognition with relatively little attention paid to polysyllabic/polymorphemic word recognition skill (Perry, Ziegler, & Zorzi, 2010; Roberts, Christo, & Shefelbine, 2011). Yet, it is unclear whether the behavioral effects reported for monosyllabic words generalize reliably to more “complex” polysyllabic words. We use the term complex here to describe a large class of polysyllabic English words, both mono- and polymorphemic, in which the relationships between orthography, phonology, and morphology are relatively opaque, making these words (e.g., mischievous, palatial, physician) difficult to learn to read for developing readers. Studies suggest that complex words place unique demands on developing readers related to word features such as syllable boundaries (Perry et al., 2010), word stress and vowel reduction (Heggie & Wade-Woolley, 2017; Ševa, Monaghan, & Arciuli, 2009), vowel pronunciation ambiguities (Venezky, 1999), larger and more complex grapheme-phoneme units (Berninger, 1994), and morphological transparency (Kearns et al., 2016) and complexity (Carlisle & Stone, 2005; Nagy et al., 2003; Nagy, Berninger, & Abbott, 2006).

The literature to date indicates that children learn to read unfamiliar words, to a large extent, by employing a phonologically-based recoding mechanism (Share, 1995). Phonological recoding is the process of translating a printed word to speech by employing orthography-to-phonology correspondences of various grain sizes to unfamiliar words (Berninger, 1994; Nation & Castles, 2017; Ziegler & Goswami, 2005). However, learning to read words in English is modulated by child-level factors beyond phonologically-based decoding skill that include, but are not limited to, vocabulary knowledge, morphological knowledge, orthographic knowledge, and print experience (see Cunningham, Perry, & Stanovich, 2001; Goodwin, Gilbert, Cho, & Kearns, 2014; Harm & Seidenberg, 2004; Kearns & Al Ghanem, 2019; Keenan & Betjemann, 2007; Nation & Snowling, 1998; Ouellette & Fraser, 2009; Plaut, McClelland, Seidenberg, & Patterson, 1996; Ricketts, Nation, & Bishop, 2007). These child-level factors interact with various word-level factors to either hasten or impede complex word reading acquisition.

Here we are concerned with child and word predictors of complex word reading that have been largely unexplored in the literature as they relate to mono and polymorphemic word reading. From the standpoint of the child, we focus on individual differences in semantic, phonological, and orthographic skills that should impact the speed at which children develop word reading skills (e.g., Nation, 2017; Perfetti, 2007). Likewise, we were interested in exploring word predictors that capture semantic, phonological, and semantic features of words that have not been previously explored. For example, word concreteness is a word-specific feature that refers to the ease with which a word can elicit a mental image in the reader (Paivio, Yuille, & Madigan, 1968). We posit, along with others (e.g., Balota, Cortese, Sergent-Marshall, Spieler, & Yap, 2004; Laing & Hulme, 1999; Pexman, 2012; Strain, Patterson, & Seidenberg, 1995), that concreteness is an important word feature associated with the semantic representation of the word that may facilitate the reading of the word beyond orthography to phonology regularity/consistency (see Sabsevitz, Medler, Seidenberg, & Binder, 2005; Steacy & Compton, 2019). Thus, we were interested in exploring predictors such as concreteness that may differentially activate some type of perceptual codes that potentially affect the ease with which words are learned by developing readers.

Contemporary models of word reading, in this case the Triangle Model of reading (Seidenberg & McClelland, 1989), provide a useful framework for considering individual differences in word reading development at both the level of the child and word (e.g., Plaut et al., 1996). The Triangle Model consists of associative mappings between the orthographic (O), phonological (P), and semantic (S) layers that operate simultaneously within a distributed network (see Plaut, 2005). The model provides a means to examine developmental mechanisms associated with reading words aloud by providing direct (O→P) and indirect (O→S→P) pathways by which orthography produces output phonology (see Harm & Seidenberg, 1999; 2004; Seidenberg & McClelland, 1989). It is important to note that attributes of both the word (e.g., orthographic, phonological, and semantic) and network (i.e., the child) support or inhibit the mediating role of semantics (i.e., O→S→P pathway) in facilitating word production.

In addition to direct connections between orthographic, phonological, and semantic layers, the model also contains various multidirectional “cleanup” units that receive input from the phonological and semantic units and provide activations back to layers in such a way that degraded patterns of activation are repaired (see Harm & Seidenberg, 1999; 2004). For instance, cleanup units allow the phonological system to learn higher order relationships between phonemic features and also allow the network to complete patterns that are missing features to repair or complete partial or noisy representations between orthography and phonology (Harm & Seidenberg, 1999). Thus, the Triangle Model predicts that semantic properties of the word and properties of the network cleanup units are likely to influence the reading of complex words.

Word Predictors of Complex Word Reading

Morphological structure.

In the case of polymorphemic words, studies suggest that morphemes function as perceptual units that influence word recognition in developing readers (e.g., Carlisle & Stone, 2005; Nagy et al., 2006). Castles, Nation, and Rastle (2018) elucidate the importance of morphological structure in reading morphologically complex words. They suggest that when the underlying regularities that occur between spelling and meaning lead to inconsistencies between spelling and sound (e.g., magician), readers begin to learn that particular groups of letters are associated with particular meanings. This knowledge can help children to understand new words that contain multiple morphemes. Importantly, the authors note that this generalization would be impossible for monomorphemic words. As children develop this morphological knowledge, they begin to perceive larger groups of letters that can support decoding of polymorphemic words. As the authors note, once students acquire knowledge of the morphological regularities between spelling and meaning, orthographic learning does not need to occur one item at a time. Instead, “for those words comprising more than one morpheme, recognizing and getting to the meaning of printed words can be based on analysis of the constituents” (p. 23). The ability to identify and use morphological units in polymorphemic words, however, is moderated by word characteristics such as the transparency of the morpheme (Carlisle & Stone, 2005; Kearns et al., 2016; Schreuder & Baayen, 1997), the consistency of the relationships between the orthographic and phonological units of the morphological units (e.g., peace to pacify; Chateau & Jared, 2003; Jared & Seidenberg, 1990; Yap & Balota, 2009), the frequency of the base morphemes that make up the word (Carlisle & Katz, 2006; Nagy et al., 1989), and morphological family size (Schreuder & Baayen, 1997).

Imageability/concreteness¹.

Results from computational modeling studies have shown that semantic features of a word can significantly influence the reading of low-frequency monosyllabic words with atypical spellings and pronunciations (i.e., angst), indicating the importance of the O→S→P pathway under certain circumstances (e.g., Grainger & Ziegler, 2011; Harm & Seidenberg, 2004; Plaut et al., 1996). Behaviorally, semantic characteristics of words have been explored by examining the effects of imageability, a word-level semantic variable, on word reading performance in skilled adult readers (Strain, Patterson, & Seidenberg, 1995) and children (Laing & Hulme, 1999; Steacy & Compton, 2019). Results from training studies in developing readers indicate that for irregular words, imageability is a significant predictor of difficulty, with high-imageability words being easier for children to learn compared to low-imageability words (see Duff & Hulme, 2012; Steacy & Compton, 2019). Results suggest that word-level semantic features influence the rate of acquisition in developing readers; however, this work has relied almost entirely on monosyllabic words without consideration for the specific demands of polysyllabic words.

Child Predictors of Complex Word Reading

Semantic knowledge.

Because the mappings between orthography and phonology in English are often opaque in complex words, there is growing evidence implicating the role of semantic knowledge (e.g., semantic & morphological) and partial-decoding in the reading of complex words (see Kearns & Al Ghanem, 2019; Nation & Snowling, 1998; Ricketts et al., 2007; Tumner & Chapman, 2012; Wang, et al., 2013). For example, Ricketts et al. (2007) found that item-specific vocabulary knowledge accounted for unique variance in irregular word reading in developing readers. Having item-specific vocabulary knowledge for a word has also been shown to be a significant predictor of orthographic learning within a self-teaching model of reading development (Wang, Nickels, Nation, & Castles, 2013). Keenan and Betjemann (2008) speculated that item-specific semantic activation may help to “fill voids” in phonological-orthographic processing in individuals with insufficient mappings (p. 193).

Set for variability/lexical flexibility.

There is growing evidence both computationally and behaviorally that lexical flexibility at the network/child-level may be an important skill when learning to read alphabetic orthographies (see Harm & Seidenberg, 2004; Venezky, 1999). Computational modeling studies have shown that phonological cleanup units are important for enabling the network to generalize quasi-regular relationships between orthography and phonology relationships (see Harm & Seidenberg, 1999; 2004). In addition, modeling studies have shown that disruptions to the phonological cleanup units result in degraded network performance mimicking that of individuals with moderate to severe phonological dyslexia; that is, severely impaired nonword reading and moderately impaired exception word reading (see Harm & Seidenberg, 2004; Steacy et al., 2021). Behaviorally, the ease by which a child can disambiguate the mismatch between the decoded form of a word and its actual pronunciation (i.e., phonological cleanup) has been referred to as a child’s set for variability (SfV). The rather awkward term “set for variability” is rooted historically in writings by Gibson (1965) and later Venezky (1999) who advocated for the use of phonics instruction for developing readers in English with the important caveat that in order for children to be successful with this approach they would need a “set for variability” in English. As Venezky put it, “if what is first produced does not sound like something already known from listening, a child has to change one or more of the sound associations (most probably a vowel) and try again” (p. 232).

The SfV task has been operationalized into an oral language task (see Tunmer & Chapman, 1998) requiring an individual to disambiguate the mismatch between the decoded form of an irregular word and its actual pronunciation². Behaviorally, the SfV task can be conceived of as drawing on similar lexical mechanisms associated with the phonological cleanup units in the Triangle Model. In the task, children are asked to identify the correct pronunciation of spoken English words that were “mispronounced” based on regular decoding rules, as they might be if they were regular words or partially decoded (e.g., /brikfəst/ for /brɛkfəst/). Elbro et al. (2012) suggest that this skill serves as a bridge between decoding and pronunciations and may be an important second step in the decoding process (see also Venezky, 1999). Here we conceptualize the SfV task as tapping a child’s lexical flexibility to be able to go from the decoded form of a word to the actual pronunciation of the word. Correlational studies indicate that this ability to go from a decoded form of a word to a correct pronunciation predicts individual differences in reading irregular words (Tunmer & Chapman, 2012), regular words (Elbro, et al., 2012), and nonwords (Steacy et al., 2019a). SfV has also been found to be a stronger predictor of word reading than phonological awareness in students in grades 2–5 (e.g., Steacy et al., 2019b). However, these studies have relied primarily on examining the relationship between the SfV task and monosyllabic and monomorphemic word reading, again without consideration for the specific demands of complex polysyllabic words.

Present Study

In the present study we explored a range of child- and word-level factors that predicted the probability that children, grades 2–5, would correctly read a diverse set of mono- and polymorphemic polysyllabic words. We were particularly interested in the unique role of child (i.e., lexical flexibility as measured by SfV and vocabulary) and word (i.e., concreteness) predictors of complex word reading in the presence of other important child-level (e.g., morphological awareness and phonological awareness) and word-level (frequency, length, morphological complexity, and transparency) predictors. We were also interested in the effects of morphological structure on item difficulty. That is, whether children would exploit morphological structure to help with decoding complex (i.e., mono- vs. polymorphemic) words. This study of polysyllabic word reading is unique in that it includes both mono and polymorphemic polysyllabic words. Diving a bit deeper into what affects polymorphemic word reading, we were also interested in the role that orthographic-phonological transparency of the base morpheme would play in the reading of polymorphemic words and also whether a word-level continuous measure of transparency (i.e., rating of recoding difficulty) contributed significantly to complex word reading. We hypothesized that students’ polysyllabic word reading accuracy would be better if they had stronger PA, vocabulary, morphological, and SfV skills. We also hypothesized that students would read transparent and higher frequency words with more accuracy. Finally, we hypothesized that students would use the morphological structure available in polymorphemic words to read polymorphemic words with more accuracy than monomorphemic polysyllabic words.

We were also interested in exploring interactions between word structure (mono- vs. polymorphemic) and predictors at the level of the word (concreteness) and the child (set for variability and vocabulary). We chose to focus on these three interactions because we wanted to capture properties of both the word and the child. Concreteness/imageability is often used as a representative semantic property of a word. Vocabulary is used as a measure of depth and breadth of semantic knowledge and set for variability is used as a measure of lexical flexibility and clean-up during word reading. This allowed us to evaluate whether these influences varied between mono- vs. polymorphemic words using multiple properties of the word and the child. We hypothesized that child skills such as vocabulary and set for variability and word properties such as concreteness would impact word reading accuracy more in monomorphemic words that lack morphological structure.

Method

Participants

Participants were 75 children in grades 2–5 from a public school in the Southeastern United States. Prior to the study, ethical approval was obtained from the Florida State University ethics committee, which conforms to the U.S. Federal Policy for the Protection of Human Subjects. Prior to participation, teacher and parent consent was sought and student assent was obtained. The students in this study were primarily African American (N = 69), with a small number of Hispanic students (N = 6). Prior to the study, ethical approval was obtained from the institution’s Internal Review Board (IRB), in compliance with the U.S. Federal Policy for the Protection of Human Subjects. Prior to students’ participation, teacher and parent consent was acquired, as well as student assent. Demographic data for participants are presented in Table 1. There was an intentional oversampling for children who were struggling with learning to decode words, which is reflected in the age-adjusted scaled and standardized scores for measures of phonemic awareness and word reading with normal age-adjusted scaled scores in vocabulary.

Table 2.

Means, standard deviations, and correlations for word features

Variable	M	SD	1	2	3	4	5	6	7
1. Mono	0.44	0.50
2. Poly Trans	0.30	0.46	−.58**
3. Poly NonTrans	0.26	0.44	−.53**	−.38**
4. SPTR	2.51	0.39	−.13	.10	.05
5. Number of letters	8.11	1.80	−.10	.19	−.08	.39**
6. Frequency	46.43	9.47	.20	−.28*	.07	−.40**	−.20
7. Concreteness	3.12	1.22	.65**	−.42**	−.29*	−.10	−.22	.04
8. Complex WR	0.66	0.21	.27*	.03	−.34*	−.59**	−.22	.38**	.33*

Note. M and SD represent mean and standard deviation. The mean of the monomorphemic, polymorphemic transparent and nontransparent can be interpreted as the proportion of the words that are classified as each. Mono = Monomorphemic; Poly = Polymorphemic; Trans = Transparent; SPTR = Spelling to pronunciation transparency rating; Complex WR = Complex word reading. Binary correlations are phi between Monomorphemic, polymorphemic transparent, and polymorphemic nontransparent; and binary-continuous point biserial between those predictors and the continuous variables.

^*p < .05

^**p < .01.

Procedures

All testing sessions were administered by trained research assistants in the fall or spring of grades 2–5. Each research assistant achieved at least 80% procedural fidelity in a mock testing session before testing participants in schools. All testing sessions were audio recorded for scoring and reliability purposes. All tests were double-scored and double-entered by a second research assistant, and any discrepancies were resolved by the project coordinator.

Dependent measure

Complex words.

The dependent measure was an experimenter-created list of polysyllabic words that varied in number of morphemes (mono- vs. polymorphemic), phonological and orthographic shifts between the base morpheme and derived word, and stress patterns. The full list of complex words is provided in the appendix. Mono- and polymorphemic words were matched on length and frequency. In general, monomorphemic words in English tend to be more concrete and the words on our list reflected this fact. Concreteness ratings are on a five-point scale (Min 1, Max 5). A rating of 1 indicates an abstract word while a rating of 5 indicates a concrete word. Data from the English Lexicon Project (Balota et al., 2007) suggests that for the larger corpus of English words, the mean concreteness of monomorphemic polysyllabic words is 3.49 (SD = 1.05) while the mean concreteness for polymorphemic polysyllabic words is 2.89 (SD = .98). Our words showed a slightly larger difference between the two at a mean concreteness of 4.00 (SD = 1.17) for monomorphemic polysyllabic words and a mean concreteness of 2.42 (SD = .71) for polymorphemic polysyllabic words. Note that the data used from the English Lexicon Project for these analyses were: number of morphemes, number of syllables, and concreteness data (Bysbaert et al., 2013).

Child measures

Morphological awareness (MA).

MA was measured using a composite score from the means of performance on three separate morphological measures. The students had all written items in front of them and the assessor read all test items and answer options to the student to ensure MA testing with minimized word reading demands. In the first test of suffix choice (Nagy et al., 2003), the child was asked to choose the derivational form of the word that completed a sentence correctly (25 items). For the second test, the examiner read the child five pseudo-derived words and four sentences using each word. The child was asked to choose the sentence in which it made sense. For the third test, the tester presented the child with 14 items (each with four nonword options containing grammatical information) and asked the child to choose the one that fit the sentence. Past studies using these measures suggest that these tasks are reliable (nonword: α = .73, Lesaux & Kieffer, 2010; combined real word and nonword: α = .77, Ramirez, Chen, Geva, & Kiefer, 2010). The reliability coefficients (ordinal alpha) for our sample for real words, nonwords, and morphological signals were .92, .65, and .75, respectively.

Phonological awareness (PA).

The Elision task from the Comprehensive Test of Phonological Processing (CTOPP-2; Wagner, Torgesen, Rashotte, & Pearson, 2013) was used to measure phonemic awareness. Students were asked to delete individual phonological units from words. The authors report test-retest reliability of >.90 for children in grades 2–5.

Set for variability (SfV).

Based on the work of Tunmer and Chapman (1998, 2012) and Steacy et al. (2019b), SfV was evaluated by participants’ ability to identify the correct pronunciation from spoken English words that were “mispronounced” based on regular decoding rules, as they might be if they were regular words or partially decoded (e.g.,/brikfəst/ for /brɛkfəst/). The ordinal alpha for our sample was .90 for children in grades 2–5.

Sight word reading efficiency (SWE).

The SWE task from the Test of Word Reading Efficiency (Torgesen, Wagner, & Rashotte, 2012) tasked students with reading a list of words in order of difficulty for 45 s. The authors report an alternate form reliability of >.90 for grades 2–5.

Vocabulary.

The expressive vocabulary subtest from the WASI (Wechsler, 2011) required students to identify pictures and define words aloud. Interrater reliability for this task ranges from .92-.94 for children in grades 2–5 (McCrimmon & Smith, 2013).

Word measures

Concreteness.

Concreteness was coded using ratings from Brysbaert, Warriner, and Kuperman (2014) for 40,000 generally known English words. People were asked to rate the concreteness of words on a scale of 1 (abstract) to 5 (concrete). Some examples from our sample include: heavenly (1.5), origin (2.03), congratulate (3.04), organist (4), and elephant (5).

Frequency.

We used the standard frequency index (SFI) from the Educator’s Word Frequency Guide (Zeno, Ivens, Millard, & Duvvuri, 1995). SFI represents a logarithmic transformation of the frequency of word type per million tokens within a corpus of over 60,000 samples of texts.

Number of morphemes.

Each was coded for number of morphemes using the English Lexicon Project (Balota et al., 2007). The number of morphemes reported for target words ranges from 1 to 4 morphemes. The number of morphemes was used to create a binary contrast between mono and polymorphemic words.

Spelling-to-pronunciation transparency rating.

To address how easy it was to arrive at each word’s correct pronunciation by applying typical decoding rules, undergraduate student participants (N = 241) rated this difficulty on a 6-point scale (1= very easy, 6= very difficult). This measure has previously been used by Steacy et al. (2017) with expert raters. Participants were given the following prompt: “... pretend that the letter string is unfamiliar to you and apply a letter-sound reading strategy to the letter string, then rate the ease of matching your ‘sounded out’ form using phonics rules of the letter string to the actual word pronunciation,” along with an example of easy vs. difficult words for matching the decoded form to the correct pronunciation (as determined previously by expert raters). Inter-rater reliability for this sample of words was .98.

Word length.

Word length was calculated as the number of letters in each word.

Morphological transparency.

To determine the morphological transparency of the polymorphemic words, we examined whether there was a shift in pronunciation or spelling when a suffix was added to create a derived word (Carlisle, 2000). Words were coded as containing a phonological shift (e.g., confess-/kən ˈfɛs/ to confession-/kən ˈfɛ ʃən/), orthographic shift (e.g., intense to intensity), both (e.g., nature to natural), or no shifts (i.e., transparent words; e.g., classic to classical). Inter-rater reliability was 100%.

Analytic Approach

Contrast coding.

Words were classified into one of three groups: monomorphemic, polymorphemic transparent, and polymorphemic nontransparent. Contrast coding was used to investigate whether there was a difference in performance on monomorphemic and polymorphemic words (transparent and nontransparent combined) and whether there was a difference between polymorphemic transparent words compared to polymorphemic nontransparent words. Words categorized as transparent were those that did not contain any orthographic, phonological, or stress shifts between root word and a derivational form, while nontransparent words were those that did present a shift of any kind. For the first contrast, monomorphemic words received the code of −2/3 while both sets of polymorphemic words (transparent and nontransparent) received the code of 1/3. The second contrast was based on morphological transparency of the root word (see above). For this contrast, monomorphemic words received the code of 0, polymorphemic transparent words received the code of 1/2, and polymorphemic nontransparent words received the code of −1/2. These contrast codes are orthogonal meaning that each can be interpreted as a test of mean differences between groups.

Cross-classified generalized random-effects models.

Item responses were modeled using cross-classified generalized random-effects (CCGRE) models (see Gilbert, Compton, & Kearns, 2011 for overview). These models allow for the estimation of variability in item responses between students as well as between words. In these models, children are crossed with items and both children and items were allowed to be random factors. Since all of the responses were dichotomous (correct/incorrect), a binomial distribution with a logit link function was used to predict the probability of a correct item response based on the set of predictors. All child and word predictors were raw scores that were grand mean centered before being entered into the cross-classified models. This centering allows for easier interpretation of the intercept and the coefficients in the model. The intercept represents the logit for a student at the mean on all child predictors reading a word at the mean of all word predictors (i.e., the average child reading the average word).

Results

Descriptive statistics and correlations for the word level predictors are presented in Table 2 and those of child level predictors are presented in Table 3. Note that in Table 2, the variables for monomorphemic, polymorphemic transparent, and polymorphemic nontransparent are binary with a 1 representing words in that category and 0 for words that do not fit in that category. We report the phi coefficient between these binary variables as well as point biserial correlations between these predictors and the continuous word features. The complex WR variable represents the proportion of the sample who read each word correctly aggregated up to the word.

Table 3.

Means, standard deviations, and correlations for child level predictors

Variable	M	SD	1	2	3	4	5
1. MA	9.04	2.72
2. PA	22.63	6.31	.53**
3. SWE	59.00	11.46	.57**	.32**
4. Vocabulary	22.37	4.98	.59**	.21	.46**
5. SfV	26.97	9.20	.63**	.54**	.51**	.47**
6. Complex WR	0.66	0.24	.69**	.64**	.72**	.51**	.73**

Note. M and SD represent mean and standard deviation. MA = Morphological Awareness; PA = Phonological Awareness; SWE = Sight Word Efficiency; SfV = Set for Variability; Complex WR = Complex word reading

^*p < .05

^**p < .01

Table 4.

Results from the unconditional model, main effects model, and interaction model without controlling for word reading.

	Unconditional Model			Main Effects Model			Interaction Model
Parameter	Logit	z-value	p-value	Logit	z-value	p-value	Logit	z-value	p-value
Intercept	1.29	3.70	<0.001**	1.27	5.99	<.001***	0.94	3.62	<0.001**
Word-Level Predictors
C1: Poly vs Mono words				0.04	0.09	.931	−0.12	−0.28	0.780
C2: Trans vs NonTrans words				1.53	3.25	.001**	1.46	3.23	0.001**
SPTR				−2.20	−4.26	<.001***	−1.92	−3.73	<0.001***
Number of letters				0.01	0.04	.968	−0.02	−0.23	0.822
Frequency				0.07	3.35	<.001***	0.07	3.66	0.003**
Concreteness				0.55	2.90	.004**	0.27	1.18	0.238
Person-Level Predicts
Grade				0.54	2.97	.003**	0.54	2.97	0.003**
Morphological Awareness				0.09	1.22	.224	0.09	1.22	0.223
Phonological Awareness				0.11	4.12	<.001***	0.11	3.88	<0.001***
Vocabulary				0.04	1.16	.246	0.04	1.16	0.247
Set for Variability				0.07	3.91	<.001***	0.07	3.91	<0.001***
Interactions
C1 × Concreteness							−0.79	−2.00	0.045*
C1 × Set for Variability							0.001	0.01	0.978
C1 × Vocabulary							−0.001	−.01	0.989
Random Effects	Variance			Variance	% var explained		Variance	% var explained
Person	3.768			0.957	74.60		.952	74.73
Word	3.719			1.399	62.38		1.281	65.56

Note. C1 = Contrast 1; C2 = Contrast 2; Poly = Polymorphemic; Mono = Monomorphemic; Trans = Transparent; SPTR = Spelling to pronunciation transparency rating.

^*p < .05

^**p < .01

^***p < .001.

Item responses for 54 words from 75 children were modeled to examine the contributions of person (i.e., morphological awareness, phonological awareness, sight word efficiency, vocabulary, and set for variability) and word features (i.e., monomorphemic vs polymorphemic, polymorphemic transparent vs nontransparent, number of letters, frequency, spelling-to-pronunciation transparency ratings, and concreteness³) to variability in item responses. To investigate which predictors were uniquely significant after controlling for the other predictors in the model, a model with all the predictors entered simultaneously was fit. Furthermore, given that monomorphemic polysyllabic words tend to be more concrete than polymorphemic polysyllabic words (based on all polysyllabic words in the English Lexicon Project), we included an interaction between contrast 1 and concreteness to see whether the impact of concreteness differed when reading polymorphemic compared to monomorphemic words.

First, an unconditional model (Table 4) was fit to estimate the variance attributable to children, the variance attributable to words, and a grand mean (intercept) that yields the probability of getting an item correct (in logits). Results revealed both variability due to children (SD = 1.94) and variability due to words (SD = 1.93), indicating enough variability to attempt to predict this variability with our predictors. The intercept of the unconditional model indicates that an average child reading an average word had a probability of .78 of reading the words correctly.

Table 5.

Results from the unconditional model, main effects model, and interaction model controlling for word reading.

	Unconditional Model			Main Effects Model			Interaction Model
Parameter	Logit	z-value	p-value	Logit	z-value	p-value	Logit	z-value	p-value
Intercept	1.29	3.70	<0.001**	1.26	6.20	<.001**	0.94	3.68	<0.001**
Word-Level Predictors
C1: Poly vs Mono words				0.04	0.46	.933	−0.12	−0.28	0.781
C2: Trans vs NonTrans words				1.53	3.25	.001**	1.47	3.23	0.001**
SPTR				−2.20	−4.26	<.001**	−1.92	−3.73	<0.001**
Number of letters				0.01	0.04	.971	−0.02	−0.23	0.819
Frequency				0.07	3.36	<.001**	0.07	3.67	<0.001**
Concreteness				0.55	2.90	.003*	0.27	1.18	0.237
Person-Level Predicts
Grade				0.22	1.29	.198	0.22	1.29	0.198
Morphological Awareness				0.05	0.74	.458	0.05	.747	0.455
Phonological Awareness				0.10	4.13	<.001**	0.10	4.13	<0.001**
Sight Word Efficiency				0.06	4.52	<.001**	0.06	4.52	<0.001**
Vocabulary				0.02	0.74	.460	0.02	.74	0.247
Set for Variability				0.06	3.56	<.001**	0.06	3.57	<0.001**
Interactions
C1 × Concreteness							−0.79	−2.00	0.046*
C1 × Set for Variability							0.001	0.06	0.951
C1 × Vocabulary							0.001	.01	0.989
Random Effects	Variance			Variance	% var explained		Variance	% var explained
Person	3.768			0.700	81.42		0.697	81.50
Word	3.719			1.406	62.19		1.288	65.38

Note. C1 = Contrast 1; C2 = Contrast 2; Poly = Polymorphemic; Mono = Monomorphemic; Trans = Transparent; SPTR = Spelling to pronunciation transparency rating.

^*p < .05

^**p < .001.

Next, we ran two sets of CCGRE models in which predictors were entered simultaneously with results reported in Tables 4 and 5. The models in Table 4 do not control for overall reading skill (SWE), while the models in Table 5 include SWE as a control variable. We chose to include both models with and without SWE to demonstrate that the results do not differ when controlling for prior reading skill. We consider the model controlling for SWE to be a strong test of additional predictors above and beyond word reading skill. The main effects model indicates that significant unique contributions were observed for grade, morphological transparency (polymorphemic transparent vs nontransparent), spelling-to-pronunciation transparency rating, frequency, concreteness, phonological awareness, and set for variability. When word reading was added to the model, it was significant while grade was not. The significance of other predictors was not impacted by including SWE in the models, suggesting these predictors to be important for complex word reading even after accounting for sight word efficiency skills. We use the model controlling for prior word reading for all interpretations of main effects moving forward. The main effects observed for morphological transparency indicated that the average polymorphemic polysyllabic word had a probability of .78 of being read correctly while the average monomorphemic polysyllabic word had a probability of .77 of being read correctly. The main effect for the spelling to pronunciation transparency rating indicated that a word 1SD above the mean on the rating (higher difficulty) had a probability of .60 of being read correctly while a word 1SD below the mean on the rating had a probability of .89 of being read correctly. The frequency main effect indicates that a word 1SD above the mean had a probability of .87 of being read correctly while a word 1SD below the mean had a probability of .65 of being read correctly. The concreteness main effect indicates that a word 1SD above the mean had a probability of .87 of being read correctly while a word 1SD below the mean had a probability of .64 of being read correctly. The main effect for phonological awareness indicates that a child 1SD above the mean on PA had .87 probability of reading a word correctly compared to .65 for a child 1SD below the mean. The main effect for SWE indicates that a child 1SD above the mean on SWE had .88 probability of reading a word correctly compared to .64 for a child 1SD below the mean. Finally, the main effect for SfV indicates that a student 1SD above the mean on SfV had a probability of .86 of reading a word correctly while a student 1SD below the mean had a probability of .67 of reading a word correctly. There were no significant main effects for polymorphemic vs monomorphemic words, number of letters, morphological awareness, or vocabulary.

When interactions between the contrast code for morphological structure (C1) and the lexical properties were added to the model, only the interaction between C1 and concreteness was significant such that more concrete words had a higher probability of a correct response for only monomorphemic words, whereas polymorphemic words showed no difference in difficulty based on the level of concreteness (see Figure 1). This interaction indicates that the main effect for concreteness in the Main Effects model is entirely driven by the monomorphemic words.

Figure 1:

Interaction between contrast 1 and concreteness rating in the prediction of word reading. The colors along the x-axis denote the observed values of concreteness for polymorphemic (dark blue) and monomorphemic words (light blue). Concreteness on the x-axis is grand mean centered (0 corresponds to the mean concreteness rating of 3.12, −1 corresponds to a concreteness rating of 2.12, 1 corresponds to a concreteness rating of 4.12, and 2 corresponding to a concreteness rating of 5.12. The maximum concreteness rating was 5. Note that polymorphemic words have a smaller range of observed values than monomorphemic words. The shaded areas represent the 95% confidence interval surrounding each line with the darker blue shaded area representing the confidence interval for polymorphemic and light blue representing that of monomorphemic words. Significant differences are represented by areas in which the line for one is outside of the shaded area for the other.

Discussion

English, a quasi-regular orthography, places unique demands on developing readers (Seymour, Aro, & Erskine, 2003), requiring advanced decoding skills that are sensitive to multiple grain sizes (Berninger, 1994; Zeigler & Goswami, 2006) and other orthographic and linguistic constraints such as syllabic (e.g., Perry et al., 2010) and morphological (Carlisle & Stone, 2005) structure. As children move through elementary school, complex words become more frequent, take on greater prominence in terms of content-specific information, and increase in decoding demands. There is clear evidence that important differences exist between monosyllabic and polysyllabic words requiring developing readers to bring to bear more advanced knowledge of orthography, phonology, and morphology to the problem of recognizing complex polysyllabic words (e.g., Bijeljac-Babic, Millogo, Farioli, & Grainger, 2004; Carlisle & Kearns, 2017; De Luca, Barca, Burani, & Zoccolotti, 2008; Kearns & Al Ghanem, 2019; Kearns et al., 2016; Yap, & Balota, 2009). However, less is known about the specific demands that complex polysyllabic words place on developing readers. The purpose of this study was to investigate child-level lexical flexibility (i.e., set for variability), word-level concreteness, and interactions between lexical properties and word structure within a comprehensive item-level model of complex word reading. The present study extends the current literature on word reading development by examining word- and child-features that predict complex word reading of polysyllabic words that are either monomorphemic or polymorphemic.

Results from the complete model indicate that child-level predictors representing phonemic awareness, word reading skill, and SfV and word-level predictors associated with the contrast between transparent and nontransparent polymorphemic words, the continuous rating of relative transparency, and word frequency were significant item-level predictors of complex word reading. The model accounted for considerable child- and word-level variance (80.9% and 65.3%, respectively) and results are generally in line with previous findings suggesting the importance of child-level phonological skills (Share, 1995) and word-level features that relate to frequency and transparency between orthography and phonology (Seidenberg, 1992; Yap & Balota, 2009). One rather surprising finding was that while morphological awareness skill was highly correlated with complex word reading it was not a significant predictor in the simultaneous prediction model. However, given the significant correlations among morphological awareness skill, phonemic awareness skill, sight word reading efficiency, and SfV it is likely that the four measures were competing for the same common variance in complex word reading, thus limiting the unique contribution of morphological awareness as a predictor. It is also possible that the lack of main effect for MA was due to task demands in our study (i.e., children had print in front of them but heard someone read the items). The same is likely true for child-level vocabulary skill which was also not a significant predictor of complex word reading skill in the simultaneous prediction model even though it was significantly correlated with complex word reading.

Results confirm that SfV is a strong and unique predictor of complex word reading accuracy, even in the presence of other important linguistic predictors. This adds to prior research suggesting that SfV skill may tap the distributed connections in the reading system between orthography, phonology, and semantics in monosyllabic and monomorphemic words (Edwards et al., in press; Elbro et. al., 2012; Steacy, et al., 2019b; Venezky, 1999) and now complex polysyllabic words. To help better understand the unique relationship between SfV and word reading we draw parallels between network components of the Triangle Model and behavioral skills of the child. We equate the skills required for a child to complete the SfV task as being the behavioral equivalent of the computational processes related to the phonological cleanup units in the Triangle Model. The cleanup units allow the network to repair or complete partial or noisy representations just as children are often faced with noisy or ambiguous phonological signals related to word decoding. For instance, in English, children who faithfully employ basic decoding skills will often generate phonological outputs that fail to match exactly with the phonological representation of the word stored in the lexicon and thus must clean-up the output phonological representation (e.g., decoding the word island results in /ɪz lənd/ and must be disambiguated/restored to /ˈaɪ lənd/).

Steacy et al. (2019b) have previously demonstrated that the ability to complete the set for variability mispronunciation task is both an important metalinguistic skill related to students’ ability to correctly read monosyllabic words and a significant item-specific skill for item-based word reading. This behavioral result is quite similar to network cleanup units facilitating the phonological system’s ability to learn higher order relationships between phonemic features (equivalent to metalinguistic skill) and also allows the model to complete patterns that are missing features to repair or complete partial or noisy representations (equivalent to an item-specific skill). We hypothesize here that this clean-up process in developing readers becomes even more important as words expand to include multiple syllables/morphemes where stress patterns and vowel reduction resulting in schwa further complicate the relationship between orthography and phonology. The lack of a significant interaction between word structure (i.e., mono- vs polymorphemic words) and SfV tends to suggest a general influence irrespective of morphological structure and is not surprising given that both sets of words are complex from an orthography to phonology standpoint. This convergence between child performance on the SfV task and the operation of network cleanup units in computational models opens new opportunities for examining individual differences in children’s development of word reading skills using computational modeling techniques and could presumably be used to examine the effects of SfV training on network performance.

At the level of the word, results indicate no main effect of word concreteness on complex word reading in the interaction model. However, a significant crossover interaction between concreteness and word structure was identified with higher probabilities of reading mono- over polymorphemic words at higher levels of concreteness and the opposite at lower levels of concreteness. We put forth the following hypothesis to help explain the interaction between word structure and concreteness. Clearly in the polysyllabic monomorphemic words used in this study there was a strong relationship between the probability of reading a word and the relative concreteness. Similarly, there is evidence that developing readers require fewer exposures to learn more concrete monosyllabic words compared to more abstract words (see Laing & Hulme, 1999; Steacy & Compton, 2019). Our results support this relationship in polysyllabic words but only for monomorphemic words. We speculate that relative to polymorphemic words, monomorphemic words may lack clear orthographic structure that allows straightforward decoding and therefore the ability to form an image of the word may be a driving feature that facilitates decoding. Whereas in polymorphemic words, the orthographic structure anchored by morphemes may support decoding of these complex words. Clearly the lack of a relationship between polymorphemic word reading and concreteness ratings supports our hypotheses. This is consistent with the literature suggesting that morphemes function as perceptual units that influence word recognition in developing readers (e.g., Carlisle & Stone, 2005; Nagy et al., 2006). Further supporting this hypothesis is our result that polymorphemic words with a transparent base morpheme (e.g., organist) were easier to read compared to words with a nontransparent base (e.g., natural).

Limitations

Our enthusiasm for the results is tempered in part by the small sample size of children, the fact that the students in the sample may not be representative of a broader population of students, and the fact that there was a relatively small set of words representing the category of complex words. We also recognize that the words were not sampled to have a parallel set of monosyllabic words and we did not account for other approaches to capturing orthographic structure of words beyond mono- vs. polymorphemic. We support future work that will further explore how students negotiate the orthographic structure of polysyllabic words. Further work that explores the relationship between morphology and other aspects of the orthographic and phonological structure is also warranted.

Future Directions

Our results suggest important lexical influences at the level of the child supporting the prominence of lexical flexibility and some type of morphological deconstruction on the part of the child to aid in complex word reading. At the level of the word, the importance of concreteness in word learning appears to be substantial but significantly moderated by the structure of the complex word with the effect being predominately in monomorphemic words. While we recognize that this is a correlational study, our results do point to the potential benefits of training children to be flexible with output phonology when attempting to read complex words and further that helping children recognize morphological structure during decoding may facilitate decoding accuracy. We support continued studies that examine these effects using causally sensitive designs.

Table 1.

Demographic statistics (N=75)

Variable	n	%	Mean	SD
Age (Years)			8.875	.953
Gender
Female	45	40
Male	30	60
Race
African-American	69	92
Hispanic	6	8

Appendix

Table 1A.

Word List

	Contrast 1	Contrast 2	Concreteness	Frequency
alligator	−0.67	0	4.96	46.4
animal	−0.67	0	4.61	62.3
anticipate	−0.67	0	1.54	45.8
beastly	0.33	0.5	2.26	39.4
capitalize	0.33	0.5	2.29	41.6
categorize	0.33	−0.5	2.60	34.0
caterpillar	−0.67	0	4.87	47.1
classical	0.33	0.5	2.18	48.9
confession	0.33	−0.5	2.40	43.3
confusion	0.33	−0.5	2.24	53.1
congratulate	−0.67	0	3.04	40.6
considerate	0.33	0.5	1.92	42.3
convention	0.33	−0.5	3.28	53.0
cultural	0.33	−0.5	2.10	54.7
disloyalty	0.33	0.5	1.50	37.3
edgy	0.33	−0.5	1.87	32.9
elephant	−0.67	0	5.00	53.3
entirely	0.33	0.5	1.81	56.7
family	−0.67	0	4.23	66.0
finality	0.33	−0.5	1.67	37.6
flowery	0.33	0.5	2.68	39.9
gallery	−0.67	0	4.44	47.1
heavenly	0.33	0.5	1.50	47.2
independence	0.33	0.5	1.87	56.3
intensity	0.33	−0.5	2.14	49.0
macaroni	−0.67	0	4.97	41.4
magician	0.33	−0.5	4.21	46.3
majority	0.33	−0.5	2.54	56.2
masterful	0.33	0.5	1.93	38.7
metal	−0.67	0	4.87	59.3
movement	0.33	0.5	3.63	60.1
mustang	−0.67	0	4.89	41.1
natural	0.33	−0.5	1.85	62.7
odorous	0.33	0.5	3.27	31.4
organist	0.33	0.5	4.00	30.1
origin	−0.67	0	2.03	53.3
paradise	−0.67	0	2.72	46.8
parent	−0.67	0	4.56	55.2
pepperoni	−0.67	0	5.00	22.1
potato	−0.67	0	4.85	51.6
precision	−0.67	0	2.41	47.9
pyramid	−0.67	0	4.96	50.0
raccoon	−0.67	0	4.67	48.2
remember	−0.67	0	2.41	63.5
rosy	0.33	−0.5	3.41	45.3
routine	−0.67	0	2.70	52.8
salamander	−0.67	0	4.70	37.6
secretive	0.33	0.5	2.25	36.3
security	0.33	−0.5	2.82	55.2
showy	0.33	0.5	2.43	38.1
stylish	0.33	−0.5	2.14	41.5
surrender	−0.67	0	2.63	48.1
tarantula	−0.67	0	4.83	36.6
unworkable	0.33	0.5	1.92	33.9