Morphological effects in visual word recognition: Children, adolescents, and adults.

Oral vocabulary

This was measured using the Vocabulary subtest of the WASI-II ( Wechsler, 2013 ) for which participants are asked to verbally define words.

Word reading

This was assessed using the Sight Word Efficiency (SWE) and Phonemic Decoding Efficiency (PDE) subtests of the Test of Word Reading Efficiency–Second Edition (TOWRE-2; Torgesen, Wagner, & Rashotte, 2012 ) for which participants read aloud a list of words (SWE) or nonwords (PDE) as quickly as they can in 45 seconds.

Procedure

The visual lexical-decision task was completed individually in a quiet room in school or at the university. Participants were instructed that they would be shown a series of words on the screen, and to indicate using a key press whether or not each was a real word that they knew, as quickly as possible. Participants were shown 12 practice items followed by the experimental items. Each trial began with a black fixation cross, which appeared in center of the screen for 1000ms, followed by the target, which appeared in lowercase Calibri font in the center of the screen until a response was made. For the practice items only, participants were given feedback on RTs and accuracy. Participants were given a short break after every 20 trials. The E-prime 2.0 program ( Schneider, Eschman, & Zuccolotto, 2012a , 2012b ) was used to present instructions and stimuli, and to record responses.

Results

Table 2 summarizes performance by age group on background measures. Mean scores indicate performance that is close to test norms. Responses (accuracy and RTs) to nonwords in the visual lexical-decision task were analyzed. Inverse transformations were carried out on RTs to correct for distribution skews and transformed data were used throughout the analyses. RTs for incorrect responses were excluded, amounting to 25%, 23%, 15% and 12% of the data for children, younger adolescents, older adolescents and adults respectively. For the analysis, outliers were removed by excluding RTs that exceeded three standard deviations from the mean for that participant. Tables 3 and 4 show mean accuracy and mean RTs respectively for each nonword type by age group.

We used R (Version 3.3.0; R Development Core Team, 2016 ) and the lme4 package (Version 1.1–12; Bates, Maechler, Bolker, & Walker, 2016 ) to perform a generalized linear mixed-effects analysis of the effect of condition (pseudomorphemic vs. control) and age group (children vs. younger adolescents vs. older adolescents vs. adults) on the log odds of accuracy, and a linear mixed-effects analysis of the effect of condition and age group on RTs. For each analysis, condition, age group, and the interaction between condition and age group were entered into the model as fixed effects. 1 We took a design-driven approach to determine the structure of random effects, starting with random intercepts by-participant and by-item, along with by-participant random slopes for the effect of condition and by-item random slopes for the effect of age group. Where a model failed to converge, or inspection of the correlations between intercepts and slopes of random effects indicated that the model was overparameterized, we simplified the random effects following recommendations from Baayen, Davidson, and Bates (2008) . In each analysis, we analyzed 9,240 observations from 154 participants responding to 60 nonwords.

The final model used for the analysis of accuracy was structured as follows: Model ← glmm (log.odds.accuracy ∼ Condition × Age group + (1|Participant) + (1|Item). Table 5 presents the output from this model.

The intercept represents the performance of the youngest age group (children) in the control condition; all other estimates are relative to this value. To determine whether the main effects of condition, age group and the condition x age group interaction were significant, pairwise likelihood ratio tests (LRTs) were used to compare the full model with simplified models in which the main effects were removed in turn. These comparisons indicated a significant effect of age group (LRT: χ 2 [6] = 61.44, p < .001), condition (LRT: χ 2 [4] = 47.48, p < .001) and a significant Age Group × Condition interaction (LRT: χ 2 [3] = 32.43, p < .001). The interaction between condition and age group was explored using the package phia ( De Rosario-Martínez, 2015 ). An examination of simple effects revealed that the effect of condition was significant for children (χ 2 [1] = 6.81, p < .01), younger adolescents (χ 2 [1] = 11.04, p < .01), older adolescents (χ 2 [1] = 33.32, p < .001) and adults (χ 2 [1] = 23.90, p < .001). Examination of interaction contrasts showed that the magnitude of the effect of condition did not differ significantly between children and younger adolescents (χ 2 [1] = 1.70, p = .38), or between older adolescents and adults (χ 2 [1] = 0.51, p = .47), but the magnitude of the effect was significantly greater for older adolescents than for younger adolescents (χ 2 [1] = 13.48, p < .01).

RTs

The final model used for the analysis of RTs was structured as follows: Model ← lmer (reaction time (RT).outliers.removed ∼ Condition × Age group + (1|Participant) + (1|Item). Table 6 presents the output from this model.

The intercept again represents the performance of the youngest age group (children) in the control condition and all other estimates are relative to this value. As before, we used pairwise LRTs to analyze the main effects of condition, age group and the condition x age group interaction. These comparisons indicated a significant effect of condition (LRT: χ 2 [4] = 70.65, p < .001), age group (LRT: χ 2 [6] = 164.00, p < .001), and a significant Age Group × Condition interaction (LRT: χ 2 [3] = 65.59, p < .001). The interaction between condition and age group was explored using the package phia ( De Rosario-Martínez, 2015 ). An examination of simple effects revealed that the effect of condition was significant for older adolescents (χ 2 [1] = 12.37, p < .01) and adults (χ 2 [1] = 29.38, p < .001), but not for children (χ 2 [1] = 0.15, p = 1.00) or younger adolescents (χ 2 [1] = 0.10, p = 1.00). Examination of interaction contrasts showed that the magnitude of the effect of condition did not differ significantly between children and younger adolescents (χ 2 [1] = 0.78, p = .38), but the effect was greater for older adolescents than for younger adolescents (χ 2 [1] = 15.31, p < .001), and greater for adults than older adolescents (χ 2 [1] = 5.84, p < .05).

Discussion

This study used a lexical-decision task to investigate the developmental trajectory of online morphological processing in nonword reading. Accuracy was lower for pseudomorphemic nonwords compared with control nonwords across all age groups; participants were more likely to incorrectly accept nonwords comprising a real stem and suffix ( earist ) than nonwords comprising a real stem and nonmorphological ending ( earilt ). This effect was greater in adults and older adolescents than in children and younger adolescents. The discrepancy in accuracy is consistent with existing adult findings ( Crepaldi et al., 2010 ; Taft & Forster, 1975 ) and provides verification of morphological sensitivity in English-speaking children aged 7 to 9 ( Burani et al., 2002 ; Casalis et al., 2015 ). The current study rectifies limitations in stimuli previously used with children (e.g., Burani et al., 2002 ; Casalis et al., 2015 ), and for the first time incorporates data from adolescent participants. Our findings are inconsistent with supralexical theories that see morphological analysis as taking place after lexical access ( Giraudo & Grainger, 2001 ). Nonwords by definition are not represented in the lexicon. Therefore, if morphological structure is analyzed following lexical access, then there should be no difference in responses to pseudomorphemic ( earist ) and control nonwords ( earilt ) because both nonword types will be treated equally. Instead, our data lend support to morpho-orthographic theories that argue that the process of decomposition takes place prior to lexical access ( Rastle & Davis, 2008 ; Taft, 2004 ), and dual-route models in which both whole-word access and decomposition are available ( Baayen et al., 1997 ).

The RT data were less clear-cut. Both adults and older adolescents were slower to reject the pseudomorphemic nonwords ( earist ) relative to the control nonwords ( earilt ), replicating previous findings with adults (e.g., Crepaldi et al., 2010 ). This is consistent with Taft and Forster’s (1975) theory that complex words are stored in their root form in the lexicon, and are stripped of their affixes during recognition. A nonword comprising an existing stem and suffix ( earist ) will result in a lexical entry being retrieved ( ear ). The process of checking the legitimacy of the stem-suffix combination will generate longer RTs compared to nonmorphological nonwords ( earilt ), which are not decomposed and can be rejected once a search of the lexicon reveals no match. However, no difference in RTs was found for children and younger adolescents, corroborating findings from Casalis et al. (2015) , which indicated that although French children were slower and less accurate to reject nonwords comprising a stem and suffix, the effect for English-speaking children was limited to accuracy.

Why might morphological effects emerge in accuracy but not RTs in children and younger adolescents? One possibility is that the types of suffixes used in the pseudomorphemic condition influenced response times. Previous studies with children have tended to include only neutral suffixes such as – y and – er (e.g., Carlisle & Stone, 2005 ; Laxon et al., 1992 ), which attach to independent words, do not alter stress in the word to which they attach, and are more productive than nonneutral suffixes such as – ic and – ary ( Tyler & Nagy, 1989 ). The pseudomorphemic nonwords in the present study contained both neutral and nonneutral suffixes (60% and 40% respectively). It has been argued that the process of decomposition may vary according to suffix type ( Hay, 2003 ) and there is some indication that children’s knowledge of these two types of suffix develops differently as they undergo a period of overgeneralization in the acquisition of neutral, but not nonneutral, suffixes ( Tyler & Nagy, 1989 ). Thus, it is plausible that for the younger age groups, the morpheme interference effect on RTs only emerged for the more predictable, rule-driven neutrally suffixed pseudowords. However, subsequent analyses did not show this to be the case: the difference in RTs did not vary between the neutrally- and nonneutrally suffixed stimuli in either age group (all p s > .05).

A second possibility is that the mechanisms driving decomposition may differ between the younger and older age groups, and that children and younger adolescents might rely more heavily on explicit morphological knowledge in their decisions than the older participants. One argument raised by an anonymous reviewer is that the younger age groups may be more sensitive than the older age groups to the presence of an existing stem across both nonword types, independent of the morphological status of the nonword (see Casalis et al., 2015 ; Giraudo & Voga, 2016 ). This would slow responses to the control nonwords as well as the pseudomorphemic nonwords, which might account for the absence of an RT effect in the younger age groups. This would not explain the observed differences in accuracy, but slower responses to all nonwords could result in greater reliance on explicit processes to determine lexical status, leading to more errors in the pseudomorphemic condition.

Following the suggestion of a reviewer, we investigated the role of semantic interpretability to explore the idea that the younger age groups were relying more on explicit morphological knowledge than the older age groups. Semantic interpretability refers to the ease with which morphologically structured nonwords can be interpreted on the basis of the meanings of their morphological components ( Longtin & Meunier, 2005 ). Nonwords such as trueness are semantically interpretable: the suffix – ness attaches to adjectives to form a noun, the stem-suffix combination is in accordance with English phonotactic rules, and there are equivalent real word examples (e.g., gentleness ). All 30 pseudomorphemic nonwords were coded as either semantically interpretable or uninterpretable on the basis of the above c

Footnotes

Incorporating performance on background measures of reading and vocabulary in models examining accuracy resulted in a failure to converge, indicating that our data lacked sufficient power to explore individual differences. Thus, our final models included just the fixed effects of condition, age and their interaction.