Statistical learning offers a compelling new account of the nature of children's learning. Broadly speaking, statistical learning refers to the process through which regularities in the input are uncovered; this term was coined by Saffran, Aslin, and Newport (Reference Saffran, Aslin and Newport1996) in the description of their seminal studies of lexicon formation. In this now classic research, 8-month-old infants’ sensitivity to the likelihood of co-occurrence of syllables allowed them to extract “words” from an artificial language. The power of the statistical learning mechanism has been demonstrated across otherwise separate domains, such as learning about auditory (Saffran, Johnson, Aslin, & Newport, Reference Saffran, Johnson, Aslin and Newport1999; Saffran, Reeck, Niebuhr, & Wilson, Reference Saffran, Reeck, Niebuhr and Wilson2005), visual (Fiser & Aslin, Reference Fiser and Aslin2001, Reference Fiser and Aslin2002), and tactile (Conway & Christiansen, Reference Conway and Christiansen2005) stimuli (for a review, see Perruchet & Pacton, Reference Perruchet and Pacton2006). The present study is designed to test the explanatory power of statistical learning in its most challenging domain: learning about meaning.

A primary area of current debate lies in the types of learning that can be explained by a statistical learning mechanism. There have been multiple demonstrations of statistical learning of surface features of stimuli. Evidence of statistical learning of co-occurrence of letters is one such example (e.g., Treiman & Kessler, Reference Treiman and Kessler2006; for demonstrations in other domains, see Gómez, & Gerken, Reference Gómez, Gerken, Tomasello and Bates2001). It has recently been suggested that the power of statistical learning might be limited to such surface features (Peña, Bonatti, Nespor, & Mehler, Reference Peña, Bonatti, Nespor and Mehler2002). It is far more contentious as to whether statistical learning can account for learning at higher levels, specifically in terms of grammar (e.g., Saffran, Reference Saffran, Colombo, McCardle and Freund2009; Seidenberg, MacDonald, & Saffran, Reference Seidenberg, MacDonald and Saffran2002). Morphemes are considered to be the foundational and universal building blocks; they are the units on which grammar is based. Testing learning in the morphological domain offers a unique test of the power of statistical learning.

In the study of spelling development, it has long been argued that learning about higher-order dimensions, specifically morphology, involves learning rules. According to several theories, it takes several years for children's spellings to reflect morphemic regularities that exist in the orthography (~age 8), progressing through stages as they do so (e.g., Frith, Reference Frith, Patterson, Coltheart and Marshall1985; Henderson, Reference Henderson1985). The prediction is that children learn rules when these exist (Nunes, Bryant, & Bindman, Reference Nunes, Bryant and Bindman1997). As an example, Nunes and her colleagues showed that children first spell past tense verbs phonetically (making mistakes such as *walkt), before using -ed consistently to represent past tense. Henderson went on to suggest that children first become sensitive to morphological regularities for inflectional morphology and later for derivational morphology, which involves more substantive changes between words. All such models predict protracted learning for the acquisition of rules.

The substantial evidence that even very young writers’ spellings reflect the morphemic structure of words (e.g., Treiman & Cassar, Reference Treiman and Cassar1996; Treiman, Cassar, & Zukowski, Reference Treiman, Cassar and Zukowski1994) contradicts the predictions of a rule-based model. For example, Treiman and her colleagues showed that young children are more likely to include the spelling of the penultimate sound in a word when the sound is a part of a root than when it is not (e.g., in comparing /n/ in rained and in brand; see also Kemp, Reference Kemp2006). Similarly, Deacon and Bryant (Reference Deacon and Bryant2005) demonstrated that young children are better able to spell the initial portions of two-morpheme than one-morpheme words (e.g., turn in turning and turnip; see also Deacon, Reference Deacon2008; Deacon & Bryant, Reference Deacon and Bryant2006a, Reference Deacon and Bryant2006b; for a review, see Pacton & Deacon, Reference Pacton and Deacon2008). These differences emerge across whole-word spelling tasks, choice tasks, and completion tasks (e.g., Deacon, Reference Deacon2008; Deacon & Bryant, Reference Deacon and Bryant2005; Treiman & Cassar, Reference Treiman and Cassar1996; Treiman et al., Reference Treiman, Cassar and Zukowski1994). It seems that even young children are better at identifying or generating the correct spellings for sections of words when these spellings are morphological determined than when they are not. This evidence counters a rule-based approach in its findings that children's spellings are influenced by the morphological structure of words at an age at which rule-learning models predict that spelling should be entirely driven by phonological regularities.

Perhaps even more problematic for rule-based approaches are the findings that adult spellers do not rely on simple morphological rules when it is possible to do so. The clearest demonstration of this emerges in a study of spelling of the plural in pseudowords. Following rules would lead adults to spell pseudowords ending in the /z/ sound that are presented as plurals with an s; instead, it seems that adults’ spelling reflects sensitivities to the multiple regularities that underlie print. Kemp and Bryant (Reference Kemp and Bryant2003) showed that both child and adult spellers were twice as likely to use -s to spell pseudowords presented as plurals when the penultimate sound was a consonant than when it was a long vowel (e.g., preens vs. smees). The substantial variation in performance based on the adjacent sound (or letter) shows that even mature spellers do not rely on a rule such as “spell plurals with an s.” The pattern of results is consistent with statistical learning of letter patterns because -s follows consonants far more frequently than long vowels in English (for comparable findings in French and Dutch, respectively, see Pacton, Fayol, & Perruchet, Reference Pacton, Fayol and Perruchet2005; and Sandra, Frisson, & Daems, Reference Sandra, Frisson and Daems2004). There was some evidence of sensitivity to the role of plurality in determining the use of -s; university-educated adults and better spelling children were more likely to use -s when the long-vowel pseudowords were presented as plurals rather than as nonplurals (e.g., “those smees” vs. “that smeese”). The clear effect of the adjacent sound on pseudoword spelling suggests that this sensitivity to morphological structure is not underpinned by knowledge of a rule.

Statistical learning offers an alternative account as to how children might learn about the morphological basis of spelling, particularly as to why we see evidence of sensitivity to morphological structure in the spelling of very young children. Specifically, it is suggested that children are sensitive to the co-occurrence of sound, letters, and meaning (or of phonology, orthography, and semantics, respectively; Pacton & Deacon, Reference Pacton and Deacon2008; Pollo, Treiman, & Kessler, Reference Pollo, Treiman, Kessler, Grigorenko and Naples2008; Seidenberg & Gonnerman, Reference Seidenberg and Gonnerman2000). Children might, for example, become sensitive to the regularity that plurality in English is generally marked with an s, despite different pronunciations (/s/ and /z/, as in cats and dogs), as a result of the co-occurrence of these sounds, with the letter s and the meaning of plurality. To take an example from derivational morphology, the words teach and teacher are likely to appear together in text, and the overlap of the initial letters teach and the sound /ti:t∫/ along with the meaning “instruct” could lead to a strong association between these forms. Similarly, children might see -er and hear /ər/ often in combination with the meaning “someone who does something.” These co-occurrences in real words in a language could lead children to use consistent spellings for these units (-s, teach, and -er), even at a young age.

Statistical learning approaches would also predict sensitivity to other dimensions of print, such as simple orthographic regularities. Several studies, including that of Kemp and Bryant (Reference Kemp and Bryant2003), demonstrate children's and adults’ sensitivity to the frequency of letter patterns and of meaning patterns (see also e.g., Pollo, Kessler, & Treiman, Reference Pollo, Kessler and Treiman2009; Treiman & Kessler, Reference Treiman and Kessler2006). The emergence of these effects in studies with pseudowords shows their independence from the frequency effects of any one individual word. Statistical learning approaches would predict sensitivity to the frequency of letter- and meaning-based patterns in print.

In the present study, we identified morphological patterns in English in which there is natural variability in the frequency of the semantic and orthographic patterns. We tested the impact of these patterns on children's performance on a spelling choice task. This offered a test of the hypothesis that the frequency with which children are exposed to morphological regularities should impact spelling acquisition.

First, we examined the effects of the frequency of meaning of specific morphemes. The letter pattern -er can be an inflection or a derivation (as in the comparative -er in warmer and the agentive -er in teacher). There are far more derivational than inflectional words ending in -er: a total of 156 derivational forms across the whole language in comparison to only 53 inflectional forms. The same pattern occurs when one looks at the frequency of occurrence of a word in young children's books; there are 65 derivational forms and 33 inflectional forms ending in -er with a frequency at Grade 2. These tallies are from an analysis of the children's frequency database (Zeno, Reference Zeno1995) that is considered to be the most representative of available databases (Lee, Reference Lee2003). Although there are clearly far more derived than inflected forms across adult and child language, average frequencies of these instances are remarkably similar (e.g., log Grade 2 frequencyFootnote ¹ values of 1.27 vs. 0.98 for inflected and derived forms, respectively, p > .05). This means that children are exposed to far more instances of words ending in -er that are derived than inflected. Spelling choices for derived and inflected words ending in -er will be contrasted to those for -er in control words for which there is no semantic basis (such as corner); the difference between choices on morphologically related and control forms is taken to reflect sensitivity to morphological structure in spelling (e.g., Deacon & Bryant, Reference Deacon and Bryant2006a; Treiman et al., Reference Treiman, Cassar and Zukowski1994). If children rely on statistical extraction of morphological spelling regularities, then repeated exposure to a specific meaning in combination with a letter pattern and a sound pattern should affect children's spelling choices; we should see children demonstrating sensitivity to the morphological basis of the spelling of the derivational affix -er earlier than for the inflectional affix -er. This offers a test of the impact of the frequency of occurrence of specific meanings.

Second, we examined the effects of the frequency of the orthographic pattern. In terms of the derivation used to form nouns, -er is far more frequent than the -or allomorph. There are more derived words that end in -er than in -or (e.g., teacher and actor); in the Zeno (Reference Zeno1995) database there are 156 derived forms ending in -er and only 73 ending in -or. This difference in frequency also emerges when one considers just the forms to which young children are likely to be exposed; there are 65 -er derived forms and 28 -or ones with a frequency at Grade 2. As with the inflectional and derivational items ending in -er, these two allomorphs are of similar frequency (e.g., log values of 0.98 vs. 0.81 for -er and -or forms, respectively, p > .05). Children are exposed to far more instances of derived words ending in -er than in -or. If children rely on statistical extraction of morphological spelling regularities, then repeated exposure to a specific spelling in combination with a specific meaning and sound should affect children's spelling choices; children should demonstrate sensitivity to the morphological basis of the spelling of the -er derivational allomorph earlier than the -or allomorph. Spelling choices in these two cases will be contrasted to that of the endings of control words for which there is no semantic basis (such as corner and tractor), with the difference taken to reflect sensitivity to morphological structure in spelling. This contrast tests the impact of the frequency of occurrence of specific spellings.

These contrasts also test the predictions of a rule-based approach. According to this approach, only older writers’ spellings (likely those of children age 8 years and up) would reflect the morphological structure of words. In the context of this study, only children aged 8 years and up should demonstrate greater accuracy in choosing the correct spelling for -er when it is a morpheme (in words such as warmer) in contrast to when it is not (in words such as corner). Henderson's (Reference Henderson1985) further specification of the rule-based model suggests that children might learn about the influence of morphological structure on spelling first for inflections and only later for derivations. There is some evidence that this is the case; children demonstrate sensitivity to the role of morphological structure in spelling inflections at a far younger age than they do for derivations (e.g., Carlisle, Reference Carlisle1988; Deacon & Bryant, Reference Deacon and Bryant2005; Sangster & Deacon, Reference Sangster and Deacon2011; Sterling, Reference Sterling1983). To our knowledge, no study has been able to control for the spelling difficulty of the inflectional and derivational suffixes. The contrast of -er as an inflection and a derivation offers a novel test of Henderson's prediction in a manner that precisely balances the spelling difficulty of the letter patterns.

To summarize, we contrast children's spelling choices for words ending in -er and -or to evaluate predictions of statistical and rule-based explanations of spelling development. This extends prior tests of the potential influence of statistical learning on spelling by manipulating the semantic and orthographic frequency of an individual spelling. We asked 6- to 10-year-old children to choose endings for inflected, derived, and control words that ended in -er. These sets of words were balanced for token frequency. We also asked these children to choose endings for derived and control words ending in -or. A statistical learning approach would predict that the type frequency of the specific meaning or letter pattern should influence even young children's spelling choices. More frequent patterns (such as -er as a derivation) should be acquired earlier than less frequent patterns (such as -er as an inflection or -or as a derivation). A rule-based approach would predict that only older children's spelling choices should be sensitive to morphological regularities and that this sensitivity should emerge first with inflections and later with derivations. This experiment tests these diverging predictions.

RESULTS

Children's answers were scored as correct or incorrect. For the words ending in -er, each child contributed three sets of scores (each out of a maximum of six): one score for the one-morpheme control words, one for the inflected, and another for the derived words. For the words ending in -or, each child contributed two sets of scores (each out of a maximum of nine): one from the one-morpheme control words and the other from the two-morpheme derived words.

Analyses showed reasonable reliability (α = 0.82) for the spelling choice task. Inspection of the data revealed six outliers. Following the guidelines set out in Howell (Reference Howell1992), we applied Winsorizing techniques to the few identified outliers; these ensured that all data could be included in the analysis without allowing undue influence from the outliers.

We present mean proportion correct in each of the conditions for the younger and older children separately in Table 1. Results for each grade are also reported here. We analyzed these results with two separate analyses of variance, both of which evaluated proportion scores for the younger and older children.

Table 1.

Mean (standard deviation) proportion of -er and -or choices for the one-morpheme, inflected, and derived words ending in -er (out of a maximum of 6) and for the one-morpheme and derived words ending in -or (out of a maximum of 9)

Note: The remaining responses reflect choice of -r.

To examine our first question about the influence of the semantic frequency, we analyzed the data from the words ending in -er. These were analyzed using a mixed analysis of variance (ANOVA) with word type (one-morpheme, inflected, or derived) as the within-subjects factor and group (younger and older) as the between-subjects factor. There was a main effect of word type, F (2, 186) = 10.83, p < .001, and an interaction between word type and group, F (2, 186) = 3.60, p = .05. We examined the word type effect at each grade to understand the interaction between word type and group. There were significant word type effects both for the younger, F (2, 112) = 10.50, p < .001, and for the older groups, F (2, 74) = 5.17, p < .01. Paired-sample t tests with Bonferroni corrections were conducted. The younger children were significantly choosing the correct spelling for the derived more often than for both the one-morpheme words, t (56) = 4.57, p < .001, d = 0.45, and the inflected words, t (56) = 3.57, p < .001, d = 0.30. Their scores for the inflected words did not differ from those for the one-morpheme words, t (56) = 0.20, p = .84. The older children were more likely to choose -er (correctly) as the spelling for the endings of both the inflected words, t (37) = 2.55, p = .01, d = 0.38, and the derived words, t (37) = 2.90, p < .01, d = 0.42, than for the endings of the one-morpheme words. Their scores for the inflected and derived words did not differ, t (37) = .00, p = 1.00.

To examine our second question as to the influence of frequency of the orthographic pattern, we contrasted likelihood of correct performance with the derived and one-morpheme control words for the -er and -or data. We analyzed data for the correct responses with a mixed ANOVA with ending type (-er or -or) and word type (one-morpheme or derived) as the within-subjects factors and group (younger and older) as the between-subjects factor. There was a main effect of group, F (1, 93) = 37.62, p < .001, showing that the older children performed better than the younger children. There were significant main effects of ending type, F (1, 93) = 2.171, p = .001, and word type, F (1, 93) = 10.67, p < .01, which were qualified by an interaction between the two variables, F (1, 93) = 24.87, p < .001. There was no three-way interaction between the three variables. The two-way interaction arose because there was clearly a difference in the effect of word type in each case. As we uncovered earlier, children's spelling choices were more accurate with derived words than with control -er words, t (94) = 6.27, p = .001, d = 0.52. There was no such difference for the -or words, t (94) = 1.20, ns.

We confirmed our pattern of results in several ways. We conducted analyses of the -er choices with analyses by items. In these analyses, as in those by participants, scores tended to be higher for the derived condition than for the one-morpheme control condition, for both the younger and the older children (ps = .04 and .09, respectively). For the older children, scores for the inflected items tended to be higher than those for the one-morpheme control condition (p = .10). Only significance in the analyses by participants is required to reject the null hypothesis given the high degree of selection of the items into conditions (Raaijmakers, Schrijnemakers, & Gremmen, Reference Raaijmakers, Schrijnemakers and Gremmen1999). It is not unusual for significant results in analyses by participants to emerge as trends in analyses by items, given the substantial reduction in power observed when conducting analyses across items. Further, inspection of Table 1 suggests that the patterns observed in the groups (each collapsed across two grades) are similar to those observed in the grades that make up each of the groups.

Given our use of a spelling choice task, we contrasted mean performance for each group in each condition against chance, and we examined performance for individual children. Mean correct choices for each group in each condition were significantly different than chance (0.33; ts > 3.60, ps ≤ .001). Both the younger and the older children's choice of -er in the -er conditions was above chance levels, with only seven individual children in the younger group and none of the older children performing at or below chance. For the -or task, choice of -or by the younger children was below chance, with 43 children performing at or below chance, and by the older children the choice was above chance, with 10 performing at or above chance. Extending this examination to incorrect choices, we see that both the younger and the older children were less likely than chance to choose -or in the -er conditions (ts > 8.95, ps ≤ .001), with only five individual children in the younger group and none of the older children performing at or above chance. This pattern parallels their greater than chance likelihood to choose the correct answer, -er. For the -or task, choice of -er by both the younger and the older children was above chance (ts > 2.65, ps ≤ .015), with 52 of the younger children and 29 of the older children choosing this option at or above chance. Results for the -or task show that the younger children were more likely than chance to choose -er and less likely than chance to choose -or. The older children were more likely than chance to choose both -er and -or.

Given the dominant choice of -er for the items ending in -or, we examined whether children's choices of this option differed on the basis of the morphological status of the words. We conducted a mixed ANOVA with Word Type (one-morpheme or derived) as the within-subjects factor and Group (younger and older) as the between-subjects factor. Only a main effect of group emerged, F (1, 93) = 19.83, p < .001, with no main effect of Word Type, F (1, 93) = 1.36, ns, or interaction between Word Type and Group, F (1, 93) = 0.52, ns.