Factors Affecting Accuracy of Past Tense Production in ......children, but especially children with...

Factors Affecting Accuracy of Past TenseProduction in Children With SpecificLanguage Impairment And TheirTypically Developing Peers:The Influence of Verb Transitivity,Clause Location, and Sentence Type

Purpose: The author examined the influence of sentence type, clause order, andverb transitivity on the accuracy of children’s past tense productions. All groups ofchildren, but especially children with specific language impairment (SLI), werepredicted to decrease accuracy as linguistic complexity increased.Method: The author elicited past tense productions in 2-clause sentences from 5- to8-year-old children with SLI (n = 14) and their typically developing peers (n = 24).The target sentences varied in the type and obligatory nature of the second clause andthe number of arguments.Results: On average, 85% of the responses across all groups and sentence typescontained 2 clauses. Fewer 2-clause sentences were produced in the complementclause condition than in the other conditions. Sentence type and clause order, butnot argument structure, influenced use of past tense. Children with SLI had a similarbut less accurate profile as compared with the age-matched group. The youngermean length of utterance (MLU)–matched group reflected decreased accuracy witheach additional source of linguistic complexity.Conclusions: Increased syntactic difficulty decreases use of morphology for allchildren, supporting the hypothesis that processing demands influence morphologicalaccuracy. MLU-matched children, but not children with SLI, were more affected bychanges in linguistic complexity. Further work on age-related changes in sentenceproduction is necessary.

KEY WORDS: complex syntax, past tense, SLI, logistic regression,argument structure

C hildren with specific language impairment (SLI) have a deficit inlanguage acquisition and use in the absence of any obvious causalfactors such as autism, hearing impairment, or mental retarda-tion (for a full description of the profile, see Leonard, 1998). In English,variable use of tense and agreement morphemes (am, is, are, –s as inplays, or –ed as in jumped) is a hallmark characteristic of the disorder, somuch so that a major theory of the problem, the extended optional infini-tive account (Rice, Wexler, & Cleave, 1995), is named to describe that pre-cise phenomenon. Difficulty with tense markers in general and with past

Amanda J. OwenUniversity of Iowa, Iowa City

Journal of Speech, Language, and Hearing Research • Vol. 53 • 993–1014 • August 2010 • D American Speech-Language-Hearing Association 993

Complimentary Author PDF: Not for Broad Dissemination

tense in particular persists into the early elementaryschool years (e.g., King & Fletcher, 1993) and can be ob-served under appropriate task demands into adolescence(Leonard,Miller, & Finneran, 2009). It is worth pointingout that variable use is observed in both typically devel-oping (TD) children and children with SLI, albeit for dif-ferent time periods and to differing degrees (Goffman &Leonard, 2000). Continued difficulty has implicationsfor success in academic settings, given that morpholog-ical deficits often reappear in complex tasks such as pro-ducing written texts even after those same deficits havebeen resolved inspoken conversational language (Windsor,Scott, & Street, 2000).

Whereas much attention has been given to the ob-serveddifferences betweenTDchildrenand childrenwithSLI, less attention has been directed to the sources thatmight influence within child variability. The purpose ofthis study was to examine linguistic factors that maycontribute to the variable use of grammaticalmorphologyand to determine whether these factors affect childrenwith and without SLI in similar ways. Working within ageneral processing framework of language production,I hypothesized that any increase in linguistic complexitywould influence children with SLI to a greater degreethan their TD peers and that the potential for errors asa result of increased processing demandswould accumu-late more rapidly as additional sources of complexity wereincluded in the sentence. Specifically, I was interestedin how changes in the syntactic planning unit (two co-ordinated main clauses, a main clause and an optionaltemporal clause, or a main clause and a complementclause), location of the clause within the sentence (first/second), andnumber of arguments (transitive/intransitive)influenced the production of regular and irregular pasttense.

Variability in Verb MorphologyPast tense production in childrenwith SLI and their

TD peers has been especially well studied. Some ac-counts of the deficits observed in children with SLI arebased on the notion that these children have particulardifficulty with the linguistic operations that underlie theuse of tense and agreement morphology (e.g., uniquechecking constraint [UCC]; Wexler, 1998; representa-tion deficit for dependent relations [RDDR]; van der Lely& Battell, 2003). For instance, the UCC account claimsthat children know about tense and agreement but havedifficulty checking both features and, thus, omit mor-phemes such as –ed or is in English. Other accounts aregrounded in a more general processing approach andassume that working memory or processing speed defi-cits affect the acquisition of past tense (e.g., Ellis Weismer,Evans,&Hesketh, 1999;Miller, Kail, Leonard,&Tomblin,2001).

Particularly consistent with the second approach,past tense production seems to be mediated by proper-ties of the individual lexical item being inflected. Rang-ing from the type and token frequency of the individualverb (Albright & Hayes, 2003; Nicoladis, Palmer, &Marentette, 2007; Plunkett & Marchman, 1996) to theease of inflecting a particular phonological form (Berko,1958;Leonard,Davis,&Deevy, 2007;Marchman,Wulfeck,& Ellis Weismer, 1999; Marshall & van der Lely, 2006;Oetting&Horohov, 1997)or the facilitative roleof prototyp-ical lexical aspect (Bloom, Lifter, &Hafitz, 1980; Johnson&Fey, 2006; Shirai&Anderson, 1995), word-level factorscan facilitate or hinder the ability of TD children to usepast tense. However, not all of these factors have thesame influence on productions by children with SLI. Forexample, the accuracy of their productions is more af-fected thanTD children’s productions by thephonologicalformof the lexical item (Leonard et al., 2009) but is less af-fected by the lexical aspect of the word (Leonard, Deevy,et al., 2007).Using a single case study of a childwithSLI,Johnson and Morris (2007) demonstrated that phono-logical and aspectual factors influence tense productionadditively. In this one child, they observed 0% accuracywhen a lexical item both ended in an obstruent coda andused nonprototypical lexical aspect, 33% accuracy whenonly one factor was included, and 100% accuracy when acontinuant codawas used and the lexical aspect of the verbwas facilitative. Typical 22-year-olds who were includedin the study showed a less graded influence—accuracydeclined only when the lexical item both ended in anobstruent coda and used nonprototypical lexical aspect.This case study suggests that children with SLI aremore vulnerable to the demands created by the absenceof facilitative lexical and phonological factors than theiryounger TD peers, but larger scale studies are requiredto verify this finding.

Most of these factors function at the level of the indi-vidual lexical item. When they extend to include the pre-dicate (as in lexical aspect) or the sentence (as with verbposition in the sentence; Dalal&Frome-Loeb, 2005), thesefactors influence past tense production primarily becausethey affect the perceptibility or prototypicality of the lexi-cal item in that specific context. I aim to extend the findingthatmultiple factors canaffect production beyond the levelof word-internal factors to include both word-internal andword-external factors. In the section that follows, I review afrequently citedmodel of sentenceanddiscourseproductionand the rather sparse literature that examines sentence-level factors in children with SLI and their TD peers.

Production Processes at the Sentenceand Discourse Levels

Many studies of sentence- and discourse-level pro-duction processes are grounded in themodel of Bock and

994 Journal of Speech, Language, and Hearing Research • Vol. 53 • 993–1014 • August 2010


Levelt (1994; e.g., Ellis Weismer & Hesketh, 1998; Guo,Tomblin & Samelson, 2008; Leonard et al., 2000). In thismodel, sentence production proceeds incrementally, withgrammatical morphology being among the last elementsincluded in the process. To produce a sentence, the speakerfirst conceives of the message he or she wishes to conveyand then selects lexical items. These lexical items areassigned to functional roles (case assignment) within thesentence based on event structure and attentional focus.Next, a sentence frame is selected that suits the lexicalitems and functional roles, and constituent assemblybegins. It is at this point that inflectionalmorphemes areinserted into the sentence. Phonological encoding is thefinal step in the model, at which point production of thesentence is presumed to begin. It is important to pointout here that incremental processing does not mean thatthe entire sentence is planned at each stage before pro-ceeding to the next stage or before phonological encodingbegins. Rather, there is evidence from pausing and hesi-tations in adults that planning occurs at the level of thephrase or clause (Holmes, 1988). Furthermore, the waythat speaking latency and utterance duration vary withplanning load suggests that phonological encoding andproductionmay begin before the entire utterance is com-pletely formulated (Ferriera&Swets, 2002). Thiswidelyadopted model provides a framework for understandingthe components of sentence production. The value comesfrom its comprehensiveness and the fact that it providesa framework for thinking about ways that one sentenceproduction component may influence another aspect ofproduction.

Within this general framework, there is evidencefrom adult priming studies that more activated or moreaccessible elements lead to greater processing efficiencyand influence production (Bock & Griffin, 2000). Just asa more accessible message, a more activated syntacticframe, or more frequent lexical items each can lead tofaster ormore fluent responses due to decreased process-ing demands, a less accessible message, less frequentlexical items, or an increase in the number of functionalroles to be assigned may each lead to an increase in pro-cessing demands. In adults and proficient language users,such a model may play out in terms of response time,latency to begin speaking, and fluency effects.

The model has been extended by other researchersto address other populations (Leonard et al., 2000). Inindividuals who are not proficient language users (e.g.,young children, individuals with language impairment)or who have other capacity limitations (e.g., individualswith aphasia), increases in processing demandsmay leadto decreases in accuracy for vulnerable elements withinthe sentence, such as morphology (Leonard et al., 2000).In addition, an increase in hesitations or dysfluenciesmay also be observed, a disruption in production that ispredicted more directly by the Bock and Levelt (1994)

model (Rispoli & Hadley, 2001). From the findings ofLeonard and colleagues (2000, 2002), I can infer that asmore resources are devoted to the early process of for-mulation, the odds of an error somewhere within thesentence may increase. Leonard and colleagues (2000)demonstrated that the benefits of priming that accruefor the auxiliary (a functionword) also accrue for simplepast tense (–ed) despite the fact that most models ofsentence production assume that inflection is retrievedearly on as part of the sentence frame, prior to the order-ing of lexical items and insertion of functionwords. Thus,the framework of the Bock and Levelt (1994) model—incombination with extensions of the model by Leonardand colleagues (2000, 2002), which assume processinglimitations and the influence of these limitations on bothfunction words and bound inflectional morphology—canaccount for a myriad of factors that affect accuracy ofmorphological production. Increasedprocessing demandsrelated to lemma selection, thematic role assignment, orother early elements of the sentence production processesshould lead to a higher likelihood that errors will occurin constituent assembly and, in particular, an increasedlikelihood of errors in tense and agreement production.

The integrity of this extension of the model is sup-ported by research inwhich themanipulation of process-ing demands early in sentence planning affectsmorphemeproduction. Although the evidence previously cited comesfrom priming studies, there is also evidence that com-plexity influences sentence production without the useof a priming paradigm to increase the accessibility of thetargets. For instance, Grela and Leonard (2000) elicitedsentences using simple intransitive, transitive, andditran-sitive verbs with and without a prepositional phrase fromthree groups of children: children with SLI, TD agematches, and TD children matched on mean length ofutterance (MLU). They reasoned (a) that as argumentstructure complexity increased, so would processing loadand (b) that auxiliary is, the most vulnerable element ofthe sentence, would be omitted. Although the overall er-ror rate was higher for the childrenwith SLI as comparedwith their age- andMLU-matched counterparts, both chil-dren with SLI and their MLU-matched peers were morelikely to omit the auxiliary when producing sentences thatrequired three arguments than when producing sentencesthat required only one or two arguments. Importantly,the addition of a prepositional phrase to lengthen thesentence did not influence the results, showing that it isthe number of arguments, rather than overall sentencelength, driving the findings. Agematcheswereunaffectedby the argument structure manipulation and performedat ceiling in all conditions (Grela&Leonard, 2000). Thesefindings demonstrate that children with SLI are moreaffected than their TD age mates by increases in syntac-tic complexity. This result has been confirmed both inspontaneous language samples (Grela & Leonard, 1997)

Owen: Production Processes and Complex Syntax 995


and cross-linguistically (Dutch: de Jong, 1999; French:Pizzioli & Schelstraete, 2008).

Thordardottir (2008) also considered the role of taskdemands on grammatical accuracy, but instead of focus-ing on the level of the sentence, she focused on the dis-course context. Grammatical accuracy in conversational,narrative, and expository samples was considered forEnglish-speaking school children with and without SLI.Both groups showed lower performance with verb mor-phology in the expository and narrative contexts than inthe conversational contexts. Thus, in English, context ap-pears to influence accurate production at multiple levels,even in childrenwho are beyond the age of initialmastery.

Motivation for the Study QuestionsIn summary, the grammaticalmorpheme productions

of children with SLI are sometimes more vulnerable toprocessing demands than those of their TD counterparts.In some studies, children with SLI do not capitalize onfacilitative factors (e.g., prototypical lexical aspect) to thedegree that same-age and younger TD peers do, whereasin other studies (e.g., phonotactic probability), childrenwith SLI are more affected than either group of TD chil-dren. In still other studies, both children with SLI andchildren at a similar MLU level are affected to the samedegree (e.g., argument structure).

This variety of relative performance may be due,in part, to the interaction between word- and sentence-level factors. That is, a given word-level factor such asnonprototypical aspect may represent a higher process-ing load in a complex sentence structure than in a sim-ple sentence structure. As of yet, it is not known howmultiple linguistic factors interact with each other toinfluence sentence production in children. One mighthypothesize that increased processing demands canaccumulate throughout the sentence and that when asufficient number of demands are placed on the produc-tion system, some elements will be lost. Thus, the pro-duction of a complement clause structure along with twoor more arguments may combine to place larger thannormal processing demands on the production systemand cause the child to omit grammatical morphemes,even when production of a simpler complement clausestructure or sentences that contain multiple argumentsmay not be problematic independently. Corroboration ofthe cumulative effects observed is necessary, given thatthe strongest evidence comes from a single case study ofa child with SLI; TD children demonstrated cumulativeeffects only when both nonprototypical aspect and the fi-nal consonant of the lexical itemwere factors (Johnson&Morris, 2007).

In the processing literature, work with childrenhas focused almost exclusively on simple sentences, bothin terms of priming studies (e.g., Leonard et al., 2000)

and in terms of the manipulation of processing load viaincreased arguments within a sentence (e.g., Grela &Leonard, 2000). Inasmuch as syntactic complexity is arelevant factor influencing processing load, complex sen-tence frames should induce more difficulties with vul-nerable elements in the production process than simplesentence frames. Furthermore, sentence frames that aredivisible and can be planned in smaller units, such ascoordinated sentences, may place fewer demands on theproduction system than sentences that function as acoherent whole (e.g., sentences containing complementclauses, in which the complement clause functions as anobligatory argument of the main verb; Holmes, 1988).In fact, extant evidence suggests that complex sentencesmay constitute a particularly high processing load forchildren with SLI. These children are reported to usefinite complement clauses and adverbial clauses laterand with reduced frequency than would be expectedfor their age (Marinellie, 2004; Schuele & Dykes, 2005;Schuele &WismanWeil, 2004). Elicited production dataalso show that children with SLI produce fewer finitecomplement clauses than their TD peers who are thesame age or who have the same expressive vocabularyskills. Furthermore, when they do produce the comple-ment clauses, they make more errors, leaving out bothverb-related morphemes and the optional complemen-tizer that to a greater degree than do their peers (Owen& Leonard, 2006). Overall, this would suggest that chil-dren with SLI have greater difficulty with and show amore protracted developmental course for complex syn-tax, although they are able to produce these sorts ofsentences by early elementary school.

Although not designed to address the effect of sen-tence complexity on grammaticalmorpheme production,a study by Thordardottir (2008) provided indirect evi-dence of this effect. Recall that in that study, childrenwith SLI produced less accurate verbmorphology in nar-rative and expository discourse contexts than in conversa-tional contexts. Importantly, the rate of complex sentencesis higher in narrative and expository discourse than inconversational discourse (Nippold, Hesketh, Duthie, &Mansfield, 2005). Therefore, sentence complexity mayhave contributed to these results.

In this study, I was interested in examiningwhethermultiple types of linguistic complexity increase process-ing load in a cumulative fashion and whether childrenwith SLI are affected by these processing demands toa greater degree than TD children. To that end, I elic-ited past tense forms from children in three differentcontexts. I chose past tense production because I antic-ipated that the morphophonological complexity associ-ated with English past tense would be highly vulnerableto changes in processing load. Two potential sources oflinguistic complexity were chosen: sentence type (coordi-nated, temporal adverbial, and finite complement) and



verb transitivity (transitive and intransitive verbs). Basedon sentence production work with adults, sentence typewas expected to reflect different units of planning. Adultsseem to plan coordinated clauses individually, pausingbetween each clause. In contrast, finite complementclauses seem to be planned with the main clause, pre-sumably because the clause is an obligatory argumentof the verb (Holmes, 1988). Temporal adverbial clauseshavenot been tested directly butmost likelywould fall inbetween, given that the two clauses are syntacticallydependent on the main clause, but the adverbial clauseremains optional. The size of the unit that is beingactively plannedmay influence the number of elementsthat have to be retrieved and held in working memorywhile the production operations proceed to speech. Thus,the number of arguments in a sentence should also af-fect the difficulty in producing the utterance, consider-ing that this influences the number of lexical items thatmust be retrieved, assigned thematic roles, and held inworking memory.

Hypotheses and PredictionsI hypothesized that linguistic complexity would ac-

cumulate over the course of sentence production. There-fore, I predicted that the likelihoodof omittingamorphemewould be higher in a sentence that was more syntacti-cally complex and/or that contained multiple sourcesof complexity. Likewise, I predicted that portions of thesentence that were planned or produced later would bemore likely to contain speech errors than earlier por-tions of the sentence because of trade-off effects betweenaccuracy and efficiency. Once speaking has begun, ifthere is no obvious place to pause to formulate the restof the sentence, then the pressure to speak—in combi-nationwith the increasedprocessing load ofmaintainingtheunuttered portions of the first clause inmemorywhilecontinuing to plan the second clause of the sentence—may lead to greater error rates. However, this is not tosay that length is the driving factor in production errors.Rather, based on prior work (de Jong, 1999; Grela &Leonard, 2000; Pizzioli & Schelstraete, 2008), I arguethat it is the production operations—such as planningmultiple clauses or retrieving additional arguments—that drive the resource demands. To verify this, a lengthmanipulation was added to half of the intransitive verbsto control for any effects of utterance length (Grela &Leonard, 2000). With this in mind, the odds of includinga correct tense marker should decline with each addi-tional source of complexity for all children. What shoulddiffer across groups are the initial levels of accuracy andthe rate of decline in accuracy as thenumber of sources ofcomplexity increase. Because younger children have hadless experience with and exposure to complex sentencesand are generally less proficient at using morphology

than older children, I predicted that younger TD childrenwould be less likely to produce accurate past tense mark-ers in all sentences as comparedwith their olderTDpeersbut that the rate of change in accuracy would be similar.

Given the hypothesis that children with SLI expe-rience working memory/processing capacity deficits(e.g., Ellis Weismer et al., 1999; Miller et al., 2001; butsee Leonard, EllisWeismer, et al., 2007, for evidence thatthese are not interchangeable concepts), I predicted thatthese children would be more sensitive to the accumula-tion of complexity within a sentence than their TD peers.That is, children with SLI should be less likely to pro-duce a tense or agreement marker than age-matchedchildren for all sentence type and verb transitivity com-binations. The odds of the children with SLI producing apast tense form might be expected to be comparable tothose of a childmatched onMLUwhen the task demandswere low, given the limited experience of younger chil-dren and the limited tense proficiency of children withSLI. However, as the task demands increased, the oddsof a correct production were expected to decline morerapidly for children with SLI than for either of the twocomparison groups, considering that the children withSLI were hypothesized to be more sensitive to the in-creased task demands than their TD peers, regardlessof age.

MethodParticipants

All research reported in the paragraphs that followwas completed in accordance with the ethical guidelinesfor human subjects research as described in theBelmontReport (Harms, 1979) and required by the National In-stitutes of Health and the Institutional Review Boardat the University of Iowa. Fourteen children with SLI(5;0–8;1 [years;months]) and 24 TD children partici-pated in this study. The TD children were divided intotwo groups: 13 of the TD children (designated as theAGE group) were matched within 3 months of age to achild with SLI; 11 of the TD children (designated as theMLUgroup) were 4-year-olds whowerematchedwithin5 raw score points on the Expressive Vocabulary Test(EVT; Williams, 1997) and within 0.35 words onMLU ina 100-utterance language sample to a child with SLI.

MLU was deemed an appropriate matching crite-rion because utterance-length restrictions may have aparticularly strong influence on the production of two-clause sentences. Although MLU is not a valid mea-sure of syntactic complexity for clinical purposes at thisage (Scarborough, Rescorla, Tager-Flusberg, Fowler, &Sudhalter, 1991), the inclusion of a youngerMLU-matchedcontrol group helps rule out absolute utterance-lengthrestrictions as a reason for low performance. Following



previous work (Owen & Leonard, 2006), the two groupswere also matched on EVT scores because expressivevocabulary may be related to the ability to use verbsflexibly in a variety of syntactic frames. I did not di-rectly match children on morphological skill level, butpost hoc tests show that the groups do not differ in thisrespect, either.

The children in the SLI group met at least two ofthe following four selection criteria. These criteria wereselected because of the psychometric strength of eachmeasure as a means of identifying children with SLIrather than froma theoreticallymotivated position aboutthe construct of SLI.

Criterion 1. Child is currently enrolled in speech,language, or reading intervention (or is enrolled in ther-apy during the prior academic year if seen during thesummer), per parent report.

Criterion 2. Child scored below the 10th percentileon the Structured Photographic Expressive LanguageTest—II (SPELT–II; Werner & Kresheck, 1974). TheSPELT–II was used because of its high sensitivity andspecificity in identifying children with SLI, particularlybetween ages 5 and 7 years (Plante & Vance, 1994).

Criterion 3.Child obtained a composite standard scoreat or below 85 on the Test of Narrative Language (TNL;Gillam & Pearson, 2004), a measure of 5- to 10-year-oldchildren’s ability to use and understand discourse-levellanguage. The TNL was included as a diagnostic cri-terion because it shows good sensitivity and specificity.Using a standard score (SS) cutoff of 85, the TNL hasyielded a mean difference score between TD and SLIchildren of more than 1.5 SD in previous validation stud-ies, and it has been validated for children older than age7 years (Spaulding, Plante, & Farinella, 2006).

Criterion 4. Child scored below a standard score of7 on theNonwordRepetition (NWR) subtest of theNEPSY(Korkman, Kirk, & Kemp, 1998). All of the NEPSY sub-tests have a mean of 10 and an SD of 3. Nonword repe-tition accuracy is a persistent deficit for children withSLI, even after grammatical impairments appear to havebeen resolved (Stothard, Snowling, Bishop, Chipchase, &Kaplan, 1998) and is considered a clinical marker for SLI(Oetting & Cleveland, 2006). Current recommendationsare thatNWRbe used in combinationwith anothermea-sure or in combination with clinical judgment becausechildren without SLI can also show deficits in NWR.

Table 1 provides information about how each childqualified for inclusion in the SLI group. In addition tothe previously described measures, the Peabody PictureVocabulary Test—III (PPVT–III; Dunn & Dunn, 1997)and the EVT (Williams, 1997) were administered to doc-ument the children’s single-word vocabulary skills, anda 100-utterance language sample was collected over the

course of the diagnostic and experimental visits witheach child for use in MLU matching (see Table 2).

All TD children hadnohistory of speech, language, orreading therapy and scored at or above the typical rangeon all of the speech and language measures. The MLUgroupwas not administered the TNL orNWR subtest be-cause normative data are not available for children underage 5;0 for these assessments. Table 2 shows the meanscores obtained by each group of children on each of thesemeasures, along with other demographic information. Re-sults of t tests showed that the children with SLI did notdiffer from the AGE group in age in months, t(26) = 0.04,p = .97, or from the MLU group on MLU in words, t(24) =0.46, p = .65, on raw EVT score, t(24) = 0.42, p = .68, or onpercent correct on theSPELT, t(24)=0.99,p= .33, using cri-teria recommended byMervis andKlein-Tasman (2004).

All children met the conventional criteria for partici-pation in a study on SLI. Each child passed a hearingscreening at 25 dBHL for each ear at 500, 1000, 2000, and4000 Hz, and obtained a standard score above 83 on theMatrices subtest of theKaufmanBrief IntelligenceTest—2(KBIT–2; Kaufman & Kaufman, 2004). According to par-ent report, no child had a history of frank neurological im-pairment or a previous diagnosis of autism or pervasivedevelopmental disorder. To ensure that utterances as longas the experimental targets werewithin the capabilities of

Table 1. Means of qualifying for inclusion in the specific languageimpairment (SLI) group.

SubID Qualifying criteria Tx w/in 12 mos SPELT–II TNL NWR

SLI 1 4 y F F FSLI 2 3 y F F PSLI 3 3 y F F PSLI 4 3 y F F PSLI 5 3 y F F PSLI 6 3 y F P FSLI 7 3 y P F FSLI 8 2 y F P PSLI 9 2 y F P PSLI 10 2 y P P FSLI 11 2 y P P FSLI 12 2 n F F PSLI 13 2 n F F PSLI 14 2 n F F P

Note. Children are sorted according to the number of qualifying criteriathat they met (as listed in article text). The table indicates whether childrenwere (y) or were not (n) enrolled in therapy within the last 12 monthsand whether the children passed (P) or failed (F) each of the followingthree qualifying tests: the Structured Photographic Expressive LanguageTest—II (SPELT–II; Werner & Kresheck, 1974); the Test of NarrativeLanguage (TNL; Gillam & Pearson, 2004); or the NEPSY (NonwordRepetition [NWR]; Korkman, Kirk, & Kemp, 1998). Tx = treatment;SubID = subject identification.



all participants, the mean length of the five longest utter-ances in a representative language sample (MLU5) wasequal to or greater than eight words for each individual.

StimuliThree different types of sentences were chosen as

stimuli: (a) two coordinated clauses (e.g., Ernie hoppedand Elmo kicked the ball); (b) a main clause and a tem-poral clause (e.g.,The alienswhistledwhenMinnie kickedthe ball); and (c) a main clause and a complement clause(e.g.,Ratty guessed that Elmo kicked the ball). These sen-tence types were chosen because they form a continuumfrom completely optional and independently plannedclauses (coordinated clauses) to clauses in which oneserves as an obligatory argument to the other and arepresumed to be planned together by adult speakers (com-plement clauses). Temporal adverbial clauses usingwhengenerally provide information about a foregrounded eventset against a background event (e.g., she carries an um-brella when it rains) or about two events that are cau-sally related (e.g., she fell down when he tripped her).Thus, the events being described are more related thanthe coordinated clauses, but the temporal clause is notobligatory as with the complement clause structures.Thirty-six items were constructed for each sentence type.

So that all three conditions were as similar as pos-sible, all of the sentences consisted of amain clausemade

up of a subject and a verb followed by another finiteclause. The first clause of the coordinate and temporalconditions always involved a regular intransitive verb.The 36 verbs used in the first clauses of the coordinateand temporal conditions were identical.1 Because thesame verbs could not be used in the first clause of thecomplement condition, the first clause of this conditionused 10 mental verbs (guess, think, believe) or commu-nication verbs (answer, say) found in the vocabulary offirst-grade children (Moe, Hopkins, & Rush, 1982). Theseverbs also all occur in the input to children found in theChild Language Data Exchange System (CHILDES;MacWhinney, 2000) database, with amedian frequencyof approximately 541 instances of the stem and inflectedforms (range goes from 18 instances of discover to 16,208instances of know). The verbs know and guess occurredsix times; all other main mental/communication verbswere used three times each. Three of these verbs wereirregular, and seven were regular.

The second clause for all three sentence types wassystematically varied tomanipulate the number of argu-ments present as a second source of complexity withinthe sentence. One third of the sentences for each sen-tence type included transitive verbs (12 items), and two-thirds included intransitive verbs (24 items). Of theintransitive verbs, half for each sentence type were elic-ited in scenarios designed to elicit an intransitive verbplus an adverb or prepositional phrase (e.g., jumpedquickly, climbed over the slide; 12 items), and half werein scenarios designed to elicit an intransitive verb alone(12 items). This served as a control to determinewhethersentence length or transitivity was the primary source ofcomplexity. Half of the second clauses for each type ofsentence used a regular past tense verb (18 items, six pertransitivitymanipulation), and half of each sentence typeused an irregular past tense verb as the target verb. Reg-ular and irregular verbs and verb transitivity typesweredistributed evenly throughout the list of each sentencetype. Table 3 contains sample sentences, and the Appen-dix contains a complete listing of verbs.

ProcedureFor each sentence type, children watched a short

elicitation scenario, and then a question prompt wasposed. Complete examples for each sentence type can befound in the Appendix. The experimenter provided thefirst word of the first clause and encouraged the child tocomplete the rest of the sentence. To reduce memory de-mands, toys and props used in the elicitation scenarios

1There is one exception to this statement: Following piloting, the item gigglewas changed to laugh because children tended to switch these verbs. Thischange did not occur in the coordinate condition, and thus, items wereelicited using the verb giggle.

Table 2. Demographic information about participants.

Demographicinformation AGE MLU SLI

N 13 11 14# in therapy 0 0 11

Age 6;7 (10.6 mos)a 4;4 (3 mos) 6;7 (11.1 mos)

MLUwb 5.42 (1.41) 5.19 (.55) 5.39 (1.16)MLU5b 11.47 (1.35) 11.56 (.89) 10.29 (1.61)FVMCb 97.2% (2.71) 95.6% (6.34) 88.1% (17.01)

SPELT–IIc 109.8 (7.8) 117.3 (12.2) 79.0 (10.9)TNL Indexc 109.5 (16.3) — 83.3 (14.9)NEPSY NWRd 12.3 (2.6) — 7.3 (1.8)KBIT–2c 99.6 (9.37) 106.2 (8.08) 97.4 (11.97)

Note. The total number of children included from each group is listed,aswell as the number of children enrolled in therapy.MLUw=mean length ofutterance in words; MLU5 = length in words of the five longest utterances foreach individual child; FVMC = Finite Verb Morphology Composite (whichis the percent correct in obligatory contexts of am, is, are, regular past –ed,and third-person singular –s); KBIT–2 = Kaufman Brief Intelligence Test2nd ed.;Matrices subtest). Dashes indicate that the taskwas not administered.aM (and SD) of age is shown in years;months. bComputed fromspontaneous language samples taken at the time of the experiment.cReported in standard scores with a mean of 100 and an SD of 15.dM = 10; SD = 3.



remained in view of the children while they answered.If a child did not respond or provided an incompleteresponse (e.g., a simple sentence, only the second clause,or only a noun), the examiner first prompted by askingthe child to say the whole sentence. If a child subse-quently produced an adequate response, this responsewas scored. If the child still did not produce a fullresponse, the examiner provided the first clause of thesentence and asked the child to complete the second clauseto encourage the child’s participation in the experiment.These truncated responses were elicited and coded butare not included in the responses reported in the sectionsthat follow because I could not be sure of the child’srepresentation of the sentence as one or two clauses.

Presentation was blocked by sentence type to en-courage children to use the target sentence types. Theorder of presentation of the blocks was counterbalanced.In general, all items from a block were administered inone visit; however, blocks could be broken in half toaccommodate fatigue or inattention on the part of the

child. If a blockwas divided, the child’s participationwasextended by one visit so that two different sentencetypes were not administered on the same day.

Response coding.The experimentersmade awrittenrecord of the child’s responses during the task. All re-sponses were also audiorecorded and transcribed, usingthe onlinewritten record as a guide. Responseswere codedfor the number of clauses in the response; type of mor-phological marking on the verb (simple past, past pro-gressive, omitted, other); accuracyof thepast tensemarkingin both clauses (correct, incorrect, over-regularized); andinclusion of the verb complement (transitive verbs) ormodifiers (intransitive verbs) in the second clause. Chil-dren’s responses sometimes differed from the intendedtarget, so responses are reported in terms of percentaccuracy based on the child’s actual response, not the in-tended responses. The response rate for two-clause sen-tences of each type is reported in Table 4 and is discussedin the first portion of the Results section. Two-clauseresponseswere defined as those inwhich a child produced

Table 3. Sample sentences from each condition.

Verb type Coordinate Temporal Complement

Intransitive Reg Lion growled and Ron coughed. The aliens smiled when Froggy hummed. Ratty remembered (that) Pumbacoughed.

Irreg Piglet listened and Frog hid. The aliens snored while Froggy hid. Ratty thought (that) Pumba hid.

Intransitive + lengthener Reg Lion counted and Ron dancedwith Pooh.

The aliens walked when/while Minniejumped really fast.

Ratty knew (that) Marty crawledunder the table.

Irreg Eyeore colored and Raccoonran to the table.

The aliens listened when Froggy sangloudly.

Ratty remembered (that) Snailsang quietly.

Transitive Reg Raccoon baked and Eyeoresquished a worm.

The aliens wiggled when Bear closedthe door.

Ratty said (that) Alex scared hisfriends.

Irreg Piglet paddled and Frog atea cookie.

The aliens sailed when Magenta brokethe chair.

Ratty imagined (that) Simba bitthe girl.

Note. Reg = regular; Irreg = irregular.

Table 4. The mean number of scorable (two-clause) responses produced by each group of children for each conditon.

Group Target

Coordinate condition Temporal condition Complement condition

Intrans Trans Total Intrans Trans Total Intrans Trans Total

24 12 36 24 12 36 24 12 36

AGE M 23.77 11.85 35.62 21.54 11.08 31.77 22.23 9.38 31.62SD 0.44 0.38 0.65 6.21 0.67 9.33 3.37 1.66 4.87

MLU M 23.36 11.64 35.00 22.00 10.55 32.55 14.82 7.40 21.55SD 0.67 0.50 1.10 3.03 1.81 4.80 8.78 3.50 12.44

SLI M 20.93 10.50 31.43 21.36 10.29 31.64 18.79 8.85 27.00SD 3.75 2.10 5.68 2.73 1.59 4.18 5.99 1.21 8.53

Note. Note that there are twice as many intransitive (Intrans) responses expected as transitive (Trans) responses, as half of the intransitive responses wereelicited in scenarios that biased the use of a prepositional phrase or adverb.



a verb in both the first and second clause and used a dis-tinct subject in the second clause.

Reliability. Three transcripts from each group wereretranscribed by an independent coder, and transcrip-tion reliability was computed for word-level accuracy(M = 90%; range = 75%–100%) and for tense marking(M = 93%; range = 81%–100%). These same transcriptswere recoded by a third independent coder on all of therelevant dependent variables. Reliability for coding wasapproximately 97% for all relevant variables (e.g., num-ber of clauses in response, tense coding for first and sec-ond clause, use of a verb complement or modifier) withthe exception of determining the number of clauses pro-duced in the complement clause condition, for which re-liability was 80% (range = 24%–100%). Discussion withthe coders demonstrated that misapplication of the cri-teria for two-clause responses in the complement clauseconditionwas responsible for themain discrepancies thatweredriving thepoor reliability.Tworesponse types shouldhavebeen scoredas one-clause responses (and, thus,wereexcluded from future analysis): (a) responses in whichthe examiner prompted the child with the first verb butthe child provided all other information (e.g., Examinersays, “Ratty guessedI”, and child responds, “I thatSimbawore a hat”), and (b) responses in which the childresponded only with the name of one of the characters(e.g., Ratty guessed Simba). In addition, one responsetype should have been scored as two clauses (e.g., Rattyguessed Zebra playing). All 1,368 responses to the com-plement clause condition were recoded using consensusscoring, and disagreements for all except three itemswere satisfactorily resolved. These three items were ex-cluded from subsequent analyses. Response coding wasrechecked for accurate application of the rules when datawere entered into a spreadsheet for further analysis.

Data Analysis: Rationale for MultilevelLogistic Regression

Multilevel logistic regression was used for exploringthe data. This method offers several advantages overanalysis of variance (ANOVA) in terms of maximizingdata retention and understanding sources of variance(Baayen, Davidson, & Bates, 2008; Dixon, 2008; Jaeger,2008; Quene& van den Bergh, 2008). Specifically, the ap-proach allowed us to include each child’s response to eachitem as a separate data point and avoid setting a mini-mum number of responses per child in each condition(Quene & van den Bergh, 2008). Unlike a mixed model(repeated measures) ANOVA, a completely balanced de-sign is not necessary, and the model accounts for unbal-anced designs by treating instances where there are fewerobservations as less reliable (Quené & van den Bergh,2008). A third advantage to this approach is that item andsubject effects are modeled simultaneously. Uncontrolled

factors such as verb frequency, the phonological composi-tion of the verb, and the likelihood of an individual sub-ject producing a correct item are thus taken into accountwithin the model (Jaeger, 2008). Finally, by modeling theresults logistically, the model treats changes between ex-treme points as more important than changes near thosepoints (Jaeger, 2008), which is expected if the underly-ing process is a probabilistic one. This is more appropri-ate than computing a pseudocontinuous result (percentaccuracy) from noncontinuous underlying data and at-tempting to apply statistics appropriate to continuousmeasures.2

ResultsAnalysis 1: Scorable Responsesand Data Retention

Approximately 15% of the responses were unana-lyzable for one reason or another, with production of one-clause responses being the primary source of exclusion.3

With this inmind, I first examined the data to determine ifall three groups of children were equally likely to producetwo-clause (scorable) responses for the three sentencetypes under consideration. Of a potential 4,104 possibleresponses, 3,994 items were available for this analysis(3 Conditions × 36 Items × 38 Subjects). Approximately3% of the responses were excluded from all analysesbecause the number of clauses could not be determinedbecause of unintelligible, off-topic, or otherwise unana-lyzable responses on the part of the child. The averagenumber of responses produced by children for each groupin each condition is shown in Table 4. A mixed modelregression fitted to a binomial logistic model was com-pleted, with participant, first-clause verb, and second-clause verb treated as untested random factors, andsentence type (Coordinate, Temporal, Complement) andgroup (AGE, SLI, MLU) tested as fixed factors. TheAGE group and coordinate clause types were set as thereference groups because these allow for the most log-ical “baseline” comparisons.

The model coefficients are shown in Table 5. Table 6shows the predicted probability and predicted odds ofproducing the target response for each group and condi-tion. If I interpret the results for each individual factor,I can see that there are disparities across groups and

2For readers familiar with hierarchical linear modeling (HLM), both HLMand multilevel regression are based on similar underlying mathematicaltechniques. Multilevel logistic regression is one way of representing theseconcepts, which are common to a variety of analysis approaches.3It is worth noting that the data retention in this study is comparable toother studies eliciting complex sentences in children from ages 3 to 9 years(Donaldson, Reid, & Murray, 2007; Eisenberg, 2004; Marinellie, 2004).Depending on the elicitation context, the age of the children, and the targetforms, experimenters tend not to elicit target two-clause responses for aroundhalf of the itemspresented to the children.Thus, the results that I observed in thecomplement clause conditionarenotunusual orunexpected for this areaofwork.



conditions. For instance, the AGE group is more likelythan both the MLU and SLI groups to produce a two-clause response in the coordinate and temporal condi-tions, but all groups are highly likely to produce thetargets. The complement clause condition seems to bewhere the majority of the nontarget responses occurred,and there weremany fewer target responses for both theSLI- andMLU-matched groups than for the AGE group.

Similar concerns arise when I consider the produc-tion of particular verb argument structures within eachsentence type. As can be seen in Table 7, only six MLUchildren produced five or more scorable utterances inthe transitive/complement condition. Similar difficulties

were observed in the intransitive + lengthener condi-tion. Thus, longer utterance targets were less likely toresult in scorable responses, particularly for the twogroups with lower language abilities.

The discrepancies in rate of production across groupsraise questions about the selection of TD control groups.Recall that age-matched children are often included fortwo reasons. First, age-matched children provide a clin-ical reference sample. Children are generally judged to beimpaired in a clinical setting relative to their age levelrather than relative to some other scale such as MLU orvocabulary size. Second, age-matched groups are includedto ensure that the task is appropriate for children of thisage. Clearly, TD 6- and 7-year-olds can complete thistask appropriately. TheMLU groupwas included to ruleout utterance-length restrictions and limitations due togeneral expressive vocabulary skills as reasons for anydifficulties observed in the SLI group. Given the difficultythat both the younger children and childrenwith SLI hadwith the complement clause condition, one cannot com-pletely rule this out. However, all three groups of childrenwere capable of producing the temporal and coordinateresponses, whichwere comparable in length and vocabu-lary frequency to the complement clause condition.

These results are not surprising: I expected the age-matched group to be generallymore proficient than eitherof the two comparison groups and the coordinate clausecondition to be the easiest condition. However, the im-plication for future analyses is that there are more datapoints from the age-matched group, which should lead tomore reliable parameter estimates in the final model

Table 5. Results of the regression model showing the likelihood of producing two-clause responses.

Factor Variance SD p Coefficient SE

Random factorsItem 0.97 0.98Subject 2.44 1.56

Fixed factorsIntercept < .0001 7.97 1.214Diagnostic group (reference

category = Age)SLI < .0001 –4.01 1.301MLU < .0001 –1.80 1.400Sentence type (reference

category = Coordinate)Temporal < .0001 –4.76 1.143Complement < .0001 –4.88 1.144Diagnostic Group × Sentence

Type (reference category = Age, Coordinate)SLI × Temporal < .001 3.87 1.157SLI × Complement .034 2.46 1.163MLU × Temporal .018 2.96 1.254MLU × Complement .617 –0.62 1.251

Note. Age and coordinate conditions are set as reference groups. SE = standard error.

Table 6. Predicted probabilities and odds of producing a two-clauseresponse, computed from the model shown in Table 5.

Predicted probabilities Predicted odds

Condition AGE SLI MLU AGE SLI MLU

Coordinate 0.999 0.981 0.998 2905.61 52.51 480.10Temporal 0.961 0.956 0.988 24.92 21.61 79.09Complement 0.957 0.823 0.662 22.13 4.66 1.95

Note. Odds close to 1 indicate that there was an equal likelihood ofproducing or not producing a two-clause response, and, thus, high oddsreflect a high likelihood of producing the targets. As the probabilitiesapproach 0 or 1, the odds become much more extreme such that a smallchange in the probability of producing a two-clause response near theextreme points of the scale leads to greater changes in odds than a changeof similar magnitude in the middle of the scale.



for this group. This is especially true in the complementcondition, which is the condition inwhich I see the great-est discrepancies across groups, and this should make itmore difficult to find differences between the groups inthis condition. Future analyses should be interpretedwiththese caveats in mind.

Analysis 2: Effects of Linguistic ComplexitySentence type and clause location. I restricted anal-

yses to the accuracy of past tense use for two-clauseresponses only. The production of any overt form of past

tense—including the use of past auxiliaries within pastprogressive responses and over-regularized forms of thesimple past—was the outcome variable. This method ofscoring is consistent with the idea that childrenwith SLIhave particular difficulty with finiteness (Rice, Wexler,Marquis, & Herschberger, 2000).

To test the role of sentence complexity and clauselocation on past tense production, a binomial logisticregression was fit, with Subject (38 individuals), First-ClauseVerb (47 types),4 andSecond-ClauseVerb (36 types)as randomfactorsandDiagnosticGroup (SLI,MLU,AGE),Clause Location (first, second), and Sentence Type (co-ordinate, temporal, and complement) as fixed factors.Percent correct use and standard deviations are shownin Table 8. Because of the particular hypotheses aboutthe accumulation of difficulty with increasing complex-ity andwith serial order, all three-way interactions wereincluded in themodel, allowing us to assess the way thatdiagnostic group, sentence type, and clause location af-fected each other.The log-likelihoodvaluesassociatedwiththe model employing three-way interactions was com-pared to simpler models using a c2 test for goodness offit, and it was found that the full model was a better fitfor the data than any of the alternative reduced models(ps < .0001). This suggests that themore complexmodelis justified and that the three groups of children respondeddifferently to the combination of clause location and sen-tence type. The model is shown in Table 9, with age andcoordinate clause as the reference group. The predictedprobabilities and odds are available in Table 10.

Visual presentation of the model, as shown in Fig-ure 1, eases the interpretation of the results, given thatcomparison across three groups and three conditionsis mathematically complex. Typically, for models withmultiple categorical predictors, the model is rerun, ref-erence variable(s) are reshifted, and then the model isreinterpreted. Note that the underlying model does notchange, but direct comparisons between groups becomemore readily observable (Jaccard, 2001). All three groupswere highly likely to produce past tense in the first clausesof the coordinate and complement conditions. The SLIgroup was less likely to produce a past tense marker inthe first clauses than the age- or MLU-matched groups(Coordinate SLI/AGE odds ratio [OR] = 0.56; MLU/SLIOR = 0.29; Complement SLI/AGE OR = 0.51, SLI/MLUOR = 0.12). There appear to be differences between theAGE and MLU groups if one inspects the ORs (Coordi-nateAGE/MLUOR=0.52;ComplementAGE/MLUOR=0.23). However, these differences were not meaningfulin terms of the model, which relies on log odds to more

4Laugh was used as a verb in the first clause in the coordinate list, andgiggle was used in the temporal list, leading to one additional main verb.This discrepancy did not affect the complement list because it used onlymental and communication verbs in the first clause.

Table 7. Mean percent correct use and SDs for the use of past tensefor short and long intransitive and transitive responses.

Group

Pasttenseuse

Coordinate

Intransitive Intransitive + lengthener Transitive

AGE Avg 92.16% 81.25% 87.82%SD 10.85% 27.59% 23.57%N 11 13 13

MLU Avg 95.10% 86.06% 87.45%SD 6.59% 13.72% 14.30%N 11 11 11

SLI Avg 64.98% 69.56% 69.48%SD 27.74% 27.91% 31.35%N 13 13 13

Group

Complement


AGE Avg 83.51% 80.65% 76.77%SD 22.10% 14.21% 21.23%N 13 13 13

MLU Avg 41.44% 45.30% 38.56%SD 32.30% 35.66% 29.43%N 8 7 6

SLI Avg 51.44% 64.08% 61.98%SD 31.59% 31.13% 32.04%N 12 11 13

Group

Temporal


AGE Avg 93.97% 97.28% 93.97%SD 6.16% 4.76% 9.50%N 12 11 11

MLU Avg 82.24% 83.45% 82.32%SD 27.84% 23.62% 22.67%N 11 11 11

SLI Avg 75.64% 67.84% 66.09%SD 27.09% 33.45% 31.38%N 14 13 13

Note. Only children who produced five or more scorable responses areincluded in each average (Avg). The number of children (N ) contributingto the average is listed below each condition.



accurately capture the dichotomous nature of the out-come variable (Coordinate, p = .21; Complement, p = .24).

In the coordinate condition, the two TD groupsmain-tained a comparable level of accuracy in the second clauseas compared with the first (AGE OR = 1.1; MLU OR =1.4), but the SLI group showed a significant decline (OR=1.9) in accuracy for the same comparison. In the com-plement clause condition, all three groups showed a de-cline in the likelihood of producing a past tense marker inthe second clause of the complement clause condition.This drop was especially pronounced for the MLU group(OR = 80.26), with this being the condition in which chil-dren were least likely to produce a past tense markeracross all conditions and groups. The SLI group demon-strated a noticeable decline from the first to the secondclause (OR = 3.68). Although it attained significance, thedropwasmuch smaller for theAGEgroup (OR=1.38) forthe same comparison.

The temporal condition generally showed the oppo-site pattern as that of the other two conditions in that thefirst clause was less accurate than the second clause forall three groups.Whereas groupdifferencesweremost ob-servable in the second clause for the coordinate and com-plement clause conditions, group differences were mostpronounced in the first clause of the temporal condition.This is reflected in the Group × Sentence Type interac-tion in themodel,with theSLI group being equally likelyto produce or not produce a past tense marker. This canbe compared to the MLU and AGE groups, who were 2and 7.5 times as likely to produce a past tense markerthan not, respectively.

Effects of transitivity and length. I also consideredwhether the number of arguments in the second clauseinfluenced the likelihood of producing an overt pasttense form. Table 7 reports the accuracy of children’sproductions in each condition. A binomial logistic regres-sion was fit, with Subject (38 individuals), Second-ClauseVerb (36 types), and Sentence Type (3 types) as randomfactors and Diagnostic Group (SLI, MLU, AGE) and

Transitivity (Transitive, Intransitive-Short, Intransitive-Long) as fixed factors. This model did not predict theproduction of past tensemore accurately than the modelthat only includeddiagnostic groupalone,c2(6,N=3,365)=6.9156, p = .32.5 These results suggest that the numberof arguments in the second clause did not influence pasttense production in a measurable way.

DiscussionI was interested in using the likelihood of producing

past tensemorphemes as ameans of determiningwhetherchildren were vulnerable to increased processing loadin sentence production from a variety of sources at theverb, clause, and sentence levels. Grounding our work ina processing capacity perspective on grammatical er-rors, I proposed that (a) sentences that were complexwould bemore difficult than compound sentences, (b) laterclauses would bemore difficult than earlier clauses withineach sentence, and (c) an increased number of argumentswould also add difficulty to the sentences. I observed thatcomplex sentences did have more errors than compoundsentences. Our predictions with regard to clause orderheld for the complement clause sentences, in which thefirst clause was more accurate than the second clause,but were not supported by the temporal adverbial sen-tences, a point to which I return in the paragraphs thatfollow. The predictions with regard to argument struc-ture influences were not supported in any of the threegroups.

I also predicted that childrenwithSLIwould bemorevulnerable to the accumulation of difficult elementsthan either of the two groups of TD children. This predic-tion was not directly supported. Although the childrenwith SLI did start lower than the TD children on all

Table 8. Average percent correct use of past tense, for group, sentence type, and clause location.

Group

Coordinate Temporal Complement

1st clause 2nd clause Total 1st clause 2nd clause Total 1st clause 2nd clause Total

AGE Avg 85.95 84.83 85.39 84.62 95.19 89.91 88.78 80.09 84.76SD 25.26 24.50 24.75 16.00 5.50 9.05 16.44 17.20 12.46

MLU Avg 91.57 89.21 90.38 67.40 81.48 75.20 97.02 43.40 73.14SD 6.55 9.10 6.78 27.74 24.16 25.96 4.04 26.29 15.54

SLI Avg 78.03 68.29 73.16 48.60 70.65 59.82 78.91 55.65 67.98SD 27.13 27.61 24.91 33.13 27.83 23.40 29.26 32.71 25.15

Note. Correct use included the use of simple past, over-regularizations, and past progressive forms. The mean number of clausesproduced by individual children from each group for each condition can be found in the “Total” columns in Table 4.

5Similar results were obtained when sentence type was treated as a fixedfactor, and the full model, including transitivity/ length, was compared toa reduced model that included Diagnostic Group and Sentence Type as fixedfactors, c2(6, N = 3,365) = 5.3707, p = .49.



three sentence types, their pattern of performance lookedmore like a depressed version of that of their age-matched peers rather than a unique response profilethat demonstrated more rapid declines in accuracy aslinguistic complexity increased. In contrast, the MLU-matched group showed a pattern of performance on thecomplement clause structures that was more consis-tent with accumulating complexity. Taken together,our results suggest that morpheme production is indeedaffected by complexity on at least two levels—sentence

type and clause position—in all three groups of children,albeit to differing degrees. This adds to a body of evidenceshowing that children with SLI, similar to their peers,are sensitive to factors that increase processing demands,ranging from low phonotactic probability to increasedsyntactic complexity. The fact that the MLU and SLIgroups have different response profiles despite being well-matched reinforces the observation made byMervis andKlein-Tasman (2004) that children may obtain similarresults on standardized tests via different means and

Table 9. Regression model for the use of an overt past tense marker.

Factor Variance SD Coefficient STD error p

Random factorsMain verb 0.059 0.242Subordinate verb 0.081 0.285Subject 1.241 1.114

Fixed factorsIntercept 2.193 0.350 < .0001

Diagnostic group (referencecategory = Age)

SLI –0.588 0.474 .215MLU 0.652 0.523 .213

Sentence type (reference category =Coordinate)

Temporal –0.188 0.202 .351Complement 0.285 0.233 .222

Clause location (reference category =First clause)

Second clause –0.101 0.200 .614

Diagnostic Group × Sentence Type(reference category = Age, Coordinate)

SLI × Temporal –1.415 0.264 < .0001SLI × Complement –0.086 0.294 .770MLU × Temporal –1.801 0.313 < .0001MLU × Complement 0.809 0.499 .105

Diagnostic Group × Clause Location(reference category = Age, First)

SLI × Second –0.531 0.265 .045MLU × Second –0.251 0.334 .453

Clause Location × Sentence Type(reference category = First Clause, Coordinate)

Second Clause × Temporal 1.547 0.335 < .0001Second Clause × Complement –0.605 0.289 .036

Diagnostic Group × Sentence Type ×Clause Location (reference category = Age, Coordinate, First)

SLI × Temporal × Second 0.167 0.410 < .0001SLI × Complement × Second –0.063 0.388 .036MLU × Temporal × Second –0.254 0.472 < .0001MLU × Complement × Second –3.422 0.586 .036

Note. This table displays the final model, with interactions between group, condition, and clause locationincluded. Age and coordinate clause are set as the reference group.



may be using different processes in their responses tothe linguistic variables under examination. It also raisesquestions about the interaction between linguistic knowl-edge and processing capacity and how these interactionschange over developmental time.

Study LimitationsBefore I discuss each of these predictions in greater

detail, three limitations of the study should be considered.

First the number of responses varied considerably acrossgroups and sentence types. Children with SLI and theirMLU-matched peers weremuch less likely than the age-matched children to produce two-clause responses in thecomplement clause condition. Although the rate of pro-duction is consistent with previouswork (Donaldson et al.,2007; Marinellie, 2004) and I chose a method of analysisthat accommodates such variability, this means that theconfidence intervals surrounding the estimates in the re-gression model are likely to be larger—and under those

Table 10. Predicted probabilities and odds for the production of an overt past tense marker.

Predicted probabilities Predicted odds

Clause 1 Clause 2 Clause 1 Clause 2

Condition SLI AGE MLU SLI AGE MLU SLI AGE MLU SLI AGE MLU

Coordinate 0.833 0.900 0.945 0.726 0.890 0.924 4.98 8.96 17.17 2.65 8.11 12.1Temporal 0.500 0.881 0.701 0.747 0.969 0.858 1.00 7.42 2.35 2.96 31.53 6.03Complement 0.859 0.923 0.981 0.623 0.896 0.390 6.07 11.91 51.37 1.65 8.63 0.64

Note. The information shown here corresponds to the model in Table 9. Odds close to 1.0 indicate that marking past was as likely as not markingpast. Odds ratios can serve as the effect size measure for comparing results under different conditions for naturally dichotomous variables (Chinn, 2000).When the two events differ only in one way, the computation is simple: The odds of one event occuring under certain conditions (e.g., Clause 1 of theCoordinate condition by children with SLI) are divided by the odds of the event occuring under other conditions (e.g., Clause 2 of the Coordinate conditionby children with SLI): 4.98/2.65 = 1.87. In other words, children with SLI are nearly twice as likely to produce past tense in Clause 1 as in Clause 2. TheMLU-matched group is about 1.5 times more likely to produce a past tense form in Clause 1 of the coordinate condition than in Clause 2 (17.17/12.1 = 1.41),and the age-matched group is nearly equally likely to produce a past tense form under each condition (8.96/8.11 = 1.10).When the comparison is made acrossmore than one variable (e.g., both group and condition change), then the computation becomes more mathematically complex (see Jaccard, 2001).

Figure 1. The likelihood of producing a past tense form as predicted by the model shown in Table 9. The panel on the left shows predictedprobabilities for the first clause of each condition, and the panel on the right shows predicted probabilities for the second clause. MLU = youngertypically developing (TD) children matched to SLI group on mean length of utterance; AGE = TD children matched to children with specificlanguage impairment (SLI) on age; SLI = children with specific language impairment.



circumstances, it may be difficult to observe differencesacross groups/conditions. Nonetheless, themagnitude ofthe observed differences was sufficiently large that sig-nificant differences were obtained. Thus, one can con-sider the estimates to be conservative estimates of whatonewould see had there beenmore observations contrib-uting to the regression model.

The second caveat is a similar onewith regard to theproduction of transitive verbs in the complement condi-tion. Again, there was great variability associated withthe likelihood that children would produce the targets,and this could have influenced our ability to observe dif-ferences, particularlywith regard to thepredictions aboutthe accumulation of difficulty. Unlike the sentence-levelcomparisons, one cannot simply dismiss this variabilityby pointing to observed differences because transitivityhad no effect on the likelihood of producing past tense.However, two considerations may allow us to infer thatthis is a real “null” finding rather than an artifact re-sulting from too few observations: First, if I remove thecomplement clause condition and only examine the tem-poral and coordinate conditions in which children morestably produced the targets across all three elicitationconditions, I see that verb transitivity still does not as-sist in predicting past tense production over and abovegroup membership, c2(6, N = 3,365) = 4.326, p = .63,suggesting that the lack of differences does not stemwholly from an unbalanced design. Second, and perhapsmore convincingly, in their work with younger children,Grela and Leonard (2000) found significant differencesonly between intransitive and ditransitive conditions.Possibly, the difference between intransitives and tran-sitives is not sufficiently large to allow the observation ofprocessing load changes in an offline production task. Inwork on French with children of a similar age as thosein our study, researchers indeed have found differencesbetween transitive and intransitive verbs in the area ofarticle agreement, but only for sentences with adjectivesadded in to increase length (Pizzioli & Schelstraete, 2008).More fine-grained measures using reaction time or eyetracking as a component of the production task mightdemonstrate differences that are not observable here(see Lee & Thompson, 2008, on the use of eye trackingin production tasks with adults with agrammatism) inmuch the sameway that eye tracking demonstrates dif-ferences in comprehension that are not observable viapointing only (Eberhard, Spivey-Knowlton, Sedivy, &Tanenhaus, 1995).

Finally, it is possible that the children’s familiar-ity with the vocabulary items may have influenced thelikelihood of childrenproducing the target response typesor of using an overt past tensemarker in their responses,particularly for the younger MLU-matched group. Thetarget verbs were selected from the book The Vocabularyof First-Grade Children (Moe et al., 1982); TD first-graders

are, on average, 2 years older than the children in theMLU group andmay havemore robust vocabulary skillsthan the SLI group. Indeed, I cannot rule out limitedexperience with some of the verbs considering I did notdirectly assess these children’s familiarity with the tar-get verbs.However, I canexaminehowverbswerematchedacross conditions and the frequency of the verbs chosen.Recall that the verbs used in the first clauses of thecoordinate and temporal conditions were identical. Thus,any differences across conditions due to verb familiarityalone should be reflected in both conditions. Instead, Iobserved that the productions in the first clause of thetemporal condition were less accurate than those ofthe coordinate and complement conditions. Likewise, theverbs employed in the second clauses for all three condi-tions were identical, and thus differences across condi-tions should not be attributable to this factor. The oneplace where the conditions were not matched on verbswas in the first clause of the complement clause condi-tion. Because of the need for complement-taking verbs,the 10 verbs used here were different than the otherfirst-clause verbs and were generally more frequent thanthe verbs used in the other two conditions, t(46) = 1.95,p = .06. However, higher frequency verbs in the mainclause of the complement clause condition work againstthe predicted and observed findings that the comple-ment clause condition would be the most difficult sen-tence type. Thus, this does not seem to be an area ofconcern with regard to the observed results.

Predictions About Sentence Typeand Clause Location

In general, our predictions with regard to sentencetype and clause location held for the complement andcoordinate conditions. Children had greater difficultywith past tense forms in the complement clause conditionthan in the coordinate condition. This was primarily re-lated to differences of accuracy in the second clauses: Thefirst and second clause of the coordinate condition andthe first clause of the complement condition were rela-tively comparable, whereas the second clause of the com-plement condition was significantly worse. This supportsthe idea put forth by Holmes (1988) that the complementclause is planned simultaneously with the main clauseof a sentence, whereas the coordinate clause is plannedsequentially, extending the findings from adults to chil-dren. Sequential planning would lead to fewer words/smaller units being retrieved and held in working mem-ory and, thus, should reduce errors associatedwith select-ing the appropriatemorphemes. It is worth rememberingthat the complement and coordinate sentence types hadthe same number of nouns, verbs, and inflections, so itis not the total content that is influencing accuracy but,



rather, the structure of the sentence and the way thosenouns and verbs are related to each other.

Early acquired forms, such as coordinated clauses,were produced with great efficiency by all three groups ofchildren. It is not clear if this is because they are wellpracticed or because they are less complex (or both).Nonetheless, this increased efficiency makes it difficultto observe changes in processing demands associatedwith sentence production via accuracy measures in sim-pler sentence types because children may demonstrateceiling-level performance very early. Although comple-ment clauses occur relatively early in children’s speechas frozen forms or unanalyzed discourse markers, dataon complement clauses analyzed within a constructiongrammar framework suggest that these structures areactually acquired much later (Diessel, 2004; Kidd, Lieven,& Tomasello, 2006). These later acquired forms appear tomake sufficient demands that morphological errors areindeed observable and, thus, may be a tool that can beused to examine morphological production in older chil-dren when online tools are not available.

The temporal clause productions did not follow myinitial predictions, in that the first clause was less accu-rate than the first clauses of the other conditions andwas less accurate than the second clause of the temporalcondition. Perhaps it is more common to start sentenceswith the temporal adverbial clause (e.g., when Erniejumped, Elmo laughed) than the opposite clause location(e.g., Elmo laughed when Ernie jumped) and, thus, chil-dren had difficulty due to the unusual clause order. Acorpus analysis examining adult use of British Englishfound that the order of main and adverbial clauseswas influenced by the temporal order of the events(simultaneous/sequential) and the nature of the adver-bial clause (temporal/causal) along with the length ofthe adverbial clause (Diessel, 2008). An elicitation studythat varied the timing of events and the semantics of theword when would be required to confirm this explana-tion. A second possibility is that temporal adverbialclauses function like the addition of adverbs to simplepast sentences. Krantz andLeonard (2007) demonstratedthat children with SLI and their MLU-matched peerswere less likely to mark tense when a time adverb waspresent in the sentence. A similar process in which chil-dren perceive that tense/time has already been encodedmay also be occurring here. In our case, the entire adver-bial clause may be taking the place of words such as justand already and influencing the likelihood of markingtense in the sentence. Elicitation of sentences with bothtemporal and nontemporal adverbial clauses would pro-vide relevant evidence and allow us to disambiguatefrequentist-based accounts and accounts related to theencoding of temporal information. Given that adver-bial clauses are common in academic texts (Loban,1976) and their appropriate use may be necessary for

academic success, this is clearly an areaworthy of furtherinvestigation.

Taken together, these results suggest that the partic-ular linguistic structure employed in a taskmay influenceaccuracy of elements sometimes thought of as beingseparable units. Even as I acknowledge the ways thatphonological form, lexical frequency, or familiarity witha concept may influence the use of language at the wordlevel, these results highlight a need for similar attentionat the phrase, sentence, and discourse level—not onlywithchildren with SLI but also with adults and TD children.

Predictions About Group PerformanceTurningmy attention to the performance of the three

groups of children on these sentence types, I see that theage-matched children showed very few differences acrossconditions and clause locations. These observed differ-ences seem to be related to increasing clausal complexityor clause location as described previously, but recall thatfactors such as length and transitivity were not influen-tial. Processing demands associated with particular lin-guistic structures continue to affect production, even forchildren who are very proficient language users.

The profile of the MLU group in the second clausesexemplifies the accumulation of complexity that I hadpredicted for the SLI group. For “easy” things, such asthe first clause of the coordinate and complement condi-tions and the second clause of the coordinate condition,they performed similarly to the AGE group. As thingsbecamemore difficult, there were greater declines in thelikelihood of including the target morphological forms.Thus, in the second clauses, the childrenperformedworsein the temporal condition than in the coordinate condi-tion and worse in the complement condition than in thetemporal condition. Indeed, on the second clause of thecomplement condition, they showed a precipitous dropin accuracy and were by far the least accurate of thethree groups of children examined.

A variety of explanations for this finding come tomind. The MLU group was younger than either of thetwo comparison groups, but they arewell matched on bothexpressive vocabulary scores and utterance length tothe SLI group. One possibility is that despite being wellmatched on language abilities, theMLU group had slowerprocessing speed or poorer working memory abilities thanthe SLI group. Unfortunately, I did not collect NWR scoreson theMLU group because theNEPSYis only normed toage 5 years. Thus, I cannot directly test this hypothesiswith this data set. Future work incorporating both ver-bal and nonverbal measures of working memory wouldhelp elucidate this possibility. Another possibility is thatsurface characteristics of language, such asMLU or rawscores ona standardizedmeasure, donot accurately reflect



depth of knowledge/experience with a syntactic form andthat these children differed in linguistic abilities fromthose in the other two groups in ways that are not cap-tured by our matching variables. Further exploration offactors that support sentence production processes inTD children is clearly necessary, as is work on how theseprocesses undergo developmental change. It is obviousthat children do not simply becomemore accurate acrossthe board as they get older, but instead, use of morphol-ogy improves faster in some areas than in others.

The changes in second-clause accuracy observed inthe SLI group as compared with the age-matched groupwere proportional across the three conditions. Thus, al-though complexity does seem to influence the produc-tions of childrenwith SLI, it is not cumulative in thewaythat I expected. Although it is possible that I did not ob-serve an accumulation of difficulty due to the limitednumber of productions of transitive and intransitive +lengthener targets, I also did not observe increasinglylarge declines in accuracy across clause locations, an areain which sufficient responses were obtained. The ob-served pattern of results is consistent with the resultsfrom studies testing the idea that children with SLI areimpaired due to slower processing speed within the gen-eralized slowing hypothesis. These studies have shownthat children with SLI have proportionally but not ex-ponentially slower linguistic and nonlinguistic process-ing (Miller et al., 2001), a result that is borne out by thefinding that the children with SLI have a uniformlylower pattern of results than the age-matched children.It has recently been argued that verbal workingmemorydeficits, rather than processing speed, are considered tobe a factor that is separable from processing speed andthat makes a greater contribution to the observed defi-cits (Leonard, Ellis Weismer, et al., 2007). Working mem-ory could be implicated if one assumes that the availabilityof lexical items, fluencywith theparticular syntactic frame,or rehearsal of sentence elements affects processing de-mands. Indeed,MacDonald andChristiansen (2002) arguethat knowledge is not separable from speed of processing/working memory. Further work is required to distin-guish between these perspectives within a processingdeficit approach to SLI.

Although this study was framed within a process-ing perspective, others have argued that the difficultieschildren with SLI face in the use of tense and agreementare related to an underlying linguistic deficit. The UCCaccount argues that children with SLI experience a pro-tracted maturational period in which a grammaticalconstraint prevents them from checking more than oneuninterpretable grammatical feature within a clause(Wexler, 1998). This constraint leads to difficulty withtense and agreement and the omission of tense-relatedmorphemes. Likewise, the RDDR hypothesis (van der

Lely & Battell, 2003) proposes that children with SLIhave difficulty withmovement, which in turn limits theirability to construct utterances that require feature check-ing, noncanonical word order, or grammatical movement.Both of these propose a difficulty with the way that tenseis represented in the grammar rather thanwith theprovi-sion of tense and agreement morphemes due to process-ing constraints. Such accountswould predict that childrenwith SLI would have difficulty with tense marking incomparison with TD children, regardless of the levelof complexity present in the sentence—predictions thatare consistentwith the observation that the childrenwithSLI show a generally similar but overall less accurateprofile than their age-matched peers.

It is more difficult to explain why children with SLIand their TD counterparts might have difficulty withtense and agreement use in the same sentence but in dif-ferent locations or across different sentence types fromthe standpoint of a strictly representational account. Ifchildren have difficulty with checking features associ-ated with tense (Wexler, 1998), why would they be ableto do so in one clause or sentence type and not the other?If children have difficulty with movement across longdistances (van der Lely & Battell, 2003), why do theyhave difficulty with tense marking in sentences wherethemovement operations are similar? TheUCCaccount,at least, is hypothesized to hold for both TD children andchildren with SLI, but for different durations.

Future accounts of the deficits observed in childrenwith SLI may need to clearly incorporate evidence thatsupports both representational- and processing-basedapproaches to language production. That is, perhaps thedata observed here are best accounted for by acknowl-edging the influences of processing demands on the per-formance of both the TD children and the children withSLI, while also incorporating representational deficitsto account for the generally depressed performance lev-els of the children with SLI. At the same time, cautionshould be exercised in saying that the results observedare attributable to a particular developmental or lan-guage level, as seen via the surprising findings from theMLU-matched group. One possibility is that differentdevelopmental trajectories may lead to similar patternsof results (Karmiloff-Smith, 1998). For instance, processing-based deficits experienced early on by children with SLImay lead to such entrenched use of particular verb formsthat they appear as representational deficits later indevelopment. Similarly, TD children clearly overcome aprofile that I believed to be consistent with processingdeficits to eventually become proficient language users.Reconciliationof these resultswith currentaccounts of typ-ical development and deficits in SLI will require a greaterunderstandingofdevelopmental processesandhowchangein language use occurs over developmental time.



AcknowledgmentsThis work was supported by an internal research grant

from the University of Iowa. I would like to thank the childrenand families who participated in this project and the supportof Augustana College (Allison Haskill) and the Scottish RiteProgram (Elizabeth Merrifield) for assistance with subjectrecruitment. Marie Christiansen and Stacy Meyers assistedwith stimuli development. Rebecca Eness, Katie Errek, LyndiHill, Talia Hindin, Kenneth Marciniak, Amanda Murphy,Laura Romey, Vicki Samelson, Li Sheng, and RachelWakefieldassisted with data collection, transcription, and coding. SusanWagner Cook provided advice on statistical analysis. Thisarticle benefited from careful reading and comments from JeanK. Gordon, Ling Yu Guo, Karla McGregor, and Vicki Samelsonand from discussion within the Language Discussion Groupat theUniversity of Iowa. Portions of this articlewere presentedat the 28th annual meeting of the Society for Research inChild LanguageDisorders held inMadison,Wisconsin, in 2007.

ReferencesAlbright, A., &Hayes, B. (2003). Rules vs. analogy in Englishpast tenses: A computational/experimental study.Cognition,90, 119–161.

Baayen,R.H.,Davidson,D. J.,&Bates,D.M. (2008). Mixed-effects modeling with crossed random effects for subjectsand items. Journal of Memory and Language, 59, 390–412.

Berko, J. (1958). The child’s acquisition of English morphol-ogy. Word, 14, 150–177.

Bloom, L., Lifter, K., & Hafitz, J. (1980). Semantics of verbsand the development of verb inflection in child language.Language, 56, 386–412.

Bock, K., & Griffin, Z. M. (2000). The persistence of struc-tural priming: Transient activation or implicit learning?Journal of Experimental Psychology, 129, 177–192.

Bock, K., & Levelt, W. (1994). Language production: Gram-matical encoding. In M. A. Gernsbacher (Ed.), Handbook ofpsycholinguistics (pp. 945–984). San Diego, CA: AcademicPress.

Chinn, S. (2000). A simple method for converting odds ratiosto effect sizes for use inmeta-analysis. Statistics inMedicine,19, 3127–3131.

Crain, S., & Thornton, R. (1998). Investigations in universalgrammar: A guide to experiments in the acquisition of syntaxand semantics. Cambridge, MA: The MIT Press.

Dalal, R. H., & Frome-Loeb, D. (2005). Imitative productionof regular past tense –ed by English-speaking children withspecific language impairment. International Journal ofLanguage & Communication Disorders, 40, 67–82.

de Jong, J. (1999). Specific language impairment in Dutch:Inflectional morphology and argument structure (Doctoraldissertation). University of Groningen, Groningen, theNetherlands.

Diessel, H. (2004). The acquisition of complex sentences.Cambridge, United Kingdom: Cambridge University Press.

Diessel, H. (2008). Iconicity of sequence: A corpus-basedanalysis of the positioning of temporal adverbial clausesin English. Cognitive Linguistics, 19, 465–490.

Dixon, P. (2008). Models of accuracy in repeated-measuresdesigns. Journal of Memory and Language, 59, 449–456.

Donaldson, M. L., Reid, J., & Murray, C. (2007). Causalsentence production in children with language impairments.International Journal of Language and CommunicationDisorders, 42, 155–186.

Dunn, L. M., & Dunn, L. M. (1997). Peabody Picture Vocab-ulary Test—III (PPVT–III). Circle Pines, MN: AGS.

Eberhard, K. M., Spivey-Knowlton, M. J., Sedivy, J. C., &Tanenhaus, M. K. (1995). Eye movements as a windowinto real-time spoken language comprehension in naturalcontexts. Journal of Psycholinguistic Research, 24, 409–436.

Eisenberg, S. (2004). Production of infinitives by 5-year-oldchildren with language impairment on an elicitation task.First Language, 24, 305–321.

Ellis Weismer, S., Evans, J., & Hesketh, L. J. (1999). Anexamin

Factors Affecting Accuracy of Past Tense Production in ......children, but especially children with...

Documents

Transcript of Factors Affecting Accuracy of Past Tense Production in ......children, but especially children with...