Measurement Properties of Mean Length of Utterance in School-Age Children

Abstract

Purpose:

Mean length of utterance (MLU) is one of the most widely reported measures of syntactic development in the developmental literature, but its responsiveness in young school-age children's language has been questioned, and it has been shown to correlate with nonsyntactic measures. This study tested the extent to which MLU shows measurement properties of responsiveness and construct validity when applied to language elicited from elementary school children.

Method:

Thirty-two typically developing children in two age groups (5 and 8 years) provided four short language samples each. Language samples were elicited in a question–answer context and a narrative context. MLU was calculated with both morpheme and word counts. Other established measures of syntactic complexity (clausal density [CD], developmental level [D-Level], mean length of clause [MLC]) and lexical diversity (lexical density, moving-average type–token ratio, number of different words) were also calculated.

Results:

Linear mixed-effects analyses revealed that MLU varied systematically with discourse context and children's age group. The syntactic measures, CD and MLC, were found to vary systematically with MLU. None of the lexical diversity measures varied systematically with MLU.

Conclusion:

Results suggest that MLU is a responsive and valid measure of children's syntactic development across age and discourse context during the early school-age years.

The ability to easily quantify language complexity and its development is vitally important for speech-language pathologists (SLPs) because the evaluation of language skills at a particular developmental stage may lead to a child's initial enrollment in, or continuing eligibility for, special education services (American Speech-Language-Hearing Association, 2016). Language sample analysis (LSA) is often used in assessment as either an alternate or supplement to norm-referenced testing (Castilla-Earls et al., 2020). One of the most widely used measures of syntactic complexity derived from LSA is mean length of utterance (MLU), which is seen by clinicians as a robust measure of children's syntactic development (e.g., Bernstein & Tiegerman-Farber, 1997; Miller & Chapman, 1981). MLU has featured prominently in studies of monolingual English language development during the early school-age years, especially in studies that seek to distinguish children with typical language from those with communication disorders (see, e.g., Charest et al., 2020; Fey et al., 2004; Moyle et al., 2011; Rice et al., 2010). However, its relevance for capturing syntactic complexity in older, elementary school-age children's language has been questioned (Bernstein & Tiegerman-Farber, 1997; Blake et al., 1993; Klee & Fitzgerald, 1985; Rondal et al., 1987). The current study therefore asks: Does MLU provide responsive and valid measures of syntactic complexity in young school-age children's language?

MLU was originally proposed by Brown (1973) to capture the early language development in his longitudinal study of three children. He counted the number of morphemes in each utterance that the children produced and then calculated the mean number of morphemes produced per utterance. Many studies on child language development have followed suit (Heilmann et al., 2010; Hewitt et al., 2005; Moyle et al., 2011), but others have used the number of words in an utterance as the unit of count (Charest et al., 2020; Fey et al., 2004; Nippold, 2009). Although we might expect that calculating MLU using morphemes rather than words would result in outcomes that provide more developmental information, Parker and Brorson (2005) argue that the unit of count is inconsequential. They found that MLUm (by morphemes) and MLUw (by words) are correlated in 3-year-old English-speaking children's language. In addition, Rice et al. (2010) indicate that MLU in either words or morphemes yields reliable and valid estimates of children's language growth.

Measurement Responsiveness and Validity

Brown (1973) argued that MLU indexed the constructional complexity of children's developing language. Yet, there is disagreement regarding the sufficiency of MLU as a measure of syntactic complexity beyond a certain level of development (Bernstein & Tiegerman-Farber, 1997). For example, the last stage of development that Brown described, Stage V, is characterized by an MLU of between 3.75 and 4.5 morphemes, which is correlated with the use of advanced forms such as third-person singular as well as contractable auxiliaries and copulas. According to Brown, this stage is typically reached by 41–46 months old. After Stage V, Brown found that children produce relative clauses attached to the subject (e.g., The dog that jumps over the pig bumps into the lion; Diessel & Tomasello, 2001), as well as embedding and conjoining within the same sentence. Because this type of complexity may not be reflected in an increase in utterance length, Brown suggested using MLU as a measure of syntactic development only up until Stage V.

Subsequent studies have also questioned whether MLU is in fact a valid measure of syntax in older, school-age children. For example, Frizelle et al. (2018) analyzed narratives obtained from a large cross-sectional sample of speakers (N = 354), aged 4 years to adult. They measured syntactic complexity in the samples using both MLU (with words as the count unit) and clausal density (CD), which was defined by Scott (1988) as the mean number of clauses per utterance. Frizelle et al. argued that although both MLU and CD increased with age and were highly intercorrelated, CD provided better evidence of developmentally related change in complexity. They argued that CD provided information about the degree of subordination; utterance length indicates nothing in particular about the syntactic structures in use. Moreover, like Brown (1973), Frizelle et al. argued that, at a certain point in development, MLU is likely to be inversely correlated with syntactic ability. Specifically, increases in phrasal complexity entail the use of subordinate clauses, which packs more information into a single sentence using the same number or fewer words.

In studies that focus on identifying language impairment, MLU is paired with lexical measures to provide a more complete description of atypical language (Fey et al., 2004; Miller et al., 1992; Paul et al., 1996). Lexical diversity measures are assumed to provide information about language ability not captured by MLU under the assumption that MLU measures syntax. For example, the number of different words (NDW) in a sample and moving-average type–token ratio (MATTR) are used to capture age- and impairment-related differences in expressive vocabulary (Charest et al., 2020; Dethorne et al., 2005; Fey et al., 2004; Watkins et al., 1995). However, some of these studies also report a correlation between measures of lexical diversity and MLU (Dethorne et al., 2005; Ukrainetz & Blomquist, 2002), which again begs the question of what exactly MLU measures. In this study, we address this question in younger and older school-age children by eliciting narrative samples across two discourse contexts.

Current Study

Disagreements exist regarding the upper age at which MLU is useful and so whether MLU is a responsive measure of language development in young school-age children. Does the measure capture language differences during these years (i.e., responsiveness), or is it too coarse-grained of a measure to do so, as Frizelle et al. (2018) might argue? In addition, correlations between MLU and measures of lexical diversity, such as NDW, suggest that data generated by MLU, although designed to assess syntax, may instead measure language development more broadly, which raises the question of construct validity. The current study therefore asks two research questions.

First is Research Question 1 (RQ1): Is MLU a responsive measure of cross-sectional age-related and context-related differences in young school-age children's language? Responsiveness is typically defined as the ability to detect change over time in the construct being measured (Mokkink et al., 2010). We operationalized measurement responsiveness in the current study as the detection of difference in language between two age groups (i.e., 5- and 8-year-olds) and two discourse contexts: narrative and question–answer contexts.

The two ages of comparison were selected intentionally. Five is the age when children enter kindergarten and may be referred by their teachers for a speech-language evaluation. A comparison group of 8-year-olds was chosen because our targeted expressive language skills will have developed beyond what is seen in 5-year-olds to the extent that group differences are expected. Younger and older school-age children were expected to differ in their language ability based on typical developmental norms (e.g., Hoff, 2014). This expectation was confirmed in this study using a norm-referenced language assessment to evaluate language ability in the 5- and 8-year-old child participants.

Two discourse contexts were chosen for this study because changes in context are known to produce different syntactic outcomes (e.g., Nippold, 2009; Nippold et al., 2015). The different contexts have different task requirements that drive language behavior in certain ways; in particular, more complex task requirements drive more complex language. For example, Nippold (2009) showed that expository language, elicited by asking school-age children (ages 7;3–15;4 [years;months]) questions about how to play their favorite game, is associated with a greater variety of subordinate clause types compared to conversational samples. MLU has also been reported to be higher for narrative-retelling tasks than for critical-thinking tasks (Nippold et al., 2015), albeit with an older group of children, adolescents. The question–answer context in the current study was expected to elicit greater complexity in language relative to the narrative context. The materials used for language sample elicitation require the children to answer “why” and “how” questions. These types of questions prompt higher level thinking and reasoning (e.g., making inferences, problem solving, persuasive argument) and word choices that force subordination.

Next is Research Question 2 (RQ2): Is MLU a valid measure of syntactic complexity in young, school-age children, or does it capture other aspects of language complexity? Construct validity is the degree to which a measurement captures the construct in question (Mokkink et al., 2010). Although MLU is assumed to be a measure of syntactic development, it has been correlated with measures of lexical diversity in studies that use both (e.g., Dethorne et al., 2005). If MLU is a valid measure of syntax, then this correlation is likely due to underlying language ability: The child who produces language that is syntactically complex also has a larger expressive vocabulary than the child who produces language that is simple. If this hypothesis is correct, then MLU should correlate more strongly with other known measures of syntax and less with measures of lexical diversity once the shared variance in language ability is accounted for. To test this prediction, three established measures of syntactic complexity and three established measures of lexical diversity were calculated for each sample (see Method section for specific measures and procedures). Multiple measures were used to robustly characterize “lexical diversity” and “syntax,” because we know of no gold standard measures for these constructs.

Method

Participants

Thirty-two children participated in the study. Participants were sixteen 5-year-old (eight female, eight male) and sixteen 8-year-old (eight female, eight male) children with typically developing speech, language, and hearing. The average age in the 5-year-old group was 5;5 (range: 60–70 months); it was 8;5 years in the 8-year-old group (range: 96–107 months). Parents reported English as the children's first/native language. The English dialect was Standard American inflected by the back-vowel fronting typical of the West Coast. Five children had one caregiver whose first language was either German (one in the 5-year-old group and one in the 8-year-old group) or Spanish (two in the 5-year-old group and one in the 8-year-old group). As per parent report, the racial and ethnic distribution in the 5-year-old group was 63% White, 13% biracial (Asian American and White), 13% Hispanic, 1% Black, and 1% Native American, and in the 8-year-old group, the distribution was 88% White and 12% biracial (Asian American and White). Because socioeconomic status is positively correlated with caregiver education (Davis-Kean, 2005), caregiver education information was collected. Children in the current sample had caregivers with mostly high educational attainment: In the 5-year-old group, 47% of caregivers had advanced degrees (i.e., master's, PhD, or MD), 28% had a college degree, 13% had some college, 3% had a high school degree, and 9% had not finished high school, and in the in the 8-year-old group, 56% of caregivers had advanced degrees (i.e., master's, PhD, or MD), 31% had a college degree, 3% had some college, 3% had a high school degree, and 6% had not finished high school. Participants were recruited from the Portland, Oregon area using fliers distributed throughout the community (e.g., libraries, schools) and through word of mouth. Institutional review board approval was granted by the University of Oregon and extended to data collection at Portland State University, which followed all protocol guidelines and used approved consent/assent materials.

Hearing was screened at 1000, 2000, and 4000 Hz at a threshold of 25 dB SPL. Typical speech and language development was determined using the articulation subtest of the Diagnostic Evaluation of Articulation and Phonology (DEAP; Dodd et al., 2002) and the Core Language Score on the Clinical Evaluation of Language Fundamentals–Fifth Edition (CELF-5; Wiig et al., 2013). Children's scaled scores on the DEAP ranged from 7 to 12 (M = 10.53, SD = 1.9). Their Core Language Scores from the CELF-5 ranged from 86 to 136 (M = 116.1, SD = 10.12) without a significant difference between the two age groups (t = 0.019, p = .493, Cohen's d = .007).

Procedure

Each of the 32 participants provided four short language samples for analysis. These were elicited in two contexts: (a) a question–answer context and (b) a storytelling context. The same materials were used in both contexts. These materials included six picture prompts for each of two stories, Dog Comes Home and Bunny Goes to School, taken from the School-Age Language Assessment Measures (SLAM; Crowley & Baigorri, 2015). For each story, children inspected the picture cards placed before them while answering standard evaluation questions (see Appendix A). The questions required the participant to deduce (e.g., “Why is she in the bathtub with a white dog now?”), infer (e.g., “Why did she come to school?”), problem solve (e.g., “What would you do if a bunny started hopping around your school?”), predict (e.g., “What do you think will happen when the boy goes home?”), and use theory of mind (e.g., “What is the girl thinking here?”). Then, with the pictures still in front of them, the children were asked to tell the story shown in the pictures (see Appendix B for sample narratives). During language sample elicitation, experimenters provided natural prompts, such as question repetition, encouragement to answer the question (e.g., Experimenter: “What do you think the mother's going to do now?” Child: “I don't know.” Experimenter: “Do you have any thoughts? What do you think?”), and encouragement to begin or continue the story narrative (e.g., Experimenter: “Now you tell me the whole story.” Child: “I don't want to.” Experimenter: “But you already told me so much of it. What's happening here?”). The language samples were audiovisually recorded in a quiet experimental room at Portland State University.

Language Sample Transcription and Coding

Audio files of the language samples were transcribed using Systematic Analysis of Language Transcripts (SALT; Miller & Iglesias, 2019) conventions to identify utterances, words, morphemes, unintelligible segments, and mazes (Miller et al., 2019). Specifically, the samples were first segmented into C-units (i.e., one main clause and any modifiers or subordinate clauses; Loban, 1976) using SALT's C-unit segmentation rules (Miller et al., 2019). Utterances that were less than a C-unit (e.g., “sorry,” “bye”) were included in the analyses as long as they were not maze behavior (i.e., false starts, repetitions, reformulations, and filled pauses, e.g., um, uh, well). Stereotypic closing (e.g., “the end”) and side comments were placed on special lines and were not included as part of the transcription. Lexical verbs and copulas were coded with “[v],” so that the number of clauses (i.e., main and subordinate clauses defined as a statement containing both a subject and a predicate) could be tallied automatically by SALT, as per Fey et al. (2004). Infinitives (e.g., “because she want/3s[v] to hide[v] it from the mom”) and bare infinitives (e.g., “and she think/3s[v] her mom won't let[v] her keep[v] the dog) were coded as clauses.

Measures

Table 1 summarizes the various measures taken in this study and how they were calculated based on the segmentation and transcriptions of the language samples. Several of the measures were automatically calculated using SALT.

Table 1.

Measure calculations.

Measure	Calculation
Mean length of utterance in morphemes	Dividing the total number of morphemes by the total number of C-units (SALT generated)
Mean length of utterance in words	Dividing the total number of words by the total number of C-units (SALT generated)
Syntactic complexity
Clausal density	Dividing the total number of clauses in a sample by the total number of C-units
Developmental level	Manually scoring each utterance on an 8-point complexity scale and calculating the average for each sample
Mean length of clause	Dividing the number of total words (SALT generated) in a sample by the number of clauses in that sample
Lexical diversity
Lexical density	Dividing the number of lexical items (nouns, verbs, adjectives, and adverbs) by the total number of words in each sample × 100
Moving-average type–token ratio	Dividing moving-average NDW by moving-average number of total words (SALT generated)
Number of different words (NDW)	Tabulation of the total number of different words in a sample (SALT generated)

Note. SALT = Systematic Analysis of Language Transcripts.

The three syntactic measures were the aforementioned CD measure, a measure known as “developmental level” (D-Level; Covington et al., 2006), and mean length of clause (MLC; Kallay & Redford, 2020). CD and D-Level have been widely used in basic and clinical research as measures of syntactic development (Cheung & Kemper, 1992; Covington et al., 2006; Fey et al., 2004; Frizelle et al., 2018; Lu, 2009; Rosenberg & Abbeduto, 1987) and have also been shown to distinguish between samples elicited under different discourse contexts (e.g., conversational vs. expository; Nippold, 2009; Nippold et al., 2005). D-Level indexes grammatical complexity with an acquisition-based sentence complexity scale ranging from simple sentences to sentences with more than one level of embedding. Each C-unit was manually scored for its D-Level using the Covington et al. (2006) scale in Table 2. The average D-Level was then calculated for each sample. MLC was calculated by dividing the number of total words (NTW) by the total number of independent and dependent clauses in a sample.

Table 2.

Developmental level scale.

Score	Description	Examples
0	Simple sentences, including questions; sentences with auxiliaries; simple elliptical sentences	“I'm not taking a bath.”
1	Infinitive or –ing complement with the same subject as the main clause	“And the boy is going to give the bunny a carrot.”
2	Conjoined noun phrases in subject position; sentences conjoined with a coordinating conjunction; conjoined verbal, adjectival, or adverbial constructions	“The mom came over and grabbed the bunny and went home.”
3	Relative clause modifying the object of the main verb; nominalization in object position; finite clause as object of the main verb; subject extraposition; raising	“And (some) some are not happy that the bunny is so sneaky.”
4	Nonfinite complement with its own understood subject; comparative with object of comparison	“I saw him walking the dog.”
5	Sentences joined by a subordinating conjunction; nonfinite clauses in adjust positions	“And ask the teacher if I could feed the bunny some carrots.”
6	Relative clause modifying the subject of the main verb; embedded clause/nominalization serving as the subject of the main verb	“Because the ones who are afraid are surprised.”
7	More than one level of embedding in a single sentence	“It is because the teacher told her that there was a bunny.”

Note. The developmental level scale was based on Covington et al. (2006). Examples are from current study data, except no. 4 for which there were no data examples.

The three measures of lexical diversity included lexical density (LDensity) and the aforementioned MATTR and NDW. LDensity has been used to analyze the complexity of spoken (Johansson, 2009) and written (Hall-Mills & Apel, 2015) narratives. This measure distinguishes between words with lexical properties (i.e., nouns, verbs, adjectives, and adverbs) and those with grammatical properties (i.e., conjunctions, articles, auxiliary verbs, interjections, determiners, and prepositions), counting only those with lexical properties (Johansson, 2009). LDensity was calculated manually for each sample by first running a Grammatical Categories analysis in SALT (i.e., list of parts of speech generated by the GramCats algorithm, which is 95.1% accurate; Channell & Johnson, 1999) to determine the number of lexical items (i.e., content words: nouns, verbs, adjectives, and adverbs) and then dividing by the SALT-generated NTW for the sample. The adverb category included adverbs of manner (e.g., “randomly”), time (e.g., “now”), place (e.g., “outside”), frequency (e.g., “usually”), and purpose (e.g., “so that”).

MATTR measures lexical diversity by calculating type–token ratios (i.e., TTRs, a ratio of unique lexical items divided by the total number of words in a sample) for successive nonoverlapping segments of a sample (Covington & McFall, 2010). This is a measure of unique lexemes but is calculated within a moving window size. Because MATTR uses length-controlled windows, it controls for the known sample length problems of other widely used measures (e.g., TTR; Charest et al., 2020). An MATTR closer to 1.0 indicates a varied vocabulary, and an MATTR closer to 0.0 represents a limited, repetitive vocabulary. MATTR was calculated in SALT with a window size of 20 words.

Finally, NDW was calculated. SALT calculates NDW based on the production of unique free morphemes, so that play, play/ed, and play/ing would be treated as one word occurring 3 times (Miller et al., 2019). Although NDW is used to calculate MATTR, the two measures provide unique information, and both are widely used in research and clinical settings.

Reliability

Measurement reliability depended on the reliability of the transcriptions, C-unit segmentation, and the manual scoring of D-Level. To ensure transcription, segmentation, and coding reliability, we used a consensus procedure (as in Guo & Eisenberg, 2015; Shriberg et al., 1984). Each sample was initially transcribed and segmented by the first author. Then, a trained research assistant reviewed the data while listening to the recorded language samples and reading the initial segmented transcriptions. Transcription, segmentation, or coding disagreements were identified and then reviewed and discussed until agreement was obtained for all transcripts (e.g., Frizelle et al., 2018).

Each sample was initially scored for D-Level by the first author. The audio files of 12 language samples (10%) were then randomly selected from the data, and a research assistant independently rated and scored these samples for D-Level. The raters' scores had a moderate level of agreement in a Cohen's kappa calculation (κ = .70; McHugh, 2012). Despite the relative difficulty of D-Level judgments, the reliability of these and all other measures was judged to be suitable for their subsequent analyses.

Analyses

Analyses were completed using R Studio (RStudio Team, 2020), a companion program to R (R Core Team, 2018). R packages were used for data management (tidyr; Wickham & Henry, 2018), analysis (lme4; Bates et al., 2015; sjstats; Lüdecke, 2021), and visualization (ggplot2; Wickham, 2016). For the first research question (RQ1), linear mixed-effects modeling was used to test for measurement responsiveness, that is, the effect of the two categorical variables, children's age (between-subjects) and discourse context (within-subject) on MLUm and MLUw. Participant was entered as a random intercept. Parent-reported sex of the participant was entered as a control variable. Tests for significance were obtained through model comparison using the likelihood ratio tests. Interaction terms were removed from the models when not significant. Partial eta squared was calculated to determine effect sizes, where .01 indicates a small effect, .06 indicates a medium effect, and .14 indicates a large effect (Field, 2013).

To test whether MLU is a valid measure of syntactic development and complexity (RQ2), the other standard measures of syntax (CD, D-Level, MLC) and lexical diversity (LDensity, MATTR, NDW) were entered as predictors into linear mixed-effects models that had MLUm or MLUw as the outcome variable. Participant was entered as a random intercept. In this way, the analyses controlled for shared variance among the measures due to language ability, allowing us to test for the predicted unique relationship between specific measures of syntax and MLU.

Results

Table 3 displays means, standard deviations, and ranges of sample lengths in duration (minutes), number of C-units, and NTW by age group for the two types of language samples. The question–answer sample durations include the time that the experimenter was asking questions. The number of C-units reference only language the children produced. One child declined to tell one of the stories, and so the total number of language samples was 127 rather than 128. The total number of utterances evaluated was 1,147 C-units.

Table 3.

Means, standard deviations, and ranges of sample lengths.

Sample	Duration (min)		C-units		NTW
	M (SD)	Range	M (SD)	Range	M (SD)	Range
Question–answer
5-year-olds	1.23 (0.02)	0.34–2.30	9.47 (2.83)	9–20	54.44 (22.87)	21–116
8-year-olds	1.13 (0.02)	0.39–2.36	9.31 (2.13)	5–14	68.16 (29.69)	23–143
Narratives
5-year-olds	0.41 (0.01)	0.21–1.37	7.81 (1.99)	4–12	54.58 (17.84)	27–104
8-year-olds	0.45 (0.01)	0.17–1.57	9.50 (3.10)	5–18	81.59 (30.37)	36–175
Overall
5-year-olds	1.03 (0.03)	0.18–2.30	8.65 (2.57)	4–20	54.51 (20.38)	21–116
8-year-olds	0.59 (0.01)	0.17–2.36	9.41 (2.64)	5–18	74.88 (30.55)	23–175

Note. NTW = number of total words.

Measurement Responsiveness of MLU (RQ1)

Descriptive data for MLUm and MLUw and complete model results for RQ1 are presented in Table 4. Descriptive data show differences by discourse context and age group for MLUm and MLUw.

Table 4.

Means, standard deviations, and fixed effects on mean length of utterance (MLU) by age group and by discourse context.

Measure	5-year-olds		8-year-olds		Estimate	SE	χ2	p	ηp 2
	M	SD	M	SD
MLUm	6.98	1.67	8.73	2.18	1.75	.45	12.05	< .001***	.12
MLUw	6.39	1.57	7.86	1.97	1.47	.42	10.34	.001**	.11
	Question–answer		Narrative
MLUm	7.11	2.05	8.63	1.93	1.50	.25	31.07	< .001***	.26
MLUw	6.41	1.86	7.87	1.70	1.44	.22	34.75	< .001***	.29

Note. Degree of freedom for all models is 1. “m” indicates morphemes as the unit of measurement; “w” indicates words as the unit of measurement. The interaction between fixed effects is not shown because it was not significant. The reference category for group was 8-year-olds and for context was narratives. MLUm = MLU in morphemes; MLUw = MLU in words.

^** p ≤ .01.

^*** p ≤ .001.

The main effects of age group and discourse context (see Table 4) were significant for both MLUm and MLUw, as illustrated by the boxplots in Figure 1. The interaction between group and context was not significant. The direction of the effects was as follows: The 8-year-old group produced samples with higher MLUs compared to the 5-year-old group (see Table 4); narratives were produced with higher MLUs compared to the question–answer samples (see Table 4). Effect sizes in partial eta squared ranged from .05 to .12 for group and from .12 to .29 for context.

Figure 1.

Mean length of utterance (MLU) by age group and discourse context: MLU in morphemes (MLUm; left) and MLU in words (MLUw; right).

Construct Validity of MLU (RQ2)

Descriptive data for the three measures of syntactic complexity and three measures of lexical diversity are presented in Table 5 by age group and discourse context. All measure values, except MATTR, were higher in 8-year-olds compared to 5-year-olds as well as higher in the narrative context compared to the question–answer context. Table 6 displays the correlations between the predictor variables. As expected, there is an especially strong correlation between NDW and two of the syntactic complexity measures (i.e., D-Level and CD).

Table 5.

Means and standard deviations for syntactic complexity and lexical diversity measures.

Measure	5-year-olds		8-year-olds
	M	SD	M	SD
Syntactic
Clausal density	1.25	0.28	1.50	0.32
D-Level	0.78	0.67	1.39	0.85
Mean length of clause	5.12	0.98	5.24	0.81
Lexical
LDensity	47.45	5.62	45.69	5.06
MATTR	0.80	0.06	0.80	0.06
NDW	33.56	10.25	43.23	15.18
	Question–answer		Narrative
	M	SD	M	SD
Syntactic
Clausal density	1.36	0.36	1.40	0.28
D-Level	1.07	0.89	1.10	0.74
Mean length of clause	4.73	0.71	5.63	0.85
Lexical
LDensity	45.44	5.82	47.70	4.71
MATTR	0.82	0.06	0.79	0.04
NDW	38.36	14.57	38.51	13.10

Note. D-Level = developmental level; LDensity = lexical density; MATTR = moving average type–token ratio; NDW = number of different words.

Table 6.

Correlations for study variable means (N = 32).

Variable	1	2	3	4	5	6
1. CD	—
2. D-Level	.87**	—
3. MLC	.09	.31	—
4. LDensity	−.49**	−.54**	−.13	—
5. MATTR	.40**	.32	.20	−.31	—
6. NDW	.76**	.76**	.36*	−.34	.62**	—

Note. CD = clausal density; D-Level = developmental level; MLC = mean length of clause; LDensity = lexical density; MATTR = moving-average type–token ratio; NDW = number of different words.

^* p ≤ .05.

^** p ≤ .01.

To answer RQ2, the syntactic complexity and lexical diversity measures were used to predict MLUm and MLUw in linear mixed-effects models. The results indicated a significant relationship between two of the syntactic measures and MLU: Both CD and MLC predicted MLUm and MLUw (CD: MLUm, t = 20.49, p < .001, η_p ² = .79; MLUw, t = 26.44, p < .001, η_p ² = .85; MLC: MLUm, t = 23.67, p < .001, η_p ² = .83; MLUw, t = 33.59, p < .001, η_p ² = .90). None of the measures of lexical diversity were significantly related to MLU in the overall model, despite a strong bivariate correlation between MLU and NDW: r(32) = .79, p < .001. This result shows the importance of controlling for shared variance in the analyses.

Discussion

Overall, the results indicate that MLU is responsive to cross-sectional age-related and context-related differences in the language of young school-age children. In addition, the results suggest that MLU is a valid measure of syntactic development in these age groups.

MLUm and MLUw were found to be responsive measures of linguistic complexity. The effect of age group on MLU was in the predicted direction: Language samples produced by 8-year-old children had greater MLUs than those produced by 5-year-old children. This result is consistent with the expected development of language—an expectation that was confirmed in this study with norm-referenced language testing (i.e., raw scores on the CELF-5 were greater in the 8-year-olds). MLU was also different across the narrative and question–answer discourse contexts, but the direction of this effect ran counter to expectations. Nippold et al. (2015) found that MLU was systematically longer in language samples obtained from children ages 12;10–14;11 during a critical-thinking task than in language samples obtained in a conversation or narrative task. This finding led to our prediction that the question–answer context would produce longer MLUs than the narrative context. We expected the question–answer context would require children to think more critically about the picture sequences than they might during a narrative and thus produce longer utterances. However, this prediction did not consider that children in the current study were much younger than those in the Nippold et al. study. It is possible that our prediction would have been born out were we to work with young adolescents who likely have more experience answering critical-thinking questions in a thorough manner. However, there are at least two other possible explanations for the unexpected direction of the effect of context on MLU.

First, the systematic difference in MLU by discourse context may reflect the particularly short and formulaic utterances elicited in the question–answer context (e.g., “I don't know,” “bye,” “sorry”). For example, there were 31 instances of “I don't know” in the question–answer samples from the 5-year-old group and eight instances in the 8-year-old group. Because the questions required higher level thinking, it is not surprising that the younger children struggled more with answering these types of questions and the result was a perfectly appropriate, but short, “I don't know” response. In addition to “I don't know,” there were nine instances of one-word responses in the question–answer context (e.g., “bye” in response to “What is the teacher thinking now?” where the picture shows the teacher waving; “sorry” in response to “What would you say to your mom if you were the girl here?” where the picture shows the mom looking angry). These one-word responses, although short, were nonetheless pragmatically appropriate responses and correct based on what is shown in the pictures. In a clinical setting, the best practice in this case would be for the tester to prompt for guesses to elicit more language because eliciting high-quality language samples is vital when conducting LSA (Miller et al., 2019). Indeed, elicitation of the best language samples requires the tester to show interest, make natural contributions, and prompt by adding supportive comments (Nippold, 2014; Shipley & McAfee, 2009). In this study, the different testers did follow up with children but varied in their level of assertiveness in the follow-up. Overall, the question–answer context did elicit different language from the narrative language context, but not in the intended direction because children engaged in pragmatically appropriate behavior by providing just the relevant information in answers to the questions given.

A second possible reason for the unexpected direction of the effect of context on MLU may have been the fixed order in which the tasks were completed. The question–answer task always preceded the narrative elicitation in our study. This order, at variance with SLAM instructions, followed from our previous work. The question–answer context allows the tester to build rapport with the child, helps familiarize the child to the pictures, and helps the child to conceptualize a more coherent and complete narrative than they might otherwise produce (Kallay & Redford, 2020; Redford, 2013). However, by tapping into children's reasoning skills and prompting them to think conceptually before producing a narrative, we may have increased the probability that narratives would be produced with more complex syntax, given the relationship between higher level thinking skills and syntactic complexity (de Villiers & Pyers, 2002; Nippold et al., 2007). This possibility is instructive clinically in that it suggests narrative language samples may provide especially good information about syntactic development if children are encouraged to think through the story they will produce with a question–answer phase preceding narrative elicitation. Therefore, both tasks of the SLAM should be used with early school-age children because they each provide information about the speaker's ability to formulate complex language and comprehend the reasoning and problem-solving types of questions.

Whether it is calculated in morphemes or in words, MLU varied systematically with the age of children who produced the language samples under analysis and with the discourse contexts in which the samples were produced. So MLUm and MLUw are responsive measures in typically developing young school-age children. Because there were no significant differences between MLU results that varied by unit of count (morpheme vs. word), we conclude that the unit of count does not matter in these age groups, as other studies have similarly shown (Parker & Brorson, 2005; Rice et al., 2010). However, when an SLP is conducting LSA as part of a language evaluation and suspects issues with morphological development, using MLUm would be beneficial for establishing baselines and identifying errors.

We conducted a linear mixed-effects analysis to assess the strength of the relationship between MLU and established measures of syntactic complexity and lexical diversity to determine the construct validity of MLU as a measure of strictly syntactic complexity. Indeed, MLU was found to be a valid measure of syntactic development in young school-age children: It covaried with measures of syntactic complexity and not with measures of lexical diversity in models that controlled for shared variance between the different measures. Strong correlations between two syntactic complexity measures (CD and D-Level) and two lexical diversity measures (NDW and MATTR) can be interpreted as the measures indexing the same construct. CD and D-Level both relate to the complexity of the utterances, mainly due to subordination, and MATTR and NDW both focus on the variety of vocabulary used by the children. In clinical practice, these findings suggest the use of MLU to measure syntactic complexity and that additional measures are needed to assess lexical diversity.

Taking a closer look at the relationship between MLU and syntactic complexity, the descriptive data for our measures indicate that as the length of an utterance increases, so too does the number of clauses in the utterance regardless of a child's age. An example, from an 8-year-old's narrative, illustrates the relationship between MLU and clause number: “and then he jumped out 'cause he saw the carrot on the chalkboard” contains just one C-unit but is composed of 14 morphemes and two clauses. Thus, although developmental increases in syntactic complexity can occur independent of decreasing MLU via subordination (Frizelle et al., 2018), we did not find this to be true in this study. Children in the older group did engage in more subordination than the children in the younger group, but their utterances were longer due to their use of more adjectives (e.g., “new,” “little”), more adverbs (e.g., “so,” “really”), and fewer sentence fragments. Overall, the result indicates an important relationship between CD and MLU and between MLC and MLU in these age groups.

The example utterance above suggests that complex syntax, associated with conceptual complexity, is also associated with a conceptual richness reflected in the expressive vocabulary. Indeed, a bivariate correlation between a measure of lexical diversity, NDW, and MLU was strong in the present data. This finding replicates previous work that has reported the correlation between NDW and MLU (Dethorne et al., 2005; Ukrainetz & Blomquist, 2002)—a correlation that suggests a strong association between the developing lexicon and the development of complex syntax. This study's analyses assessed the hypothesis that this association is due to shared variance—specifically, to overall language ability—rather than to a causal or identity relationship between lexical diversity and syntax. This hypothesis was supported: The full model results indicated that, once shared variance is accounted for, only measures of syntactic complexity correlate with MLU. We take the resulting correlations between other standard measures of syntax and MLU to indicate an identity relationship and so again conclude that MLU is a valid measure of syntax in young school-age children.

Limitations and Future Directions

We find that MLU is responsive to age and discourse context, but we acknowledge the small sample may limit the generalizability of our results to other populations. For example, younger children may require longer samples to show this type of effect because their language production is more variable than that of older children (Guo & Eisenberg, 2015). This is true for children with delayed speech and language skills as well.

More research is required to validate the use of short samples like those analyzed in our study in a variety of populations, particularly with children with disorders, because the primary clinical purpose of LSA is to identify language disorders in children. Although the traditional recommendation for LSA is to use samples of 50 utterances or more to assess children's development (Eisenberg et al., 2001; Miller et al., 2019), the results from some studies that investigate the effect of sample size on measurement reliability suggest that shorter samples may also yield good-quality assessment information. For example, shorter language samples have been shown to be reliable for certain measures of productivity (i.e., number of total utterances and words per minute) and lexical diversity (NDW) and for MLUm (Casby, 2011; Heilmann et al., 2010). It is certainly more practical to elicit short language samples, so if these can provide an effective measure of language skills in school-age children, busy clinicians will benefit (Casby, 2011; Heilmann et al., 2010, 2020; Heilmann & Malone, 2014).

Although the order of the question–answer sample always preceding the narrative samples was intentionally set to build rapport and prompt the children to think about the stories before providing a narrative, the lack of counterbalancing the elicitation of the two language sample types is a limitation. The order utilized in this study may have contributed to the narrative samples being longer than the question–answer samples. Asking the questions after the narrative elicitation task may have produced a different result (i.e., longer answers from the children). Future work requires counterbalancing the language sample types to eliminate these types of effects.

Finally, it may be of interest to reconsider our conclusion that the unit of count (word vs. morpheme) did not make a difference with regard to the findings of responsiveness and construct validity with further research. It is tempting to suggest the use of MLUw with early school-age children because it requires less time and effort than MLUm; however, it seems likely that morpheme counts provide important additional information about developmental maturity that is not captured by MLUw alone (e.g., systematic deletion of third-person markers).

Data Availability Statement

The data sets generated and/or analyzed during the current study are available from the corresponding author on reasonable request.

Appendix A

Question Prompts for Each Story

Questions for Dog Comes Home

1. What do you think the girl is thinking here?

2. Why do you think she's putting the dog in her bag?

3. Why do you think the girl's getting so dirty?

4. Why is there a white dog in the bathtub now?

5. What do you think the mother's going to do now?

6. What would you say to the mom if you were the girl here?

Questions for Bunny Goes to School

1. Why do you think the bunny jumped out of the backpack?

2. Why do you think some students are afraid and some students are laughing?

3. What would you do if a bunny came to your school?

4. What was the boy's idea?

5. How did the mom know to come to the school?

6. Why do you think the mom came to the school?

7. What do you think will happen when the boy when he goes home?

8. What do you think the teacher's thinking now?

Appendix B

Sample Narratives

Narrative by a 5-year-old participant for Dog Comes Home

She found a little dog under there.

And then (she) it started licking her.

And then it got her clothes all dirty.

And then she put it in her bag.

And then she told the dog to be quiet.

And then she goed in bathtub.

And then they were washing off.

Narrative by an 8-year-old participant for Dog Comes Home

A girl found a dog under a porch.

She found it.

and she made friends with it.

but she thought she wouldn't be able to bring it home.

Because her mom didn't want her to maybe.

I don't really know probably cause someone in her family's allergic to it.

I don't really know.

(um tries to she tries to) she smuggles the dog in in her backpack.

(um) her mom tells her to have a bath cause she's so dirty.

The dog comes out of the backpack and hops in the bath (or bath).

(um) and the mom is angry.

And it looks like she's swearing.

Funding Statement

This work was supported by the Eunice Kennedy Shriver National Institute of Child Health & Human Development (NICHD) under Grant R01HD087452 (Principal Investigator: Redford). The content is solely the authors' responsibility and does not necessarily reflect the views of NICHD. Portions of these finding were presented as a poster at the 2021 Symposium on Research in Child Language Disorders, Madison, WI.