The Utility and Robustness of Sentence Repetition as a Marker of Developmental Language Disorder: An Analysis of the Surrey Communication and Language in Education Study Project Data

Abstract

Purpose:

This study examined the diagnostic power and robustness of sentence repetition (SR) in distinguishing monolingual English-speaking children with developmental language disorder (DLD) from typically developing (TD) children, using a large-scale secondary analysis of the Surrey Communication and Language in Education Study data set. We evaluated SR's effectiveness across different scoring methods in age-matched groups across three different time points.

Method:

A total of 407 children were included. Bayesian beta regression models compared SR performance between groups across four scoring methods. SR's diagnostic accuracy was assessed using area under the curve (AUC), sensitivity, specificity, and likelihood ratios.

Results:

Children with DLD consistently performed less accurately than TD peers on SR tasks, with differences persisting across ages 5–11 years and across all scoring methods. AUC values indicated good to excellent discriminative ability, especially in younger groups, though sensitivity and specificity fell below clinical thresholds.

Conclusions:

These findings support SR's diagnostic potential as a reliable language measure. SR may serve as a valuable tool due to its ease of use and scoring flexibility, though additional assessments are recommended for comprehensive language evaluation.

Supplemental Material:

https://doi.org/10.23641/asha.31431517

Developmental language disorder (DLD) is characterized by impairments in acquisition, use, and understanding of spoken language (Bishop et al., 2017). DLD can have significant long-term impacts on academic achievement, social interaction, and overall communication skills (Conti-Ramsden et al., 2001; Durkin & Conti-Ramsden, 2007). There is a critical need for accurate and reliable tools that can identify children with DLD or identify those who may require further diagnostic evaluation (McGregor, 2020). One promising marker is performance on sentence repetition (SR) tasks. The evidence base for the diagnostic usefulness of SR as a marker for DLD is somewhat problematic, due to the small size of the samples and diverse procedures used in studies. However, a recent meta-analysis (Ward et al., 2024) found the collective evidence to be robust. The current study seeks to build on this by investigating the diagnostic power and robustness of SR tasks in a large-scale, secondary data analysis of the Surrey Communication and Language in Education Study (SCALES; Norbury, 2022). Through a series of simulated case–control studies, we assess the effectiveness of SR tasks in detecting DLD across different scoring methods and across children aged 5–11 years.

The Use of SR Within the Literature

The use of performance on SR tasks as a marker for DLD has gained traction since Conti-Ramsden et al. (2001) demonstrated that SR achieved better sensitivity (true positives; 90%) and specificity (true negatives; 85%) for identifying DLD than a third-person singular task, a past tense marking task, or a nonword repetition task. Archibald and Joanisse (2009) found SR tasks to be a better predictor of DLD than nonword repetition for English-speaking children, with sensitivity and specificity values both exceeding 80%. A meta-analysis by Pawłowska (2014) looked at the use of three marker tasks (verb tense, nonword repetition, and SR) of language impairment in English-speaking children across 13 studies (four of which used SR). SR was found to be the best marker of the three, with Pawłowska noting that this may be due to the advantages of SR in being sensitive to a wide range of language and processing deficits. Their conclusion is that SR is promising as a screening tool for DLD but that it needs further research validation. This hesitancy is justified by the high levels of heterogeneity and small samples sizes seen across the studies reviewed. There is also evidence of SR's utility beyond English, with studies showing sensitivity and specificity above 80% for monolingual speakers of Hebrew (Armon-Lotem & Meir, 2016), Russian (Armon-Lotem & Meir, 2016), French (Fleckstein et al., 2018; Leclercq et al., 2014), German (Hamann & Abed Ibrahim, 2017), Mandarin (Wang et al., 2022), and Turkish (Kaçar-Kütükçü & Topbaş, 2024).

The methodological variation in the application of SR tasks across the literature has been considered in a couple of recent reviews. A scoping review by Rujas et al. (2021) looked at the broad use of SR across 203 studies published in English or Spanish. There were at least 50 original SR tasks seen across the studies with considerable variation seen in the uses of these tasks. From their review, SR is concluded to be useful for assessing language ability and specifically presents a “suitable” task for detecting the presence of DLD in children. A recent systematic review and meta-analysis by Ward et al. (2024) explored the use of SR tasks in the literature for distinguishing monolingual DLD and typically developing (TD) groups of children. There were 66 studies identified, which involved 19 different languages, 39 different SR tasks, and 4,447 children. Across the studies, children with DLD were found to perform an average of 2.08 SDs worse than TD children on SR tasks. A key finding was that this power to differentiate the groups was robust to differences in (a) whether the tasks were standardized or not, (b) whether sentences were prerecorded or not, (c) whether scoring was conducted during the task or after, (d) the type of scoring method used, (e) whether clinical or population DLD samples were used, and (f) the language the task was conducted in. This supports the clinical application and usefulness of SR as a task that can be adapted to contextual demands.

Ward et al. (2024) note key limitations of their review, similar to those mentioned by Pawłowska (2014). These include small sample sizes within individual studies, as well as comparisons of individual aspects of methodological variation being limited by high heterogeneity and potential confounding between studies. The current study seeks to address these limitations by conducting a secondary data analysis within a single large-scale, population-based study. This approach allows for consistent methodology, larger sample sizes, and controlled comparisons within the same data set, thereby reducing between-study confounds and increasing interpretability.

Further, while reviews such as these provide valuable insight into the methodological variability of SR tasks, few studies have directly examined age-related effects on SR performance or diagnostic accuracy. The meta-analysis by Ward et al. (2024) included studies with participants ages between 3 and 16 years, and although encompassing a broad developmental range, age effects were not specifically examined. Similarly, Rujas et al. (2021) reported that most SR studies have focused on children aged between 5 and 8 years, with fewer including younger or older participants. In the current study, age was controlled for within each time point to reduce potential confound. Although we did not formally assess age-related changes across time, examining SR performance at three developmental stages provides an informal indication of how performance on a specific SR task may change across the age range for which it was designed.

Methods of Scoring SR Performance

One key methodological variation in the application of SR tasks is in how the productions are scored. These measures of SR performance can, for example, involve binary scoring where the whole sentence production is marked as either correct or incorrect. For instance, for the target sentence “The dog is running,” a binary approach would assign a score of 1 only if the entire sentence was produced correctly. They can also be more detailed and involve a score per production in relation to the number of errors made or whether specific targets within the sentence were produced correctly or not. Again, for a target such as “The dog is running,” scores may be assigned based on whether specific elements (e.g., the, dog, is, running) are produced correctly. Each method can bring with it different potential benefits. A binary method of scoring is the simplest method and has been suggested to be the most efficient method for use in clinical contexts because of this (Hamann & Abed Ibrahim, 2017). In addition, Ebert et al. (2019) found that binary scoring methods can be reliably implemented, even by those with no background in language assessment. On the other hand, scoring that considers specific target structures within sentences provides the most detailed window onto a child's language ability (Komeili & Marshall, 2013).

In terms of which scoring method can better identify children with DLD, there have been mixed research findings. Some research (e.g., Hamann & Abed Ibrahim, 2017; Taha et al., 2021; Theodorou et al., 2017) indicates that scoring method does not affect an SR task's ability to distinguish children with DLD from TD children (when looking to achieved sensitivity and specificity). However, other studies (e.g., Pham & Ebert, 2020; Wang et al., 2022) have found that binary scoring methods less reliably distinguish children with DLD from TD children (lower sensitivity and specificity values) when compared to more detailed methods of scoring.

One important consideration in interpreting these findings is the language in which the SR tasks were administered. The studies cited above span multiple languages, including German, Arabic, Greek, Vietnamese, and Mandarin. These languages vary in features such as syntactic structure and morphological richness, which may influence how reliably different measures capture language impairment (Leonard, 2017). For example, languages with rich morphology may benefit more from specific target- or error-based scoring, whereas in languages like English, which is less morphologically complex, binary scoring might be just as effective.

In the meta-analysis by Ward et al. (2024), a subgroup analysis comparing four types of scoring method (binary, subsentence binary, error, and target structure scoring) revealed no effect of scoring decision on the ability of SR to distinguish DLD and TD groups across the studies. The current study directly compares how choice of scoring method influences the diagnostic power of SR for identifying DLD in English, within a single large-scale sample.

Current Study

In the current study, we compare SR performance in DLD and TD groups via a secondary data analysis using data from the SCALES: Intensive Data T2–T5, 2012–2020 (SCALES; Norbury, 2022). SCALES is a longitudinal U.K. population study of language development involving state-maintained schools across Surrey, England. SR performance within SCALES was measured using the School-Aged Sentence Imitation Test (SASIT-E32; Marinis et al., 2011), an SR task designed to assess a range of morphosyntactic and syntactic structures in school-aged children. The SASIT-E32 has been shown to be valuable in identifying DLD in both monolingual and bilingual children (e.g., Chiat & Roy, 2013; Komeili & Marshall, 2013). We compare performance on this measure across four scoring methods: whole sentence accuracy (binary scoring), correct content words, correct function words, and correctly inflected verbs.

Using these data, children identified as meeting criteria for DLD were matched to TD children through age-matching across three separate time points (ages 5–6, 7–8, and 10–11 years). In doing so, we aim to assess differences in SR performance across scoring methods. This study addresses this in two ways. First, Bayesian beta regression models, and specifically Bayes factors, are used to compare SR performance between groups. The primary advantage of this Bayesian approach is that it allows us to quantify the strength of support for both the experimental and the null hypothesis (meaning we would be able to distinguish evidence of the absence of an effect from mere absence of evidence due to sample size). Second, we assess the ability of SR to detect DLD with measures of diagnostic accuracy (sensitivity, specificity, and likelihood ratios [LRs]).

We aim to address the following questions:

1. Do monolingual children with DLD differ from age-matched groups of TD children on SR tasks?

2. What diagnostic accuracy does the SR task achieve?

3. Does the type of scoring used to measure performance on the SR task affect the difference in performance between DLD and TD groups and the diagnostic power of the SR task?

Based on the results of the Ward et al. (2024) meta-analysis, we hypothesize that children with DLD will perform with less accuracy on each method of scoring on the SASIT-E32 in comparison to TD children. We predict that calculated Bayes factors will indicate support for this hypothesis. We also predict that each method of scoring on the SASIT-E32 will show good diagnostic accuracy, consistent with conventional benchmarks (including sensitivity and specificity exceeding .90, a positive LR greater than 3, a negative ratio below .30, and an area under the curve [AUC] of at least .70), in differentiating between children with DLD and TD children.

Method

Our study, including research questions (RQs) and methods of analyses, was preregistered at https://osf.io/ue9qb/. We address our RQs through a secondary data analysis. We use data collected as part of the SCALES (Norbury, 2022), which is publicly accessible through the UK Data Service with identifier https://doi.org/10.5255/UKDA-SN-8968-1. SCALES is a longitudinal U.K. population study of language development involving state-maintained schools across Surrey, England. SCALES protocol and consent procedures were developed in consultation with Surrey County Council and approved by the Royal Holloway Ethics Committee in Year 1 of the study and the University College London Research Ethics Committee (9733/002) in Year 6 of the study.

Data Collection Procedures

Data collection for SCALES occurred in two phases. In the first phase, all state-maintained schools in Surrey were invited to participate in 2011. This resulted in the screening of 7,267 children in reception classes (4 or 5 years of age) across 161 schools. This is referred to as Time Point 1 (T1) of the study. As part of this screening, teachers completed the Children's Communication Checklist–Short (CCC-S; Bishop, 2003a) resulting in a score per child out of 39. A subsample was then selected for in-depth assessment in the second phase of the study using a stratified random sampling process. All children reported to have “no phrase speech” (the teacher responded “no” to the question “Is the child combining words into phrases or sentences?”) were selected for inclusion. The remaining children were stratified by sex, season of birth, and risk for DLD (higher scores on the CCC-S) for selection. Specifically, there was an oversampling of girls and those with CCC-S scores that indicated a low language level. Those from larger schools were also more likely to be selected. Those attending special schools for learning disabilities and those with English as an additional language were excluded.

This resulted in 529 children selected for intensive testing as part of Phase 2 of SCALES. This testing occurred at four different time points—when the children were in Year 1 (T2; 5–6 years old), Year 3 (T3; 7–8 years old), Year 6 (T4; 10–11 years old), and Year 8 (T5; 12–13 years old). T5 data collection, however, coincided with COVID-19 school closings, resulting in greater than expected attrition. Testing involved an assessment battery that included measures of speech and language, nonverbal cognitive skills, motor skills, executive function, speed of processing, social cognition, emotion processing, and literacy. Teachers and caregivers also completed questionnaires at each time point. In the current research, a subset of this data set is used.

Participants

A total of 407 monolingual English-speaking children from Surrey are included in the current research. Specifically, this included 407 children when they were in Year 1 (T2), 391 children when they were in Year 3 (T3), and 294 children when they were in Year 6 (T3). This sample represents a subset of the initial data set of 529 children who underwent intensive testing in Phase 2 of SCALES, based on the following exclusions. Children were excluded if they were reported to have a biomedical condition (based on caregiver and teacher reports) at any of T2, T3, or T4. Additionally, children who failed the audiometric screening test (hearing assessment) at T2 were excluded. All included children scored above −2 SDs on a nonverbal composite ability score at T2, T3, and T4. This composite score was derived from performance on the Block Design and Matrix Reasoning subtests of the Wechsler Preschool and Primary Scale of Intelligence (Wechsler, 2003). Finally, for inclusion at each specific time point, children were required to have SR scores for that time point.

Whether or not children met criteria for DLD differed between SCALES and the current study. For both, the criteria for DLD were based upon core language battery measures assessing vocabulary, grammar, and narrative skills across expressive and receptive language. These included the Expressive and Receptive One-Word Picture Vocabulary Test (EOWPVT and ROWPVT, respectively; Martin & Brownell, 2011), the Test of Reception of Grammar–Short Form (TROG-S; Bishop, 2003b), the Assessment of Comprehension and Expression 6–11 Narrative Recall (ACE-Recall; Adams et al., n.d.), and a narrative comprehension task derived from the ACE-Recall made by the SCALES team (ACE-Comp). These measures were selected to replicate existing epidemiological work (Tomblin et al., 1997) that has informed current diagnostic algorithms for language disorder in the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5). A full outline of the core language battery measures, including their purpose and known reliability, is provided in the supplementary materials of Norbury et al. (2016). DLD criteria within SCALES also included the SASIT-E32. However, as SR performance is the variable of interest in the current research, it is not used for DLD classification here.

Inverse probability weighting was utilized in SCALES to produce weighted z scores of performances (for each of the above core language measures). This weighting process was derived during SCALES (as outlined by Vamvakas et al., 2019) and was conducted to correct for sampling bias (as described earlier), ensuring that the tested sample reflected the full cohort in terms of factors such as gender, school size, and risk for language disorder. These weighted z scores of language performance were used solely for DLD identification.

Five composite scores were created from these measures by averaging the z scores of the individual tasks within each domain: expressive language (EOWPVT and ACE-Recall), receptive language (ROWPVT, TROG-S and ACE-Comp), vocabulary (EOWPVT, ROWPVT), grammar (TROG-S), and narrative skills (ACE-Recall and ACE-Comp). At T4, only four composite scores were created (all but the grammar composite), due to the TROG-S not being administered. At each time point, children were identified as meeting criteria for DLD if they scored 1.5 SDs below the mean on at least two of these composite scores. This threshold aligns with the −1.5 SDs cutoff used in SCALES and approximately mirrors the widely reported 7.6% prevalence estimate for DLD in the U.K. population (Norbury et al., 2016). Children identified as meeting criteria for DLD at any of the time points were classified as DLD across all the time points. This resulted in 67 children meeting criteria for DLD at T2, 66 at T3, and 50 at T4. Children who did not meet this criterion for DLD based upon the core language measures were classified as being TD. Based on this, there were 340 TD children at T2, 325 at T3, and 244 at T4.

The consequences of this decision to classify children meeting criteria for DLD at any one time point as DLD across all the time points were examined through a sensitivity analysis. This involved carrying out the same analysis outlined below with DLD groups where children are only identified as DLD at a time point if they met the criteria at that specific time point. The results of this, including a table summarizing how many children met criteria for DLD at each time point specifically, are outlined in Supplemental Material S1, and it was concluded that this methodological decision did not influence results.

Samples

There were three age-matched groups corresponding to the SCALES data collection time points. The time points are based on U.K. school years, with T2 being Year 1 (ages 5–6 years), T3 being Year 3 (ages 7–8 years), and T4 being Year 6 (ages 10–11 years). Table 1 provides a summary of the sample characteristics for the samples.

Table 1.

Characteristics of the TD children and children identified as meeting criteria for DLD.

Group	Time point 2
	n	Age in months	Male	Female	White ethnic origin (%)	IDACI
TD	340	71.83 (4.93)	162	178	90.88	3.89 (1.12)
DLD	67	71.10 (3.97)	37	30	83.58	3.33 (1.19)
	Time point 3
TD	325	95.38 (4.21)	152	173	90.46	3.87 (1.13)
DLD	66	95.36 (5.26)	37	29	83.33	3.35 (1.18)
	Time point 4
TD	244	134.08 (4.16)	111	133	91.80	3.94 (1.11)
DLD	50	133.18 (3.73)	26	24	86.00	3.32 (1.28)

Note. Age and IDACI are reported as means and standard deviations. IDACI is postcode indicator of neighborhood deprivation, reported here as a quintile rank with higher ranks indicating lower deprivation. Those not defined as “White ethnic origin” were defined as “other ethnic origin.” TD = typically developing; DLD = developmental language disorder; IDACI = Income Deprivation Affecting Children Index.

SR Task

The SR task used within SCALES was the SASIT-E32 (Marinis et al., 2011). The SASIT-E32 consists of 32 sentences and is designed to assess a range of morphosyntactic and syntactic structures, including auxiliaries, passives, object wh- questions, sentence complements, temporal and conditional sentence adjuncts, and object relative clauses. It was developed based on principles from the LITMUS (Language Impairment Testing in Multilingual Settings) framework (Marinis & Armon-Lotem, 2015). Full details of the task can be found here: https://www.citystgeorges.ac.uk/__data/assets/word_doc/0010/118864/SASIT-E32-Protocol,-Structures-and-Scoresheet.doc.

This was administered as per manual instructions. It is a short, structured SR task, taking about 5–10 min to complete. Using a PowerPoint presentation, children listened to 32 prerecorded sentences through headphones, verbally repeating each sentence upon hearing it. Responses were audio-recorded and later scored.

The SASIT-E32 manual outlines multiple scoring options, allowing for flexibility in scoring and level of linguistic analysis. The current study looks at the following four methods of scoring, as used within SCALES:

• Binary scoring: A simple measure of whole sentence accuracy. Each sentence was scored as correct (1) or incorrect (0), with a maximum total score of 32.

• Function word scoring: Each function word was scored as correct or incorrect/omitted. The total number of correctly produced function words was calculated (maximum score of 176). For example, in the sentence, “The sandwich was eaten by the postman,” there would be a total of four function words: “the,” “was,” “by,” and “the.”

• Content word scoring: Each content word was scored as correct or incorrect/omitted. The total number of correctly produced content words was calculated (maximum score of 121). For example, in the sentence, “The sandwich was eaten by the postman,” there would be a total of three content words: “sandwich,” “eaten,” and “postman.”

• Verb inflection scoring: Each verb inflection is scored as correct or incorrect. The total number of correctly inflected verbs was calculated (maximum score of 44). For example, in the sentence, “The sandwich was eaten by the postman,” 1 point would be awarded for the correct inflection of “was.”

Statistical Analysis

All analyses were run using R software (Version 4.3.1; R Core Team, 2023). This study looked at SR performance in those identified as meeting criteria for DLD and TD children.

RQ1: Do Monolingual Children With DLD Differ From Age-Matched Groups of TD Children on SR Tasks?

We first calculated effect sizes using Hedges's g, a measure of standardized mean difference used in the meta-analysis by Ward et al. (2024). This allowed us to compare our results with existing literature. These were calculated for each method of scoring on the SR task for each of the age-matched groups.

To formally examine performance differences, Bayesian beta regression analyses were conducted. A beta regression was chosen due to its flexibility in modeling distributions of values between 0 and 1 (such as proportions). They are flexible to distribution, being able to handle multiple instances of very high or very low scores (such as patterns of floor/ceiling levels of performance) and thus reducing the risk of Type 1 or 2 error. A Bayesian approach was chosen to measure the strength of differences between group performance.

Proportion correct on the SR task was the dependent variable, and group (DLD or TD) was the independent variable. Separate models were run for each method of scoring SR performance and time point. Since the TD and DLD groups were matched for age by school year rather than individually, age was included in the analyses at each time point individually as a potential covariate on the effect. If cases of floor performance (proportion = 0) or ceiling performance (proportion = 1) were present, models were fit to accordingly account for the extreme scores (zero-inflated, one-inflated, or zero–one inflated beta regression).

The Bayesian models were fit using the brmspackage in R (Bürkner, 2017). Bayesian analysis requires us to set priors on our parameters, reflecting our knowledge state before observing the data. Our knowledge is then updated using the data in order to give the posterior parameter estimates. As we do not have any existing identical data, we cannot assume that there is an effect of group, and so, we set the mean of our prior on the group term to be 0. We assume only that we might see values as extreme as those observed in previous work by chance and so set our prior on the group term to be a normal distribution with a mean of 0 and a standard deviation corresponding to the largest plausible group difference in log odds seen in previous work (see Dienes, 2014, for a recommendation of such an approach). This was derived (in log odds) from the meta-analysis conducted by Ward et al. (2024), which reported an average effect size of g = −2.08 (95% confidence interval [CI] [−2.32, −1.84]). We took the upper limit of this CI (−2.32) as the standard deviation of our prior. We have no basis on which to set informative priors on the other parameters (intercept and age), and so, these were set to be wide and uninformative as normal (0,10).

To assess the strength of evidence, Bayes factors were computed for each model using the method of bridge sampling. A Bayes factor of 1 indicates ambiguous evidence; a Bayes factor of > 1 indicates evidence in support of the hypothesis that there is a difference in performance between the DLD and TD groups; and a Bayes factor < 1 indicates evidence in support of the null hypothesis. We interpreted the Bayes factors using guidance by Lee and Wagenmakers (2014), adjusted from Jeffreys (1961).

RQ2: What Diagnostic Accuracy Does the SR Task Achieve?

We calculated diagnostic accuracy measures to evaluate the ability of the SR task to identify the presence or absence of DLD in the groups. These measures were derived from receiver operating characteristic (ROC) curves, using the cutpointrpackage in R (Thiele & Hirschfeld, 2021). The ROC curve plots the true positive rate (sensitivity) against the false positive rate (1 − specificity) across various performance thresholds. The optimal cutoff was identified using Youden's Index as the threshold providing the best balance between sensitivity and specificity, this specifically being the point where the ROC curve was closest to the upper left corner (which would indicate perfect classification).

Sensitivity and specificity values were interpreted using guidelines by Plante and Vance (1994), whereby values of 90% and above demonstrate good classification accuracy, values of 80%–89% demonstrate acceptable classification accuracy, and values below this indicate poor classification accuracy. LRs were then calculated to assess how a change in SR score changes the probability of a child having DLD. LRs were interpreted using recommendations by Dollaghan (2007). LR+ over 10 and LR− under 0.1 would indicate that the SR test could rule in and out DLD with high confidence. LR+ over 3 and LR− below 0.3 would imply a moderate confidence, and LR+ under 3 and LR− over 0.3 would indicate an uninformative test.

AUC is also reported, summarizing the overall accuracy of the SR task. Following guidelines by Hosmer and Lemeshow (2000), AUC values above .9 suggest outstanding discriminative ability, values between .8 and .9 suggest excellent discriminative ability, values between .7 and .8 suggest an acceptable ability, and values below this indicate an unacceptable ability.

Again, as groups were matched by school year, the influence of any residual age differences was assessed. This was done by plotting a separate ROC curve using age as the predictor. The resulting curve was observed to be very close to the diagonal, indicating that age alone did not discriminate between DLD and TD groups at that time point.

RQ3: Does the Type of Scoring Used to Measure Performance on the SR Task Affect the Difference in Performance Between Age-Matched DLD and TD Groups and the Diagnostic Power of the SR Task?

To answer this, Bayes factors were compared for each regression model for each scoring method and time point, assessing the relative strength of evidence for group differences. Interpretations of Bayes factors (as outlined in RQ1) were directly compared to determine whether the level of support for the presence or absence of an effect differed across scoring methods.

To compare the diagnostic performance, we compare AUC values for each scoring method at each time point using the pROC package in R (Robin et al., 2011). We used bootstrap methods with 10,000 bootstrap replicates to test whether one scoring method produced a significantly different AUC compared to another. This helped us to determine whether any of the scoring methods were better at distinguishing DLD from TD children.

Results

Performance on the SR task for each group, expressed as proportion correct across all methods of scoring, is shown in Table 2. As shown, mean scores for some scoring methods, and particularly for the TD group at later time points, began to approach ceiling levels, with proportion correct nearing 1.00.

Table 2.

Summary of SR performance for every matched TD–DLD grouping.

Time point	Binary scoring		Content word scoring		Function word scoring		Verb inflection scoring
	TD	DLD	TD	DLD	TD	DLD	TD	DLD
T2	0.54 (0.24)	0.24 (0.23)	0.92 (0.08)	0.80 (0.12)	0.88 (0.10)	0.71 (0.17)	0.78 (0.15)	0.53 (0.22)
T3	0.71 (0.17)	0.47 (0.24)	0.97 (0.04)	0.91 (0.08)	0.94 (0.04)	0.86 (0.11)	0.89 (0.08)	0.75 (0.17)
T4	0.85 (0.10)	0.70 (0.19)	0.99 (0.02)	0.94 (0.07)	0.97 (0.02)	0.92 (0.08)	0.94 (0.05)	0.86 (0.12)

Note. Performance is reported as group means and standard deviations of raw scores. SR = sentence repetition; TD = typically developing; DLD = developmental language disorder.

In their meta-analysis, Ward et al. (2024) found an average effect size of g = −2.08 (95% CI [−2.32, −1.84]) for the difference in performance between DLD and TD groups across the studies reviewed. Effect sizes resulting from the present study for each group comparison are shown in Table 3.

Table 3.

Summary of effect sizes for SR performance differences between each matched TD and DLD groups.

Time point	Binary scoring	Content word scoring	Function word scoring	Verb inflection scoring
T2	−1.28 [−1.56, −1.00]	−1.33 [−1.60, −1.05]	−1.57 [−1.85, −1.29]	−1.47 [−1.75, −1.19]
T3	−1.34 [−1.63, −1.06]	−1.26 [−1.54, −0.98]	−1.46 [−1.74, −1.17]	−1.44 [−1.72, −1.15]
T4	−1.24 [−1.56, −0.92]	−1.11 [−1.43, −0.79]	−1.13 [−1.45, −0.81]	−1.19 [−1.51, −0.87]

Note. Effect sizes represent a calculation of standardized mean differences using Hedges's g (Hedges, 1981). Effect sizes are reported with 95% confidence intervals. SR = sentence repetition; TD = typically developing; DLD = developmental language disorder.

Group Comparisons

Comparisons of performance between DLD and TD groups on the SR task are visualized in Figure 1, for proportion correct by each of the four methods of scoring. Bayesian beta regression models were fit to compare the difference in performance between the groups. Overall, 12 models were fit for each of the four scoring methods for each of the three age-matched groups. All models were fit with a one-inflated beta regression model due to the presence of ceiling levels of performance in some children. This is except for the comparison of binary scores with children at T2, where the model was fit with a zero–one inflated model due to the presence of both ceiling and floor levels of performance. Parameters on all the final models converged successfully (all R^ values < 1.00), and posterior predictive checks indicated good model fit. Supplemental Material S2 includes time-series plots of Markov chains showing parameter convergence for all models and plots for posterior predictive checks comparing the observed and simulated distributions of task performance.

Figure 1.

Density plot of the difference in SR performance between DLD and TD groups for each method of scoring. Each row represents each group, being those matched at Time Points 2, 3, and 4, respectively. DLD = developmental language disorder; TD = typically developing; p_score = proportion correct on the SR task; SR = sentence repetition.

Table 4 presents the average marginal effects summarizing the difference in SR performance between TD and DLD groups in interpretable terms of proportion correct. A full outline of the posterior parameters for each model is provided in Appendix A. There was found to be no confounding influence of age across any of the time points, with 95% credible intervals for the age parameter overlapping zero across all models.

Table 4.

Average marginal effects of being TD (as opposed to meeting criteria for DLD) on SR performance, along with resulting Bayes factors for the strength of evidence in support of an effect of group on SR performance.

Time point	Scoring method	Average marginal effect	95% CIs	BF 10
T2	Binary	0.25	[0.19, 0.31]	1.01 × 1016
	Content word	0.10	[0.07, 0.14]	1.43 × 1012
	Function word	0.17	[0.13, 0.21]	2.48 × 1019
	Verb inflection	0.23	[0.18, 0.29]	4.17 × 1017
T3	Binary	0.25	[0.19, 0.31]	1.87 × 1018
	Content word	0.06	[0.04, 0.08]	4.82 × 1010
	Function word	0.09	[0.07, 0.12]	2.34 × 1023
	Verb inflection	0.15	[0.11, 0.19]	2.73 × 1019
T4	Binary	0.15	[0.10, 0.21]	1.87 × 1013
	Content word	0.04	[0.03, 0.07]	4.07 × 1012
	Function word	0.05	[0.03, 0.07]	5.31 × 1021
	Verb inflection	0.08	[0.05, 0.11]	5.85 × 1010

Note. Average marginal effects refer to the proportion difference in SR performance from being TD as opposed to meeting criteria for DLD. A positive effect means better SR performance. Bayes factors were calculated using a bridge sampling procedure. TD = typically developing; DLD = developmental language disorder. SR = sentence repetition; CI = confidence interval; BF 10 = Bayes factor value giving the evidence for H₁ over H₀.

At T2 (U.K. School Year 1, ages 5–6 years), the found marginal effects suggest that, on average, children with DLD scored 25% lower when compared to those who were TD using the binary method of scoring, 10% less using content word scoring, 17% less using function word scoring, and 23% less when using verb inflection scoring. At T3 (U.K. School Year 3, ages 7–8 years), the found marginal effects suggest that, on average, children with DLD scored 25% less when compared to those who were TD using the binary method of scoring, 6% less using content word scoring, 9% less using function word scoring, and 15% less when using verb inflection scoring. And at T4 (U.K. School Year 6, ages 10–11 years), the found marginal effects suggest that, on average, children with DLD scored 15% less when compared to those who were TD using the binary method of scoring, 4% less using content word scoring, 5% correct using function word scoring, and 8% less when using verb inflection scoring.

Table 4 also outlines the Bayes factors that resulted from each model for the strength of evidence in support of an effect of group on SR performance. Following interpretation guidance from Lee and Wagenmakers (2014; adjusted from Jeffreys, 1961), there was extreme evidence for there being a difference in SR performance between the age-matched DLD and TD groups. This was true for all time points, and there was no difference in this classification by method of scoring the SR productions.

Diagnostic Accuracy

Table 5 shows the optimal cutoff points for each method of scoring the SR task, along with the resulting AUCs, sensitivity, specificity, and LRs for each of the time points. Figure 2 plots the ROC curves for the scoring methods at each time point.

Table 5.

Diagnostic accuracy metrics for groups.

Time point	Method of scoring	Optimal cutoff	AUC	Sensitivity	Specificity	LR+	LR−
T2	Binary	.31	.82	.75	.80	3.73	0.32
	Content word	.87	.80	.70	.78	3.18	0.38
	Function word	.80	.82	.72	.84	4.35	0.34
	Verb inflection	.68	.81	.78	.76	3.30	0.29
T3	Binary	.63	.79	.77	.68	2.38	0.34
	Content word	.96	.75	.71	.67	2.16	0.43
	Function word	.90	.80	.62	.86	4.47	0.44
	Verb inflection	.84	.80	.68	.75	2.76	0.42
T4	Binary	.78	.76	.72	.73	2.70	0.38
	Content word	.96	.73	.48	.89	4.50	0.58
	Function word	.94	.75	.56	.92	6.83	0.48
	Verb inflection	.89	.76	.58	.82	3.29	0.51

Note. Optimal cutoff points refer to the proportion correct on the task. AUC = area under the curve (calculated from receiver operator curves); LR+ = positive likelihood ratio; LR− = negative likelihood ratio.

Figure 2.

Receiver operating characteristic curves plotting sensitivity by specificity values for the diagnostic accuracy of the sentence repetition task for each group for each method of scoring. Each column represents each group, being those matched at Time Points 2, 3, and 4, respectively.

Time Point 2

At T2 (U.K. School Year 1, ages 5–6 years), all scoring methods resulted in AUC values above .80, indicating that the SR task had an excellent overall ability to distinguish between children with and without DLD (using interpretation guidelines by Hosmer & Lemeshow, 2000). However, when examining individual diagnostic metrics, the picture was more mixed.

Sensitivity (correctly identifying DLD) and specificity (correctly identifying TD) values were above .70 for all scoring methods. No scoring method resulted in both values being above the .80 threshold, typically considered “acceptable” for diagnostic use (Plante & Vance, 1994). Only specificity levels for binary scoring and function word scoring were above .80, showing “acceptable” levels of being able to correctly identify TD children.

Looking at LRs, all scoring methods resulted in LR+ values over 3.00, implying the SR task (at T2) was able to rule in DLD with moderate confidence (Dollaghan, 2007). Only verb inflection scoring resulted in an LR− score under 0.30, which indicates this scoring method was able to rule out DLD with moderate confidence.

To compare scoring method at T2, bootstrap tests were used to compare the AUCs of the ROC curves. No statistically significant difference was found between any of the scoring methods, indicating that the overall diagnostic accuracy of the SR task was not influenced by the choice of scoring.

Time Point 3

At T3 (U.K. School Year 3, ages 7–8 years), the SR task showed excellent overall discriminative ability for function word scoring and verb inflection scoring (AUCs above .80) and an acceptable overall discriminative ability for binary scoring and content word scoring (AUCs above .70). Sensitivity and specificity values for all methods of scoring ranged from .62 to .86. Only the specificity for function word scoring was above the .80 threshold, showing an “acceptable” level of identifying TD children. No pair of sensitivity and specificity were above this .80 level.

For LRs, only the LR+ value for function word scoring was above the 3.00, indicating a moderate confidence for ruling in DLD. No methods resulted in an LR− below .30 for ruling out DLD with moderate confidence.

Bootstrap comparisons of ROC curves for each scoring method revealed the AUC for content word scoring (AUC = .75) was significantly lower than that for binary scoring (AUC = .79; D = −2.25, p = .025), function word scoring (AUC = .80; D = −2.68, p = .007), and verb inflection scoring (AUC = .80, D = −2.22, p = .027). This suggests that content word scoring was less effective at distinguishing DLD and TD groups at T3.

Time Point 4

At T4 (U.K. School Year 6, ages 10–11 years), all scoring methods resulted in AUCs above .70, indicating an acceptable discriminative ability of the SR task. Sensitivity and specificity values for all methods of scoring ranged from .48 to .92. While content word scoring, function word scoring, and verb inflection scoring resulted in specificity values above .80 (an “acceptable” ability to identify TD), this came at a cost to sensitivity, which were all below .60 (a poor ability in identifying DLD). Binary scoring provided a better balance of sensitivity and specificity values, with both being above .70.

For LRs, content word scoring, function word scoring, and verb inflection scoring resulted in LR+ values above 3.00, but again this came as a cost to the corresponding LR− values. No statistically significant difference was found between the AUCs for the different scoring methods at T4, suggesting that the overall level of diagnostic accuracy achieved by the SR task did not differ as a result of the scoring method.

Discussion

This study evaluated the effectiveness of SR in differentiating between English-speaking monolingual children with DLD and TD children in data from a large-scale cohort study. The diagnostic utility of SR tasks was assessed through the calculation of sensitivity, specificity, LRs, and AUCs. Comparisons between methods of scoring the SR task were also made to consider differences in effectiveness.

Group Differences in Performance

Our first RQ asked whether children with DLD performed differently on the SASIT-E32 SR task compared to their TD peers. Consistent with our hypotheses, our findings revealed that children with DLD consistently performed with lower accuracy on the task across each time point. At T2 (ages 5–6 years), those meeting criteria for DLD performed on average 10%–25% worse on the task than their TD peers (depending on the scoring method). At T3 (ages 7–8 years), this difference was 6%–25% depending on the scoring method used. And at T4 (ages 10–11 years), this difference had narrowed slightly to 4%–15% depending on the scoring method used. This reduced difference between the groups at the later time points was likely a reflection of the groups approaching ceiling levels of performance, limiting between-group differentiation.

Crucially, Bayes factors demonstrated extreme evidence for the hypothesis of a difference in performance on the SASIT-E32 between the DLD and TD groups across all scoring methods and time points. This finding is in line with previous literature that suggests that children with DLD have difficulties with SR and will perform with lower accuracy then their peers within a targeted age group (e.g., Archibald & Joanisse, 2009; Conti-Ramsden et al., 2001; Pham & Ebert, 2020; Vang Christensen, 2019). Although the magnitude of the performance gap decreased over time, there remained strong evidence that the SASIT-E32 could consistently result in different performance between DLD and TD groups at each time point. This sustained difference over a 6-year span supports the claim that SR performance varies as a result of DLD and this effect is generally robust.

We calculated effect sizes to enable direct comparisons to those reported in the review by Ward et al. (2024). Effect sizes in the present study ranged from −1.11 to −1.57, which, although smaller than the literature's average (−2.08), still reflect substantial differences in SR performance between DLD and TD groups. These differences in effect sizes could reflect differences in the nature of the DLD samples, as the present study utilized a population-based sample, whereas many studies included in the meta-analysis were drawn from clinical samples. However, note that in the meta-analysis of Ward et al. (2024), there was no evidence that population samples had smaller effect sizes. It should though be noted that the Hedges's g method of effect size calculation assumes a normal distribution and equal variances (Turner & Bernard, 2006). Neither of these properties universally held for our data, with multiple instances of performance approaching ceiling. Therefore, while the calculation of effect sizes allows for easy comparison, there may be bias in its estimation. This supports our use of beta regression modeling, which is better suited for bounded, proportional data.

Diagnostic Accuracy

Our second RQ examined whether SR tasks could effectively distinguish between children with DLD and their TD peers with acceptable diagnostic accuracy. At T2, the SASIT-E32 demonstrated excellent discriminative ability (AUC values between .80 and .82), with sensitivity and specificity ranging between .72 and .84. At T3, there was a reduction in discriminative ability with AUC values ranging from .75 to .80 (acceptable to excellent level), and this reduced again by T4 with AUC values ranging from .73 to .76 (acceptable level). No scoring method achieved both sensitivity and specificity levels above .80 at any time point, a threshold often used as an “acceptable” clinical benchmark for diagnostic accuracy (Plante & Vance, 1994). Across all time points, specificity values tended to be higher than sensitivity, indicating the SASIT-E32 was more effective at correctly identifying TD children than identifying children with DLD.

A similar mixed pattern emerged when examining LRs. At T2, LR+ values exceeded 3.00 across all scoring methods, indicating moderate confidence in the task's ability to “rule in” DLD. However, only the verb inflection scoring method produced an LR− below 0.30, showing moderate confidence in the task's ability to “rule out” DLD. This pattern only became stronger across time points, with LR+ values remaining relatively strong, while LR− values worsened.

While there is research demonstrating that SR can result in high diagnostic accuracy, such as the findings from Christensen and Hansson (2012), Hamann and Abed Ibrahim (2017), Tuller et al. (2018), and Wang et al. (2022), who all reported instances where both sensitivity and specificity values reached 1, the literature is generally mixed. Indeed, nearly 25% of studies reviewed by Ward et al. (2024) that reported such outcomes failed to achieve both sensitivity and specificity values above .80. Similarly, in their meta-analysis comparing marker tasks, Pawłowska (2014) noted that while LRs for SR appeared “promising,” there was no conclusive evidence that SR can reliably diagnose DLD.

There is, then, precedent for the finding that SR falls short of the clinical diagnostic thresholds. Ward et al. (2024) found little evidence that task or experimental design systematically affects group-level performance differences (i.e., a difference in the size of the performance between TD and DLD groups). However, the absence of evidence of an effect in that study does not allow us to rule out the possibility that design features might contribute to variability in diagnostic accuracy (which was not directly studied). For example, of the studies referenced to above, Christensen and Hansson (2012) and Wang et al. (2022) each presented original SR tasks developed for their studies. The first involves 18 sentences developed to assess the use of past tense with specific verbs, and the second involves 16 sentences, half with passives and half with aspect markers as core elements. In contrast, the SASIT-E32 used in the current study comprises 32 sentences spanning a broader range of morphosyntactic and syntactic structures. The SASIT-E32 was developed following principles from the LITMUS framework (Marinis & Armon-Lotem, 2015), similar to the SR tasks used by Hamann and Abed Ibrahim (2017) and Tuller et al. (2018). Such differences in item composition and linguistic complexity might be worth further exploration.

It is also the case that, of the studies that reported very high sensitivity and specificity, all but Wang et al. (2022) drew DLD participants from clinical samples, involving children with preexisting diagnoses and often receiving language intervention. This is in contrast to the current study, which used a population-based sample. These sampling and diagnostic differences could of course also contribute to the variation reported in diagnostic accuracy. However, as Ward et al. (2024) reported, there was no evidence that group-level performance differences (TD vs. DLD) were smaller in population samples than in clinical ones. Indeed, in some cases, population samples showed clearer group distinctions.

From a practical perspective, while we found clear and strong evidence of performance differences between the DLD and TD groups, it is important to understand that this does not necessarily translate into clinical utility. To reliably diagnose DLD, SR tasks like SASIT-E32 should be used in conjunction with other assessment tools and information sources (e.g., case history and observations from different contexts). These considerations align with the diagnostic criteria proposed by Bishop et al. (2016) in the CATALISE-2 consortium, which emphasize the importance of a comprehensive assessment approach, including the use of multiple data sources to diagnose DLD accurately. This does not mean that SR tasks lack clinical value. We argue that SR tasks like the SASIT-E32 may serve as a useful complementary screening tool. Although sensitivity was lower than typically expected for screening purposes, the high specificity observed indicates that children performing below threshold were likely to represent true cases of DLD. This pattern suggests that the measure may work as an initial screening tool for identifying children who warrant further assessment, particularly in large-scale or resource-limited contexts where false positives could place undue strain on services. Nonetheless, the moderate sensitivity indicates that some children with DLD may not be detected, highlighting the need for future work to examine whether combining this measure with complementary indicators such as parent reports could enhance overall screening efficiency.

We observed a slight decline in diagnostic accuracy across the time points, with accuracy generally being strongest at T2 compared to T3 and then T4. This might suggest that, clinically, the SASIT-E32 is more effective when used on younger children. The structures targeted by the SASIT-E32 may still be in the process of being acquired in early childhood but may be mastered by older children and even by those with DLD. As a result, the task may become less sensitive to impairment. It is also possible that because scores on the task began to approach a ceiling level of performance as children got older, there was simply less room to differentiate performance between the groups.

This trend may be specific to the difficulty level of the SASIT-E32. A more challenging SR task containing more complex structures might maintain diagnostic accuracy in older children, and there is evidence that SR can be useful for identifying DLD in adults (e.g., Poll et al., 2016). This highlights the importance of aligning the linguistic complexity of a SR task with the target age group to achieve optimal diagnostic accuracy, particularly in contexts where standardized tests of age norms are not available. This pattern of decreasing diagnostic accuracy also reflects broader challenges in identifying reliable screening tools for school-age populations. For example, Bao et al. (2024) note the scarcity of tools with robust diagnostic accuracy for this age range. While our findings show moderate utility of SR tasks as a marker of impairment, the reduction in accuracy over time is consistent with the challenges identified in the broader literature for maintaining diagnostic precision in older children. This emphasizes the importance of further developing and refining SR tasks for this age group.

Methods of Scoring

Our third RQ asked whether the different scoring methods used with the SASIT-E32—binary scoring, function word scoring, content word scoring, and verb inflection scoring—would produce different outcomes. Across all time points, content word scoring yielded the highest levels of child SR performance (in terms of proportion correct) compared to the other methods. Performance on this measure approached ceiling across the time points, with mean proportion correct ranging from 92% to 99% for TD groups and from 80% to 94% for DLD groups. In contrast, the lowest child SR performance levels were observed with the binary scoring method, with mean proportion correct ranging from 53% to 85% for TD groups and from 24% to 70% for DLD groups. Despite these differences, all methods of scoring resulted in lower performance in the DLD groups compared to TD groups, and high Bayes factors indicated strong evidence for a group difference. Likewise, diagnostic accuracy did not vary meaningfully in most cases.

At T3, content word scoring resulted in a significantly lower AUC (.75) than the other three scoring methods. Although not statistically significant, content word scoring also produced the lowest AUC at both T2 and T4. These findings align with Seeff-Gabriel et al. (2010), who reported that both TD children and children with language impairments performed best on content word scoring when compared to function word and inflection scoring. They similarly observed the smallest group differences observed for content word scoring. This pattern may be due to children with DLD having more pronounced difficulties with function words and inflections (Chiat, 2001). However, again, the ceiling effects observed with content word scoring especially may have limited the task's ability to differentiate between groups for this method of scoring.

In practice, scoring on a detailed level (such as scoring function words/inflections) may not be practical, given the scoring time and training required. Binary scoring provides a simpler alternative, being the quickest and easiest method to administer and requiring little training to apply reliably (Ebert et al., 2019), and, as our results show, retains comparable diagnostic accuracy to more detailed methods. Indeed, in the literature, binary scoring has often been recommended as being most useful for application within a clinical context (e.g., Pham & Ebert, 2020; Taha et al., 2021), due to its efficiency and scalability. In contrast, detailed scoring has been used and recommended in studies aiming to detect specific linguistic deficits or to gain finer-grained profiles of children's language abilities (e.g., Christensen & Hansson, 2012; Hamann & Abed Ibrahim, 2017). Although a full cost–benefit analysis comparing the feasibility of different scoring approaches would be valuable, such data are not currently available. Nonetheless, these distinctions do reinforce that different scoring methods may serve different goals, and the best approach depends on whether the priority is broad screening or research precision.

It is important to note that the task of picking a scoring method by weighing up the additional time and training against linguistic insight gained may soon be reframed by technological advances. Progress in automated speech recognition and natural language processing means it may soon be feasible to automate the scoring of SR tasks. Automated scoring systems are already being developed and trialed (e.g., Albudoor & Peña, 2022; Liu et al., 2024) and show promise for producing reliable, fine-grained scoring outputs with minimal input from clinicians. As these tools become more widely available and accurate, the time and training barriers currently limiting the use of detailed scoring may be significantly reduced. In this context, more detailed scoring may become a viable and efficient option in the future, and thus any performance advantages they offer may become more easily attainable.

Strengths and Limitations

The current study involved an analysis of the SCALES data set (Norbury, 2022), which allowed for not only a large sample of TD children but, because there was an oversampling at those “at risk” for language disorder, a large DLD sample. It also allowed for a one-gate design to be used, whereby children identified as meeting criteria for DLD came from the same sample as those who were TD, which Dollaghan and Horner (2011) note is important for studies assessing diagnostic accuracy. Additionally, all children were recruited from schools within the same authority, which may have offered some consistency in language exposure.

However, the SCALES sample, in sampling school pupils from Surrey, was likely more affluent than the national average (Norbury et al., 2016). Within the work by Norbury et al. (2016), a significant association was found between socioeconomic background and language disorder (classified using the methods described earlier in this article), with children from lower socioeconomic backgrounds being more likely to meet criteria for DLD. The underrepresentation of lower socioeconomic neighborhoods in the sample limits the generalizability of the current findings.

Finally, while using primarily the same core composite measures and classification criteria as the SCALES project, classification of DLD in the current study differed slightly in that the SASIT-E32 was excluded. In contrast, the original SCALES classification included the SASIT-E32 as one of the core measures. This reclassification ensured that SR performance was not part of the criterion, but it may have led to small discrepancies in group membership relative to the original SCALES groupings (e.g., some children misclassified as TD or DLD).

A strength of the current study was the use of a Bayesian analytical framework, which allowed us to assess the strength of evidence for both the presence and absence of effects. This is important in language research, where small sample sizes and reliance on null-hypothesis significance testing can make it difficult to distinguish true absence of an effect from insufficient statistical power. The Bayesian approach also provides an informative measure of uncertainty by estimating the probability of different effect sizes given the data.

Implications

These findings have important implications for both research and clinical practice. SR tasks such as the SASIT-E32 can provide a valid and flexible tool for detecting group-level differences in language ability. Our results showed consistent performance differences between DLD and TD groups across multiple scoring methods, supporting SR as a promising marker of language difficulty. It reinforced binary scoring as a practical and valid option clinically and in research.

Clinically, while the SASIT-E32 did not meet thresholds for standalone diagnosis, its strong specificity and LR+ values do suggest value as a screening tool. The task was most accurate in younger children, highlighting its potential role in early identification. Given its efficiency and potential for automation in the future, SR may serve as a useful component of a broader assessment battery in both research and applied settings.

Conclusions

This study provides promising evidence that SR tasks, specifically the SASIT-E32, consistently differentiate performance between English-speaking children with DLD and TD children. While the task's diagnostic accuracy did not consistently meet clinical thresholds for standalone use, its efficiency, ease of administration, and adaptability make it a valuable tool in both research and screening contexts. The flexibility to select scoring methods based on specific needs, along with the potential to automate scoring, further enhances the adaptability and utility of SR as a practical and scalable approach to assessing language ability.

Data Availability Statement

The data set is publicly available (http://doi.org/10.5255/UKDA-SN-8968-1). A UK Data Service account is required to download the data set. Some variables are not archived, and the weighted z scores for each core language battery measure for T2, T3, and T4 were provided by Courtenay Norbury.

Appendix A

An Overview of All the Posterior Parameters From Each of the Bayesian Beta Regression Models Fit to Compare SR Performance Between All the Age-Matched Groups Across the Four Scoring Methods. Raw Parameters Are Provided for Reproducibility Purposes

Table A1.

Summary of raw posterior parameters outputted from Bayesian beta regression models fit to compare SR performance between all the age-matched groups across the four scoring methods.

	Time point 2
Parameter	Binary scoring	Content word scoring	Function word scoring	Verb inflection scoring
Intercept	−0.93 [−1.19, −0.65]	1.45 [1.26, 1.64]	0.92 [0.72, 1.12]	0.18 [−0.03, 0.40]
diagnosisTD	1.06 [0.77, 1.34]	0.91 [0.69, 1.13]	1.10 [0.88, 1.31]	1.06 [0.83, 1.31]
age	0.23 [0.14, 0.32]	0.18 [0.10, 0.26]	0.19 [0.12, 0.26]	0.21 [0.14, 0.29]
zoi_Intercept	−2.04 [−2.82, −1.38]	−6.43 [−11.15, −3.45]	−7.14 [−12.93, −3.62]	−6.85 [−12.21, −3.56]
zoi_diagnosisTD	−2.55 [−3.90, −1.30]	3.90 [0.90, 8.73]	2.13 [−1.67, 7.91]	2.46 [−1.01, 7.87]
coi_Intercept	−10.64 [−27.57, −1.89]
coi_diagnosisTD	10.62 [0.94, 28.34]
phi_Intercept	1.23 [0.87, 1.56]	2.19 [1.84, 2.50]	1.82 [1.49, 2.13]	1.41 [1.09, 1.69]
phi_diagnosisTD	0.14 [−0.21, 0.52]	0.48 [0.12, 0.86]	2.13 [−1.67, 7.91]	0.62 [0.29, 0.97]
	Time point 3
Intercept	−0.17 [−0.40, 0.07]	2.19 [1.95, 2.41]	1.76 [1.56, 1.97]	1.04 [0.83, 1.23]
diagnosisTD	1.07 [0.82, 1.32]	0.93 [0.68, 1.20]	1.04 [0.82, 1.26]	1.03 [0.81, 1.25]
age	0.09 [0.01, 0.17]	0.10 [0.01, 0.19]	0.08 [0.01, 0.14]	0.07 [−0.01, 0.14]
zoi_Intercept	−7.17 [−12.83, −3.65]	−2.54 [−3.51, −1.72]	−7.02 [−12.51, −3.65]	−6.54 [−11.56, −3.50]
zoi_diagnosisTD	1.63 [−2.24, 7.16]	1.35 [0.46, 2.37]	2.04 [−1.52, 7.55]	3.58 [0.50, 8.57]
phi_Intercept	1.21 [0.90, 1.49]	2.54 [2.13, 2.90]	2.38 [2.02, 2.72]	1.81 [1.46, 2.12]
phi_diagnosisTD	0.66 [0.34, 1.02]	0.90 [0.49, 1.35]	1.22 [0.85, 1.63]	0.97 [0.60, 1.36]
	Time point 4
Intercept	0.82 [0.58, 1.07]	2.62 [2.30, 2.93]	2.47 [2.22, 2.71]	1.78 [1.53, 2.02]
diagnosisTD	0.89 [0.63, 1.15]	1.09 [0.77, 1.44]	0.90 [0.65, 1.17]	0.88 [0.62, 1.14]
age	0.05 [−0.03, 0.13]	0.06 [−0.04, 0.15]	0.04 [−0.03, 0.10]	0.04 [−0.03, 0.12]
zoi_Intercept	−6.38 [−11.16, −3.21]	−1.69 [−2.47, −0.95]	−6.27 [−11.36, −3.13]	−4.07 [−6.47, −2.45]
zoi_diagnosisTD	3.02 [−0.24, 7.83]	1.15 [0.35, 2.02]	3.48 [0.32, 8.57]	1.82 [0.14, 4.25]
phi_Intercept	1.60 [1.22, 1.98]	2.69 [2.19, 3.15]	2.80 [2.36, 3,19]	2.32 [1.89, 2.71]
phi_diagnosisTD	0.92 [0.51, 1.34]	1.60 [1.08, 2.15]	1.81 [1.37, 2.27]	1.12 [0.70, 1.58]

Note. Values show estimates of the parameters surrounded by 95% confidence intervals. The parameters resulting from a model depend on whether a beta regression was fit, or whether a zero-, one-, or zero–one-inflated beta regression was fit. Intercepts represent the central tendencies (on the link scale) of the main response variable (proportion correct), zero–one inflation (zoi), continuous-one inflation (coi), zero-inflation (zi), or precision (phi). Covariate effects represent the effects of predictors (diagnosis (TD/DLD) or age) on the corresponding intercepts. Rhat values for all parameters was 1.00. SR = sentence repetition; TD = typically developing; DLD = developmental language disorder.

Funding Statement

Funding for the Surrey Communication and Language in Education Study was provided by Wellcome Trust (WT094836AIA) and Economic and Social Research Council (ES/R003041/1) grants to Courtenay Norbury. This article was completed as parts of Leah Ward's PhD research at The University of Manchester. Sponsorship and funding for the PhD and all resulting publications is provided by the UK Research and Innovation Economic and Social Research Council.