Predictors of Treatment Response for Preschool Children With Developmental Language Disorder

Abstract

Purpose

Enhanced Conversational Recast treatment is an effective intervention for remediating expressive grammatical deficits in preschool-age children with developmental language disorder, but not all children respond equally well. In this study, we sought to identify which child-level variables predict response to treatment of morphological deficits.

Method

Predictor variables of interest, including pre-intervention test scores and target morpheme production, age, and mother's level of education (proxy for socio-economic status) were included in analyses. The sample included 105 children (M = 5;1 [years;months]) with developmental language disorder who participated in 5 weeks of daily Enhanced Conversational Recast treatment. Classification and regression tree analysis was used to identify covariates that predicted children's generalization of their trained grammatical morpheme, as measured by treatment effect size d.

Results

Our analysis indicates that the Structured Photographic Expressive Language Test–Preschool 2 (SPELT-P 2) scores and the Peabody Picture Vocabulary Test–Fourth Edition scores significantly predicted the degree of benefit a child derived from Enhanced Conversational Recast treatment. Specifically, a SPELT-P 2 score above 75 (but still in the impaired range, < 87) combined with a high Peabody Picture Vocabulary Test–Fourth Edition score (> 100) yielded the largest treatment effect size, whereas a SPELT-P 2 score below 75 predicted the smallest treatment effect size. Other variables included in the model did not significantly predict treatment outcomes.

Conclusions

Understanding individual differences in response to treatment will allow service providers to make evidence-based decisions regarding how likely a child is to benefit from Enhanced Conversational Recast treatment and the expected magnitude of the response based on the child's background characteristics.

Conversational recasts are repetitions of a child's utterance that maintain the child's meaning while changing or adding grammatical information (Baker & Nelson, 1984). Recasts occur naturally in adult–child interactions with both typically developing children and children with developmental language disorder (DLD; children whose language fails to develop normally in the absence of other handicapping conditions). However, parents who are interacting with children with DLD tend to recast less frequently compared to parents talking with typically developing children (Conti-Ramsden, 1990; Conti-Ramsden et al., 1995; Nelson et al., 1995; however, see also Fey et al., 1999; Proctor-Williams et al., 2001). Additionally, children with DLD require a higher rate of recasting relative to children with typical language to benefit (Proctor-Williams & Fey, 2007). Proctor-Williams & Fey (2007) provided conversational recast intervention to language-matched children with and without DLD at varying recast densities in order to teach them novel irregular past tense verbs. Typically developing children benefited most from the “conversational” rate of 0.2 recasts per minute, whereas the high-density rate of 0.5 recasts per minute was insufficient for the DLD participants, leading the researchers to suggest that even more frequent recasts are required for positive responses in children with impaired language. Thus, although children with DLD can utilize conversational recasts, they are likely receiving fewer recasts in their natural linguistic environment than are required to be effective. However, within the therapy context, the dose of conversational recasts can be optimized to suit the needs of children with DLD.

Conversational recasting has been identified as an effective therapy approach for increasing the complexity and/or accuracy of expressive morphology and syntax in children with DLD and children with typical language development. Most conversational recast studies have reported positive results for groups of children who receive this treatment (Cleave et al., 2015). However, effectiveness at the group level does not always translate to effectiveness at the individual level, as some children do not demonstrate significant language improvement following recast therapy. Determining which children will likely benefit from conversational recasting versus those who will not is useful information for clinicians to have at the outset of treatment in order to make informed decisions regarding whether a particular child should receive conversational recast therapy versus an alternative treatment.

Effectiveness

In a systematic review and meta-analysis, Cleave et al. (2015) concluded that recast intervention is effective based on the fact that the majority of studies reported positive effects of recast intervention, significant positive correlations have been established between parents' use of recasts and children's language abilities, and recasting leads to more and/or earlier spontaneous target productions compared to imitation therapy (treatment in which children are asked to repeat a clinician model). The average effect size for early efficacy treatment studies was d = 0.96 (i.e., posttreatment increase of SD = 0.96). These are treatment studies designed to establish a cause–effect relation between treatment and response, and included proximal outcome measures that were similar to the treatment itself (Fey & Finestack, 2009). All of the early efficacy studies reported positive effects of the intervention. The average effect size for later efficacy and effectiveness treatment studies utilizing recasting was d = 0.76, and again, all studies reported positive effects. Like early efficacy studies, later efficacy studies are designed to establish cause–effect relations, but typically use conditions that are closer to typical clinical practice and include larger sample sizes and include distal outcome measures (Fey & Finestack, 2009). Effectiveness studies seek to understand the extent to which treatment methods are successful under real-world conditions.

Importantly, the positive effects of conversational recast treatment last beyond the treatment period, indicating the treatment produced real rather than transient changes. Several studies of conversational recast treatments have provided follow-up data (Eidsvåg et al., 2019; Leonard et al., 2004, 2006; Meyers-Denman & Plante, 2016; Plante et al., 2019). In each of these studies, children were treated for a fixed time, rather than trained to criterion. Treatment was then discontinued and children's use of the linguistic features targeted by conversational recast was reassessed later (between 1- and 2-month delay). Regardless of their relative mastery of the linguistic target trained immediately postintervention, children tend to retain the level of performance that they demonstrate at the end of treatment following a delay.

Despite the overall success of this treatment method, it is not the case that all versions of conversational recast therapy are equally effective. Cleave et al. (2015) reported that focused recasts that were restricted to changing the child's utterance to reflect a specific targeted grammatical form were more effective compared to broad recasts in which adults recast a variety of the child's errors without specific focus on a targeted form. A repeated finding for conversational recast treatment is that the effects are specific to the element(s) of language targeted by the recasts and do not extend more broadly to untreated aspects of the child's language (Eidsvåg et al., 2019; Leonard et al., 2004, 2006; Meyers-Denman & Plante, 2016; Plante et al., 2019, 2014, 2018).

The form of the recast itself also affects treatment outcomes either positively or negatively. Fey and Loeb (2002) found that recasts that were in the form of yes/no questions did not have significant effects on modal or auxiliary production for participants with DLD or those with typical language abilities. The authors concluded that question recasts may have been too grammatically complex for the participants to derive benefit. Plante et al. (2014) demonstrated that input that was highly linguistically varied, such that the morpheme targeted for each child was the only consistently occurring aspect of the clinician's recasts (i.e., 24 linguistically unique recasts), yielded better outcomes compared to a condition in which the child heard the same recasts repeated just once each (i.e., 12 unique recasts). Little progress was seen in the group who received less linguistically varied recasts. Likewise, Owen Van Horne et al. (2017) varied the recasts provided to children with DLD based on the relative complexity of verbs. Complex verbs were those with higher frequency of the uninflected form, higher phonological complexity, and lower telicity. Using recasting combined with imitation and modeling, the researchers found that participants who received therapy with complex verbs had higher posttreatment accuracy on trained and untrained verbs relative to participants who received the same treatment with less complex verbs.

In contrast to these findings, other aspects of recasts have been manipulated and found to have no effect on treatment outcomes. For example, recast effectiveness is not influenced by whether recasts are provided in response to the child's ungrammatical utterances (corrective recast) versus grammatical utterances (noncorrective recasts). Additionally, there are no significant differences in effectiveness of recasts provided in response to a child's spontaneous utterance as compared to prompted utterances (Cleave et al., 2015; Hassink & Leonard, 2010), nor does the spacing of recasts over time appear to influence treatment effectiveness (Meyers-Denman & Plante, 2016; Plante et al., 2019), as long as a therapeutic number of recasts are provided.

Additionally, evidence suggests that the positive effects of conversational recast therapy may be enhanced when combined with other therapy techniques. For example, Plante et al. (2018) reported that the addition of auditory bombardment of target morphemes immediately following recast treatment resulted in a higher percentage of children responding positively to the treatment relative to auditory bombardment immediately preceding recast treatment. Likewise, the results of Smith-Lock et al. (2015) support effectiveness of recasting within the context of hierarchical cuing when combined with additional clinician responses including request for clarification, imitation of the child's error, forced choice alternatives, and child imitation of the correct target.

In sum, focused conversational recasting that targets a specific language form is an effective therapy that leads to sustained effects on target production. This treatment can be further enhanced via manipulating properties of the linguistic input (e.g., variability and verb complexity) and can be successfully combined with other therapy approaches (e.g., auditory bombardment and cuing) to maximize outcomes.

Predictors of Therapy Outcomes

Beyond establishing the effectiveness of conversational recast treatment, researchers have sought to identify individual child characteristics that may make the therapy approach more or less successful for individual children. Yoder et al. (2011) assigned half of their preschool participants with DLD to complete broad recasting therapy and the other half to milieu language teaching and then assessed the relation between children's pretherapy language complexity, as measured by mean length of utterance (MLU), and their response to language therapy. Those children with lower MLUs (< 1.84) benefited more from milieu language teaching, whereas those participants with MLUs above 1.84 showed equivalent outcomes regardless of therapy condition, which suggests that conversational recast therapy, specifically broad recasting, may be better suited for children with more complex language abilities prior to therapy. However, because broad recast is a less effective version of conversational recast treatment (Cleave et al., 2015), it is unclear whether the same relationship between MLU and language outcomes would result if a more effective version of conversational recast therapy such as targeted recasting were used.

Pawłowska et al. (2008) considered the extent to which children's successful production of the third person singular –s morpheme following conversational recast therapy could be predicted by their age, their pretherapy plural –s marking accuracy, and their production of subject and verb constructions prior to therapy. All three variables were significant positive predictors of response to therapy, although plural marking and subject/verb construction did not account for unique variance when either was entered into the model after the other. The findings suggest that older children and those with more advanced grammatical agreement abilities as evidenced by plural marking and subject/verb constructions showed larger morphosyntactic gains in response to therapy.

Conversely, other research has found no relationship between child-level characteristics and response to conversational recast therapy. For example, Leonard et al. (2004) reported no correlation between children's age and their response to conversational recast therapy, and splitting the participants into those with relatively low receptive vocabularies (Peabody Picture Vocabulary Test–Third Edition [PPVT-III; Dunn & Dunn, 1997] < 85) and those with higher vocabularies (PPVT-III > 85) resulted in no group differences in treatment outcomes.

Present Study

The purpose of treatment research is the identification of therapy techniques that are effective. However, these approaches, which focus on group-level outcomes, largely overlook individual differences in response to treatment. Indeed, even within a study that reports a particular treatment/therapy to be effective, there are likely to be individual participants for whom the treatment was ineffective. Based on previously published recasting research, authors have reported a wide range of percentages of positive treatment responders from 12% at the low end (Smith-Lock et al., 2015) to 100% at the high end (Nelson et al., 1996). Although, these studies use different outcome measures, making direct comparisons difficult, it is evident from previous work that positive treatment response for all participants is rare, regardless of how positive responding is defined. Identifying which individuals would and would not benefit from a therapy has clear clinical relevance as it could allow clinicians to select the best treatments for an individual. The purpose of the current study is to determine which variable or variables significantly predict children's response to Enhanced Conversational Recast treatment, a version of conversational recast treatment that requires high-input variability (Plante et al., 2014) and attentional cuing at the time the recast is delivered (Meyers-Denman & Plante, 2016). Additional details about Enhanced Conversational Recast treatment and methodology used for the children who participated in the treatment research included in the current study is described in more detail in the Method section.

Our goal is to create a data-driven decision tree that can be used to guide treatment decisions based on a child's background characteristics. Yoder and Compton (2004) delineate three levels of possible predictors of treatment response: child, parent, and community. Here, we investigate whether the child-level variables of pretreatment performance on standardized measures of language (the Structured Photographic Expressive Language Test–Preschool 2 [SPELT-P 2], Peabody Picture Vocabulary Test–Fourth Edition [PPVT-4]), articulation (the Goldman Fristoe Test of Articulation—Second Edition; the Goldman Fristoe Test of Articulation—Third Edition), and nonverbal intelligence (the Kaufman Assessment Battery for Children, Second Edition [KABC-II]), pretreatment target morpheme production accuracy, age, and/or sex significantly predicted changes in participants' production of targeted morphemes following Enhanced Conversational Recast treatment. We also assess the potential predictive role of maternal education, which serves as a proxy for socioeconomic status (Hoff & Tian, 2005), and may also indirectly reflect parent interaction style (Yoder & Compton, 2004). These specific predictors were chosen because they were available for a large sample of children who have undergone Enhanced Conversational Recast treatment and because of the likelihood that clinicians could obtain this information via standardized testing and/or parent report in order to implement these findings in practice, which is our ultimate goal.

Because the analytic approach we apply to our data is exploratory and data-driven, we did not make specific hypotheses because we were creating data-driven models as opposed to using the data to test hypothesized models. However, some variables were expected to be linked with children's response to therapy based on previous research. Age was anticipated to positively predict outcomes (Pawłowska et al., 2008), whereas receptive vocabulary was not expected to be related to outcomes (Leonard et al., 2004). Based on Yoder et al. (2011) one would expect that expressive language abilities, indexed by the SPELT-P 2, would be positively associated with therapy outcomes since children with higher MLUs showed better response to recasting. Less is known about the relationship between other variables included in our model (e.g., nonverbal IQ, articulation, and socioeconomic status) and response to language therapy, so we had no clear expectations about these variables predicting outcomes.

Understanding which child-level variables predict response to treatment will make treatment more effective by identifying children who are both likely and unlikely to benefit from Enhanced Conversational Recast treatment. This will allow service providers and parents to make evidence-based decisions regarding whether this treatment is the best option for a given child.

Method

Participants

One hundred seven children (43 girls) with DLD participated in Enhanced Conversational Recast language treatment for morphological deficits. Children were between the ages of 4;0 (years; months) and 6;4 (M = 5;1) at the time of enrollment. All children were native English speakers, and no child was receiving outside language therapy during the time of these studies. Parents or legal guardians of all children provided informed consent and the research was approved by The University of Arizona Institutional Review Board. Of the original sample of 107, two participants (both female) were excluded from our analyses because maternal education, which is one of the predictors in our models, was unreported. This resulted in a final sample of 105 children included in our models.

Previous publications have reported on treatment outcomes from 81 of these 105 participants (nine children in Plante et al., 2014; 16 children in Meyers-Denman & Plante, 2016; 28 children in Plante et al., 2018; 10 children in Eidsvåg et al., 2019; and 18 children in Plante et al., 2019). Therapy outcome data for 24 of the participants have not been published previously. Previous publications have focused on effects of therapy manipulations on treatment outcomes and none have examined the role of child characteristics in predicting therapy outcomes, which is the focus here. All participants received Enhanced Conversational Recast treatment. For details about the variations of Enhanced Conversational Recast treatment included in these data, refer to Appendix.

Procedure

Standardized Testing

All children were assessed by a certified speech-language pathologist with extensive experience evaluating preschool age children and met the criteria for DLD diagnosis. Consistent with Bishop et al. (2016, 2017), all participants demonstrated language impairment not associated with other handicapping conditions, such as hearing impairment or cognitive impairment. Specifically, all children passed pure-tone hearing screening (20 dB at 1000, 2000, and 4000 Hz) and had standard scores above the range for intellectual disability with a minimum score of 75 (70, + SEM = 1) on the Nonverbal Index of KABC-II (Kaufman & Kaufman, 2004). All participants also scored below 87, an empirically derived cutoff score indicative of language impairment (Greenslade et al., 2009), on SPELT-P 2 (Dawson et al., 2005).

Single-word vocabulary knowledge was assessed with PPVT-4 (Dunn & Dunn, 2007). Speech production was assessed using the Goldman Fristoe Test of Articulation-Second Edition (Goldman & Fristoe, 2000) or the Goldman Fristoe Test of Articulation–Third Edition (Goldman & Fristoe, 2015), primarily to ensure that children had sufficient articulatory skills to produce the target and control morphemes that served as targets. Demographic information and standardized test scores for all participants are presented in Table 1.

Table 1.

Demographic and treatment information for participants.

Variable	Age	Mother's education a	KABC-II	SPELT-P2	PPVT-4	GFTA	Pre-Tx target use (%) b	Post-Tx target use (%) b	Target Tx effect size d c	Pre-Tx control use (%) b	Post-Tx control use (%) b	Control Tx effect size d c
Mean	61.09	14.02	100.14	65.74	93.38	73.15	6.9	54.4	5.48	7.0	8.0	−0.10
SD	6.24	1.68	12.44	13.29	12.80	17.58	7.4	36.6	6.29	8.3	18.0	2.69
Min.	48	11	77	25	62	40	0	0	−3.81	0	0	−5.19
Max.	76	17	132	87	129	113	33.3	100	17.33	30.0	100	17.33

Note. KABC-II = nonverbal index of the Kaufman Assessment Battery for Children–Second Edition; SPELT-P2 = Structured Photographic Expressive Language Test-Preschool, Second Edition; PPVT-4 = Peabody Picture Vocabulary Test–Fourth Edition; GFTA = Goldman-Fristoe Test of Articulation, Second and Third Editions; Tx = Enhanced Conversational Recast treatment. All standardized test scores are scaled with a mean of 100, SD of 15.

^a Expressed in years with 17 representing a postgraduate degree.

^b Percentage use averaged across 3 days.

^c Calculated as (posttreatment use−pretreatment use)/SD of posttreatment use.

General Treatment Context

Children participated in a half-day preschool program at a University of Arizona clinic 5 days per week for 5 consecutive weeks. The preschool program facilitated daily attendance because it permitted parents to leave their children in the program for a half day. As a result, parents remained blind to the children's treatment targets. Preschool staff was not aware of the treatment targets and did not address or correct children's grammar. The treatment consisted of approximately 30 min of individual language therapy each day in a separate therapy room. Treatment addressed the single morphological target selected for each child throughout the duration of the program.

Pretreatment Probes (Baseline)

Children participated in baseline probe sessions the first 3 pretreatment days. During these sessions, up to six grammatical forms were elicited at least 10 times each. The set of morphemes probed was determined based on errors observed in informal conversational speech and test responses for each child. Therefore, the morphemes probed reflected each child's errors rather than a predesignated list of possible morphemes. A narrower set of potential treatment targets was then identified based on grammatical forms the child produced with low accuracy levels during the three probe sessions. Low accuracy was operationalized as 33% or less on all of the 3 probe days, although actual baseline performance was much lower for most children (M = 6.9%, SD = 7.4%). From the set of grammatical forms with low accuracy, two were selected. These were typically the two with the lowest pretreatment use. One was assigned to serve as the control to be tracked throughout the intervention period and the other served as the targeted morpheme in treatment (see Table 2 for a list of target morphemes and the number of participants in the current study assigned each target form). The control morphemes are not analyzed in the current study. Treatment targets (and control forms) had to include phonemes that were within the phonological inventory of the child to assure they could be produced.

Table 2.

Number of participants assigned each morpheme as target of Enhanced Conversational Recast treatment.

Target morpheme	Participants (n)
Regular past tense –ed	39
Third person singular –s	30
Auxiliary is or are	11
Present participle –ing	8
Singular nominative pronoun he or she	7
Singular negation doesn't	4
Possessive –'s	2
Third person singular has	2
Wh-questions	2

Enhanced Conversational Recast Procedures

Treatment for all children consisted of Enhanced Conversational Recast treatment, a variation of conversational recast therapy with previously established efficacy (Plante et al., 2014). Conversational recasts were defined as clinician utterances that reflected a child's utterance in which use of the morpheme targeted in treatment was obligated. The clinician recast followed this child utterance immediately and retained the content while providing a correct grammatical model of the target morpheme. Recasts could follow an elicited or spontaneous child utterance as previous research has indicated both are effective (Hassink & Leonard, 2010) and could follow a correct or incorrect child utterance.

This variation of conversational recast therapy includes three key features: focused recasts, high variability, and attentional cues. Focused recasts occur when a single grammatical form is targeted throughout the intervention. The clinicians chose to recast child utterances that were most likely to be attempts at the target form and did not provide recasts of nontarget morphology.

The second feature, “high variability,” was provided two ways. High variability in the linguistic context was created by providing 24 “linguistically unique” recasts (Plante et al., 2014). When verb morphology was the target, unique recasts were constructed using 24 different lexical verb stems featuring target morphology in each treatment session. When the target was a pronoun, possessive, a free morpheme (e.g., pronouns), or contracted negative, high variability was achieved by assuring that the words proceeding and/or following the target morpheme were unique (e.g., 24 different verbs for pronoun she + verb constructions; 24 different nouns for possessive targets). High variability was also created by using a wide variety of treatment activities (typically three different activities per session) and infrequently repeating an activity within the 25-day treatment period. A variety of activities, such as games, art projects, book reading, and playing with toys, facilitated high variability in the linguistic context as different activities naturally promoted use of different vocabulary.

The third feature is the use of “attentional cues”—methods such as light touch, a short verbal direction such as “look at me,” or moving in front of the child to establish eye contact for the recast. For the majority of the participants reported on here (i.e., Eidsvåg et al., 2019, Meyers-Denman & Plante, 2016; Plante et al., 2018, 2019), clinicians were required to use attentional cues prior to providing each recast. The attentional cues were used along with techniques to promote child attention such as following the child's lead, encouraging the child's participation, and incorporating motivating activities in the treatment session. Cues were individualized for a child depending on how the child best responded.

Intensity Parameters

Using the intensity terminology provided by Warren et al. (2007), the intensity parameters are described as follows: the “treatment dose” (conversational recasts containing the morphological targets) was 24; the “dose frequency” was 5 days per week; the “total intervention duration” was 5 weeks. Thus, the highest possible “cumulative intervention intensity” (the product of “dose” × “dose frequency” × “total intervention duration”) was 600 recasts. The average cumulative intervention intensity was slightly less due to occasional child absences. However, children were required to complete make-up sessions if they exceeded three absences. Therefore, all children received at least 504 recasts.

Enhanced Conversational Recast Probe Sessions

Generalization probes were used to assess progress 3 times per week by the treating clinician through elicitation procedures utilizing untrained verbs. Probe sessions preceded treatment sessions so that performance reflected prior learning rather than learning from that same day. Probe procedures involved 10 elicitations of the target morpheme and 10 elicitations of the control morpheme in obligatory grammatical contexts. Probes always occurred prior to treatment that day. The final three probe sessions were used to calculate the average performance at the end of treatment. “Probe kits” containing thematic items not used during treatment and a set of 20 verbs not incorporated in recasts during treatment were used to assess generalization of the target to novel verbs and contexts. For example, the verbs “roll,” “jump,” and “fill” were all excluded from use during treatment sessions, as were farm toys, ocean toys, and race car toys. Therefore, a clinician might demonstrate a cow rolling on the ground, a whale jumping through a hoop, or a mechanic filling a gas tank to elicit these words in inflected form (e.g., Clinician: “Watch the cow roll! What did the cow do?”). The use of untrained verbs, probe kits, and timing of probes prior to treatment ensured that probe performance reflected learning rather than immediate performance effects (cf. Kamhi, 2014).

Fidelity and Reliability

Procedural fidelity and reliability for the coding of probe sessions was collected at least twice for every participant, and these sessions were spread throughout the full 5 weeks of the research period. This resulted in reliability and fidelity data on a minimum of 15% of all sessions across the studies from which these participants were drawn. The details involved in establishing procedural fidelity and reliability are published with the original studies from which the participants were drawn (Eidsvåg et al., 2019; Meyers-Denman & Plante, 2016; Plante et al., 2019, 2014, 2018). We report the aggregate data here for the participants included in this study. Fidelity metrics were calculated related to aspects of the treatment dose number and dose form, as well as assuring that words and materials withheld for testing generalization were not used during treatment. Treatment fidelity metrics ranged from 83.3% to 100% across all aspects of treatment fidelity. Likewise, fidelity of probe session procedures tracked the number and methods used to elicit target morphemes and the use of required words and materials. Probe fidelity ranged from 84.2% to 100%. The point-to-point reliability of the clinician's recording of children's use of morphemes during probe sessions ranged from 80.0% to 100%.

Outcome Measures

Children's generalization of their trained grammatical morpheme during probe sessions immediately posttreatment (i.e., no delay) served as the dependent measure in this study. These data were used to calculate a single-subject treatment effect size d for each child (see Beeson & Robey, 2006, for discussion of single-subject effect sizes). The treatment effect size d, originally used in Plante et al. (2014), was calculated as an expression of change in performance at treatment end based on the mean correct responses on the final three generalization probes (A₂ ) relative to baseline (pretreatment) performance quantified by the mean correct responses to the first three generalization probes (A₁ ), divided by the standard deviation of the values of the final three generalization probe sessions (A₂ ).

$$d=\frac{{\overline{X}}_{A_2}-{\overline{X}}_{A_1}}{{\mathit{SD}}_{A_2}}$$ (1)

In cases when there was no variance in the final three generalization probe sessions (i.e., they were all the same value), then the minimum possible standard deviation value (σ = .58) was used (a difference of one response among the 3 days) to avoid dividing by zero. Group data on pre- and posttreatment morpheme use and treatment effect sizes are reported in Table 1.

Analyses

In order to test multiple hypotheses simultaneously to identify the covariate(s) that predict participants' response to language therapy, classification and regression tree (CART; Breiman et al., 1984) analysis was employed. This method partitions the covariate space into mutually exclusive subgroups where subgroup heterogeneity is maximized. In other words, each predictor becomes binary at the value at which the two groups resulting from this split are most different from each other. This method can be utilized to identify groups of participants who may benefit differentially from an intervention versus the other participants because those that benefit more from the treatment will be grouped together and those that benefit less will be grouped together.

Because the outcome of interest in this case, treatment effect size d, is a continuous variable, regression trees were used. In the tree structure, observations were clustered in “nodes” with two daughter nodes emanating from a common parent node. The parent node is the predictor of interest with a binary value, and the daughter nodes are the groups of participants split between the two values of the parent predictor. The goal of the tree-growing procedure is to identify the value of covariates yielding the “best” split between daughter nodes. This split is deemed best when the value of the predictor variable identifies two groups for which the predictor variable yields different treatment outcomes between groups, but relatively similar treatment outcomes within-group.

An illustrative example of parent and daughter nodes can be drawn from health research predicting cancer-related deaths among women. The highest parent node in the tree is age, which divides all women participants into two daughter nodes: a younger cohort of 25–50 years versus an older cohort of 51 years and older. Participants in the two daughter nodes have different outcome values as the younger group has 0.9% cancer-related deaths, whereas the older group has 6.6% cancer-related deaths (Kitsantas & Wu, 2013).

For regression trees, the best split is defined statistically as the one resulting in the largest decrease in sum of squared error. In the current study, tree validation was done by dividing the 105 participants into training and validation sets (2:1 ratio). The tree structure was trained using the training set (n = 78) and the final tree structure was selected based on performance in the validation set (n = 27), which confirmed that the trained structure applied to data that were not used to create the tree.

One potential shortcoming of CART approaches is that if all possible splits between nodes were tested, the resulting decision tree may have a small number of observations in each terminal node, leading to good performance in the training data, but possibly subject to overfitting. This would limit the predictive utility of the tree as it would explain the data used to create the tree well but would be too specific to generalize to new data, which is the ultimate goal of these analyses. To address this problem in our analyses, a random forest procedure was employed (Breiman, 2001), which used bootstrapping to sample subsets of participants from full sample. Observations not included in the bootstrap, the “out-of-bag” (OOB) sample, were set aside, and a large tree was grown using the bootstrap sample. At each split, a random sample of covariates was selected from which the split may be made. The error rate was then estimated using the OOB sample with the overall error rate represented by the mean sum of squared error across all OOB samples. The resulting forest consists of several highly uncorrelated predictor trees, which are aggregated to yield a model that has better overall predictive power than a single tree. This approach allowed us to create the tree with part of our data and then test the tree with another part of our data to ensure that the tree fit new data that were not used in its creation.

Another advantage of the random forest approach is the “variable importance measure,” which can be calculated for each covariate (Breiman, 2001). To calculate the variable importance measure, a tree was grown from a bootstrap sample and the OOB sample impurity (misclassification or error rate) was calculated. Then the values for each covariate in the OOB sample were randomly reassigned, or permuted, to another observation and the OOB sample was run down the tree again. The difference between the original OOB impurity measure and the permuted OOB impurity measure is the “variable importance contribution” for a particular covariate. If a potential predictor were actually important in predicting the outcome, then permuting its values in the predictor space would cause the impurity measure to change in a notable way. If the predictor were not important, then permuting its values would have little effect on the impurity measure. This procedure was repeated for each tree in a random forest and the mean change in impurity between OOB sample and permuted OOB sample was computed for each covariate to give a variable importance measure. Higher values of the variable importance measure indicate the covariate is more predictive of the outcome.

Results

Random forest and CART regression analyses were used to create trees from our total sample of 105 participants with the following possible covariates included in the model: maternal education; participant age; sex; standard scores on KABC-II, PPVT-4, SPELT-P 2, and GFTA; and pretreatment target morpheme production accuracy.

Figure 1 displays the final tree structure derived from the CART analyses. Each box represents a node. Node level information is provided inside each box, including the mean treatment effect size d and the proportion of the training samples in the node. The first node splits participants based on SPELT-P 2 standard score. Participants with SPELT-P 2 less than 75 had the smallest average treatment effect size (d = 4.1, n = 74). Participants whose SPELT-P 2 scores were 75 or above were then partitioned by PPVT-4 standard scores. Participants with SPELT-P 2 scores of at least 75 and PPVT-4 scores less than 100 had smaller average treatment effect sizes (d = 5.1, n = 18). Participants with SPELT-P 2 scores of at least 75 and PPVT-4 scores of at least 100 had the largest average treatment effect size (d = 12, n = 15).

Figure 1.

The tree structure for the complete cases (n = 105 observations) using eight predictors. The box represents a node and contains the mean outcome on the top line and the proportion of data in the node on the second line. SPELT = Structured Photographic Expressive Language Test; PPVT = Peabody Picture Vocabulary Test.

Figure 2 presents the variable importance measure for each of the covariates included in the random forest analysis. Covariates with a higher increase in node purity are more predictive of the outcome relative to covariates with lower values for increase in node purity. Consistent with the tree structure, the variables with the highest increase in node purity (i.e., the most important predictors of treatment response) were PPVT-4 and SPELT-P 2 standard scores.

Figure 2.

The variable importance measures for the complete cases (n = 105 observations) using eight predictors. The x-axis corresponds to the percent decrease in mean squared error, that is the variable importance measure. The y-axis corresponds to the predictors. Each point represents the variable importance measure for a single predictor. Each variable importance measure was derived from a random forest with 5,000 trees. PPVT = Peabody Picture Vocabulary Test; SPELT = Structured Photographic Expressive Language Test; Pre_Treat_Prob = children's pretreatment production accuracy of treated morpheme; KABC = Kaufman Assessment Battery for Children; GFTA = Goldman-Fristoe Test of Articulation.

Discussion

Our analysis indicates that SPELT-P 2 and PPVT-4 scores significantly influenced the degree of benefit a child derived from Enhanced Conversational Recast treatment. If a clinician knows nothing about a child's background, the average benefit that they can expect from using this treatment is a change of SD = 5.4 from a child's initial morpheme use (see Figure 1). This treatment effect size can represent a range of start and end points because it represents a relative change from pre- to posttreatment production. For one example, a child who starts treatment using his target 20% before treatment would likely show approximately 52% correct use after 25 treatment sessions (e.g., a treatment effect size d of 5.4 = (52% end treatment use−20% pretreatment use)/5.77% as the minimum possible end treatment standard deviation). However, when the child's SPELT-P 2 scores are considered, the prognosis for change can be made more accurate. For children whose SPELT-P 2 scores are above 75, the average treatment effect size is 8.4. The hypothetical child could now be expected to show about 69% correct morpheme use posttreatment, whereas this child receiving a low SPELT-P 2 score (< 75) would lead to a predicted 43% correct use as a result of Enhanced Conversational Recast treatment in 25 treatment sessions (e.g., a treatment effect size d of 8.4 = (68.5% end treatment use−20% pretreatment use)/ 5.77% as the minimum possible end treatment standard deviation).

The SPELT-P2 standard score of 75 best differentiates participants with relatively larger versus smaller treatment responses when considered with PPVT-4 scores. The effect of SPELT-P2 can be evaluated in the broader context of typical scores for children with DLD on this particular test. Although the cut score for this test that differentiates between children with DLD and their peers with normal language is 87 (Greenslade et al., 2009), the mean performance by children with DLD was 72 in that study. Therefore, the average severity score on the SPELT-P 2 for children with DLD is nearly equivalent (within a standard error of measure) to the standard score of 75 that differentiates between children with larger versus smaller treatment responses. Children with the largest treatment response whose SPELT-P 2 scores are 75 or above reflect the upper half of the DLD continuum, and participants with smaller treatment responses fall into the lower half on this test. This is not to suggest that good treatment response can only be expected from children whose language deficits are mild. Compared with their peers with typical language development, children who score below the identification cut score (87) for this test show multiple morphosyntactic errors at ages where these should be relatively rare or absent. Within this context, however, progress using Enhanced Conversational Recast methods is likely to be slower for the more severely affected children within the range of DLD.

The marked effect of the SPELT-P 2 scores on the treatment effect size likely reflects the role of initial language severity on treatment response. This is consistent with previous reports that measures of language form predict response to conversational recast treatment. Yoder et al. (2011) reported children with a higher MLU (> 1.84) showed better response to broad recasts (recasting of any language error) than children with low MLU. Likewise, Pawłowska et al. (2008) reported that use of subject-verb constructions predicted progress with recast treatment that targeted either third person singular markers or “be” verb forms. Unlike these previously used measures, the SPELT-P 2 assesses a wide array of language forms. Therefore, the SPELT-P 2 likely reflects severity more broadly than previously reported measures of morphosyntax. The collection of available results, however, suggests that measures of language form are prognostic for treatment results using conversational recast procedures. This should not be particularly surprising, given this treatment is used to remediate errors of language form.

Our finding that single word receptive vocabulary scores also mediate Enhanced Conversational Recast treatment results is unique in the literature. Interestingly, vocabulary levels did not operate in the same way for children with different morphosyntactic skill levels as measured by the SPELT-P 2. For children with higher SPELT-P 2 scores, vocabulary scores represent a further prognostic indicator. A relatively high, but still impaired SPELT-P 2 score, combined with a high PPVT-4 score (> 100) yields the best treatment effect size at d =12. This is akin to progressing from 20% use pretreatment to 90% use in 25 sessions over 5 weeks. However, poorer vocabulary scores reduce the expected treatment effect size to d = 5.1, or progress from 20% use to 50% in the same period. It is important to note that this second group includes children whose vocabulary scores would be considered grossly within normal limits, at or just below the standard score mean of 100. Therefore, it is not the case that only frank deficits in vocabulary are associated with poorer learning. Instead, relatively strong vocabulary skills appear to assist performance during Enhanced Conversational Recast treatment. We note that the significance of PPVT-4 was somewhat unexpected given that Leonard et al. (2004) found no correlation between PPVT performance and therapy response. However, we did not find a direct predictive relationship between PPVT-4 and response to therapy, but rather one that is moderated by SPELT-P 2 performance, which may account for the discrepancy between studies.

It is difficult to know if the lowest performing subgroup of children would respond better to a different morphosyntactic treatment method. To date, direct comparisons of conversational recast treatments with alternate treatments tend to show better generalization with recast treatments than with other methods (Camarata & Nelson, 1992; Camarata et al., 1994; Gillum et al., 2003; Nelson et al., 1996, but see also Smith-Lock et al., 2015). One notable exception is Yoder et al. (2011), which showed children with low MLUs responded better to milieu therapy than recast treatment. Milieu therapy tends to target language skills broadly, and the procedures described in Yoder et al. (2011) may have supported vocabulary growth as well as morphosyntactic growth. This may have led to expanded utterances by virtue of strengthened vocabulary, accounting for the effect. Children with very low vocabulary scores may benefit from initial efforts to expand their lexicon, followed by more targeted morphology treatment. If so, there are many evidence-based approaches to vocabulary treatment for young children for clinicians to choose from (e.g., Alt et al., 2014; Goldstein et al., 2016; Kouri, 2005; Leonard et al., 2019; Riches et al., 2005; Storkel et al., 2016).

The current model differentiates between three groups of participants who vary in the magnitude of their treatment responses. However, even the group with the lowest treatment effect has, on average, a positive response to treatment with d = 4.1. In other words, it is not the case that a SPELT-P 2 score under 75 predicts that a child will be a nonresponder, but instead that their response to Enhanced Conversational Recast treatment will be relatively weaker in comparison to a child with higher SPELT-P 2 and PPVT-4 scores. Although each of the three groups has an average positive treatment effect, there were differences in the proportion of children in each group that would be categorized as nonresponders (i.e., d < 1). Among participants with SPELT-P 2 scores below 75, 32% were nonresponders. In the group with SPELT-P 2 scores above 75 and PPVT-4 scores below 100, 28% of participants did not respond to therapy. Finally, among the participants with the strongest average treatment effect that included children who scored above 75 on the SPELT-P 2 and above 100 on the PPVT-4, only one participant (6%) did not show a positive effect of Enhanced Conversational Recast treatment.

Several variables did not predict treatment response and are worth noting. First, articulation skills did not predict treatment progress. There are interactions between phonology and language skills (e.g., Benham et al., 2018; Heisler et al., 2010; Saletta et al., 2018), but this variable did not add prognostic information beyond that of the SPELT-P 2 and PPVT-4. Two versions of the GFTA were used in this study, and subtle differences in the two versions may have weakened any effect for this variable. It is more likely that the lack of effect reflects the requirement that children were able to produce the sounds that composed the morpheme selected as their treatment target. This precondition stands in contrast to the Smith-Lock et al. (2013), who reported poorer outcomes for children whose articulation errors impacted morpheme production. The child's sex and mother's education also did not contribute beyond the variables already discussed. Although each is known to exert an effect on natural language development (e.g., Hoff & Tian, 2005), the treatment may well have leveled these differences because of its intensity. Finally, nonverbal IQ did not impact treatment results. This has been reported elsewhere for studies that have used conversational recasts (Fey et al., 1993; Finestack & Fey, 2009) as well as for other treatment methods (Bishop et al., 2006; Bowyer-Crane et al., 2011; Boyle et al., 2007; Justice et al., 2017). It is worth noting that the subjects in each of these studies, and in the present analysis, had IQ scores sufficient that ruled out intellectual disability. Therefore, it is not clear that IQ would not predict treatment effects if the range of scores extended into the range associated with intellectual disabilities.

The repeated reports of a lack of influence of nonverbal IQ as treatment predictor is one factor that led to the consensus statement that DLD should be defined as a language impairment in the absence of an intellectual disability, rather a discrepancy between nonverbal IQ and language (Bishop et al., 2016). This consensus statement concluded that “The key consideration in identifying impairment is whether the child is likely to benefit from intervention and that is not determined by IQ” (p. 15). Our findings are consistent with this statement in that language factors, but not nonverbal IQ, predicted response to treatment.

Finally, pretreatment use of the child's grammatical morpheme did not predict the gains they experienced in treatment. However, it is important to note that, for the children in this study, pretreatment use was constrained to be 33% or less, and most children had very little or no pretreatment use of the morpheme targeted. Therefore, the lack of predictive value for this variable may reflect this experimental constraint. Whether children should be treated for morphemes they do not use at all, or use seldomly, has long been debated (Nelson et al., 1996). Nelson et al. (1996) noted that children showed improvement on morphemes that were sometimes used correctly even without treatment, but showed significantly stronger gains on morphemes treated using conversational recast procedures. Treating these morphemes can help “close the gap” faster than waiting for these emerging morphemes to be mastered without assistance. Four- and 5-year old children will shortly head to school, where morpheme skills will be needed not only for effective conversation, but as a building block for literacy (e.g., Apel et al., 2012; Kim et al., 2013; Ortiz et al., 2012). Therefore, maximizing morphological performance during the preschool years might be a reasonable treatment priority. It also may be the case that success with an initial morpheme target will help the child “learn-to-learn” within the conversational recast context, leading to faster learning of subsequent language targets. If this is the case, then quick success with an initial morpheme target that is emerging could have efficiency benefits when a number of morphemes must be treated for the same child. This possibility requires empirical evidence.

There are two important caveats to keep in mind when using child profiles to predict likely treatment benefits. For every evidence-based guide, like that provided here, there will be exceptions to the rule. Although we tested eight variables for their prognostic power, there may be other unmeasured variables that either moderate or mediate treatment effects. Our results could vary if different or additional predictors were included in the model, which may in turn cause violations to the patterns presented in Figure 1. Slower than expected progress should cause a clinician to reconsider whether Enhanced Conversational Recast treatment is the best choice for that child and unexpectedly poor progress may result either from particular child characteristics or a less than optimal delivery of the treatment method.

Clinicians should also be mindful that there are multiple versions of conversational recast, which do not appear to work equally effectively or efficiently. For example, broad recasts (Yoder et al., 2011) do not appear to be as effective as recasts that focus on a single morpheme error. Providing few recasts per session is not as effective as providing more recasts (cf. Leonard et al., 2004, at 12 recasts per session with Meyers-Denman & Plante, 2016, or Eidsvåg et al., 2019, at 24 per session). Repeating the same recast more than once is much less effective than assuring that each recast used in a session contain different words than the others, making them each unique (Plante et al., 2014). Encouraging children to produce a subject + verb utterance to be recasted is more effective than recasting sentences without subjects (Hassink & Leonard, 2010). Using highly common verbs is less effective than using less common when recasting verb morphology (Owen Van Horne et al., 2017). These factors are known to alter the success of conversational recast procedures. They are also all under the control of the clinician and can enhance or detract from the effectiveness of the general treatment method.

We note that our results would likely differ if a different treatment approach were used and/or if treatment were provided for a longer time. Our current conclusions only reflect children's response to this particular treatment. Some children included in our sample showed no or modest gains over the 5-week Enhanced Conversational Recast Treatment (see Table 1), whereas other children in the sample show quite significant gains over the relatively short, 5-week treatment. However, variability in treatment outcomes was key to the current analyses, which sought to predict differential response, so the current data, as opposed to a treatment approach in which all children continued until a certain criterion of performance was reached, were most appropriate for our purpose.

Conclusions

Our goal was to create a data-driven decision tree to help clinicians determine the likelihood of a particular child responding positively to Enhanced Conversational Recast. We included child-level variables as possible predictors of response to therapy, as measured by effect size d, which captured the difference in children's production of their target morpheme in three pretreatment probes compared to three posttreatment probes. The two significant predictors of children's treatment response were their SPELT-P 2 and PPVT-4 scores with participants with relatively higher SPELT-P 2 scores (> 75) and PPVT-4 scores (> 100) showing the most positive response to therapy and those with lower SPELT-P 2 scores (< 75) showing the smallest gains. Using this information, clinicians can assess whether a specific child would likely fall into the former or latter groups based on their test scores and make treatment decisions accordingly. Based on our analyses, other possible factors including a preschooler's age, sex, maternal education, articulation, nonverbal IQ, and pretreatment target morpheme accuracy were not significant predictors of treatment response, and therefore, may be less important factors for a clinician to consider when choosing whether or not to use Enhanced Conversational Recast treatments. Making evidence-based treatment decisions for individual children will allow service providers and parents to make informed decisions about a child's treatment based on current scores on widely available standardized language measures.

Appendix

Summary of Treatment Procedure Variations

Treatment Conditions. Variations of Enhanced Conversational Recast treatment provided to different sets of children included in this study of predictors of treatment outcomes were: massed versus spaced dosage schedules (Meyers-Denman & Plante, 2016), sparse versus dense dosage schedules (Plante et al., 2019), the addition of a brief auditory bombardment component to the recast treatment (Plante et al., 2018), an expressive practice condition (unpublished data), and short versus extended recasts (unpublished data).

Treatment Group Assignment. As a series of early efficacy studies with small sample sizes, true random assignment was not used in the studies from which participants were drawn, but rather participants were matched for study-critical variables across the study conditions. However, clinicians were never involved in assignment of children to experimental groups and assignments were made independently of the predictor variables examined in this study. Instead, children were assigned to balance participant sex and the specific morphemes treated across treatment conditions. Mean standardized test scores and mean age did not significantly differ between children in different conditions in any of the individual studies.

Massed Versus Spaced. The massed versus spaced treatment (Meyers-Denman & Plante, 2016) evaluated whether there were different group-level treatment effects when 24 doses of Enhanced Conversational Recast doses were provided in one versus three sessions daily. This manipulation involved spacing of the 30-min of daily treatment over three separate 10-min sessions during the half-day or providing it one 30-min session. Rate of eight recasts per 10 min was controlled (0.8 per minute on average).

Sparse Versus Dense. The sparse versus dense manipulation (Plante et al., 2019), involved the standard Enhanced Conversational Recast treatment of 24 unique recasts provided in daily treatment in either a 30-min session (sparse/low density) or a 15-min session (dense/high density). The average rate of recasts in the sparse condition was 0.8 per minute and 1.6 per minute in the dense condition.

Auditory Bombardment. The auditory bombardment variation (Plante et al. 2018) involved 24 additional unique models of the target form provided in quick succession for approximately 3–4 min either before recast treatment or after recast treatment. Auditory bombardment refers to high-density models—clinician productions of linguistic targets presented in short, grammatical sentences without the “conversational” component—that is, the child was not required to talk during that time (e.g., Encinas & Plante, 2016). Although there were no significant differences in the elicited use of morphemes during probes, the auditory bombardment after recast treatment condition resulted in more treatment responders.

Expressive Practice. The expressive practice condition (unpublished data) involved an elicited repetition of a short sentence containing the child's target structure. These repetition opportunities occurred twice each session, immediately after the 12th and 24th recast, and involved a request to repeat those recasts back to the clinician (e.g., Child: “Joey run.” Clinician: “Joey runs.” Clinician: “Tell me, Joey runs.”). Twenty-four recasts were provided in 30 min with an average dose density of approximately 0.8 recasts per minute.

Short Versus Extended. The short versus extended recast manipulation (unpublished data) contrast recasts that were four words or fewer (e.g., Spiderman jumped over.) with recasts that were longer than four words (e.g., Spiderman jumped over this tall building). Twenty-four recasts were provided in 30 min with an average dose density of approximately 0.8 recasts per minute.

Outcomes. The relative spacing of doses did not affect treatment outcomes in either study in which this was manipulated (Meyers-Denman & Plante, 2016; Plante et al., 2019) and there was a minimal difference between the two auditory bombardment conditions with slightly more treatment responders when bombardment was provided after recasting than before recasting. Bombardment did not influence the treatment effect sizes, which serve as the dependent variable in this study (Plante et al., 2018). Although unpublished, an analysis of the expressive practice condition did not indicate an advantage of this condition over Enhanced Conversational Recast alone. There is insufficient data available at the time of writing to evaluate the relative effects of short versus extended recasts. Given the relative comparability of the conditions contrasted within studies (see Table A1 for effect size means by condition); we combined samples across studies for the current analysis.

Table A1.

Means (standard deviations) and ranges of target morpheme production and treatment effect size d for each Conversational Recast Therapy condition.

Condition	Number of participants	Pretreatment probe accuracy	Posttreatment probe accuracy	Treatment effect size d
24 exemplar	19	.05 (.07) 0–.22	.51 (.35) 0–1.0	4.94 (6.45) −0.29–17.33
Bombard first	14	.13 (.10) 0–.33	.44 (.38) .03–1.0	3.72 (6.39) −3.81–15.60
Bombard last	14	.09 (.07) 0–.23	.51 (.31) 0–.9	3.80 (4.09) −1.04–12.23
Spaced	8	.05 (.07) 0–.20	.61 (.40) 0–1.0	7.46 (7.76) −3.46–17.32
Massed	8	.10 (.11) 0–.33	.69 (.33) .13–1.0	7.85 (7.39) 0.62–16.74
Sparse	12	.06 (.05) 0–.13	.68 (.32) .13–1.0	6.58 (5.88) 0.58–17.33
Dense	8	.06 (.05) 0–.13	.64 (.37) 0–1.0	6.99 (6.65) −1.16–17.33
Expressive Practice	8	.03 (.03) 0–.13	.48 (.44) 0–1.0	5.69 (7.24) 0–17.33
Short	7	.02 (.03) 0–.07	.58 (.42) 0–.97	5.80 (5.96) 0–16.74
Extended	7	.03 (.03) 0–.10	.41 (.43) 0–.97	3.88 (5.4) −0.58–15.02

Funding Statement

This work was partially funded by National Institute on Deafness and Other Communication Disorders Grant R01C015642 (awarded to E. Plante) and donations from Cécile Moore for the Talk MOORE Summer Camp program.