Feedback Processing During Probabilistic Learning in Children With Developmental Language Disorder: An Event-Related Potential Study

Abstract

Purpose:

The purpose of this study was to examine feedback processing within the context of probabilistic learning in children with and without developmental language disorder (DLD).

Method:

The probabilistic category learning task required 28 children ages 8–13 years old to classify novel cartoon animals that differed in five binary features into one of two categories. Performance feedback guided incremental learning of the stimuli classifications. Feedback processing was compared between children with DLD and age-matched children with typical development (TD) by measuring the magnitude of feedback-related event-related potentials. Additionally, the likelihood of each group to repeat a classification of a stimulus following positive feedback (“stay” behavior) and change a classification following negative feedback (“switch” behavior) served as a measure of the consequence of feedback processing.

Results:

Children with DLD achieved lower classification accuracy on all learning outcomes compared to their peers with TD. Children with DLD were less likely than those with TD to demonstrate “stay” behavior or to repeat a correct response following positive feedback. “Switch” behavior or changing an incorrect response following negative feedback was found to be at chance level in both groups. Electrophysiological data indicated that children with DLD had a smaller feedback-related negativity effect (i.e., smaller differential processing of positive and negative feedback) when compared to children with TD. Although no differences were found between the two groups in the amplitude of the P3a, strong positive correlations were found between “stay/switch” behavior and the P3a for children in the TD group only.

Conclusions:

Children with DLD do not appear to benefit from incremental corrective feedback to the same extent as their peers with TD. Processing differences are captured in the initial stages of feedback evaluation and in translating information carried by the feedback to inform future actions.

The ability to learn from feedback is critical in the school environment as most teaching methods require that learners make use of external feedback to advance their learning. This is also the case for children with developmental language disorder (DLD), a significant impairment of language expression and comprehension without obvious causal factors such as intellectual disability or neurological impairment (Bishop et al., 2017; Leonard, 1998) that affects roughly 8% of school-age children (Law et al., 2000; Norbury et al., 2016; Tomblin et al., 1997). Children with DLD are exposed to different forms of feedback at school and during most clinical interventions targeting their language skills (Baron & Arbel, 2021). Feedback is commonly used in explicit teaching approaches where rules are taught in isolation and in a structured, highly predictable, and controlled environment (e.g., Calder et al., 2020). Feedback is also part of more implicit teaching approaches such as corrective recasting, where children's erroneous responses are immediately followed by a corrective response (e.g., Plante et al., 2014). In this context, feedback is given in a less overt manner, and it is expected that the child would gradually internalize the information carried by the feedback to adjust their behavior.

There is limited but growing evidence that school-age children with DLD exhibit less efficient feedback processing when compared with their peers (e.g., Arbel & Donchin, 2014; Arbel et al., 2021; Lee, 2017). For example, Lee (2017) reported that adolescents with DLD who performed the balloon analogue risk task, in which their risk-taking behavior and responses to positive and negative outcomes were evaluated, demonstrated an overall risk avert behavior that was not affected by previous outcomes. In other words, adolescents with DLD were not using feedback from prior trials to modulate their risk-taking behavior. It is unclear, however, to what extent the behavioral indications of limited use of feedback have been confounded by the DLD participants' tendency to avoid risks. Arbel et al. (2021) compared children with and without DLD (aged 8–12 years) on a feedback-based declarative paired-associate learning task that required linking correct pseudoword names with novel objects. At the behavioral level, the efficiency of feedback processing was measured as the likelihood of repeating a correct choice following positive feedback (“stay” behavior) and correctly switching an erroneous choice following negative feedback (“switch” behavior). In comparison to children with typical development (TD), children with DLD had fewer repeated actions following positive feedback, defined as a stay behavior. Interestingly, switch behavior, or the likelihood of changing a behavior following negative feedback, was poor among children with and without DLD, indicating inefficient use of negative feedback in both groups. Although these behavioral measures reflect the consequence of feedback processing, they do not reveal how and how well feedback is processed by the brain and how this processing is related to subsequent actions that may lead to learning.

Feedback-related event-related potentials (ERPs), which are electrical brain responses time-locked to the presentation of the feedback, capture the processing of positive and negative feedback with a temporal precision of milliseconds. Such information can only be inferred from behavioral data. More specifically, the feedback-related negativity (FRN) ERP component is generated in response to the provision of feedback. Its amplitude reflects the differential processing of positive and negative feedback as it changes throughout the learning process and in relation to learning outcomes. In the study by Arbel et al. (2021), children with DLD demonstrated less efficient processing of feedback, evidenced by reduced brain activation in response to negative feedback when compared to their peers with TD. Furthermore, children with DLD had no differential responses to positive and negative feedback at the electrophysiological level. Taken together, these results point to atypical feedback processing in children with DLD (Arbel et al., 2021).

It is important to note that the atypical feedback processing found by Arbel et al. (2021) was detected within the context of declarative learning. In declarative learning tasks, the feedback is deterministic such that it provides the learner with clear information on performance that can be translated on a trial-by-trial basis into actions that will lead to favorable outcomes. Positive feedback for correct responses solidifies hypothesized association, whereas negative feedback for incorrect responses indicates the need to change responses. For example, in the task used by Arbel et al. (2021), participants selected one of two novel objects in response to a spoken word. Stimulus sets (i.e., pairs of objects and the spoken word) were held constant throughout the task. Within this context, performance feedback was highly instructive and informative on a trial-by-trial basis. If selecting one object from the pair resulted in negative feedback, selecting the other object would result in positive feedback.

In nondeclarative learning, on the other hand, the information provided by the feedback is incremental, and learning from feedback requires the gradual accumulation of knowledge over many trials (Knowlton, Mangels, & Squire, 1996; Knowlton et al., 1994). Often, nondeclarative learning is examined in the context of probabilistic tasks or those in which statistical regularities and rules account for complex stimulus–response mappings or the provision of correct or incorrect feedback. Thus, intact feedback processing is essential for learning in this context. Nondeclarative feedback-based learning relies on the intactness of basal ganglia, particularly the striatum (Batterink et al., 2019; Poldrack et al., 1999, 2001). Wilkinson et al. (2014) observed striatal activation during the performance of a feedback-based weather prediction task, which is a probabilistic classification learning (PCL) task, suggesting that the striatum plays a crucial role in regulating procedural learning with corrective feedback. Additionally, individuals with impairments of the striatum, such as those with Parkinson's disease, have been found to perform poorly on tasks that require implicit learning guided by feedback (Knowlton, Squire, et al., 1996; Shohamy et al., 2004; Wilkinson et al., 2014). Relevant to the inquiry of this study is evidence of abnormalities in the striatum (Badcock et al., 2012; Jernigan et al., 1991; Lee et al., 2013; Liegeois et al., 2002; Soriano-Mas et al., 2009; Tallal et al., 1994; Vargha-Khadem et al., 1998; Watkins et al., 1999) and frontal cortex (Clark & Plante, 1998; Cohen et al., 1989; Denays et al., 1989; Gallagher & Watkin, 1997; Gauger et al., 1997; Jernigan et al., 1991; Kabani et al., 1997) in individuals with DLD and mixed reports on whether children with DLD are impaired in their ability to learn implicitly.

Lee and Tomblin (2012) evaluated feedback-based nondeclarative learning in adults with DLD using a two-choice selection task guided by probabilistic win–loss feedback. They reported that, although adults with DLD took longer to reach a learning criterion and achieved lower accuracy scores on a subsequent test when compared with typical adults, their behavior did not indicate a different response to positive and negative feedback (i.e., there were no group differences in stay/switch behavior) during the initial stages of the task. Because reinforcement learning relies on the continuous processing of feedback, evidence of poor reinforcement learning may indicate inefficient processing of feedback. However, the study was conducted with adults whose behavior may differ from that of children due to more extensive opportunities to develop compensatory strategies. Kemény and Lukács (2010) examined probabilistic learning in children with DLD using a weather prediction task with feedback. They found that 11-year-old children with DLD demonstrated significantly lower classification accuracy than children with TD and adults. Furthermore, they found that children with DLD used strategies less often and less effectively than either control group. The authors did not examine specific responses to positive and negative feedback (i.e., stay/switch behavior), which makes it difficult to discern whether or to what extent inefficient feedback processing affected performance. Nonetheless, these results suggest there are differences in probabilistic learning among school-age children with and without DLD.

This Study

The goal of this study is to evaluate the extent to which incremental processing of feedback in a nondeclarative learning paradigm with probabilistic stimulus–response mappings is impaired in children with DLD. This inquiry is motivated by previous reports of impaired processing of performance feedback in a declarative learning task that prompted the question of whether weaknesses in feedback processing are unique to declarative learning or are manifested in other types of learning that call for different feedback processing demands. Unlike declarative learning tasks that require performance evaluation on a trial-by-trial basis and rely on memory to maintain or alter a specific choice following feedback, a nondeclarative task relies on implicit learning over the course of many trials. Because children with DLD have impaired language abilities, it is also notable that the nondeclarative task in this study utilizes visual stimuli and feedback with minimal need for verbal instruction or response. By limiting the amount of language needed to complete the task, differences between groups of children with and without DLD can be attributed to more specific mechanisms of feedback processing and implicit learning. To investigate whether the neural mechanisms underlying feedback processing driven by implicit learning are impaired in children with DLD, behavioral and electrophysiological measures were examined in these children and age-matched peers with TD. Behaviorally, accuracy was examined from a few perspectives. Overall accuracy provided a general indicator of learning outcome, whereas accuracy across the various probabilistic stimulus types and old versus new stimuli provided more specific measures of nondeclarative or implicit learning. The study also compared the proportion of stay behavior (i.e., repeating a correct behavior following positive feedback) and switch behavior (i.e., changing a behavior following negative feedback) between groups. These behaviors were examined within each stimulus to evaluate the consequence of feedback processing, which is an indirect measure of the efficiency with which positive and negative feedback is processed.

At the electrophysiological level, the study focused on two ERP markers of feedback processing—the FRN and the P3a. The FRN is an ERP component with a frontocentral scalp distribution and a peak between 200 and 300 ms following the provision of feedback (Miltner et al., 1997). The FRN amplitude is larger (more negative) following negative feedback than positive feedback. Findings from neuroimaging studies suggest that the FRN originates from the anterior cingulate cortex (ACC; Foti et al., 2015; Gehring & Willoughby, 2002; Hauser et al., 2014; Holroyd & Coles, 2002; Miltner et al., 1997), whereas other studies have also found a striatal contribution (Becker et al., 2014; Foti et al., 2011). According to the reinforcement learning theory of the FRN, this component reflects the activity of dopaminergic neurons in the ACC following negative and positive feedback, which indicates outcomes in relation to current expectations (Cohen & Ranganath, 2007; Holroyd & Coles, 2002). Another important ERP component associated with feedback processing is a positive deflection that emerges around 350 ms postfeedback at frontocentral sites (Polich, 2007). The P3a is observed to be larger in response to negative feedback than positive feedback (Cavanagh & Frank, 2014; West et al., 2014). The P3a is suggested to have neural generators within the ACC (Cavanagh & Frank, 2014; Foti et al., 2011; Soltani & Knight, 2000). Evidence from the source localization literature proposes that the FRN and the P3a are the manifestations of temporally distinct processes, which reflect the engagement of the ACC at different times during performance monitoring (West et al., 2012, 2018). Within the context of feedback-driven reinforcement learning, it has been suggested that the P3a is involved in postfeedback updating of action–outcome contingencies (West et al., 2012). The postfeedback frontocentral P3a is suggested to reflect attention allocation (Polich, 2007) and has been found to be sensitive to feedback valence (i.e., positivity and negativity) and learning outcomes (Arbel & Fox, 2021; Arbel & Wu, 2016).

This study also evaluated the feedback-related ERP components in relation to behaviors associated with the consequences of feedback processing. More specifically, children's stay and switch behaviors were examined in relation to brain responses to positive and negative feedback. Impaired feedback processing in children with DLD could be reflected in less frequent stay and switch behaviors and in smaller brain activation associated with positive and negative feedback. We further hypothesize that stronger relationships between neural correlates of feedback processing and subsequent behavior (stay/switch performance) will be found in children with TD, indicating a breakdown in communication between brain circuits responsible for the initial processing of feedback and those charged with translating feedback into appropriate action. By improving our understanding of feedback processing, we aim to improve learning outcomes in children with DLD.

Method

Participants

Twenty-eight children from the Boston area participated in the study. There were 14 participants with DLD and 14 age-matched controls with typical language development (TD). All participants were right-handed individuals between the ages of 8 and 13 years (M = 10.75, SD = 1.47), with normal or corrected vision, who reported that they had no history of head injury or other neurological deficits and that English was their predominant language. Inclusion criteria for the DLD group were as follows: (a) Core Language score on the Clinical Evaluation of Language Fundamentals–Fifth Edition of at least 1.2 SD below the mean (SS < 82) or (b) Identification Core score on the Test of Integrated Language and Literacy Skills of less than 34 if 8–11 years old or less than 42 if 12–18 years old. All children in the DLD group had a reported history of a delay in language development and persistent difficulties with verbal and/or written language. All participants in the DLD and TD groups obtained a nonverbal intelligence score above the range of intellectual disability (SS > 80) on the Matrices subtest of the Kaufman Brief Intelligence Test. Table 1 presents demographic and descriptive assessment data by group as well as results of paired t tests and chi-square analysis. The DLD and TD groups did not significantly differ in terms of age or proportion of males and females. There were significant group differences favoring those with TD on all other descriptive assessments.

Table 1.

Participant data and group comparisons.

Variable		TD	DLD	Paired-samples t test
		(n = 14)	(n = 14)	T	df	p
Age (in months)		124.57 (19.78)	124.71 (20.23)	−0.23	13	.824
KBIT matrices score		114.29 (12.05)	103.21 (12.99)	2.64	13	.020
CELF-5 Core Language score		110.15 (13.15)	82.85 (7.58)	6.11	12	< .001
TILLS Identification Core score a		—	70.80 (12.15)	—	—	—
				Chi-square test
				χ 2	df	p
Gender	Female	7	9
	Male	7	5	0.583	1, N = 28	.445

Note. Descriptive assessment data are standard scores. TD = typical development; DLD = developmental language disorder; df = degrees of freedom; KBIT = Kaufman Brief Intelligence Test (Kaufman, 1990); CELF-5 = Clinical Evaluation of Language Fundamentals–Fifth Edition (Wiig et al., 2013); TILLS = Test of Integrated Language and Literacy Skills (Nelson et al., 2016).

^a The TILLS was administered to five participants who did not meet criteria for DLD based on CELF-5 performance but whose parents reported a history of language delay or impairment.

Procedure

This study was approved by the Mass General Brigham Institutional Review Board. Parental consent and participant assent were obtained before data collection was initiated, and participants were paid for their participation. Participants were enrolled in a larger study on feedback processing containing two to three sessions lasting 90 min each. Standardized assessments were administered in one to two sessions prior to completing experimental tasks. This study took place during a single session in which experimental tasks were administered with electroencephalography (EEG) recording. Participants were fitted with a 32-channel EEG HydroCel net by Electrical Geodesics, Inc., and then seated in a quiet room, at a comfortable distance (60 cm) from a computer monitor that was adjusted in height to align the center of the screen with the participant's eye level. Each participant completed a PCL task, lasting about 15 min, while their EEG data were recorded from the scalp.

Task

The objective of the PCL task was to correctly classify novel cartoon animals into one of two categories using a trial-and-error approach guided by feedback. In other words, participants had to guess the category assignments initially and then use feedback to verify or change their future responses. Simple verbal instructions and cartoon drawings were provided in a child-friendly manner to facilitate understanding, increase motivation, and maintain an appropriate level of engagement. The framework for the task involved a short story: “Welcome to my birthday party! I need your help serving cupcakes to my guests. Animals from the Bloom family like chocolate cupcakes. Animals from the Smith family like vanilla cupcakes.” The task contained a training phase, immediately followed by a testing phase. Items (stimuli and feedback) were presented on a computer monitor, and responses were recorded on a Chronos response box. No reading or verbal responses were required during the task, which limited the potential advantage of typical language skills for children in the TD group.

Probabilistic Design of Stimuli

The framework and stimuli for the task were borrowed from Zeithamova et al. (2008) and modified to create a simpler version for children. See Table 2 for an illustration of sample stimuli and the probabilistic distribution of their features. In the original task (Zeithamova et al., 2008), participants learned to classify novel cartoon animals that differed on 10 binary dimensions (i.e., features), the probabilistic combination of which determines classification. In the modified version presented in this study, stimuli differed on five instead of 10 binary features (i.e., head position, tail shape, feet shape, body shape, and body pattern). Each of the five features had two possible presentations, one of which was associated with each of the two classification categories. For example, for the binary dimension “body shape,” the two options were “round” and “square,” with round body being a Category A feature and square body being a Category B feature. Feature distribution was based on a continuum, such that when an animal had four out of the five classification features of Category A (80% of classification features), it had one out of the five classification features of Category B (20% of classification features). When an animal had three out of the five classification features of Category A (60% of classification features), it had two out of the five features of Category B (40% of classification features). Animals with 60% or more features of a prototype were deemed as belonging to that prototype. The probabilistic nature of the task made it impossible to determine category membership based on a single feature, with ideal classification accuracy expected to match the percentage of feature overlap with each prototype. Across training and testing phases, an equal number of animal stimuli could be classified as belonging to Category A and Category B. All animal stimuli were equal in size and presented as a full-color, two-dimensional line drawing in the center of a 17-in. monitor screen.

Table 2.

Probabilistic features of stimuli in the classification learning task.

Example stimulus
Stimulus type	Prototype A	A-1	A-2	B-2	B-1	Prototype B
Classification category	A	A	A	B	B	B
Number of features (distance) from Prototype A	0	1	2	3	4	5
Number of features (distance) from Prototype B	5	4	3	2	1	0
Percentage of features shared with Prototype A	100%	80%	60%	40%	20%	0%
Percentage of features shared with Prototype B	0%	20%	40%	60%	80%	100%

Training Phase

Each trial consisted of a single animal exemplar presented in the center of the computer screen with the two category options represented by chocolate and vanilla cupcakes in the bottom right and left corners of the screen. Participants were instructed to press a button on the right side of the response box to choose chocolate and to press a button on the left side of the response box to choose vanilla. Each response was followed by visual feedback in the form of three “√”s for correct classification and three “X”s for incorrect classification. An illustration of the training trial procedure is presented in Figures 1a and 1b. During training, participants were presented with eight novel animals 20 times each. Animal exemplars were presented one at a time, in random order, for a total of 160 trials. These trials were distributed across four training blocks containing 40 trials each, with a short rest break provided between each block. A fixed number of trials, as opposed to setting a learning criterion, allowed us to better control the frequency of exposure to each type of stimulus. Out of the presented items, 75% were one feature away from prototypes A or B (80% of features shared with one prototype and 20% of features shared with the other prototype), and 25% were two features away from the prototypes (60% of features shared with one prototype and 40% of features shared with the other prototype). The emphasis on highly probable exemplars helps build a mental representation of the prototypes, which were not presented during the training phase. In other words, the boundary between categories is clearer when a larger variety of strong versus weak exemplars is provided.

Figure 1.

Illustration of individual trials during the training (a and b) and testing (c) phases of the probabilistic classification learning task.

Testing Phase

Immediately following the training phase, participants were tasked with classifying the eight animals that were presented during the training phase (i.e., old items) and eight new animals that followed the same feature probabilities. Each of the 16 unique old and new items was presented twice during the test for a total of 32 trials. In addition, animals that represented the two prototypes (i.e., animals with five unique features for each prototype) and that were not presented during training were included in the test phase as new items. The two prototypes were presented 4 times during the test. Participants were presented with a total of 40 test trials presented in random order in a single block. During the testing phase, participants' responses were not followed by performance feedback. An illustration of the testing trial procedure is presented in Figure 1c.

Behavioral Data Analysis

Learning outcomes (i.e., accuracy during training and testing) were examined to evaluate whether feedback-based learning during a probabilistic classification task differed between groups (i.e., children with DLD and TD). Within the training phase, we also examined accuracy by block to capture incremental learning from repeated exposure to the probabilistic stimuli. Within the testing phase, accuracy was evaluated by stimulus type and exposure. Stimulus type (i.e., prototypes, exemplars that are one feature away, and exemplars that are two features away) measures the impact of the stimuli's statistical properties on learning. Exposure (i.e., old exemplars vs. new exemplars and prototypes) measures the depth of learning and the ability to apply the classification rule to new items. Additionally, we evaluated participants' stay and switch behaviors following positive and negative feedback during the training phase. Stay behavior was defined as the proportion of positive feedback trials followed by a stay response (i.e., repeating the classification choice for a specific stimulus). Switch behavior was defined as the proportion of negative feedback trials followed by a switch response (i.e., choosing a different classification for a specific stimulus). It is important to note that stay and switch behaviors were calculated for each stimulus as opposed to trial-by-trial. Results reflect the average occurrence of stay/switch behavior for all stimuli. High proportions of stay and switch behaviors reflect effective use of feedback to guide performance and reinforce learning. The procedure for calculating stay/switch behavior was based on that of Arbel et al. (2021).

Behavioral data were analyzed using a series of one-way and mixed analyses of variance (ANOVAs). The dependent variable for all analyses of classification accuracy was proportion correct. The dependent variable for the stay/switch behavior analyses was proportion of trials, referring to how often each behavior occurred. The between-subjects variable was always group (TD and DLD). The within-subject variables included block during the training phase (four blocks with 40 trials in each block) and stimulus type (prototype, one feature away, and two features away), exposure (old and new item), or stay/switch behavior during the testing phase. For variables containing three levels, we examined variances using Mauchly's test of sphericity. When indicated by a significant interaction effect, pairwise comparisons were analyzed with Bonferroni correction. All error bars represent the standard error of the mean.

EEG Data Acquisition and Signal Processing

The 32-channel GES 400 System by Electrical Geodesics, Inc., was used to obtain dense-array EEG data using 32-channel HydroCel Geodesic sensor nets, composed of Ag/AgCl electrodes attached to an elastic net following the international 10–20 system. EEG was continuously recorded at a 1000-Hz sampling rate using the vertex as the reference electrode. Impedances were kept below 50 kΩ as per the manufacturer's recommendations (Ferree et al., 2001). The presentation of stimuli was controlled by programmable experiment generation software (E-Prime 2.0, Psychological Software Tools), and signals were acquired across all electrodes. The EEG data processing and analysis were performed using custom MATLAB (The MathWorks, Inc.) scripts operating in conjunction with the open-source EEGLAB toolbox (Delorme & Makeig, 2004; http://sccn.ucsd.edu/eeglab). Continuous data were filtered using a bandpass filter (0.1–30 Hz) for ERP analyses. The processed data were segmented into 1,200-ms epochs extending 200 ms before and 1,000 ms after feedback presentation. Each trial was visually inspected for movement artifacts, which were manually removed. Baseline correction was performed on the averaged data, based on signal in the 200-ms preceding the feedback stimulus (i.e., −200 to 0 ms). Data were re-referenced to the average reference (Winkler et al., 2015). An adaptive mixture ICA was applied separately to single-subject data set (Palmer et al., 2012) to detect and correct for eye movement and eye blinks. The mean number of artifact-free trials following negative feedback (NF) and positive feedback (PF) indicated no group effect (i.e., p = .259) between the TD group (NF mean = 58.07, SD = 15.09; PF mean = 91.50, SD = 15.93) and the DLD group (NF mean = 64.15, SD = 13.96; PF mean = 77.77, SD = 16.98). We further evaluated the percentage of lost trials per condition in each group. A 2 × 2 mixed ANOVA was conducted between groups (TD and DLD) on the number of trials lost in each condition (negative feedback, positive feedback). No group effect, F(1, 25) = 1.332, p = .259, η² = .051; condition effect, F(1, 25) = 1.568, p = .222, η² = .059; or a Condition × Group interaction, F(1, 25) = 32.346, p = .138, η² = .086, was observed. The data were then sorted into two categories based on negative and positive feedback for each group (i.e., DLD and TD). Note that ERP analysis was performed on trials from the training session only, and both correct and incorrect trials were part of the analysis as the purpose of the study was to compare the processing of positive and negative feedback (i.e., performance feedback following correct and incorrect responses).

ERP Data Analysis

ERP data from the frontocentral electrode of interest (FCz) were extracted using EEGLab's statistical tools for each participant and condition. To reduce the temporal dimensionality of the data set and to disentangle ERP components that overlapped in time, a temporal principal component analysis (TPCA) with Promax rotation (Dien, 2010) was conducted (e.g., Arbel et al., 2017; Arbel & Wu, 2016). The analysis used the covariance between time points and resulted in a set of eight temporal factors accounting for 86.82% of the total variance. Temporal factor 5 (TF5) with a maximal peak around 250 ms and temporal factor 3 (TF3) peaking around 400 ms were found to capture the FRN and P3a activation, respectively. The factor scores of TF5 and TF3 served as the amplitude measures of the FRN and P3a and were subjected to statistical analysis using IBM SPSS Statistics 24.0.

Results

Accuracy During Training

To evaluate accuracy during training as a function of group, a mixed ANOVA with group (TD and DLD) as a between-subjects variable and block (1, 2, 3, and 4) as a within-subjects variable was conducted. Mauchly's test was not statistically significant, χ²(5) = 2.49, p = .778; therefore, results are reported with sphericity assumed. There was a significant main effect of group, F(1, 26) = 4.93, p = .035, η_p ² = .159, indicating that children in the TD group achieved higher accuracy during training than children in the DLD group. The main effect of block was not significant, F(3, 78) = 0.97, p = .414, η_p ² = .036. However, a significant interaction between group and block was found, F(3, 78) = 4.67, p = .005, η_p ² = .152. Pairwise comparison indicated that children in the TD group were significantly more accurate than children in the DLD group in Blocks 2 (p = .012) and 4 (p = .001). For children in the TD group, there was a significant difference in accuracy between Blocks 1 and 4 (p = .043), but all other comparisons between blocks were nonsignificant (p > .190 for all other comparisons), which suggests a slow, steady increase in learning outcomes. For children with DLD, there were no significant differences between Blocks 1 and 4 (p = 1.0) or any other training blocks (p > .082 for all other comparisons), reflecting no change in learning outcome during training.

Accuracy During Testing

To evaluate overall accuracy during testing, a one-way ANOVA was conducted with group (TD and DLD) as a between-subjects variable. There was a significant main effect of group, F(1, 26) = 17.14, p < .001, η_p ² = .387, indicating that overall accuracy was significantly higher for children in the TD group (M = 0.70, SD = 0.11) than for children in the DLD group (M = 0.56, SD = 0.07). Figure 2 presents accuracy during each training block and testing.

Figure 2.

Training accuracy by block and group and testing accuracy by group. TD = typical development; DLD = developmental language disorder.

Probabilistic Learning

To evaluate the extent to which children extracted the probabilistic properties of the stimuli, we evaluated the accuracy of test items as a function of distance from the prototypes. Accuracy was evaluated for the prototype from each category, exemplars that were one feature away from the prototype, and exemplars that were two features away from the prototype. The prototypes and exemplars from each category (A and B) shared the same statistical properties and should have yielded similar results; however, it is possible that participants could demonstrate a significant preference for learning one category with greater accuracy than the other. A mixed ANOVA with group (TD and DLD) as a between-subjects variable and category (A and B) as a within-subject variable found no significant differences in accuracy for prototypes, F(1, 26) = 0.26, p = .612, η_p ² = .010; exemplars that were one feature away, F(1, 26) = 0.00, p = .987, η_p ² = .000; and exemplars that were two features away, F(1, 26) = 0.22, p = .645, η_p ² = .008. Because there were no significant differences in accuracy for Categories A and B, we combined the data from each category into single variables when analyzing stimulus type. To maintain the interpretation of the data as proportions, we averaged the values for Categories A and B together. As shown in Figure 3, the data for the TD group trend in the predicted direction with accuracy being greatest for prototypes and then for exemplars that are one feature away, and finally, accuracy was lowest for exemplars that are two features away. This pattern is not as clear for children with DLD.

Figure 3.

Test accuracy for prototypes and exemplars by group and stimulus type. TD = typical development; DLD = developmental language disorder.

A mixed ANOVA with group (TD and DLD) as a between-subjects variable and stimulus type (prototype, one feature away, and two features away) as a within-subject variable was conducted. Mauchly's test was not statistically significant, χ²(2) = 3.54, p = .170; therefore, results are reported with sphericity assumed. There was a significant main effect of group, F(1, 26) = 20.56, p < .001, η_p ² = .442, in which accuracy for the TD group was significantly higher than that of the DLD group. There was a significant main effect of stimulus type, F(2, 52) = 4.43, p = .017, η_p ² = .146, and pairwise comparisons indicated that accuracy for prototypes (M = 0.70, SD = 0.24) was greater than accuracy for exemplars that were two features away from the prototype (M = 0.58, SD = 0.15, p = .017). There was no interaction between group and stimulus type, F(2, 52) = 1.60, p = .212, η_p ² = .058.

Next, to compare the effects of old versus new items, a mixed ANOVA with group (TD and DLD) as a between-subjects variable and exposure (old and new) as a within-subject variable was conducted. There was a significant main effect of group, F(1, 26) = 15.24, p = .001, η_p ² = .370, in which accuracy for the TD group was significantly higher (66% accuracy for old items and 70% accuracy for new items) than the DLD group (58% accuracy for old items and 54% accuracy for new items). The main effect of exposure was not significant, F(1, 26) = 0.25, p = .623, η_p ² = .009, and the interaction between group and exposure was not significant, F(1, 26) = 3.07, p = .092, η_p ² = .106.

Stay and Switch Behaviors Following Feedback

To evaluate participants' response to feedback during training, we computed the proportion of positive feedback trials that was followed by a stay behavior and the proportion of negative feedback trials followed by a switch behavior. This was done on an item-by-item basis, meaning that values were computed for each of the eight stimuli and then averaged together for each participant. The proportion of stay and switch trials by group is presented in Figure 4. A mixed ANOVA with group (TD and DLD) as a between-subjects variable and stay/switch behavior (positive-stay and negative-switch) as a within-subject variable was conducted. The main effect of group was not significant, F(1, 26) = 4.06, p = .054, η_p ² = .135. There was a significant main effect of stay/switch behavior, F(1, 26) = 13.08, p = .001, η_p ² = .335, indicating that stay behavior following positive feedback (M = 0.57, SD = 0.11) was more prevalent than switch behavior following negative feedback (M = 0.50, SD = 0.07). There was also a significant interaction between group and stay/switch behavior, F(1, 26) = 7.65, p = .010, η_p ² = .227. Pairwise comparison indicated that children with DLD and TD did not differ in their switch behavior (p = .941). However, they differed in their stay behavior (p = .007), with children in the TD group showing significantly more stay behavior (M = 0.63, SD = 0.11) than children in the DLD group (M = 0.52, SD = 0.09).

Figure 4.

Proportion of stay/switch behavior by group. TD = typical development; DLD = developmental language disorder.

Electrophysiological Measures of Feedback Processing

To examine differences in brain activation associated with feedback valence (i.e., positive and negative feedback) within and between groups (i.e., TD and DLD), a two-way repeated measures ANOVA was conducted on the amplitudes of the FRN and P3a yielded by TPCA from electrode FCz. Figure 5 presents the grand average waveforms associated with positive and negative feedback at FCz in the two groups. Figure 6 presents the amplitude factor scores for each component by feedback valence and group.

Figure 5.

Grand average event-related potentials elicited by positive and negative feedback in children with typical development (TD) and children with developmental language disorder (DLD) at frontocentral electrode FCz. FRN = feedback-related negativity.

Figure 6.

Amplitudes (in factor scores) of the feedback-related negativity (FRN) and frontocentral P3a to positive and negative feedback in children with typical development (TD; on the left) and children with developmental language disorder (DLD; on the right). FB = feedback.

FRN

A main effect of feedback valence was observed, F(1, 25) = 39.77, p < .001, η_p ² = 0.61, indicating that negative feedback elicited a larger FRN amplitude when compared to positive feedback. Although there was no group effect, F(1, 25) = 0.74, p = .40, η_p ² = .03, an interaction between group and feedback was found, F(1, 25) = 4.39, p = .047, η_p ² = .15. A follow-up t test revealed that the difference of the FRN amplitudes between negative and positive feedback was more pronounced in the TD group when compared to the DLD group, t(29) = 2.095, p = .047.

P3a

A significant main effect of feedback was found, F(1, 25) = 6.08, p = .021, η_p ² = .20, specifying that the P3a amplitude was larger (more positive) in response to negative feedback than to positive feedback. There was no significant main effect of group, F(1, 25) = 0.03, p = .865, η_p ² = .001. Although the grand average data suggest amplitude differences between the two groups in relation to negative feedback, there was no interaction between group and feedback valence, F(1, 25) = 1.90, p = .180, η_p ² = .07.

Correlations Between Electrophysiological Measures of Feedback Processing and Stay/Switch Behavior

We examined the relationships between stay/switch behaviors and the feedback-related ERP components. More specifically, the amplitudes of the FRN and P3a in response to positive feedback were evaluated in relation to stay behaviors, and the amplitudes of the FRN and P3a in response to negative feedback were evaluated in relation to switch behaviors. The results indicated no relationships between the FRN and stay/switch behaviors in both groups. Strong positive correlations, however, were found between the P3a and stay/switch behaviors but only in the TD group. Correlations are presented in Table 3. In the TD group, the P3a amplitude to positive feedback was correlated with stay behaviors, r = .66, p = .01, and the P3a to negative feedback was correlated with switch behaviors, r = .73, p = .003.

Table 3.

Correlations between stay/switch behavior and event-related potential measures.

Variable	TD		DLD
	Positive feedback	Negative feedback	Positive feedback	Negative feedback
	Stay	Switch	Stay	Switch
FRN	.33	.19	.15	−.19
P3a	.66*	.73**	.22	.36

Note. TD = typical development; DLD = developmental language disorder; FRN = feedback-related negativity.

^*Correlation is significant at the .05 level (two-tailed).

^**Correlation is significant at the .01 level (two-tailed).

Summary of Results

The results indicated that children with DLD achieved lower accuracy scores during training and testing when compared with their age-matched peers with TD. They also did not demonstrate patterns of probabilistic learning that were found in children with TD. Although both groups were as likely to switch or repeat a behavior following negative feedback, repeating a correct behavior following positive feedback was less frequent in the DLD group than in the TD group. The feedback-related electrophysiological markers (i.e., the FRN and P3a ERPs) indicated reduced response differences between positive and negative feedback in children with DLD and no statistically significant group differences in the amplitude of the P3a. Although the FRN was not found to be related to stay and switch behaviors in both groups, the P3a amplitude was associated with stay and switch but only in the TD group.

Discussion

The performance of children with DLD on the feedback-based probabilistic learning task was found to be inferior to that presented by children with TD. The stay and switch behaviors and the electrophysiological data indicate that the poor performance was driven, at least in part, by ineffective processing of feedback by children with DLD. More specifically, the stay and switch behaviors demonstrated that although both groups of children were inconsistent in their response to negative feedback, the benefit of positive feedback to learning was more pronounced in the TD group. These patterns of responses to positive and negative feedback are consistent with previous findings by Arbel et al. (2021), who employed a feedback-based declarative learning task. The similarity across studies is interesting given the different role the feedback plays in declarative and nondeclarative probabilistic learning. Recall that feedback in the declarative learning task is informative on a trial-by-trial basis, whereas in the probabilistic category learning task, feedback is only useful across multiple trials that allow the accumulation of probabilistic information carried by the feedback. The current results strengthen the hypothesis that feedback processing is inefficient in children with DLD regardless of the paradigm in which it is presented. It is important to note that Lee and Tomblin (2012) reported that adults with and without DLD who performed a probabilistic reinforcement learning task did not differ in their stay and switch behaviors. Both groups were more likely to repeat a behavior following positive feedback (stay) than to change a behavior following negative feedback (switch). It is possible that the relatively intact stay behavior in adults with DLD reported by Lee and Tomblin can be attributed to an age-related increase in working memory capacity. A longitudinal evaluation of the relationship between stay and switch behavior and working memory in school-age children supports this hypothesis (Arbel et al., 2021) by finding age-related changes in stay behavior and a positive correlation between working memory scores and stay behavior.

The electrophysiological data suggest that, in children with DLD, there was a smaller processing difference between positive and negative feedback than in children with TD. These findings may be interpreted to suggest that children with DLD were less efficient in distinguishing between the information carried by positive and negative feedback. These results are in line with previous reports from our lab (Arbel & Donchin, 2014; Arbel et al., 2021). Interestingly, the amplitude of the FRN was not found to be correlated with learning outcomes in either group. These results are in contrast with previous reports of a relationship between the FRN and learning in adults (e.g., Arbel et al., 2013; Arbel & Wu, 2016; Krigolson et al., 2009; Luft, 2014; Pietschmann et al., 2008; Sailer et al., 2010; Van den Bos et al., 2009). Reports on the FRN and learning in children are limited and point to no relationship between the processing of negative feedback and learning outcomes. However, there are indications that FRN amplitude to positive feedback is associated with learning progress and strategy adjustments in children (Shephard et al., 2014) and that there are developmental changes in this relationship between the processing of positive feedback and learning outcomes reflected in the FRN to positive feedback (Arbel & Fox, 2021). When comparing the ERPs elicited in this study with the ERPs reported in the declarative learning task by Arbel et al. (2021), some noteworthy differences emerge for children with DLD. More specifically, in children with DLD, no FRN was detected during declarative learning (Arbel et al., 2021), but a clear FRN in response to negative feedback was observed during nondeclarative learning in this study. One possible explanation for the discrepancy in the results is that, in a feedback-based declarative learning task, two systems are activated simultaneously, the medial temporal lobe, which supports declarative learning, and the striatum, which is involved in feedback processing. In feedback-based probabilistic learning, on the other hand, there may be an overlap in the recruitment of brain pathways, as both probabilistic learning and feedback processing rely on striatal activation. It is, therefore, possible that, for children with DLD, the processing of feedback is affected, to some extent, by the learning paradigm and that the differences are driven by whether one or two learning systems need to be activated at once.

This study found no differences in the amplitude of the P3a between the two groups. However, the P3a was found to be correlated with stay and switch behaviors only in the TD group. We suggest that, in the context of the probabilistic classification task, the FRN amplitude reflects an initial evaluation of the outcomes of an action communicated through feedback and that the P3a triggers the process of updating reinforcement information. Our results indicate that the P3a was elicited in children with DLD, suggesting that learners were aware of the action–feedback contingencies and distinguished between positive and negative outcomes within this context. The lack of correlation between the P3a and subsequent behaviors in the DLD group may be interpreted to suggest that this information was not maintained in memory or that representations of action–outcomes contingencies were not stable enough to allow for it to affect future actions. The strong correlations between stay/switch behavior and P3a elicited by positive/negative feedback in children with TD may reflect their ability to use the information carried by the feedback to update and maintain reinforcement information and guide future actions. These results are in accordance with previous work indicating a relationship between the amplitude of the P3a and learning outcomes in adults and children with TD (Arbel & Fox, 2021; Arbel & Wu, 2016).

Beyond Feedback Processing—Probabilistic Learning in DLD

Although much of the current evidence points to inefficient feedback processing in children with DLD, it is possible that children with this disorder have an inherent deficit of probabilistic or implicit category learning. An evaluation of accuracy as a function of stimuli distance from the prototype in this type of classification task indicates whether participants' accuracy reflected the probabilistic affiliation of its items with each of the two categories. Across groups, accuracy was greatest for prototypes, followed by exemplars that were one feature away and then exemplars that were two features away. This pattern was predicted based on the probabilistic nature of the task design, and it was most prominent in the TD group. Although there was not a significant interaction between group and stimulus type, it appears that children in the DLD group demonstrated a similar level of accuracy across all stimulus types, which was significantly below that of the TD group and just slightly above chance. Similarly, in Vallila-Rohter and Kiran (2013), adults with aphasia demonstrated a pattern of learning that was similar across stimulus types and generally less accurate than that of the control group, whereas healthy adults exhibited a significant linear trend for distance (i.e., stimulus type) whereby accuracy decreased as distance from the prototype increased.

Few studies have examined feedback-based probabilistic learning in children with DLD with a notable exception by Kemény and Lukács (2010). In their study, children with DLD performed a weather prediction task that has many similarities to the probabilistic classification task in this study. Both tasks require the use of corrective feedback to determine the correct classification for stimuli with complex probabilistic qualities. Kemény and Lukács found that children with DLD demonstrated significantly lower classification accuracy than children with TD and adults. Over the course of three blocks, the performance of children with DLD remained at near-chance levels with a maximum accuracy of 58% by the final block. Our sample of children with DLD demonstrated similar learning outcomes during the training phase with some evidence of consolidation during the subsequent testing phase. It should be noted that Kemény and Lukács did not focus in their analysis or discussion on the possible contribution of feedback processing to learning within this paradigm.

Limitations and Future Directions

This study has a few limitations worth discussing. One limitation is the small number of participants. It is noteworthy that effect sizes for group differences were quite large for behavioral learning outcomes, with η_p ² ranging from .135 to .442. It is possible that the experimental task was too difficult for some children with DLD whose averaged performance during training remained near-chance levels. Future work should seek to examine probabilistic learning in larger samples of children with and without DLD using a task that provides learners with more exposure opportunities to training exemplars. It is also important that future work strives to dissociate the effects of feedback processing and learning ability for children with DLD. One approach involves comparing the same learning paradigms presented with and without feedback. By manipulating the degree to which feedback is required for learning, the function of a given learning system can be evaluated more precisely. In declarative learning tasks, there is evidence that children with DLD benefit from a feedback-free or “errorless” learning environment (Arbel et al., 2021); however, not much is known about nondeclarative or probabilistic learning with and without feedback in this population. In young adults with developmental dyslexia, a disorder that shares some foundations with DLD, probabilistic learning during a weather prediction task was found to be impaired regardless of whether feedback was presented or not (Gabay et al., 2015). Future studies should examine probabilistic learning paradigms with and without feedback to better isolate learning outcomes from feedback processing abilities.

Conclusions and Clinical Implications

This study examined feedback processing in the context of a PCL task and provides behavioral and electrophysiological evidence that children with DLD do not benefit from performance feedback to the same extent as their peers in the context of implicit learning. A similar conclusion was reached for a declarative task that provided more direct, informative feedback within a paired-associate learning task (Arbel et al., 2021). Therefore, it is important to consider the role of performance or corrective feedback, which is commonly part of general instruction for school-age children as well as many interventions provided to children with DLD. Corrective feedback is often part of structured and explicit teaching strategies that engage the declarative learning system. The feedback is deterministic and highly informative, much like the task and feedback examined by Arbel et al. (2021). There are other approaches, deemed implicit interventions, that do not overtly rely on the use of corrective feedback. However, these implicit interventions often still engage feedback processing mechanisms to varying degrees. For example, corrective recasting (e.g., Wada et al., 2020) is a common implicit intervention technique in which a clinician repeats and/or modifies a child's productions. This type of response, although not a traditional form of feedback, is likely to increase the child's awareness of their accuracy and engage feedback processing mechanisms. An alternative approach is to employ interventions that are less likely to engage feedback processing mechanisms such as auditory bombardment (e.g., Plante et al., 2018). Bombardment interventions utilize a specific type of modeling that involves brief but high-density presentations of grammatical targets within short and variable sentences. Importantly, this procedure does not provide any instructions about the activity or require any productions from the child. A growing number of studies suggest that this type of feedback-free learning environment can benefit language development (Encinas & Plante, 2016; Michalek et al., 2021; Plante et al., 2018). Because feedback is a natural and often necessary component of instruction for school-age children, it is important to understand the factors that affect the ability to learn from feedback and the conditions under which learning from feedback is optimal.

Funding Statement

This material is based upon work supported by the National Institute on Deafness and Other Communication Disorders under Grant R15DC016438-01A1 awarded to Y. Arbel.