Teaching Children with Autism Spectrum Disorder Desire-Based Emotion Prediction and Cause

Abstract

This study aimed to expand current research in one area of perspective taking related to teaching children with autism spectrum disorder to predict others’ emotions. The current study evaluated a behavioral teaching procedure on predicting and inferring the cause of emotions based on another’s desires. The procedure included a training package including multiple-exemplar training, rules, modeling, prompting, and reinforcement across scenarios in which children with autism were asked to predict how others may feel given a met or unmet desire or nondesire. Three children with autism, who did not already demonstrate this skill at baseline, were included in the study and learned a repertoire of emotion prediction and cause that generalized to untrained novel scenarios. Generalization to situations in which it was necessary to apply information about another’s desires during play activities was not observed until direct in-vivo training was implemented. Future directions and implications of this research are discussed.

Perspective taking is referred to as inferring how others may think, feel, or perceive stimuli in the environment in order to understand and predict how they may respond to these stimuli (Howlin et al., 1999). Being able to take another’s perspective is crucial to social functioning, and deficits in this area could presumably hinder the development of social relationships. Research on perspective taking has documented deficits among some children with autism spectrum disorder (ASD; Baron-Cohen et al., 2000), and it is possible that these deficits may relate to one of the core characteristics of ASD, i.e., deficits in social communication and interaction (American Psychiatric Association, 2013).

An area of perspective taking that is presumably important for making and maintaining friendships is understanding emotional cause and effect. Developmental research in the area of theory of mind has identified that in typical child development this begins with basic emotion recognition. In particular, by 3 years of age, children use emotion labels and can discriminate among happy, scared, and mad in themselves and others (Stein & Levine, 1989). Also, at this age, they understand situation-based emotions wherein stereotypical emotions are likely to occur under common situations, e.g., most everyone is happy when given a gift (Harris et al., 1989). They also understand desire-based emotions wherein one is likely to be happy when one’s desires are met (receive what one wants for their birthday) and unhappy when they are unmet (receive something one already has or dislikes for their birthday; Yuill, 1984). Around the ages of 4–6, children of typical development understand belief-based emotions wherein one’s emotions change when they realize that something they believed to be true is actually false and vice versa (e.g., one believes their bike is in the garage, but it has been moved and is no longer there; Howlin et al., 1999). In contrast, compared to same-aged peers of typical development, some children with ASD demonstrate difficulty identifying emotions evoked by particular situations (Downs & Smith, 2004; Harris et al., 1989).

Several studies have evaluated procedures for teaching individuals with ASD emotion-related skills. For example, studies utilizing computer programs designed to improve emotion recognition and prediction show encouraging results. Silver and Oakes (2001) used the Emotion Trainer software to teach emotion recognition and prediction for situation-based, desire-based, and belief-based emotions. This software also facilitated improvement to generalization tasks; however, generalization to a wider range of tasks and natural environment situations were not evaluated. Likewise, LaCava et al. (2010) evaluated the use of the Mind Reading software for teaching children emotion recognition based on facial expressions, voice recordings, and situations. Results indicated that, in addition to increases in emotion recognition, improvements were observed in social interactions with peers. However, the changes in social behavior were small and it was unclear whether the intervention was responsible for them. Furthermore, Petrovska and Trajkovski (2019) found that the use of Ucime Emocii (learning emotions) demonstrated improved recognition of situation-based emotions in children diagnosed with ASD with and without intellectual disabilities. Question-and-answer format with pictures and feedback has also been shown to improve emotion recognition and prediction for facial expression recognition, situation-based, desire-based, and belief-based emotions but without evidence of generalization to distant tasks and natural environment situations (Hadwin et al., 1996).

Behavior analytic approaches to perspective-taking began with Skinner’s analysis of the likely ways in which the verbal community teaches children to respond to their own private events. Skinner stated that, since the verbal community does not, by definition, have access to a child’s actual private events, they must infer those private events on the basis of likely correlated overt stimuli (e.g., the overt stimulus of a bleeding wound is likely correlated with the covert behavior of feeling pain) and/or likely correlated overt behavior (e.g., the overt behavior of eating rapidly may be correlated with the private event of feeling hunger; Skinner, 2011). Upon seeing those (one hopes) correlated overt events, the verbal community can then prompt the child to respond to their own private event, for example, tacting pain (e.g., “Oh, you hurt your knee, does that hurt?”). For children to later take the perspective of others, they must then do what the verbal community did when teaching them about their own private events: They must infer the private events of others by responding to conventionally correlated overt stimuli and behaviors. Thereby, a child may infer that when mommy is frowning, she feels sad or mad, for example.

A few studies have used Skinner’s analysis of private events described above to construct behavioral intervention techniques to teach individuals with ASD to tact situation-based emotions. McHugh et al. (2011) used multiple-exemplar training, prompting, and reinforcement while showing video clips of events to participants wherein the video was paused just before a puppet character reacted emotionally and participants were asked how the event will make the character feel (e.g., “How do you think Shrek will feel when his candy gets stolen?”). Generalization was observed across people, settings, and video clips; however, generalization was not tested in natural environment situations. Conallen and Reed (2016) taught children with ASD to match pictures of emotions to pictures of situations and pictures of situations to pictures of emotions using prompting and praise. Generalization to untrained picture cards was observed and participants were also observed to choose pictures of situations that made them feel given emotions; however, generalization to situations in the natural environment was not measured. Schmick et al. (2018) used relational training such as stimulus equivalence and transformation of stimulus function to teach three children with ASD to tact situation-based emotions demonstrated in video-based scenarios. Generalization to untrained video-based situations and natural environment situations was not evaluated. Belisle et al. (2020) evaluated the use of public accompaniment of stimuli (e.g., a bandage on one’s arm to indicate “hurt” or laying one’s head on a pillow to indicate “tired”) when teaching three children with ASD to tact the emotions of others in vivo with a most-to-least intrusive prompting procedure. Generalization to untrained situations was not assessed.

Although previous research has demonstrated that participants learned to tact others’ emotions using behavioral procedures, responding was under stimulus control of facial expressions (Conallen & Reed, 2016; McHugh et al., 2011; Schmick et al., 2018) or accompanying stimuli (Belisle et al., 2020) depicting common situations in which people would display stereotypical emotions (e.g., one usually feels happy at their birthday party), thus one could arguably make the case that situation-based emotions do not necessarily require children to engage in perspective taking per se. Rather, a learner can respond correctly in these situations merely by identifying culturally common relations between situations and the emotions they commonly cause, without taking the perspective of any particular person. For example, getting what one wants generally results in a positive emotion, regardless of what the particular thing is or who the particular person is, therefore it does not require another person to take the unique perspective of that person.

Desire-based emotions may involve a different relational repertoire because desires for any given person may be different than one’s own desires. For example, if I am aware that my friend Sam loves basketball and thinks we are going to play basketball, but I forgot to bring my basketball to the park, I will need to make an inference about how Sam will feel when I tell her that I forgot to bring the basketball, predict how she will behave, and adapt my own behavior based on these inferences and predictions. But if I have another friend, Jimmy, who hates basketball, I will have to make a new set of inferences and predictions and adapt accordingly.

Perspective taking that involves desire-based emotions seems to involve multiple component responses. First, one must understand another’s desires and how these may differ from their own desires. Understanding the unique desires of a friend may involve hierarchical relational responding, wherein one may frame multiple positive reinforcers in a hierarchical relation with one’s friend (e.g., “Sam loves basketball, dolls, and peanut butter sandwiches”). In addition, one may need to deictically frame (e.g., “they/me”) these same stimuli in terms of other relations, including relations of distinction, with one’s own hierarchical relations, e.g., “Sam loves peanut butter sandwiches but I hate them.” Second, one must predict how another’s emotions may be affected based on if the desire is met or unmet. Such behavior seems to involve relational framing in terms of causation or conditionality, for example, “If we have peanut butter sandwiches for lunch, then that will make Sam feel happy.” Finally, one must predict how another will respond (e.g., smile, show excitement, frown, cry) based on the predicted emotion, for example, “If Sam feels happy, she will smile and want to eat lunch together.” For many typically developing adults, these relatively simple social behaviors seem to occur with little effort, but it seems plausible that they entail relatively complex relational responding. Individuals who do not possess these repertoires of perspective-taking behavior will likely not be able to choose social activities that balance their own desires with the desires of their peers. In particular, if one is not able to discriminate one’s own desires from a peer’s different desire, then one can only be expected to choose activities that are maximally positively reinforcing to oneself. In such cases, one may be judged by others to be selfish or aloof when actually it may merely be a perspective-taking skill deficit. Teaching these skills to children early in their social development may be beneficial to a child’s relationships with peers and potentially influence making and maintaining friendships.

The current study aimed to extend previous research in several ways. First, given its success in previous research targeting teaching the acquisition of perspective-taking skills, we employed a treatment package consisting of multiple-exemplar training, rule following, modeling, prompting, and reinforcement (e.g., Najdowski et al., 2017; Najdowski et al., 2018; Persicke et al., 2013; Ranick et al., 2013; Suarez et al., 2021). We taught children with ASD the generalized skill of emotion prediction based on another’s desires when given information about their specific individualized met or unmet desires or nondesires. Thus, participants would learn to take the perspective of each person and that person’s desires during given situations and adjust their responses accordingly. In particular, participants were taught to respond to a desire-based prediction question (e.g., “How will she feel if she wants/does not want chocolate ice cream and gets/does not get chocolate ice cream?”) upon hearing the desired-based emotion scenario (e.g., “She likes/does not like chocolate ice cream and wants/does not want to get chocolate ice cream after school.”). In addition to teaching participants to predict another’s emotion, the current study also taught them to identify the cause of the emotion (“Why will she feel happy/sad?”). Second, generalization to novel, untrained scenarios was trained and assessed throughout the study. Finally, the study included an evaluation and training of the application of predicted desire-based emotions to real-life play situations in which the participants would need to apply the emotion prediction and cause information to have a successful social interaction.

Method

Participants and Settings

Three participants diagnosed with ASD were recruited from a behavioral treatment agency specializing in applied behavior analytic interventions for children with ASD. Participants were referred by clinical treatment teams that deemed desire-based emotion prediction and cause to be a clinically necessary skill for the participants to learn as part of their ongoing treatment program. All participants' primary language was English and they communicated in full sentences, had well-developed language skills, including listener behavior, echoics, mands, tacts, and intraverbals. As assessed using the Skills (n.d.) assessment, they displayed prerequisites including tacts and intraverbals associated with objects, people, actions, categories, “wh-” questions, and had an established repertoire of identifying their own and others’ emotions. Because questions related to emotions and desires were asked throughout the study’s procedures, it was ensured that participants identified what they and others wanted when asked (e.g., “What do you want?” and “What does [person] want?”).

Riley was an 8-year-old Caucasian girl who had been receiving intensive 1:1 behavioral intervention 25–40 hr per week beginning at the age of 4. At the time of the study, she was receiving 8 hr per week of behavioral services in the home setting and 2 hr per week in a clinic setting. Riley was attending second grade at a private school and had a trained behavioral 1:1 aide providing support in the classroom for 30 hr per week.

Isabel was a 6-year-old girl of a Caucasian mother and Filipino father who had been receiving intensive behavioral intervention 25–35 hr per week of 1:1 intervention in the home and community settings beginning at the age of 4. Isabel attended kindergarten at a private school without support during the course of the study.

Devin was a 6-year-old Caucasian boy who had been receiving intensive 1:1 behavioral intervention 30–40 hr per week in the school and home settings beginning at the age of 4. Devin was attending kindergarten at a private school and had trained behavioral 1:1 aides providing support in the classroom.

Sessions for all participants occurred 1–3 days per week in the home, clinic, or school (Riley only) settings. Sessions lasted no longer than 1 hr and no more than three sessions were conducted on a given day and were often interspersed with ongoing behavioral treatment goals.

Data Collection and Interobserver Agreement (IOA)

Data were collected on the percentage of correct responses to the presentation of desire-based emotion scenarios across a trial block consisting of 4 trials in baseline and posttraining sessions, and 10–20 trials during training sessions. Scenarios were predetermined and randomly selected from a random number generator for each session from a pool of scenarios developed by the experimenters. Correct responses included: (1) accurate prediction of the desire-based emotion presented in the scenario and (2) accurately inferring causality of the predicted emotion. Correct responses to both of the two questions was required in order for the participants to receive a correct score on any given trial. In addition, responses initiated after 3 s of each question were scored as incorrect.

A second trained observer collected interobserver agreement (IOA) data on 33%, 37%, and 40% of sessions for Riley, Isabel, and Devin, respectively. IOA was calculated on a trial-by-trial basis by counting the number of agreements and dividing by the number of agreements plus disagreements and multiplying by 100 to obtain the percentage of agreement. Mean agreement for correct responses equaled 100% for Riley and Isabel and 99% for Devin.

Procedure

A nonconcurrent multiple baseline across participants design was employed to evaluate the effect of the training package on the acquisition of desire-based emotion prediction and cause.

Baseline/Posttraining

The baseline and posttraining phases were procedurally identical. Each baseline session included four trials containing different scenarios in which desires were met or unmet (i.e., one scenario for each of the four conditions was presented; see Table 1 for examples of scenarios, SDs, and responses). In order to measure stimulus generalization during posttraining, the same scenarios used in baseline were repeated in posttraining but were never presented during training. Each two to three sentence scenario depicted a third person’s met/unmet desire. After reading the scenario to the participant, the experimenter would present the desire-based emotion prediction and cause questions: (SD 1) If (individual) gets/does not get (desire/nondesire) how will (pronoun) feel?; and (SD 2) Why will (individual) feel (emotion)? A trial was scored as correct if the participant responded correctly to both questions within 3 s. If either one or both questions were answered incorrectly the trial was scored as incorrect. Differential consequences were not provided for correct or incorrect responding during baseline or posttraining phases.

Table 1.

Examples of desire-based emotion scenarios, SDs, and responses across each condition

	Scenario	SD/R 1	SD/R 2
Desire–Met	Sara loves Skittles. On Halloween, Sara goes trick-or-treating and gets skittles at the first house she visits.	SD 1: If Sara gets Skittles, how will she feel? R 1: Happy (or other positive emotion).	SD 2: Why will Sara feel happy? R 2: Because she wanted Skittles and she got Skittles.
Desire–Not Met	Sara loves Skittles. On Halloween, Sara goes trick-or-treating and does not get Skittles at the first house she visits.	SD 1: If Sara does not get Skittles, how will she feel? R 1: Sad (or other negative emotion).	SD 2: Why will Sara feel sad? R 2: Because she wanted Skittles and she did not get Skittles.
Nondesire–Met	Sara hates M&Ms. On Halloween, Sara goes trick-or-treating and gets M&Ms at the first house she visits.	SD 1: If Sara gets M&Ms, how will she feel? R 1: Sad (or other negative emotion).	SD 2: Why will Sara feel sad? R 2: Because she did not want M&Ms and she got M&Ms.
Nondesire–Not Met	Sara hates M&Ms. On Halloween, Sara goes trick-or-treating and does not get M&Ms at the first house she visits.	SD 1: If Sara does not get M&Ms, how will she feel? R 1: Happy (or other positive emotion).	SD 2: Why will Sara feel happy? R 2: Because she did not want M&Ms and she did not get M&Ms.

A correct response to the desire-based emotion prediction question (i.e., SD = 1) included known positive (e.g., happy, excited, fantastic) and negative (e.g., sad, disappointed) emotion labels. It should be noted that although several positive and negative emotion labels were considered correct, the majority of participant responses were happy or sad throughout the study.

Training

Training was initiated for each participant in a stepwise fashion contingent on stable responding during baseline. Prior to each session, brief informal preference assessments (i.e., giving the participant a choice between two items or activities across two trials) were conducted to determine potential reinforcers presented for correct responses. The reinforcement schedule for each participant was determined based on their current behavior intervention plans and ranged from 10-item token boards to time-based activity completion schedules. Participants were taught to predict emotional responses and cause across four conditions: (1) desire–met; (2) desire–not met; (3) nondesire–met; and (4) nondesire–not met.

During the first three sessions of each condition, a rule was presented at the beginning of each session. For example, for Condition 1 (desire–met) the following rule was presented: “When someone wants something and gets it, they will feel happy, excited, or fantastic, because they got what they wanted.” After the first three sessions of each phase, the rules were only presented as part of the prompting hierarchy or if incorrect responding was observed across two trials as an error correction procedure.

Next, the experimenter began reading the scenarios that were randomly selected for each session. After reading the scenario, the experimenter asked attending questions related to the contents of the scenario to ensure that the participants were able to recall the contents of the scenario before proceeding to the desire-based emotion prediction and cause questions (e.g., “What was the girl’s name in the story?”; “What was the thing that she wanted?”; “What did she get?”; and “Did she get what she wanted?”). Social praise was provided for correct answers to attending questions. If participants did not answer all three attending questions correctly, the story was reread. There were no situations in which the stories were reread more than twice throughout the study due to lack of attending.

After the participant answered the attending questions, the experimenter presented the desire-based emotion prediction and cause questions. The experimenter immediately modeled the correct responses on the first presentation of each new scenario only during the first session of each condition (e.g., “Sara would feel happy because she wanted Skittles and she got Skittles.”). During all sessions thereafter, a least-to-most prompting hierarchy was used as follows. First, the scenario was immediately presented again and the experimenter prompted correct responding after a 3-s prompt delay. In particular, if the participant did not respond correctly within 3 s, the experimenter prompted the correct response using a least-to-most prompting hierarchy by first providing a rule reminder (e.g., “Remember, when a person wants something and gets it they will feel happy or surprised. because they got what they wanted.”) followed by a leading question prompt, a partial echoic prompt, and finally a full echoic prompt, as necessary. That is, if the rule reminder prompt was unsuccessful in producing the desired response, a leading question (e.g., “What did Sara want? Did she get what she wanted? So, how will Sara feel if she gets what she wants?”) was presented, followed by a partial echoic (e.g., “Sara will feel h . . .”) and finally a full echoic (e.g., “Sara will feel happy because she got what she wanted, Skittles”). Once independent correct responding was observed with the first scenario across two consecutive trials, a second, novel scenario was presented and the experimenter repeated the same procedure as above by first modeling the correct response (first session only), then reading the scenario and implementing the least-to-most prompting strategy as needed until independent correct responding was observed with the second scenario across two consecutive trials.

Finally, both scenarios were presented semi-randomly (i.e., no scenario could be presented more than twice consecutively) until correct responding was observed across four consecutive trials. Sessions concluded after 4 consecutive independent correct trials or 20 trials total, whichever occurred first. In particular for Devin, the first three sessions required more intrusive prompting (full echoic prompts) and independent correct responding was not observed until session 4 of training; therefore, the first three sessions of Condition 1 for Devin ended after 20 trials on a prompted correct response.

Once a minimum of 80% accuracy was achieved for the first two scenarios, two novel scenarios were added to the session rotation which began with first-trial probes (Tarbox et al., 2011) for the two new scenarios. Similar to baseline, responses to first-trial probes were not differentially consequated until direct teaching began. Responses to the first trial of novel scenarios were graphed separately from responses on trained trials to assess for generalization to novel stimuli across training sessions.

After data were recorded for first-trial probes with the novel scenarios, these scenarios were interspersed with previously trained scenarios during the following teaching trials throughout the session until independent correct responding was observed across 4 scenarios or after a total of 20 trials were conducted. Each time a minimum of 80% correct was achieved with the existing scenarios, two additional novel scenarios were added into subsequent sessions for the current condition.

The mastery criterion for advancing to the next condition was 100% accuracy across two to three sessions on first-trial probes with novel scenarios. Once the participants met the mastery criterion for each condition, the next condition was introduced beginning with a generalization probe to test for generalization across conditions (e.g., desire–met to desire–not met) and included two first-trial probes of new scenarios related to the new condition. For example, during the first session of Condition 2 (desire–not met), the experimenter presented two new scenarios that depicted situations in which the character in the story did not get what they wanted. Thereafter, training began for the exemplars within that condition as described previously. This method was repeated for all subsequent conditions (i.e., nondesire–met and nondesire–not met).

After participants completed all four conditions, a final phase of training (labeled as “all conditions” on Fig. 1) included the presentation of four novel scenarios per session (one scenario related to each condition). Once 100% accuracy was observed across three sessions in this phase, the posttraining phase began, as described above. As a reminder, none of the scenarios presented during baseline/posttraining were present in any other phase of the study.

Fig. 1.

Percentage of correct responses to emotion prediction and cause scenarios across sessions during baseline/posttraining, multiple exemplar training, and application training conditions. Note. Open circles represent first-trial responses to novel emotion prediction and cause scenarios. Open squares depict responses to novel application probes. Closed squares during the application training phase depict percentage of correct responses to application training scenarios. Closed triangles represent responses to scenarios only presented in baseline and posttraining. Closed circles represent responses to scenarios introduced during multiple exemplar training

Application Phase

Application Probes

Throughout the course of the study, probes were conducted on the application of desire-based emotion prediction and cause to play-based situations with a peer, sibling, or adult to test for generalization to real-life scenarios in which this skill would be applied.

At the beginning of these sessions, the experimenter gave each participant information about the desire or nondesire of the third person in a naturalistic way. For example, during a session with a peer, the experimenter would ask the peer to go get something from another room. While the peer was gone, the experimenter would tell the child, “Megan came to play with you today and she really loves playing Legos and wants to play Legos today.” Then the experimenter would ask the participant the desire-based emotion prediction and cause questions (e.g., “If Megan gets to play with Legos​​ how will she feel?” and “Why will Megan feel happy?”).

Responses to desire-based emotion prediction and cause questions were not differentially consequated during these probes. After a few minutes, the peer would enter and say to the participant, “Let’s play. What should we do?” The experimenter would then record the item/activity that the participant suggested and allow the child and peer to play for 5–10 min. If the participant chose an item/activity that was not indicated as desired by the peer, the peer was instructed to act sad by lowering their head, frowning, and not engaging in the activity. If the participant suggested an item/activity that was desired by the peer, then the peer was instructed to act happy by smiling, laughing, and engaging in the activity. After 5–10 min of the activity, the experimenter asked the participant, “When Megan asked you to play, why did you choose (activity)?” A response was scored as correct if the participants made statements related to the peer’s desire (e.g., “Megan wanted to play Legos​​,” or “Megan likes playing with Legos​​”). A response was scored as incorrect if the participant made a suggestion for a different item/activity or a nondesired item/activity or if they did not respond within 3 s. Approximately 30–60 min later in the session, the same procedure was repeated but the experimenter presented a nondesire of the peer instead (e.g., “Megan really does not like puzzles and does not want to play puzzles today.”), and the same procedure was followed as described for the desired item/activity above.

Activities chosen for the application probes were selected if they were neutrally preferred by the participant. During these probes, the most highly preferred or nonpreferred activities were not selected to ensure that the motivation to engage in an activity was not confounded by the participant’s preferences. In addition, all activities were familiar activities that participants would play with periodically. Application probes occurred at the end of baseline and posttraining phases, and at the end of Condition 2, Condition 4, and the final phase of training.

Application Training

Two of the three participants were included in this phase. Devin was not included, as he was unavailable to continue with the study at this time. This phase was included if participants did not demonstrate the application of desire-based emotion prediction and cause during the application probes, which was the case for both Isabel and Riley.

During this phase, the experimenter directly trained participants to engage in activities desired by others. The following rule was presented: “Megan is your friend and you want her to be happy and have fun playing with you so she will want to play with you again. So, you should choose to play with things that she wants to play. If you choose to play with things she wants to play, she will choose to play with things that you want to play too.” After presenting the rule, the experimenter stated the desire or nondesire of the peer and then asked the emotion prediction and cause questions. For example, the experimenter would state, “If Megan wants to play Legos and she gets to play Legos, how will she feel?” The participant should reply with “happy” or “She will feel happy.” The experimenter would then ask, “Why will she feel happy?” and the participant would then reply, “Because she wants to play Legos and she gets to play Legos.” Correct responses received social praise. For incorrect responses, the experimenter initiated the least-to-most prompting procedure as described in the training conditions above. In this phase, one additional question was presented after the desire-based emotion prediction and cause questions: “What should you choose to play when Megan asks you to play today?” This question was included to prompt the participant to identify a specific activity before the peer entered the room. If the participant responded incorrectly to this question, the experimenter presented leading questions (e.g., “What did Megan want?”; “What can you offer her to make her happy?”). If the participant continued to respond incorrectly, the experimenter presented another leading question, “If you don’t take turns playing things that Megan wants to play, do you think she will come back and play with you later?”; and “So if you want Megan to be happy, what should you do?” The participants did not need more intrusive prompting beyond these two leading questions.

If the participant did not select the desired activity (or selected the nondesired activity), the peer was instructed to act unhappy by frowning and not engaging with the activity. The experimenter then asked the participant to look at the peer and identify their emotional response and its cause. For example, the experimenter asked, “Look at Megan, how do you think she feels?” and the participant replied with the correct emotion, “Sad.” Immediately following, the experimenter asked, “Why does she feel sad?”, and the participant would respond correctly with, “Because she wanted to play with Legos​​ and didn’t get to play with Legos.” The experimenter then said, “What can you do to make Megan feel better?” and, if needed, used least-to-most prompting to help the participant to select the activity that the peer desired (e.g., “That’s right, now ask Megan to play Legos with you.”). Once the participant engaged in the peer’s desired activity for 5–10 min, the peer was instructed to say the following, “Thank you for playing [activity] with me. It made me so happy. Since you are my friend I want you to be happy too, so let’s play something you want to play.” Allowing the participant access to their preferred activity contingent on engaging in the peer’s desired activity was assumed to act as positive reinforcement. In addition, verbal praise was also delivered contingent on the participant choosing the desired activity of the peer.

Results

Percentages of correct responses during each session were calculated and graphed for emotion prediction/cause questions, first-trial emotion prediction/cause probes, application probes, and application training phase responses.

Isabel

Isabel responded with 25% accuracy across baseline sessions. Isabel quickly learned emotion prediction and cause given a met desire during Condition 1 of training, but did not demonstrate generalization to unmet desire scenarios in Condition 2 or Condition 3 until these scenarios were directly trained. Isabel did demonstrate generalization to unmet nondesire scenarios and did not require direct training on these scenarios during Condition 4. Isabel required seven sessions in which novel scenarios of all conditions were presented during each session before meeting mastery criteria for training to advance to the posttraining phase. During posttraining, Isabel achieved 100% accuracy with all scenarios presented demonstrating acquisition of a generalized repertoire of desire-based emotion prediction and cause. Although Isabel learned this generalized repertoire when presented with novel, untrained scenarios, she failed to apply this information during play with peers until directly trained during the application training phase. Isabel often responded to questions during application probes with her own desires. For example, during application probes when asked why she chose to play with her dolls, she responded, “Because I wanted to play with my dolls.”

Riley

During baseline, Riley responded between 0% and 50% correct. Riley displayed an immediate increase in emotion prediction skills when given a met desire during Condition 1 of training and she required few training sessions with unmet desire scenarios in Condition 2. Despite this quick acquisition, the skill did not generalize to unmet desires in Condition 3, therefore the same number of training sessions (eight sessions) as Condition 1 were necessary for Riley to acquire the skill of emotion prediction and cause for met nondesires (e.g., someone gets what they do not want). Generalization occurred more rapidly to scenarios depicting unmet nondesires in Condition 4 and Riley only required four sessions to reach mastery of the skill in this condition. Riley met the mastery criterion during the final phase of training, which included novel scenarios related to all previously trained conditions in three sessions and advanced to the posttraining phase. During posttraining, Riley responded with 100% accuracy on three of the four scenarios that were probed during the baseline phase. Although Riley demonstrated acquisition of a generalized repertoire of desire-based emotion prediction and cause when listening to emotion scenarios, Riley was unable to use apply this skill to real-life play situations, thus displaying no correct responding during application probes with her peers, siblings, or adults. During the application training phase, Riley required five sessions of training for both desire and nondesire conditions.

Devin

Devin exhibited low rates of responding across baseline sessions, ranging from 0% to 25% correct. Devin required a few training sessions before he responded correctly to emotional predication and cause questions with a met desire in Condition 1 of training. Generalization occurred more rapidly to unmet desire scenarios in Condition 2, and in Condition 3, he responded with near 100% accuracy within two sessions. Devin also was quick to generalize to the unmet nondesire scenarios in Condition 4. During the All Conditions phase of training, Devin responded with 100% accuracy in all three sessions and advanced to the posttraining phase. During posttraining, Devin responded with 100% accuracy consecutively for four trials and then displayed a slight decline in responding. This may have been due to the removal of consequences for responding during this phase. Devin responded with 50% accuracy during application probes with peers, siblings, or adults following Conditions 3 and 4 and during posttraining; however, he was unavailable to continue with the final application training phase.

Discussion

Results indicated that all three participants learned to predict the emotions of others and the cause of such emotions given scenarios depicting met or unmet desires and nondesires. In addition, correct responding was observed across untrained novel scenarios indicating that participants acquired a generalized repertoire of desire-based emotion prediction and cause based on the presented scenarios. Responses on the first trials of novel scenarios demonstrated the establishment of this generalized repertoire (i.e., responding observed in the absence of direct training; Tarbox et al., 2011) for this skillset. Because the goal of this study was to teach desire-based emotion prediction and cause with novel exemplars never directly trained, accuracy on first-trial responses demonstrates the acquisition of a generalized repertoire of this skill.

Previous research on emotion prediction has evaluated situation-based emotions (Belisle et al., 2020; Conallen & Reed, 2016; McHugh et al., 2011; Schmick et al., 2018). This study extends upon previous research on teaching individuals with ASD to predict others’ emotions by the addition of the desires component, suggesting that emotion prediction is teachable when based on less conventional situations where the individual’s opinion, in particular one’s desires, must be considered. In addition, this study adds to the body of literature in support of using applied behavior analytic procedures to teach various skills categorized as perspective-taking or theory of mind skills (e.g., Charlop-Christy & Daneshvar, 2003; LeBlanc et al., 2003; Najdowski et al., 2018). Finally, the study adds to the growing body of literature that demonstrates the potential of multiple-exemplar training to establish generalized operant behavior specifically related to perspective-taking skills with individuals with ASD (e.g., Dhadwal et al., 2021; Lovett & Rehfeldt, 2014; Najdowski et al., 2017; Persicke et al., 2013; Ranick et al., 2013; Suarez et al., 2021; Welsh et al., 2019).

This study, although encouraging, has several limitations. First, it is unknown which components of the multicomponent training package are necessary to obtain accurate responding. The components included in this training package were multiple-exemplar training, rule-following, modeling, prompting, reinforcement in the form of verbal praise, and access to preferred items/activities. It is possible that a component analysis would identify that some of these may be more important than others for the acquisition of a generalized repertoire of desire-based emotion prediction and cause.

Second, although responding may be established in the context of answering questions presented about scripted scenarios, this type of training did not result in generalization to real-life situations in which the participants had the opportunity to apply the skills. During the application probes in this particular study, none of the participants demonstrated the skill by suggesting activities to engage in based on another’s desires. All participants required direct training in the final application training phase of the study in order to apply the skill to real-life scenarios. In particular, Isabel and Riley only applied information about another’s desires after they received direct in-vivo training in applied play situations, even though they correctly predicted what another’s emotion would be if their desire/nondesire was met/unmet immediately before the other individual entered the room. It should be noted that the emotion prediction and cause questions taught during the previous conditions were used during the application training phase when giving the participants information about another’s desires, and therefore, the previous training with contrived scenarios may have expedited acquisition during the application training phase. This limitation should be addressed in future research on desire-based emotion prediction and cause to determine if training with contrived scenarios is necessary to include before teaching the application of this skill to real-life scenarios. It may be that learning the social language associated with the social skills is not necessary for the acquisition of the social skill itself (Peters & Thompson, 2018).

It is also unknown whether incorrect responding observed during the application probes was due to a deficit in the skill or lack of motivation to respond under real-life play conditions. It is possible that motivation to engage in the activity desired by another was very low or nonexistent during the application probes, especially if other highly preferred activities were available. The motivation to play a different, less preferred activity to make another happy may have only increased once engaging in another’s desired activity was directly consequated with subsequent access to their own preferred activity. In this study, the consequences of verbal praise for suggesting and engaging in another’s desired activity and subsequent access to a preferred activity were not provided during application probes throughout the study until the application training phase was implemented. These limitations warrant a discussion and further research on how another’s happiness may or may not become a conditioned reinforcer and under what conditions this may or may not occur.

Both the participants who participated in the application training phase demonstrated acquisition in correctly responding to the desires and nondesires of others. However, whether or not their responses appeared sincere was not measured. Future research could conduct a social validity measure that involves asking parents, peers, and teachers to rate the sincerity of the response.

Another important consideration is that engaging in an activity just to make another person happy can have downfalls, and even dangerous consequences. Although in general, considering another’s happiness can be beneficial in relationships, there may be situations in which this same behavior can be toxic and/or dangerous. It would be important to teach children that making others happy should not result in their own despair and it should not put them in situations that are unsafe or illegal. Therefore, it would be a critical next step to teach relevant and necessary skills like assertiveness, compromise, negotiations, and conflict resolution. In addition, autistic self-advocates have raised concerns with social skills training programs that intend to make autistic people appear or behave in conformity with their typically developing peers (Autistic Self Advocacy Network, n.d.). If a desires-based perspective-taking training program was used to teach autistic children to deny their own “more-autistic” preferences, in favor of their peers “less-autistic” preferences, this would be concerning. However, like all social skills that ABA programs target, desire-based perspective-taking skills should be taught to empower autistic children to achieve goals that they personally value, e.g., enhance relationships with friends they care about having.

In summary, the results of this study are encouraging and expand the current research related to emotion prediction and cause based on another’s desires. In addition, this study poses further important questions for future directions in perspective-taking research conducted with individuals with ASD.

Data Availability

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

Declarations

Conflict of Interest

The authors declare that they have no conflict of interest.

Informed Consent

Informed consent was obtained from participants’ caregivers and all procedures were conducted in accordance with internationally recognized standards for research with human participants.