Mutual Gaze: An Active Ingredient for Social Development in Toddlers with ASD: A Randomized Control Trial

Parents and toddlers aged 18–38 months were recruited through local infant toddler programs, community centers, advocacy groups, physicians’ offices, social media, and word of mouth. Eligibility criteria included: (a) having a chronological age of less than 38 months at the start of the study; (b) receiving an autism classification on the Autism Diagnostic Observation Schedule, Second Edition (ADOS-2; Lord et al. 2012), administered by an ADOS-2 reliable examiner; (c) having no other medical, neurological, or genetic concerns or disorders; and (d) having a primary home language of English or Spanish.

Among the 110 toddlers screened, 92 met the inclusion criteria and were enrolled in the study. The families were randomly assigned into the Pathways (n = 39), communication (n = 26), and SAU control (n = 27) groups (Fig. 1). Fourteen families did not complete the intervention or receive post-assessment because they had a prolonged illness, moved out of the area/country, or an intensive service opportunity became available to them, and they could not be reached to schedule post-intervention assessments. The remaining 78 children received the full intervention and post-assessments. There were no significant differences in the background variables between the families who did not complete the intervention and post-assessment procedures from those who did (Table 1); however, age in months, nonverbal IQ, and parenting stress have small effect sizes.

Descriptions of the 78 children who completed the post-intervention assessments are presented in Table 2. The constitution of the sample mirrored the statewide ECI estimates (Texas Health and Human Services, n.d.). Toddlers, on average, had significant cognitive challenges and came from culturally and socioeconomically diverse families. The sample had a low percentage of Caucasian families and a high percentage of Children’s Health Insurance Program (CHIP) Medicaid-eligible families.

At baseline, prior to the 12-week intervention phase, a clinical researcher asked the parent to report the number of hours that the toddler received which type of community-based services. The mean number of hours of community-based intervention received prior to the start of the study was 1.38 h/week, (SD = 0.95, range = 0–3.13) for the Pathways group; 1.05 h/week (SD = 0.69, range = 0–2.50) for the communication group; and 1.92 h/week (SD = 2.17, range = 0–9) for the SAU group. A one-way analysis of variance (ANOVA) revealed no significant differences in hours of intervention prior to the start of the study (F(2,73) = 2.5, p = .09, ɷ² = –0.04). Although the majority of toddlers in all groups were receiving ECI services, proportionately more children who were later randomized into the Pathways group received no intervention before the study (Table 3). All parents agreed to participate in the study, using an informed consent procedure approved by the university’s Institutional Review Board. There was no cost for participating in the study.

This study used a randomized waitlist control design. The principal investigator used a computer-generated list of random numbers to allocate participants to treatment conditions. Sealed envelopes were used for allocation concealment and were opened after all baseline assessments were completed. Clinical researchers administered the assessments and provided the intervention for the Pathways and communication groups. Recruiting, intake, and pre- and post-intervention testing procedures were identical for the three conditions. The Pathways and communication groups received 12 weeks of project-related home visits. The SAU control group received 12 weeks of intervention from community early-intervention providers. Families in the SAU control group could elect to receive the Pathways intervention at no charge when they completed the study.

Four clinical researchers participated in the study. They were responsible for administering the assessments and providing the intervention. Three had a master’s degree and were certified speech-language pathologists. The fourth had a bachelor’s degree in education and 15 years of experience as an early interventionist. Two of the speech pathologists were bilingual (English–Spanish) and, when requested, provided assessments and interventions in Spanish.

Prior to the start of the study, the clinical researchers achieved reliability on the ADOS-2 (Lord et al. 2012) with one of the authors, who is an ADOS-2-reliable examiner. Specially, each research clinician independently coded and scored video-recorded and live administrations of the ADOS-2, then checked point-by-point reliability until 90% inter-examiner agreement was achieved with the ADOS-2-reliable examiner. This was repeated until 90% reliability was achieved on three consecutive ADOS-2 administrations, which could be video recorded or live. Taking into account the age and ability level of the children in the study, each research clinician demonstrated standardized administrations on three separate occasions for the Toddler Module and Module 1 only. During the study, two clinical researchers were present for each assessment. One administered the ADOS-2 while the second assisted. The two researchers scored the assessments together and discussed any issues that might have come to light. Finally, each researcher clinician independently coded three separate ADOS-2s, every 3 months, to check that there was 80% reliability among the group, which was the case in each instance.

A similar procedure was followed to obtain reliability on other standardized assessments. Specifically, the clinical researchers reviewed administration and scoring procedures and subsequently practiced administering and scoring the test with non-project children and adults until there was 90% reliability with one of the authors. Two clinical researchers were present for child assessments (one administered while the second assisted). Only one clinical researcher administered the adult standardized interviews. All tests were independently scored by two people to check accuracy.

With regard to intervention, the clinical researchers were trained to fidelity, using practice children prior to the start of the study. In addition, they participated in weekly supervision with the first author where they reviewed video recordings of treatment sessions.

The Pathways intervention (Campbell and Hoffman 2012) was developed by practitioners in a community-based Part C program. The Pathways and communication interventions are targeted, manualized programs with English and Spanish versions of manuals in both print and digitized audio formats. Parents received family-centered coaching for 1.5 h/week, using multi-modal strategies to enhance the family’s capacity to promote their child’s development within the natural environment (Table 4). Each unit of the intervention introduced information about ASD and explicit instructions on two or three interactional strategies (Table 5). The interventions utilized play activities and daily routines appropriate to a family’s culture and encouraged parents to use the interactional strategies throughout much of the child’s day.

An important feature of the intervention is that the units are cumulative. Each unit builds on the preceding units to slowly move the parents toward increasingly more sophisticated levels of interaction with their toddler. Once an intervention strategy was introduced, the parent was expected to use that strategy for the remainder of the study. Each week, parents rated their perceived level of competence on each interactional strategy from the current unit and all previous units. Parents had the opportunity to discuss where they felt they were successful and where they felt they had challenges. A new unit was introduced when the parent was comfortable with applying all current and previous learned strategies, and the interventionists rated the parent as implementing the strategies with fidelity, using a 4-point Likert scale (Rollins et al. 2019). Consequently, some units were presented only one time, whereas other units were presented over two or more sessions, according to the parent’s abilities. Overall, the parents’ fidelity ranged from 86 to 100% (M = 96%, SD = .04).

Both the Pathways and the communication interventions promoted dyadic face-to-face reciprocal interactions within social sensory and daily routines. The difference between the two interventions were the strategies presented in Unit 3 (Table 5). Parents in the Pathways group were coached on strategies to engage their toddler in mutual gaze during motivating face-to-face activities and routines, followed by contingent natural reinforcement. The strategy requires the parent to actively establish eye gaze without verbal, visual, or physical prompts when the toddler was not spontaneously engaged in mutual gaze. In contrast, parents in the communication group were coached on strategies to facilitate communication. Strategies included creating situations for communication during motivating routines and activities of daily living, using a core vocabulary, modeling, and prompting communicative attempts. The toddler’s communication was followed by contingent natural reinforcement. Parents in both of the groups were introduced to Unit 3 between Weeks 3 and 5 of the study (M = 4.04 weeks, SD = .64). Consequently, although the difference between the two interventions was only the strategies taught in Unit 3, parents implemented the Unit 3 strategies with their children for a period of seven to nine weeks.

Families in the Pathways and communication groups had to suspend other speech, developmental, or applied behavior analysis (ABA) services for the duration of the study. These families received only 1.5 h/week of project-related intervention for 12 weeks. The SAU control group received intervention from community early-intervention providers during the 12-week intervention period. A clinical researcher asked the parents in the SAU group to report the number of hours that their child attended (a) early intervention, (b) speech/ language services, (c) ABA services, and (d) other community-based services during the post-intervention assessment. Parents reported that the SAU group received an average of 6.65 h (SD = 8.90) of therapy per week, with a range of 1–28 h/week. Of the 24 children, 13 were enrolled in a state-sponsored, local ECI program and received all of their services from that agency (M = 1.69, SD = .79 h/week). Six children had their state services supplemented with speech language and ABA services (M = 15.58, SD = 10.74 h/week). Finally, five children received all of their services privately or as part of other community-based services (M = 7.9, SD = 9.88 h/week).

The clinical researchers collected video recordings of parent-child interactions in the families’ homes at baseline, prior to randomization, and at post-intervention. All parent-child dyads were digitally recorded for 10 min, using an iPad 2 for a wide-angle recording of the interaction (Video Stream 1) and with hidden-camera eyeglasses worn by the parent to capture the child’s eye contact (Video Stream 2). Parents were instructed to play with their child as they typically do. The interventionists gave the parents instructions on where to place the hidden-camera glasses to assist with data collection but made no suggestions or recommendations about the interactional strategies.

For each recording, the two streams of digitized videos (i.e., iPad and glasses) were segmented into 2-second intervals and time-linked, using the conventions of the Child Language Data Exchange System (CHILDES; MacWhinney 2000). This allowed for data reduction, using both partial interval coding (Yoder and Symons 2010) and CHILDES transcription and coding methodology (MacWhinney 2000). Transcription was conducted at the level of the utterance and included all verbal, vocal, and gestural behaviors bounded by a pause or change in a conversational turn (Pan et al. 2005).

For each recording, each measure (defined below) was coded by a different team of research assistants during separate passes through the transcript. All coders received a multimedia coding manual with definitions and examples of coding categories. Coders trained on practice videos until they achieved substantial inter-rater agreement, measured by obtaining a Cohen’s kappa coefficient of .80 or above. To address coder drift, all coders attended weekly lab meetings to discuss coding, and their reliability was checked every 3 months. Most coders maintained good reliability throughout the study. When coder drift was identified, the unreliable coder was retrained on practice videos until a kappa coefficient of .80 was achieved. In addition, a master coder reviewed the unreliable coder’s files from the previous 3 months and recoded when necessary. All coders were blind to group assignment and time (i.e., pre- or post-intervention).

Standardized and video assessments were conducted at baseline prior to randomization and post-intervention, within 2 weeks of the start and stop of the intervention phase. The clinical researchers were blind to group assignment pre-intervention but not post-intervention. A clinical researcher who did not coach the target family conducted post-intervention assessments. All assessments were administered in the family’s home or at a convenient location.

G-Power (Faul et al. 2009) was used to calculate the sample size necessary to find a medium effect (f² = .25) for α = .05 and power (1-β = .80) when conducting R²-increment testing in a fixed model linear regression with one test predictor and between two and 14 total predictors. Effect size estimates were based on the author’s previous studies (Rollins 2016; Rollins et al. 2016), which used similar measures to those proposed here. The results indicated that a sample size of 55 participants is required to find a medium effect with a .2 probability of failing to detect a genuine effect.

IBM SPSS statistical package version 26 was used to analyze the data. Preliminary analyses were conducted to establish pre-intervention group equivalencies on all baseline characteristics found in Table 2 and to identify potential covariates for subsequent regression models. One-way ANOVAs were used to test group equivalencies for the continuous measures, and Chi square analyses were used for the categorical variables. To identify potential covariates, we performed Pearson product correlations between continuous baseline variables and outcome measures and one-way ANOVAs between categorical baseline variables and outcome measures. Covariate analyses were performed separately for the Pathways/SAU group comparison and for the Pathways/communication comparison. Following the principle of parsimony, baseline variables were retained as covariate control variables in subsequent regression models only when they were found to relate to an outcome variable of interest.

To address the research questions, we conducted a series of regression analyses for each question, using group as a dummy-coded variable. The first set of analyses compared the Pathways and the SAU groups, and the second set concerned the Pathways/communication group comparison. For each regression model, covariates, if applicable, were entered into the model first, followed by the dummy-coded group variable. Consequently, the unstandardized regression coefficient (b₁) for the dummy-coded group variable estimates the mean difference between the change (adjusted for covariates when applicable), from pre-to post- intervention, for the Pathways and the respective control group (SAU or Communication). A positive coefficient indicates that the mean change for the Pathways group is b₁ units of measurement more than the mean change for the respective control group. Treatment effects were assessed by the magnitude of the effect obtained from the coefficient of determination (i.e., R²) when only the dummy-coded group variable was in the model or the change in R² (i.e., ΔR²) from the base model that contained the covariates to the full model that contained the covariates and the dummy-coded group variable. The specific effect size calculation was f² = R²/1 − R², and interpretation was based on Cohen (1992), whereby f² = effect size, for which ≥ .02 suggests a small effect; ≥ .15, a medium effect; and ≥ .35, a large effect. In addition, the unadjusted treatment effects are reported as the standardized mean difference in change scores between two groups, or d (Table 6). Following Feingold’s (2013) recommendations for calculating d for pre-test/post-test designs, the standard deviation of the baseline scores, rather than that of the change score, on each measure was used in the calculation. Interpretation of d pre-test/post-test designs also is based on Cohen (1992), whereby d = .2 suggests a small effect; d = .5, a medium effect, and d = .8, a large effect. Further, when covariates were present, we tested for their interaction with group assignment to identify predictors of treatment. All assumptions of regression were analyzed, and no model violations were present.

Pre-treatment group equivalence was analyzed using one-way ANOVAs for continuous variables and Chi square analyses for categorical variables. There were no group differences in baseline variables, and the effect sizes for these comparisons ranged from miniscule to small (Table 2). Pearson product correlations were used to identify continuous covariates for subsequent analyses. For the Pathways and SAU group comparison (n = 56), Δ Comm was related to baseline VABS Social (r = .31, p = .022) and baseline VABS Adapt (r = .44, p = .001); ΔSCB also was related to baseline VABS Social (r = .30, p = .023) and baseline VABS Adapt (r = .38, p = .004); and ΔVABS Adapt was related to age (r = .34, p = .011), MSEL (r = .48, p = .0001), and baseline VABS Adapt (r = .42, p = .001). For the Pathways and communication group comparison (n = 54), ΔSocial Eye Gaze was related to baseline MSEL (r = − .28, p = .038), ΔComm was related to baseline VABS Adapt (r =.29, p = .031), and ΔVABS adapt was related to baseline MSEL (r = .32, p = .018). One-way ANOVAs found no significant categorical covariates for either the Pathways/SAU group comparison or for the Pathways/communication group comparison.

The regression coefficients on the dummy-coded group variable and the associated partial F-statistic and effect size estimates for the outcome variables are presented in Table 8. Any covariates that were entered into the model before the group variable are listed in the first column of the table. When covariates were entered in the model, as they are for ΔComm, ΔSCB, and ΔVABS adapt, the regression coefficients reflect the adjusted mean difference in change on the outcome measures. The results suggest that the difference in change from baseline to post-intervention was significantly greater in the Pathways group for social eye gaze and VABS Social, the two social measures. The magnitude of the effect for both measures is large. The large effect (f²) obtained from the regression models associated with these two measures also are reflected in the d statistics reported in Table 6.

The difference in change on the communication composite, adjusted for baseline VABS Adapt and VABS Social, was not significant, and the effect was miniscule, as it was in the unadjusted effect size reported in Table 6. Together, the two covariates accounted for 21% of the variation in the ΔComm (R² = .21, f(2,52) = 6.85, p = .002). In contrast, the difference in change for the SCB measure, adjusted for baseline VABS Adapt and VABS Social, was statistically significant, and the effect size approached medium size. It is noteworthy that the unadjusted large effect size presented in Table 6 is untrustworthy due to the specification error, as it does not adjust for the two covariates that, together, accounted for 15% of the variation in ΔSCB (R² = .15, f(2,53) = 4.599, p = .014). Finally, adjusting for age, baseline MSEL, and VABS Adapt, the Pathways group made significantly more change in VABS Adapt, and the magnitude of the effect was considered medium, as was the unadjusted effect size seen in Table 6. Together, the covariates accounted for 34% of the variation in the change in adaptive functioning (R² = .34, f(3,52) = 8.90, p = .0001). No significant interaction effects were found when comparing the Pathways and SAU groups.

The regression coefficients on the dummy-coded group variable and the associated partial F-statistic and effect size estimates for the outcome variables are presented in Table 9. Any covariates that were entered into the model before the group variable are listed in the first column of the table. When covariates were entered in the model, as they are for ΔSocial Eye Gaze, ΔComm, and ΔVABS Adapt, the regression coefficients reflect the adjusted mean difference in change on the outcome measures. The results suggest that, adjusting for baseline MSEL, the Pathways group made statistically more change on social eye gaze compared to the communication group, and the effect was of medium size.

The unadjusted large effect presented in Table 6 is untrustworthy due to the specification error, as it does not account for MSEL. Alone, MSEL accounted for 8% of the variation in ΔSocial Eye Gaze (R² = .08 f(1,52) = 4.52, p = .038). The Pathways group also made significantly more change on VABS Social, and the magnitude of the effect approached medium size, which is consistent with the effect size reported in Table 6. There were no significant differences between the Pathways group and the communication group on the two communication measures. Interestingly, the results in Table 6 suggests that there may be a small effect of Pathways on change in SCB; however, our study was not powered to find small effects. Finally, adjusting for baseline MSEL, the Pathways group made significantly more change on VABS Adapt, and the effect was considered small to medium. The covariate, baseline MSEL, accounted for 10% of the variation in ΔVABS Adapt (R² = .10, f(1,52) = 5.95, p = .018). No significant interaction effects were found when comparing the mutual-gaze and communication groups.

The difference between the Pathways and communication group interventions was only one unit (i.e., Unit 3). Due to the similarities between the interventions, we might have expected the effect size for the Pathways/communication group comparison to be small, as was the case for change in SCB. One limitation of this study is that it was not powered to find small effects. It is noteworthy, however, that the two project-related interventions focused on developmentally informed early social skills and that the units were cumulative. The progressive nature of the intervention allowed parent-toddler dyads to develop increasingly more sophisticated interactions before the parent added new and perhaps more difficult elements into the interaction. Therefore, although the difference between the groups was just a single unit, parents embedded the unit’s strategies into their interactions for 5 to 7 weeks.

Another potential limitation is that all of the outcome measures were obtained during a parent-toddler interaction or by parent report. Because we were evaluating a parent-mediated intervention, parents implemented the intervention and, therefore, were not blind to the intervention condition. This suggests that the Pathways and communication groups had an advantage over the SAU group, as they could utilize facilitative strategies learned during the intervention phase (Yoder and Crandall 2019). Similarly, it is not known whether the results found for parents would generalize if unfamiliar adults were interacting with the toddlers (Yoder and Crandall 2019). Future research should include standardized assessments, such as the Communication and Symbolic Behavior Scales, Developmental Profile (Wetherby and Prizant 2002), implemented by unfamiliar interventionists to address the consistency of measurements across groups and generalizability.

Another limitation was the short duration of the 12-week intervention, coupled with the lack of a follow-up assessment several months after families finished the study. It may be that 12 weeks is too short a time for parents to maintain the interactional strategies learned in the study, especially if they receive conflicting messages from professionals when returning to SAU programs. The results need to be replicated and include follow-up assessment 3 and 6 months post-intervention to determine whether the effects are sustained. Finally, although interventionists were blind to group assignment at intake, a limitation of this study is that interventionists were not blinded when administering post-intervention assessments. Fuller and Kaiser (2019) found the effects from this risk to be low; nonetheless, to reduce bias, an interventionist who was not familiar with the family conducted their assessments.