The intervention takes the form of an interactive plush toy (Figure 3), which was designed to be handed over to the child and support in-the-moment soothing (see Theofanopoulou et al [31] and Slovak et al [32] for the design and data from previous deployments).

The toy is introduced to the child as an anxious creature that needs attention from humans, such as soft stroking and hugging. Embedded electronics enable the toy to produce vibration patterns that simulate a heartbeat (ranging from frantic to slow and steady). When picked up, the toy emits a frantic heartbeat that slows down if the child uses calm stroking movements, as registered by the embedded sensors. If the toy is soothed for long enough, the prototype transitions into a purring vibration indicating a calm, contented state. The minimum time for this transition is less than 1 minute, but the transition can take longer depending on the child-specific interactions with the prototype.

The logic model underlying the intervention is assumed to operate on 3 levels building on each other—see the study by Theofanopoulou et al [31] for more details:

Deployments with the research prototypes underpinning the current Purrble [31,32] show that, across all 25 families, children reported that the smart toy was incorporated into the children’s ER practices and engaged with naturally in moments the children wanted to relax or calm down. Specifically, the data from [31] shows that the children interacted with the toy throughout the week-long deployment (eg, average active use for 74.9, SD 64.1 minutes per day; median 60.5), they found the experience enjoyable, and all children requested to keep the toy longer. Children’s emotional connection to the toy appears to have driven this strong engagement. Parents reported satisfaction with and acceptability of the toy. No quantitative data on changes in ER were collected in previous studies because of the small sample size and the focus on feasibility and understanding of appropriation within families.

When compared with a traditional noninteractive plush toy, the intervention model underpinning Purrble includes 2 possible pathways through which the effects should occur:

For these reasons, we argue that a comparison with a nonactive control—such as waiting list or treatment-as-usual (ie, nothing)—would not allow us to distinguish the hypothesized impact on in-the-moment soothing of interactivity versus the emergence of new family routines and would also be open to unequal social desirability bias. However, from the perspective of the hypothesized logic model, it is not necessary for the active control to have exactly the same form factor as the active toy, as long as it is comparable in size, shape, and appeal. In fact, we have explicitly decided not to use deactivated Purrble units as active controls because of the increased risk of unblinding, whereby the participants search for or come across Purrble on the web (or notice the plastic enclosure with electronics inside the toy) and assume that their unit is malfunctioning.

The selected active control toy is the Wild Republic 8″ Hedgehog animal. The selection process was guided by the following requirements: the plush toy needed to have analogous size, weight, and quality of materials, and at least similar (if not higher) visual appeal. We also made sure to include the design characteristics that our previous work suggested were important for the narrative around the toy [31,32]. These included selecting a similarly stylized animal (to enable emotion projection and feelings of care), as well as no visible mouth on the toy (to prevent setting an expectation about the toy’s emotional state as a mouth would imply an emotional expression). In addition, we have adapted the one-page parent-facing descriptions of the narrative that come with Purrble also for the active control unit; as such, the active control families will receive the same general narrative—including that the creature is anxious and needs human care, but without the explicit mentions of the toy interactivity—and the same suggested activities for parents.

To validate that the active control is at least as visually appealing as the intervention, we ran a web-based experiment in which participants were randomly assigned to rate either the Hedgehog or the Purrble images. In both cases, the prompts were professional photos from the front and side on a white background, presented at equal size (Figure 4). The experiment was powered to detect a medium-sized effect (d=0.4) at 80% power for a comparison on a single measure, resulting in a sample size of 200 (1:1 allocation ratio). Participants were recruited through the web-based research platform Prolific, with the survey hosted by Qualtrics (including blocked randomization). Inclusion criteria for parents were the eldest child born in 2010-2013 (approximately aged 8-10 years), country of residence in the United Kingdom or the United States, and above 95% acceptance of tasks on Prolific. The participants were prompted to imagine that their oldest child had received the plush toy pictured above as a present. We then asked 3 questions, with the first question—appeal—preselected as the primary measure: (1) How appealing do you yourself find the toy? (2) How appealing do you think your child would find the toy? and (3) How likely would you be to recommend this toy to another parent?

The results show that the hedgehog was consistently rated higher than Purrble on all 3 questions (Figure 5). This suggests that, if anything, the active control should be more appealing to our study participants than the intervention units: it is a particularly stringent control condition to test the effect of a visually appealing, but noninteractive, plush toy. In other words, if it was purely the visual appeal of the units that would drive child or parent engagement and the resulting ER effects (as opposed to the interactivity of the intervention units), we would expect the active control to show at least as good if not better engagement and reports of changes in child ER from the families.

The primary outcome measure will consist of a composite end-of-day 4-item parent report measure of the perceived child ER ability throughout the day. The specific items are listed below, all measured as a visual analog scale [62,63] with not at all and very much so as the anchors. The composite score for each day will be computed as the mean value across the 4 items:

Specifically, this EMA item composite aims to indirectly tap into day-to-day changes in a child’s ability to downregulate their emotions after (and thus cope with) triggering situations they routinely experience in their daily life. Our expectation is that contact with the Purrble will lead to the following:

By being measured repeatedly over time and within subjects, the items capture the changes in emotional outcomes that would indicate changes in the child’s ER ability, assuming that the trait-based reactivity to daily stressors remains stable. The items were thus selected to tap into the proximal outcomes of ER behavior that (1) would be affected if the intervention is effective, (2) is directly observable by parents, (3) is state-based (rather than trait-based) to enable daily measurement, and (4) is connected to the intervention theory of change.

The item selection drew on a range of established measures of emotion dysregulation (Strengths and Difficulties Questionnaire, SDQ [64], Emotion Regulation Checklist, ERC [65], Difficulties in Emotion Regulation Scale (DERS) [66], and Children’s Emotional Management Scales [67]), as well as qualitative data from previous deployments [31,32] with parental reports of increased emotion regulation after potentially triggering events being the common theme.

We will collect semistructured interview data with parents of up to 40% of the experimental group sample (20-25 families), and approximately 25% of the control group (15 families). The interviews will be conducted within 2 weeks following the primary data collection period. We will specifically aim to recruit families who show the highest or lowest change in the outcome data over the primary period to qualitatively understand the potential moderators of intervention responses for future research.

Following previous work [31], the semistructured interview guide will explore the engagement with the toy, any qualitative changes in child or family behavioral patterns that parents notice, appropriation (ie, how the intervention ended up being used by different participants), and use trajectory over time. In addition, we draw on the data from daily questionnaires as part of the interviews, such as discussing the trajectories of daily parental reports on child ER with the parent (eg, asking about specific instances where there is a spike or as a way of referring to particular times in the deployment).

We will use the following protocol to encourage participants’ adherence to the data collection schedule. All decision points are based solely on data collection, rather than any indication of the intervention use or nonuse. The protocols for daily and weekly data collection were independently run.

The daily measures adherence protocol will be as follows: when a participant misses their end-of-day questionnaire, the system automatically generates a reminder next morning. If a participant has already received 2 automated reminders in a row and again misses daily measures, a research assistant will call the participant (in addition to an automated reminder) on the next workday, following a predetermined call script. If the participant does not respond, they will receive one more reminder the next day, and a second call on the following workday. If no data are received, the participant will be marked as dropped out as they will have missed at least 6 subsequent daily measures (ie, more than 20% of the overall data points). The protocol resets when the participant submits a daily questionnaire. In summary, the daily adherence protocol was as follows: 2× reminder, 1× call+reminder, 1× reminder, and 1× call+reminder, dropped from the study.

The weekly measures adherence protocol will be as follows: the survey links will be sent on Saturday midday, with an automated email reminder on Sunday morning. If data are missing by the end of Sunday, a research assistant will call the participant on Monday, following a predetermined call script. The participants will be sent a new link to the survey, with the possibility of submitting their response for the week by the end of Monday. We will not use adherence to weekly surveys as a decision to drop participants from the study, as weekly surveys do not collect the primary outcome.

All calls and other communications with participants will be logged by the research assistant on the web. Although the research assistant will be able to unblind the participants’ condition if necessary, we do not expect this will be needed in most of the calls.

All analyses of daily EMA outcomes (the primary outcome of daily ER and secondary outcomes of engagement with the toy and mood), as well as the analyses of the weekly outcomes, will be conducted using random-effect models for longitudinal regression. These models consider the nested nature of repeated-measures data and are robust to data missingness and violations of the normality assumption. The regression models will examine the difference in the outcome as a function of the assigned condition (Purrble vs active control). The models will adjust for the baseline levels of the score on the total difficulty scale in the SDQ to decrease noise. To account for participant-to-participant variability in EMA scoring as well as in trajectories of change in EMA scores over time, we will include a random intercept and random slope for time. All analyses will be conducted on an intent-to-treat basis and will use data from all randomized participants regardless of their level of participation.

A separate model will be fitted to each outcome. Given that this is an exploratory trial, we will not formally adjust for multiple comparisons. However, we will be cautious in our conclusions about any significant findings and will interpret all results in light of all performed analyses.

To examine the link between engagement with the intervention and ER outcomes, we will conduct 3 types of exploratory analyses: first, given that we hypothesize that Purrble will be more engaging than the control-condition stuffed animal, we will examine whether intervention engagement is influenced by the treatment condition. To do so, we will regress our daily measure of engagement on the treatment condition, controlling for the previous day’s engagement to reduce noise. Similarly, we will regress the weekly TWEETS scores on the treatment condition, controlling for the previous week’s engagement. Second, we will examine whether engagement moderates the intervention impact on both our primary outcome, the daily measure of emotional regulation, and the weekly ERC and SDQ measurements. For the daily model, we will use our primary measure, the EMA assessment of ER, as the dependent variable, and will include, as regressors, treatment indicators and the EMA measure of engagement, as well as a term for their interaction. To decrease noise, the model will also include as a covariate the previous day’s score on the ER measure. Similarly, for weekly models, we will regress the weekly ERC and SDQ scores on terms for the treatment indicator, TWEETS score, and their interaction. As in the daily model, we will also include a term for the ERC or SDQ score of the previous week to reduce noise. As an alternative approach, we will consider using the weekly average of the daily EMA engagement scores rather than TWEETS for the weekly analyses of the moderating influence of engagement, as these 2 scales tap into different aspects of the engagement experience.

Finally, as our logic model postulates that engagement may also mediate the impact of intervention, we will also conduct an analysis of the mediating role of engagement. Given the ambiguities of mediation analyses in longitudinal settings and the lack of consensus on best practices, for this analysis, we will follow the original Baron and Kenny approach to establishing mediation [75]. To do so, we will use, as our outcome variables, the change scores for the emotion-regulation measures (ERC and SDQ) from baseline to the end of study (end of week 4). To measure engagement, we will use the average of the weekly TWEETS assessments over the course of the study. To examine the mediating role of engagement, we will conduct 3 sets of analyses: first, we will estimate the impact of the treatment condition on the change in ER by regressing the emotion-regulation change score on the indicator of the treatment condition. Second, to examine whether treatment had an impact on engagement, we will regress the average of the weekly TWEETS assessments on the indicator for the treatment condition. Finally, we will estimate a model that regresses the ER change score on both the treatment indicator and engagement. To assess preliminary evidence on the mediating role of engagement, we will examine the magnitudes of effect coefficients in all models, as well as statistical significance, to determine whether the inclusion of engagement in the model has reduced the impact of the treatment condition. Given that the ERC and SDQ tap into different aspects of emotion regulation, we will conduct these analyses for both ERC and SDQ change scores.

We calculated the sample size requirements to be able to detect a difference between the two arms on our primary outcome measure: the daily parent report of the child’s ER throughout the day. On the basis of the data from our preliminary studies, we expect to see a medium effect size for this measure; therefore, to be conservative, we used a Cohen d of 0.3 in our sample size calculation. With this assumption, we calculated the sample size to be able to detect the main effect of the condition (Purrble vs active control) with 90% power and an α level of .05. Under the conservative assumption that the correlation between repeated measures will be 0.75, the required sample size is 92 participants. We inflated this number by 10% to help power the exploratory analyses of the secondary outcomes. Assuming a 20% dropout rate, our final sample size is 120 families.