Every child learns
differently.
So should their
preparation.
We studied whether a short 10-minute warm-up before a learning session helps children with autism engage more — and tried to find the right warm-up for each individual child. The study is now complete. Read the results →
The Setting
Clinical Context: Maya Care and Grow
In Agartala, India, early autism intervention requires a highly individualized approach. Our study is integrated directly into the daily routines of a center serving 25 children with diverse developmental profiles.
Children aged 3 to 14 attend 2–16 hours of learning sessions per week. In this setting, the first 15 minutes are critical; practitioners currently use a variety of activities to 'prepare' a child for learning.
While these preparations—ranging from massage and bubbles to puzzles and music—are delivered with expert judgment, there is currently no data-driven way to know which specific activity maximizes engagement for a specific child.
The Opportunity
- 📍 Real-world setting with routine delivery
- 📊 Feasible data collection via practitioner involvement
- 🔄 Repeated-measures design for highly variable responses
- 🎯 Personalized results for every enrolled child
"Every child responds differently. Our goal is to replace intuition with evidence."
The Challenge
Why does preparation matter?
Children with autism often experience challenges regulating their sensory systems before a learning session. A child who arrives overwhelmed, under-stimulated, or dysregulated may struggle to engage — not because of the content, but because of how they feel. A brief, targeted warm-up can help bridge that gap. But the right warm-up isn't the same for everyone.
Individualized by design
Each child serves as their own control. We compare all four conditions within the same child across sessions, so results aren't muddied by differences between children.
Bayesian learning
Our statistical approach gives probabilities, not just yes/no answers. We can say "there's an 85% chance Condition A is best for this child" — and stop as soon as we're confident enough
Adaptive & humane
We stop testing a child the moment we have a clear answer, or if any condition consistently distresses them. No child is kept in an ineffective routine for longer than necessary.
The Interventions
Four preparation conditions
Each child is assigned one fixed activity per condition before the study starts. The 20-minute learning session that follows is identical each time — so any difference in engagement can only be due to the warm-up.
High-Energy Input
⚡ Jumping & Body Massage
Targets the vestibular and proprioceptive systems with movement and deep pressure. Ideal for children who are under-aroused or need activation before learning.
⏱ 10 minutesQuiet & Fine Motor
🌿 Clay Modelling
Gentle tactile input designed to reduce arousal. Well-suited for children who arrive hyperactive, anxious, or sensory-overloaded and need to settle before learning.
⏱ 10 minutesAutonomy & Choice
🎯 Choice Board
The child selects their own warm-up from a fixed menu of options. Builds agency and intrinsic motivation — a key factor in engagement for many children with autism.
⏱ 10 minutesNo Preparation
— Standard Session Start
The session begins immediately with no warm-up. This is the current standard of care and serves as our baseline for comparison. It may turn out to be optimal for some children.
⏱ 0 minutesProcess
How does a session work?
A typical session takes about one hour. The order of conditions across sessions is randomized — a bit like shuffling cards — so no child gets the same warm-up two sessions in a row.
Child arrives at the centre
The rater (a trained observer who scores engagement) notes the child's arrival state: calm, tired, or hyperactive. This is recorded but does not change the assigned condition.
Observation10-minute warm-up (or none)
The practitioner delivers the assigned preparation condition — Stimulating, Calming, Child-Led, or skips it entirely. The activity is one the child already does at the centre.
InterventionStandardized learning block
The same structured curriculum — specific Discrete Trial Training tasks or tabletop work — is used in every session. The rater observes from a separate room, scoring engagement at 5, 15, and 30 minutes.
Learning sessionEngagement is scored (1–10)
The rater assigns a BRES-10 score. The three scores are averaged into a single composite for that session. 20% of sessions are double-scored for reliability checks.
MeasurementWeekly Bayesian check
Every Sunday, the model evaluates each child's data. If we're ≥90% confident in a winner — or confident that no prep truly makes a difference — we stop and implement the personal recommendation immediately.
Stopping ruleMeasurement
What does the engagement score mean?
The BRES-10 (Blinded Rater Engagement Scale) is a 1–10 scale scored by a trained observer who does not know which warm-up condition was assigned. Higher scores mean more independent, sustained engagement.
Your Child's Wellbeing
Safety comes first
This study uses only activities already practiced at Maya Care and Grow. Every safeguard below is active from the first session.
Automatic condition removal
If a child scores ≤ 2 (active refusal/distress) for three consecutive sessions under any condition, that condition is immediately removed from their rotation. Parents are notified.
Voluntary participation
Participation is entirely voluntary. Withdrawing at any time has zero effect on your child's access to services or quality of care at the centre.
Early stopping
As soon as the model is confident in a recommendation, randomization stops immediately — your child doesn't continue in an experimental condition longer than needed.
No novel interventions
All four conditions are existing activities at the centre. We are not introducing anything new — only studying the timing and type of familiar activities.
Ethical framework
The study follows the Declaration of Helsinki. Written parental consent is required, and children's assent is sought in an age-appropriate way before enrolment.
Blinded scoring
The person scoring engagement does so from a separate room without knowing which warm-up was given — reducing the risk of biased observations.
Child Archetypes
What response patterns do we expect?
Based on clinical knowledge of this setting, we anticipate children will fall into four profiles. The study is designed to detect meaningful differences for each type.
One preparation type is clearly and substantially better. The model typically resolves these children within 20 sessions.
SuperiorityA best condition exists but the effect is modest. Requires more sessions to separate signal from noise.
Superiority or AIPETwo or more warm-up types are equally beneficial. The recommendation is: 'use whichever is convenient.' Most children are expected to fall here.
AIPE criterionThis child arrives well-regulated and doesn't need a warm-up. No Preparation is perfectly valid — and affirmed as a clinical choice.
ROPE criterionSimulated Power Analysis
How confident will we be, and when?
Before collecting any data, we ran computer simulations to confirm the study can deliver reliable answers. These charts show how resolution rates grow with more sessions — and how they differ by child type.
Resolution Rate by Sessions
Proportion of simulated children with a confirmed recommendation
Resolution by Child Profile at 32 Sessions
Performance varies by how distinct the best preparation is
Illustrative Group-Level Results
Simulated model output (900 observations, 25 children) — estimated benefit of each condition vs. No Preparation
Methodological Rigor
Stability & Sensitivity
We performed 'stress-tests' on our statistical models to ensure clinical recommendations are driven by the child's actual progress, not by background noise or mathematical assumptions.
🌪️ Environmental Robustness
We tested the model by removing information about the child's age and the specific teacher.
Result: Clinical recommendations remain identical even when ignoring environmental covariates.
⚖️ Decision Stability
We re-ran the study using other mathematical assumptions (Bayesian priors) to see if the outcome changed.
The Qualitative Safety Net: In cases where a recommendation is mathematically "fragile" (sensitive to assumptions), the Principal Investigator and Lead Practitioner conduct a manual review of session notes before finalizing the clinical assignment.
Results
No preparation effect detected
Data collection ran from 20 April to 10 June, 2026, with 31 children and 292 rated sessions. Neither the group as a whole nor any individual child showed a detectable benefit from any of the three warm-ups.
Group-level effect of each preparation
Bayesian model estimate across all children, adjusted for practitioner and age
Engagement points vs. No Prep. Dot = best estimate; thick bar = 80% credible interval; thin line = 95%. Shaded band = ±1 point, the pre-registered zone of “no meaningful difference”.
Raw average engagement
Mean BRES-10 score per condition (1–10), before modelling
All four conditions average between 5.2 and 5.5, while scores within any single condition spread widely (standard deviation ≈ 2 points).
What this means: none of the three warm-ups changed engagement in a detectable way. The best estimates range from −0.24 (Calming) to +0.14 (Stimulating) points on a 10-point scale, and every 95% interval includes zero. It does not show that warm-ups never help: the study could not tell a small or child-specific benefit apart from noise.
Individual Results
Child by child
The study was designed to find the best warm-up for each child. Use the arrows to browse all children, identified by anonymous number.
29 sessions (21 Apr – 10 Jun). Every warm-up is very likely within ±1 point of No Prep. For this child, the choice of warm-up does not appear to matter, so skipping it is a reasonable default.
Estimated effect for this child
Group effect plus this child's own deviation
Engagement points vs. No Prep. Dot = best estimate; thick bar = 80% credible interval; thin line = 95%. Shaded band = ±1 point, the pre-registered zone of “no meaningful difference”.
Probability each condition is best
Share of posterior draws in which it scores highest
A recommendation needed one condition to reach 90% (marked line).
Every rated session
Each dot is one session's engagement score (1–10); the dark tick is the average
Discussion
Why we didn't detect an effect
We did not detect any effect of the pre-session warm-up, for the group as a whole or for any individual child. We believe this says more about how the trial was run than about whether warm-ups work. Several practical problems each added noise to the engagement scores. After accounting for differences between children and practitioners, the scores for the same child still varied from one session to the next by about 1.5 points (standard deviation). That is larger than the effects we were trying to find.
Ratings were not precise enough
Engagement was meant to be scored live at 5, 15 and 30 minutes. In practice, many ratings were written down from memory, often hours after the session.
In the data: 35% of ratings were submitted more than an hour after the warm-up was logged (26% more than three hours). In 36% of sessions, all three time points got the exact same score. This pattern suggests an overall impression was recalled, not three separate observations.
The warm-up was often pushy
Instead of meeting the child where they were, the warm-up often turned into getting the child to do a task within minutes of arriving. Being made to act too soon tended to put the child in a worse mood, which carried over into the learning session.
Well-meant extra help overwhelmed children
When a child was not following the warm-up, a second assistant would sometimes step in to help. Two adults directing one child was often too much, and the extra pressure made things worse.
Activities were too repetitive
Each child had one fixed activity per condition, and sensory needs vary hugely from child to child. Doing the same activity again and again made children bored within a few days. A better design would keep the category fixed (e.g. 'stimulating') but change the activity from day to day.
Child choice led to hard transitions
When allowed to choose, children usually picked the same favourite object or activity they were fixated on. They then struggled to let go of it when the warm-up ended, and that frustration spilled into the session. This may have cancelled out any benefit of having a choice.
Too few sessions per child
The design called for 12–32 sessions per child. Irregular attendance and the end of the collection window meant most children never reached that range.
In the data: 20 of 31 children (65%) had fewer than 12 rated sessions. Several had only one or two.
Taken together: imprecise ratings, warm-ups that sometimes stressed children rather than preparing them, repetitive activities, and too few sessions per child all added noise. Any real signal from the preparation conditions was lost in it. The model also estimated that children differed very little in how they responded to each warm-up. As a result, each child's estimates were pulled strongly towards the group average, which is why the individual results above look so alike.
What we would do differently
- →Score engagement live, on a phone or tablet, at the exact time points, and reject late entries.
- →Make the warm-up invitational: offer the activity, and let the child ease in or opt out without pressure.
- →One adult per child during the warm-up. No stepping in.
- →Keep a pool of activities for each category and rotate them, so 'Stimulating' is not the same jump every day.
- →For child-led warm-ups, exclude fixation objects and use a visual timer to prepare the transition out.
- →Secure consistent attendance, or plan a longer collection window, before starting a per-child design.
The result is null and mostly reflects limits of the process, so we will not write it up as a journal paper. This page is the final report of the study. The data, model and analysis code are open in the repository below, as pre-registered.
Open Science
Research & Transparency
We are committed to full transparency. Below you can find the source code, pre-registered protocols, and the technical reports that validate our statistical approach.
Questions & Answers