Question

Compared with the same amount of time spent reading standard material on child development, does a brief training for parents (noticing their own automatic reactions under stress and modelling a “pause” in front of the child) reduce emotionally dysregulated behaviour in children aged 6–10 during a frustrating cooperative puzzle task?

Design

Coding scheme for the primary outcome

The task is run once at baseline (before randomisation) and once at week 8, with two puzzles of equal difficulty. Parent and child work together against a timer to complete a picture puzzle in 10 minutes, of which 4 minutes form the frustration phase: a necessary piece “cannot be found”, and the timer warns that time is running out. At the end of the frustration phase the researcher “finds” the piece, the child always completes the puzzle, and the task ends in success.

Coders watch only the video of the frustration phase, divided into 24 ten-second intervals. An interval is coded “dysregulated” if any of the following occurs: (a) negative affect of intensity 2 or more on a 0–3 scale (frowning, whining, shouting, crying, rated from face and voice together); (b) disruptive behaviour (throwing or hitting puzzle pieces, verbal aggression towards the parent); (c) leaving the task for more than 5 seconds with negative affect. The primary outcome is the number of dysregulated intervals divided by 24.

Reliability requirements: the coding manual is finalised and published before enrolment; coders practise on 30 training videos and must reach an intraclass correlation of 0.80 or more with the master codes before coding study videos; a random 20 % of study videos are coded independently by a second coder and assessed with a two-way random-effects, absolute-agreement, single-measure intraclass correlation, ICC(2,1) (Koo and Li 2016), with 0.75 as the threshold for the primary analysis to be interpretable; drift is checked after every 50 videos. The primary analysis uses the first coder’s scores.

Sample size

# Sample size for the Chapter 8 trial (one parent-child dyad per family).
# Primary outcome: share of 10-second intervals of child dysregulation in a
# frustrating cooperative puzzle task, coded blind from video. ANCOVA on the
# baseline task shrinks the variance by (1 - rho^2). d in SD units of the
# observed score; normal approximation; two-sided alpha .05.
import math
from statistics import NormalDist
N = NormalDist()

def per_arm(d, rho, power=0.90, alpha=0.05):
    z = N.inv_cdf(1 - alpha / 2) + N.inv_cdf(power)
    return math.ceil(2 * z * z * (1 - rho * rho) / (d * d)) + 1

print("dyads per arm for 90 % power")
print(" effect d | baseline-outcome rho = 0.3 | 0.4 | 0.5")
for d in (0.25, 0.30, 0.35, 0.40):
    print(f"   {d:.2f}   | {per_arm(d, 0.3):26d} | {per_arm(d, 0.4):3d} |"
          f" {per_arm(d, 0.5):3d}")
n = per_arm(0.30, 0.4); rec = math.ceil(n / 0.85)
print(f"chosen: d = 0.30, rho = 0.4 -> {n} analysable dyads per arm;")
print(f"  allowing 15 % loss (drop-out, unusable video): recruit {rec} per arm,")
print(f"  {2 * rec} families in all")
print("if the effect were 0.30 SD on the TRUE (error-free) score,")
print("coder unreliability would shrink the observed effect:")
for icc in (0.9, 0.8, 0.7):
    d_obs = 0.30 * math.sqrt(icc)
    print(f"  coder reliability {icc}: observed d = {d_obs:.3f}"
          f" -> {per_arm(d_obs, 0.4)} per arm")
# dyads per arm for 90 % power
#  effect d | baseline-outcome rho = 0.3 | 0.4 | 0.5
#    0.25   |                        307 | 284 | 254
#    0.30   |                        214 | 198 | 177
#    0.35   |                        158 | 146 | 130
#    0.40   |                        121 | 112 | 100
# chosen: d = 0.30, rho = 0.4 -> 198 analysable dyads per arm;
#   allowing 15 % loss (drop-out, unusable video): recruit 233 per arm,
#   466 families in all
# if the effect were 0.30 SD on the TRUE (error-free) score,
# coder unreliability would shrink the observed effect:
#   coder reliability 0.9: observed d = 0.285 -> 219 per arm
#   coder reliability 0.8: observed d = 0.268 -> 247 per arm
#   coder reliability 0.7: observed d = 0.251 -> 282 per arm

With a smallest effect of interest of 0.30 standard deviations, a baseline–outcome correlation of 0.4 and 90 % power, 198 analysable dyads are needed per arm; allowing for about 15 % loss through drop-out and unusable video, 233 families are recruited per arm, 466 in all. If the 0.30 is understood as an effect on the error-free true score, coder unreliability shrinks the observed effect: at a reliability of 0.8 the observed effect is about 0.268 and 247 dyads per arm would be needed. This is why coding reliability is treated in the protocol as a hard threshold.

Decision rule

# Decision rule, fixed in advance. d = arm A minus arm B (primary outcome,
# pooled-SD units, from the pre-specified ANCOVA); NEGATIVE d = less child
# dysregulation in arm A. SESOI 0.30 SD. Coding gate: ICC(2,1) on the 20 %
# of videos coded twice must be at least 0.75.
SESOI, ICC_MIN = 0.30, 0.75

def verdict(d, lo, hi, icc, fidelity_ok, harms):
    if harms:
        return f"d = {d:+.2f}: PAUSE for safeguarding review - {harms} event(s)"
    if icc < ICC_MIN:
        return f"d = {d:+.2f}: NOT INTERPRETABLE - coder ICC {icc:.2f} < 0.75"
    if not fidelity_ok:
        return f"d = {d:+.2f}: NOT INTERPRETABLE - < 80 % of sessions as planned"
    if hi <= -SESOI: v = "A reduces dysregulation by at least the SESOI"
    elif hi < 0:     v = "A reduces dysregulation; size not shown to reach 0.30"
    elif lo > 0:     v = "A INCREASES dysregulation"
    elif lo >= -SESOI and hi <= SESOI: v = "no difference worth acting on"
    else:            v = "unclear: the interval is too wide to decide"
    return f"d = {d:+.2f} ({lo:+.2f} to {hi:+.2f}): {v}"

for case in ((-0.50, -0.70, -0.30, 0.84, True, 0),
             (-0.22, -0.42, -0.02, 0.84, True, 0),
             (-0.05, -0.25, 0.15, 0.81, True, 0),
             (-0.15, -0.35, 0.05, 0.86, True, 0),
             (-0.45, -0.65, -0.25, 0.68, True, 0),
             (-0.40, -0.60, -0.20, 0.83, False, 0),
             (-0.35, -0.55, -0.15, 0.82, True, 1)):
    print(verdict(*case))
print("secondary outcomes (parent-rated Emotion Regulation Checklist, the")
print("parent's observed pauses) are reported with intervals, never decisive")
# d = -0.50 (-0.70 to -0.30): A reduces dysregulation by at least the SESOI
# d = -0.22 (-0.42 to -0.02): A reduces dysregulation; size not shown to reach 0.30
# d = -0.05 (-0.25 to +0.15): no difference worth acting on
# d = -0.15 (-0.35 to +0.05): unclear: the interval is too wide to decide
# d = -0.45: NOT INTERPRETABLE - coder ICC 0.68 < 0.75
# d = -0.40: NOT INTERPRETABLE - < 80 % of sessions as planned
# d = -0.35: PAUSE for safeguarding review - 1 event(s)
# secondary outcomes (parent-rated Emotion Regulation Checklist, the
# parent's observed pauses) are reported with intervals, never decisive

The example results are invented. Only if the upper limit of the confidence interval is no higher than −0.30 standard deviations will the report say that “arm A reduced children’s emotional dysregulation by at least the smallest effect of interest”. If the double-coding intraclass correlation is below 0.75, or fewer than 80 % of arm-A sessions were delivered as specified, the report will say “not interpretable” whatever the result; if there is a child-protection event related to the study, the trial is paused for a safety review and the event is reported.

What this trial cannot show

First, it involves no genes or epigenetics: even a positive result would say nothing about any biological mechanism of “intergenerational transmission”, only whether a short training aimed at parents’ behaviour affects children’s behaviour in one task. Second, the outcome is a laboratory task measured once at week 8; it cannot show how children behave at home, at school or years later. Third, parents know which training they received and may deliberately behave better on video; the primary outcome is children’s behaviour coded by blinded coders, which partly mitigates this, but the parents’ own behaviour in the task may itself be part of an expectation effect. Fourth, arm A combines two components, awareness and modelling a pause, and the trial cannot show which one works. Fifth, families who volunteer and are not in crisis may be very different from the families who most need help. Sixth, this is not a substitute for psychotherapy or family therapy.

Ethics, child protection and conflicts of interest

The trial requires approval by an independent ethics committee, and the protocol, coding manual and analysis plan are registered before the first family is enrolled (Nosek et al. 2018). Parents give written informed consent; children are told about the study in age-appropriate words and pictures and give verbal assent, and a child who does not assent is not enrolled. Every child is given a “stop card” that can be raised at any moment to stop the task. The frustration phase has pre-specified stopping rules: if the child cries continuously for more than 30 seconds, clearly asks to stop or shows marked distress, the researcher ends the frustration phase at once and “finds” the piece.

The consent form states the limits of confidentiality in advance: if researchers observe or hear anything during the study suggesting that a child may be at risk of harm, a designated child-protection lead will handle and refer it according to local law and institutional child-protection procedures. All researchers receive child-protection training. Videos are stored encrypted, used only for coding and destroyed after a fixed period. After the study, families in arm B can receive the arm-A training materials.