Game On—Complier Average Causal Effect Estimation Reveals Sleeper Effects on Academic Attainment in a Randomized Trial of the Good Behavior Game

Same trial as Humphrey et al. 2018 — 77 English primary schools, N = 3,084 children aged 6-7 at baseline

England; secondary analysis of the EEF Good Behaviour Game cluster-randomised trial, with reading measured at post-test and at one-year follow-up.

The steelman for the GBG. Uses complier average causal effect (CACE) estimation against objectively logged dosage from the programme's own online scoreboard, plus a one-year post-intervention follow-up the main report did not have. CACE is an instrumental-variable-style estimate and is only as good as the exclusion restriction; the compliance categories are constructed post hoc from the dosage distribution, which is where the "optimal range" claim comes from.

At post-test there was no effect on reading in either the ITT or the CACE model. At one-year follow-up the ITT result remained null, but a significant effect appeared among MODERATE compliers only (delta = 0.10). Neither low nor high dosage produced an effect, which the authors interpret as an optimal-dosage window and a sleeper effect. The honest reading is that a 0.10 SD estimate in a post hoc compliance subgroup, one year after a trial that was null on every pre-specified outcome, is a hypothesis rather than a finding.

Low — randomised trial; the threat is selection into compliance, which CACE is designed to but cannot fully address.

A re-analysis of a null trial rather than an independent replication. Its own logic cuts against the programme as sold - benefit is claimed only for a middle band of dosage, and one quarter of intervention schools could not sustain any dosage at all.