We computed power by performing the same statistical analysis that we would apply to the real data, on 1000 simulated data sets. If the average rate of offences is 0·05/subject per d(corresponding to 3·85 offences over the 77 mean d of the study),the power to detect an improvement of 28 % (corresponding to αA=0·25 in our statistical model) at a threshold corresponding to 0·05 significance would be about 0·80. (We took the variance parameter – measuring the variation among children in their offence rates – to be 0·0040, but substantial changes in this parameter were found to produce only minor changes in the power.) If the mean offence rate were only 0·02, the power would drop to 0·454; if the mean offence rate were raised to 0·10,the power would go up to 0·97.