This study examined theoretical, methodological, and statistical problems involved in evaluating the outcome of aggression on the playground for a universal prevention intervention for conduct disorder. Moderately aggressive children were hypothesized most likely to benefit. Aggression was measured on the playground using observers blind to the group status of the children. Behavior was microcoded in real time to minimized potential expectancy biases. The effectiveness of the intervention was strongly related to initial levels of aggressiveness. The most aggressive children improved the most. Models that incorporated corrections for low reliability (the ration of variance due to the true time-stable individual differences to total variance) and censoring (a floor effect in the rate data due to short periods of observation) obtained effect sizes 5 times larger than models without such corrections with respect to children who were initially 2 SDs above the mean on aggressiveness.