Author's reply

We take issue with Gregori's statement that methodological concerns render the overall findings of our report ¹ questionable. Gregori correctly observes that ‘the number of violent people in their cohort is very low’ and goes on to suggest that reporting results on such small samples should be discouraged. We are interested in life-course factors that predict adult violence in the hope that such research might inform early life-course interventions. We therefore have two options. Either recruit violent offenders and enquire of their childhoods, or follow a cohort of individuals recording information on their circumstances to assess associations with later problem behaviour. Unfortunately, compounding the vagaries of human memory are the particular difficulties many offenders have with recalling what they did the previous day, let alone several decades ago. It is thus unfeasible to conduct retrospective studies; this leaves cohort studies as the only realistic and robust methodology. We are fortunate in the UK to have one of the most highly regarded cohort studies in the world, but despite its large initial sample size the rarity of violence means that only a small number of respondents demonstrate the behaviour of interest. Should we, as Gregori counsels, simply not consider using the British Cohort Study to look into childhood factors predicting adult violence because violence is rare? We suggest that this would be a valuable and informative resource squandered if that advice were followed. Gregori also suggests that models on rare data should not involve too many covariates. In our short paper we reported that we considered various configurations including the unadjusted association between confectionery and violence and that the strength of association was consistent across models – analytically we did as much as we could to test this association. We chose not to report simpler models and hardly mentioned the extensive analyses assessing the impact of attrition simply because we felt this paper suited a short-report format and including this additional information would only detract from what was a perfectly well-articulated finding. We therefore maintain that we analysed some of the best cohort data available to assess childhood predictors of an important outcome and found a robust association. We were honest with regard to the sample size, concluding in the paper that this is one area that should be addressed before firm conclusions can be drawn.