familial environmental risks. Thirdly, we did a series of logistic regressions to test which risk factors could help to identify which bullied children would engage in self harm. We evaluated the following risks: family adversities (socioeconomic deprivation, family history of attempted/completed suicide, physical maltreatment), children’s concurrent mental health problems, and children’s low IQ. Finally, we used the “punaf” command in Stata (v11.2) to calculate the population attributable fraction for self harm based on exposure to bullying victimisation (reported by either mother or child) in this sample. We corrected all analyses (except the twin discordance analysis) for the non-independence of the twin observations by using the Huber-White variance estimator. We did a sensitivity analysis using multi-level mixed models in SAS with the PROC GLIMMIX procedure and obtained almost identical results (data not shown).
