We do not want our finding of a sixfold increase in reporting five or more violent incidents in persons with psychosis to give a misleading impression regarding the association of violence with mental illness. This was the only finding suggesting increased risk and means that there is a small subgroup of people with psychosis who are repeatedly violent. The real message of our paper should have been that the true risks of violence from people with psychosis, at the population level, are exceedingly small.

Professor Persaud's impression might be owing to the space in our paper devoted to discussing the public health impact of alcohol misuse and antisocial personality disorder on violence. In an additional paper published recently in the American Journal of Epidemiology we make the point about psychosis more strongly (Reference Coid, Yang and RobertsCoid et al, 2006). Researchers with an interest in violence and psychosis often emphasise that relative risks of violence are greater for individuals with psychosis but they ignore the fact that illnesses such as schizophrenia are rare and that persons with psychosis account for an exceptionally small number of violent incidents at the population level. Detaining more persons with psychosis in hospital would have a very small effect in reducing violent crime (Reference Fazel and GrannFazel & Grann, 2006).

Misleading impressions based on relative risks are typical for homicides perpetrated by people with psychosis. These are often based on Scandinavian countries where the base rate is exceptionally low (Reference Hodgins and JansonHodgins & Janson, 2002). In locations where the base rate is very high, for example certain areas in the USA and South American countries, people with psychosis hardly feature in criminal statistics.

Careful reading of our paper will reveal how we dealt with confounding from comorbid conditions. We agree with Professor Persaud's point about residents in violent neighbourhoods entirely, but the sampling frame was intended to exclude bias from factors such as socioeconomic deprivation. We used two-level hierarchical models throughout the analysis to take account of clustering from these areas. We would concede, however, that our study did not adequately explore the important issue of neighbourhood effects.