A two-level cross-sectional model using grafted polynomials

Multilevel generalized linear models for modelling age-related gender difference in violent behaviour and associated factors in the general household population

It is preferable to use longitudinal data when studying patterns of violence and antisocial behaviour over the lifespan together with the associated risk factors in the general population. From the statistical modelling perspective, random samples of cross-sectional data, representative of the population, can be a reliable alternative. Sampling, weighting, and possible geographical clustering of the behaviour must be considered in the analysis together with correct choice of model as a function of age, although cohort effects and age effects are not separated from the analysis. This paper demonstrates the use of multilevel generalized linear models in the British National Survey of Psychiatric Morbidity in 2000. A multilevel logistic model as a special case of a generalized linear model with individual weightings was adapted for a dichotomous measure of violence and extended to Poisson and negative binomial outcomes. Three types of age function, discrete age effects, continuous age effects, and piecewise polynomial function of age intervals were evaluated for goodness of fit, and for their practical advantages and disadvantages. Models were developed for possible risk factors in relation to specific age groups of interest.