We analysed data according to fracture sites. Where it was possible to obtain overall estimates of the probability of abuse for a particular fracture type from cross-sectional studies, we undertook a meta-analysis and presented the result as forest plots.
We pooled estimates from individual studies using the method of De Simonian and Laird1, deriving a confidence interval for the pooled estimate and testing for heterogeneity between studies1. Given the varied nature of the studies, our work was inevitably a pooled estimate of different populations rather than a single one. Insufficient detail was shown in most papers to be able to analyse by age and so in the forest plots the studies were ranked in increasing order of mean age in the abused cases, as far as was possible. The results were generally summarised as the proportion of children with a given fracture type who are classed as abused – that is the predictive value of the fracture type for identifying abuse. Proportions were compared between groups using the chi-square test or Fisher’s exact test where that was appropriate.