In meta-regression, a random-effect meta-analysis is extended to evaluate the association between methodological characteristics of studies and heterogeneity in prevalence estimates [76]. First, a model with no covariates was estimated, which is equivalent to a random effects meta-analysis, and a pooled estimate is computed with a correspondent estimate of heterogeneity. Second, a model including all covariates was estimated and analyzed with a backward procedure. The covariates with the higher p value were progressively deleted until all remaining had a correspondent p value of 0.1 or less. The effect of the following study level covariates over the heterogeneity of results was estimated by: i) continent; ii) age group: below or above 12 years old; iii) diagnostic criteria: DSM-III-R or DSM-IV; iv) source of information for the diagnosis: in order of priority for selection: best-estimate, and rule, or rule, parent, teacher, child (because only one study derived the diagnosis based on ‘and rule’, but also on ‘or-rule’, we selected the rate associated with the latter); v) requirement or not of impairment for the diagnosis: yes or no. Although sex of the child
