Background Child conduct disorders have important long-term public health sequelae, including higher risk of substance use and violence in adolescence and higher risk of mental health problems in adulthood. Parenting interventions informed by social learning theory have been consistently shown to be effective for reducing child disruptive behaviour, but the diversity and variety of interventions prevents a clear understanding of why interventions are differentially effective. Intervention components rarely exist in isolation, and interventions may be best understood in terms of the combinations of components they include rather than any one component. We tested if components of social learning theory-based parenting interventions for child disruptive behaviour outcomes formed underlying groupings across trials, and analysed how these groupings were associated with differential intervention effectiveness.

Methods We searched existing systematic reviews across a variety of platforms and updated searches in recent reviews. We included randomised controlled trials of parenting interventions for child conduct disorder measuring child disruptive behaviour outcomes. We developed and applied an exhaustive 19-component scheme against all trials. We estimated a latent class model for components in each trial arm and chose the best-fitting model. We subsequently took 20 draws from the probability distribution of the latent classes for each arm, network meta-analysed each draw, and combined findings from each draw using Rubin’s rules.

Results We included 151 trials with 329 arms comparing relevant parenting programmes against each other, against treatment as usual or against no treatment. Across 19 components, our latent class analysis identified five intervention types (four active, one no treatment). As compared to no treatment, ‘relationship-building’ was most effective at reducing child disruptive behaviour (Cohen’s d −0.71, 95% CI [−0.90,–0.53]) but ‘parental self-management’ (−0.54, [-0.76,–0.32]), ‘breaking coercive cycles’ (−0.53, [-0.74,–0.33]) and ‘all components’ (−0.49, [-0.68,–0.29]) were effective as well. There was some evidence that ‘relationship-building’ was more effective than other active intervention types but confidence intervals for these comparisons were broad.

Discussion We brought together two types of methods, latent class modelling and network meta-analysis, to examine how intervention classes are associated with differential intervention effectiveness. This is the first analysis of its kind. While all active intervention classes were better than no treatment and no active intervention class was better than any another, we did not find that using all possible components in the same intervention was associated with increased effectiveness as compared to more spare groupings of components.