Article type
Year
Abstract
Background and objective: Methodological work on network meta-analysis in complex interventions has highlighted the ability of this method to examine the effectiveness of individual intervention components, as well as interactions with other components to estimate additive and multiplicative effects. However, intervention components rarely exist in isolation, and variables, including interactions, derived from an exhaustive component scheme may well exceed the number of variables appropriate for a meta-regression. Using insights from latent class modelling with distal outcomes, we combined latent class models with network meta-analysis to examine how empirically derived component classes (EDCCs) could be used to estimate the relative effectiveness of interventions.
Methods and results: We systematically reviewed social learning theory-based parenting interventions for child conduct disorders and located 195 eligible trials. Two expert researchers in the field developed an exhaustive component scheme, and applied it with a third systematic reviewer against all trials. To develop the EDCCs, we estimated a latent class model for components in each trial arm using robust standard errors to account for non-independence of observations, and chose the best-fitting model as judged by scaled relative entropy. We subsequently took 20 draws from the probability distribution of the latent class for each arm. We entered each draw into a network meta-analysis model, and combined findings from each model using Rubin’s rules. We then bootstrapped the combined estimates to rank the EDCCs using the surface under the cumulative ranking curve method.
Discussion: We brought together two types of methods, latent class modelling and network meta-analysis, to examine how EDCCs are associated with differential intervention effectiveness. EDCCs account for the potential interactions between components in those classes, and provide an alternative approach to theoretically-derived intervention classes. Moreover, using EDCCs overcomes the ‘small-n’ problem in high-dimensionality component schemes and offers information on ‘best bet’ combinations of components.
Methods and results: We systematically reviewed social learning theory-based parenting interventions for child conduct disorders and located 195 eligible trials. Two expert researchers in the field developed an exhaustive component scheme, and applied it with a third systematic reviewer against all trials. To develop the EDCCs, we estimated a latent class model for components in each trial arm using robust standard errors to account for non-independence of observations, and chose the best-fitting model as judged by scaled relative entropy. We subsequently took 20 draws from the probability distribution of the latent class for each arm. We entered each draw into a network meta-analysis model, and combined findings from each model using Rubin’s rules. We then bootstrapped the combined estimates to rank the EDCCs using the surface under the cumulative ranking curve method.
Discussion: We brought together two types of methods, latent class modelling and network meta-analysis, to examine how EDCCs are associated with differential intervention effectiveness. EDCCs account for the potential interactions between components in those classes, and provide an alternative approach to theoretically-derived intervention classes. Moreover, using EDCCs overcomes the ‘small-n’ problem in high-dimensionality component schemes and offers information on ‘best bet’ combinations of components.