ABSORBENTES PARA INCONTINENCIA (AIU)

Specification of outcomes and effect measures (PRISMA-IPD #13)

To assess potential harms and benefits of the IY intervention we consider the main effect of the

intervention (IY vs. control) on each secondary outcome within T1 (post treatment). As an initial step in a series of independent, univariate models (i.e. multilevel analyses of covariances), we tested main effects of

the intervention on 12 prespecified outcomes: children’s ADHD symptoms, children’s emotional problems,

self-reported positive parenting behaviour (use of praise, rewards and monitoring), self-reported negative parenting behaviour (corporal punishment, threatening, laxness and shouting), parental depression, parenting stress and parental self-efficacy. In the future, we plan to carry out a multivariate analysis that simultaneously assesses the effects on the 12 variables.

For positive variables (i.e. those for which a higher value is a positive outcome) a statistically significant positive effect of treatment can be seen as a benefit, whereas a negative coefficient may be interpreted as a harmful effect. For negative variables (i.e. those for which a higher value represents a more negative outcome), the regression coefficients are in the opposite direction for harms and benefits. We condition on

baseline values of the outcome of interest because these are known to influence post-treatment outcomes. In addition, random effects are added to the analysis models to account for between-trial heterogeneity both in the intercept and in the coefficient of treatment. Additional random effects for IY group are added only in the IY arm, to allow for the fact that individuals within the same therapy group may be more correlated than individuals in different therapy groups. As for the moderator modelling, fixed effects are added to account for trial design features, including variables used for stratification and randomisation batch when the randomisation ratio was varied within a trial.

Fixed effects included in the analysis model for harms and benefits are:

l baseline values of the outcome of interest

l a dummy variable coding for treatment condition (IY vs. control)

l child age and child gender (used for stratification in some trials)

l dummy variables coding for relevant trial design features (i.e. other stratification variables and randomisation ratio batches).

Random effects included in the model are: 1. Cluster structure of the pooled data:

i. Random intercept varying at the level of trial (up to 14 levels) to account for predictive effects of trial characteristics (e.g. differences in trial target populations or general service organisation contexts affecting control groups) on outcome under the control condition.

ii. Random intercept varying at the level of treatment cluster when cluster randomisation was used (trial specific).

iii. Random intercept varying at the level of IY training group within the IY arm of a trial to account only for predictive effects of the training group/therapist environment within the active

treatment arm.

2. Random coefficients representing effect heterogeneity:

i. The regression coefficients representing treatment effects (of trial arm dummy variables) are allowed to vary with trial to model treatment–effect heterogeneity (e.g. because of differences in treatment implementation or target population) not already captured by fixed baseline × trial arm interaction terms. To estimate a standardised effect size, we ran the analysis for each variable using a standardised outcome. The outcome was standardised by dividing it by the SD of the variable at baseline. This expresses the treatment effect size in units of baseline SDs, meaning that the size of this effect can be compared across outcomes. In addition, we analysed ECBI-I in the same way, so as to compare the treatment effect on the secondary outcomes with the effect size on the primary outcome.

Most of the secondary outcomes are available within only a subset of the trials and so the sample size is limited for this analysis.

We addressed these issues by using multilevel modelling (random-effects modelling) to capture the hierarchical structure of each variable, accommodated design features by trial-specific fixed effects (e.g. conditioned the model for trial 10 on school year strata) and fitted resulting models by maximum likelihood which is valid under a MAR assumption regarding the process that generates the missing data. In our context MAR implies that variables that are included in the analysis model can drive missingness without this leading to bias. We also intended to analyse the outcome variables in a single multivariate model to further relax missingness assumptions, by allowing missingness of the values of one variable to be driven by other observed outcome values. However, it was not technically feasible to simultaneously model the hierarchical structure as well as fitting a covariance matrix for the multiple outcome measures

per participant. We thus ran two sets of analyses, each having to make some further assumptions to become feasible: (1) a set of univariate analyses that fully capture the hierarchical structure but require the more restrictive MAR assumption and (2) a simplified multivariate analysis that deals with all the outcome variables in one model but does not account for all the trial features (in particular IY group and cluster effects have to be assumed absent).

The single multivariate model (2) includes dummy variables for each outcome and interactions between these dummy variable and each covariate. We estimated the model with an unstructured covariance structure, which allows correlations to vary between each pair of outcomes. We also included our measure of disruptive child behaviour (i.e. the ECBI-I) to the model, the primary outcome of the trials, to allow for further relaxation of the missing data assumptions.

Synthesis methods (PRISMA-IPD #14)

Similar to analytic plan part 1, we used a one-stage approach (this is PRISMA-IPD terminology). This means that analyses were conducted in a pooled data set of harmonised data from each trial, rather than

analyses being conducted within the individual trials and aggregate trial results combined afterwards. Clustering of participants within studies was accounted for by analysing data in a hierarchical structure. The pooled data set has a hierarchical structure with families (level 1 units) nested within therapy groups (level 2 units) within the intervention arm and therapy groups nested within trials (level 3 units).

Exploration of variation in effects (PRISMA-IPD #14)

Exploring variation in effects of the intervention by participant characteristics was not part of this research question.

Additional analyses (PRISMA-IPD #16)

Chapter 3 Results and discussion of moderator

analyses