Multivariate multilevel analysis: An elegant alternative for the separate analysis of correlated outcomes

Introduction: Surprisingly, multivariate multilevel analyses, which is often used to analyse longitudinal data, is seldom used to analyse multiple, correlated outcomes, such as questionnaire subscales or biological risk factors. Therefore, the aim of the study was to compare multivariate multilevel analysis with separate, per outcome, analyses in data with correlated outcomes retrieved from the same individual, in both cross sectional and longitudinal studies.
Method: To compare the results of multivariate multilevel analysis with the separate analyses of each outcome, both real life example datasets and simulated data in which various characteristics are systematically varied, were used.
Results: Multivariate multilevel analyses produced substantially more accurate estimates and standard errors than separate analyses for each outcome in incomplete datasets and when analysing longitudinal relations with time-dependent covariates. In complete datasets, for time-independent covariates (e.g., baseline characteristics) and when the development over time is analysed, results of the two analytical approaches were highly similar. It was also found that similarity of standard deviations between the outcomes is needed for a proper estimation of the standard error of the regression coefficient in a multivariate multilevel analysis. Similarity of standard deviations can be obtained by using z-scores. Furthermore, before using longitudinal multivariate multilevel analyses it has to be evaluated which type of clustering provides the best model fit.
Conclusion: Multivariate multilevel analysis is an elegant way to analyse multiple correlated outcomes, which is superior to separate analyses for each outcome in case of missing data and/or when timedependent covariates are analysed.