Artikel
Graphical models illustrate complex associations and causal relations between variables describing human functioning
Suche in Medline nach
Autoren
Veröffentlicht: | 6. September 2007 |
---|
Gliederung
Text
Introduction: Human functioning is serving as a main outcome in health studies. Due to the complex nature of human functioning, studies often concentrate on single aspects of human functioning such as mobility. While it is important to gain knowledge about the determinants of single aspects, a better understanding of the association structure of all components of human functioning might be beneficial for the design and planning of interventions. The analysis of multiple correlated variables, however, leads to the problem of multiple testing and variance inflation. To overcome this problem, graphical models have been used, albeit not for the analysis of functioning.
The aim of this study was to illustrate the application of graphical models and to gain insight into human functioning in patients after an acute disease or injury.
Material and Methods: We collected information on 113 second level categories of the International Classification of Functioning, Disability and Health (ICF) in a convenience sample of patients undergoing rehabilitation.
Meinshausen and Buehlmann [Ref. 3] used the LASSO for linear models to obtain a neighbourhood estimator. For this task, we applied constrained logistic regression [Ref. 2]. The outcome of this neighbourhood selection process was further enhanced by bootstrap aggregating [Ref. 1] resulting in more stable graphs.
To overcome the problem of missing data, we generated multiple imputations for incomplete multivariate data by Gibbs Sampling.
Results: 616 patients (mean age 63, 54% female) were included in the analyses. The resulting subgraphs were stable and meaningful.
Discussion/Conclusion: In order to understand the complex structures of human functioning, the different aspects of functioning have to be decomposed. Modelling these many aspects may become a methodological challenge. Graphical modelling can be the basis for the identification of confounders, intermediate variables, meaningful dimensions, subsets for regression analysis, and intervention targets.