Dimension reduction in human functioning and disability outcomes research: graphical models versus principal components analysis.

Purpose: The complexity of rehabilitation outcomes in terms of functioning and disability leads to the need for dimension reduction in relation to specific research or clinical problems. Several statistical procedures are at hand. This article examines whether and to what extent principal component analysis (PCA) and graphical models differ in the dimension reduction of data based on the International Classification of Functioning, Disability and Health (ICF).

Methods: Using a data set of 1048 persons with spinal cord injury from 14 different countries as a case in point, this article compares the solutions in dimension reduction generated with a graphical model based on least average shrinkage selection operator (LASSO) regression on one hand and with a PCA on the other.

Results: Some factors extracted with the PCA properly match the clusters found with the graphical model, while in others less commonality can be found. In general, overlap ranges from 9 to 100% with 75% on average. Many of the discovered clusters or factors, i.e. dimensions, are compatible with the ICF structure, particularly in activity and participation.

Conclusion: Functioning is a multidimensional, complex and dynamic outcome. We recommend being careful with dimension reduction based on statistical procedures alone. Theoretical considerations and clinical significance should always guide statistics. Stability of emergent dimensions that are compatible with theoretical or clinical considerations is a most important point. At least two different statistical procedures for dimension reduction, for instance PCA and LASSO regression, should be applied to conservatively select those dimensions that stay stable with both procedures.