Introduction
Many regard health as a multidimensional construct. Correspondingly, a number of health-related quality-of-life (HRQOL) instruments are built around a framework of subscales,– with each subscale intended to capture a particular dimension (e.g., physical, social, emotional) of the overall construct (HRQOL). The validity of the total instrument is based on the strength of the validity of the underlying subscales. Not infrequently, the instrument's scoring algorithm will also allow the derivation of summary scores. That is, the items used to construct the instrument's N subscales are further aggregated to yield M (<N) summary scales; when M = 1, the instrument yields an overall summary score.
A fundamental assumption in both Classical Test Theory (CTT) and Item Response Theory (IRT) is the unidimensionality of the latent trait. In the case at hand, this means the unidimensionality of each of the HRQOL instrument's subscales and, for that matter, the unidimensionality of any summary scales. If the analyst insists that health is a multidimensional construct, at least two important questions arise.
First, must we therefore derive and apply each subscale using information collected only from the items on that subscale, so that each subscale in the multidimensional construct essentially “floats on its own bottom”? Or, alternatively, is there some way to strengthen each subscale by drawing strength from the ensemble of information available across all related subscales?