Multi-dimensional Models for Outcomes Measurement

Patient-reported outcomes (PRO) data have contributed enormously to health services research. However, they are complex to conceptualize and use, and investigators have expressed the need for a flexible set of analytic and scoring methods that can account for the multi-dimensionality of PRO data as well as provide interpretable scores that are useful for various research and policy applications.

Existing methods have a limited ability to capture the multi-dimensional characteristics of PRO data, which include a range of broad domains, such as health-related quality of life (HRQOL, such as physical functioning and social role participation) and symptoms (fatigue, pain, depression). Each domain is considered multi-dimensional in that it includes a variety of heterogeneous indicators.

Currently, both the classical and modern measurement approaches commonly used to analyze and score PRO data assume that one dominant factor underlies each domain and accounts for most of the variation in scores. This is called the assumption of unidimensionality. If more than one factor exists, the domain must be divided into sub-domains to apply these methods. Under this restrictive set of assumptions, efforts to summarize these data into broader constructs suffer from the lack of clear statistical and analytical decision-rules. Measures are often simply added together to create a combined score, or, alternatively, patient or expert judgment is used to weight these factors within or across domains.

More complex psychometric modeling methods exist that take advantage of the correlations among PRO domains to improve the measurement of patients at the sub-domain level. Other new models, such as the bi-factor model, can account for the hierarchical nature of the data, thereby improving the estimation of summary scores of various HRQOL and symptom domains in cancer and other diseases.

The Outcomes Research Branch funded Dr. Robert Gibbons (Director, Center for Health Statistics and Professor of Biostatistics and Psychiatry at the University of Illinois-Chicago) to assess the value of using these innovative psychometric models to improve PRO measurement. These models create a profile of scores as well as an overall summary score, an improvement over standard approaches used in PRO measurement.

The results of this study can be found in The Added Value of Multidimensional IRT Models (PDF).