Objective: To illustrate newly developed statistical methods in analysis of correlated binary outcome data in musculoskeletal (MSK) disease.
Methods: We applied 3 alternative statistical approaches to evaluate the relation of several risk factors to presence of knee osteoarthritis using data from the Framingham Osteoarthritis Study. The methods were (1) an ordinary logistic regression model using each knee as an independent unit of observation; (2) an ordinary logistic regression model treating each person rather than the knee as the unit of analysis; and (3) generalized estimating equation (GEE) and polychotomous logistic regression (PCHLE) using each knee as the unit of analysis but accounting for the correlation between fellow knees. We discuss the advantages and disadvantages of each method with respect to validity, precision, and interpretability.
Results: The GEE and PCHLE models had clear advantages. They simultaneously evaluated the effects of person specific and knee specific risk factors, increased precision, enhanced the interpretability of variables, and provided new insights about how risk factors act.
Conclusion: While the choice of statistical approach depends critically on the scientific question of interest, the GEE and PCHLE approaches will often be optimal in assessments of factors associated with MSK conditions affecting multiple correlated sites within the body, especially when the interest of the study focuses on site specific risk factors.