| Disease misclassification is a common issue in many areas of research, particularly in arthritis research. Specifically, disease misclassification occurs when the presence and absence of arthritis is misdiagnosed by readers based on radiological information. Most investigations in arthritis research have assumed that misclassification is non-differential; i.e., the propensity of misclassification is not affected by physiological characteristics of the individual. Few, if any, studies have evaluated whether disease misclassification is truly non-differential, and to what extent the effect estimate of a risk factor is biased by such assumption. To understand the nature of disease misclassification, sometimes a 'gold standard', a measurement which can classify the disease status with complete accuracy or is widely accepted as being the best available, is available. In arthritis research, film readings adjudicated by a panel of multiple experienced readers to decide whether disease is present or has progressed are commonly referred as the 'gold standard' radiographic readings. When the 'gold standard' is available, the level of misclassification can be estimated for exposed and unexposed group separately, and the differences across risk factor categories can be compared. In this dissertation, these questions were addressed using data from the MOST (Multicenter Osteoarthritis Study), in which associations of various risk factors to the risk of knee osteoarthritis (assessed with knee radiographs) are evaluated. The present work introduces a Bayesian approach for logistic regression with a binary response that is subject to misclassification. The modeling approach assumes that a small validation sample with 'gold standard' information is available, and this information is incorporated as informative priors on the levels of misclassification, or utilized to augment the imperfectly measured outcome. Extensive simulations are carried out to evaluate the performances of the models under both non-differential and differential misclassification scenarios. Models for differential misclassification are illustrated using the MOST data in order to analyze the relationship between obesity status and radiographic knee osteoarthritis. More general models are proposed to deal with the correlated response (e.g. two knees within one subject) under differential and non-differential misclassification settings, and the performances of these models were evaluated. The modeling approaches developed in this dissertation provide new insight to important questions related to the effect of individual characteristics on the likelihood of disease misclassification, both substantively and methodologically. |
