Many applications of biomedical science involve unobservable constructs, from measurement of health states to severity of complex diseases. In this dissertation I utilize joint latent variable methods to combine item selection and validation to identify significant items in a symptom scale and determine how these symptoms relate to "gold standard" diagnostic measures. Joint latent variable models eliminate bias inherent in traditional two-stage methods and provide a global test of the association between the underlying construct and a clinical measure. In Chapter 1, a review of latent variable methods for multivariate outcomes is provided. Chapter 2 proposes a Multiple Indicator Multiple Cause (MIMIC) model to perform item reduction and validation simultaneously. A modified Score test for individual factor loadings in the MIMIC model is derived. The methods are motivated by an example from a premenstrual syndrome (PMS) clinical trial in which one objective was to determine a reduced number of core symptoms in the diagnosis of severe PMS and to compare patient-reported symptom information to a clinician-rated "gold standard" diagnostic measure. Chapter 3 applies an extension to the MIMIC model to patient-reported outcomes (PROs) from the Physical Activity and Lymphedema (PAL) clinical trial. PROs are a potentially less expensive and time-consuming measure of diagnosis than some clinical measures. An extension of the MIMIC model for ordered categorical outcomes determines which symptoms are important indicators of lymphedema and how these symptoms compare to clinical endpoints. Finally, in Chapter 4, a multivariate zero-inflated proportional odds (MZIPO) model is proposed to account for excess symptom non-response at baseline. This model adds a latent class component to the traditional MIMIC model. The MZIPO model is applied to the PAL data to obtain more accurate estimates of the latent construct and its association with current measures of lymphedema severity.