Exploring heterogeneity of stated preferences through latent class analysis
MetadataShow full item record
Patient preferences have been increasingly incorporated into clinical and regulatory decision-making. It leads to a growing interest in advancing and applying methods to study preference heterogeneity. Latent class analysis (LCA) is an emerging technique used in stated-preference studies to segment people by preferences instead of observed characteristics (e.g. demographics). The objective of this dissertation is to examine and advance the application of LCA in stated-preference studies in the context of health to support medical decision-making. A systematic review was first conducted to document segmentation methods used in the health-focused stated-preference studies in current literature. It identified current practices and knowledge gaps. LCA is then applied to empirical stated-preference data generated by two most commonly used stated-preference methods identified in the systematic review, namely discrete choice experiment (DCE) and best-worst scaling (BWS). Model specifications were modified in both applications to better serve policy and clinical decision-making. Latent class logit (LCL) is the most commonly used segmentation model that has been applied in both DCE and BWS. However, both applications have limitations. LCL sometimes over-fits the DCE data and leads to too many classes that are difficult to incorporate in policy or clinical decision-making when there is substantial within-class preference heterogeneity or significant overlap between classes. Random effects are incorporated in LCL models as a remedy in this dissertation. With more flexible model specification, random effect LCL is shown not only reducing the number of classes but also better capturing the complex and dispersed preference pattern among patients than LCL, leading to improved model fit and prediction accuracy. When LCL is applied to BWS data, information criteria also often fail to identify the best-fitted model and parsimonious segmentation results due to cross-task constant utility assumption in LCL. A standard LC model was used to relax the constraint in this dissertation. It dramatically reduced the number of classes and generated practical segmentation results. Given that regulatory and clinical decision-makers often prefer parsimonious results due to limited resources and capacity to accommodate too many preference types, more flexible model specifications should be used in stated-preference studies to generate more practical results to support decision-making.