Date of Award

Spring 2026

Document Type

Open Access Dissertation

Department

Epidemiology and Biostatistics

First Advisor

James W. Hardin

Abstract

A discrete choice experiment is usually conducted in economics background to measure how consumers would like to trade off attributes of a product, service, or policy. In this experiment, respondents are provided different hypothetical combinations of varied components by using an experimental design such as the full factorial design, which fully represents the all combinations of attributes and levels. The mathematical background of a discrete choice experiments is based random utility theory, in which either positive or negative values are given each characteristic and the weighted sum values of each choice given in the experiment will be compared. This experiment has been used across different fields such as health economics, marketing, transportation, public policy and so on to reflect respondents’ real preference.

At first, choices provided to decision makers are considered as independent and a multinomial model is fit to data to extract information. However, the independence assumption can be violated in real life. For example, some alternatives may have inherent connections due to some observable or unobservable common parts. Thus, using a traditional multinomial model may cause bias. As a result, a nested logistic model is used to address this issue by incorporating the common parts into the model. Also, sometimes different groups people may hold different preferences, which means individual behavior toward the same level of attribute may vary. In other words, there may exist more than one set of parameters. Accordingly, latent class logit model were invented to capture individual behavior by including latent or observable factors into model. In this dissertation research, we will try to explore a new model called latent class nested logit model, which can handle both situations at the same time.

Another important issue that needs to be addressed is the opt-out alternative. In real life situation, decision makers may not choose any alternatives given to them due to different reasons. If an opt-out alternative is not given in the experiment, severe bias could happen because the experiment does not align with what happen in reality. In order to extract as much useful information from decision makers as possible, some variations of traditional discrete choice experiment have been designed such as Dual Response 2Max design and Dual Response AnyMax design. In these designs, each decision maker is asked to choose their preference in two steps. In the first step, the choice is made among forced alternatives that do not include optout while in the second step, opt-out alternative will be combined either the most preferred alternative from the first step or the all other alternatives. A previous study has explored how these new designs would affect the accuracy and power of model fitting under independent alternatives. In this dissertation research, the similar comparisons were made but under the nested assumption.

Finally, in a previous study, a best-worst experiment design was conducted to capture decision makers’ preference over different levels of attributes. In such a design, each decision maker was asked to choose their most favored alternative, then least favored alternative and finally were asked if they would opt-out, which means they would like to choose none of provided alternatives. One potential issue is that some respondents who choose opt-out for all choice sets were excluded from the experiment, although they did choose most and least favored alternative in the first two steps. These exclusions are problematic because the information from these people can be useful if their decision to opt out all chocies was made not intentionally but coincidentally. In order to fully capture this useful information, we developed novel methods to fully utilize opt-out information, enhancing the conclusiveness of research results derived from real data.

Rights

© 2026, Dai Fang

Download

Included in

Biostatistics Commons

Share

COinS