Discrete choice experiments in practice

There are 5 stages to conducting a discrete choice experiment (DCE). Before reading into the components of each of these stages and how they can be used, take a look at the following infographic for a summary.

Stage 1: Identifying attributes and levels

This stage is essential to the quality of the study findings. It includes evidence and rationale, input from target respondents and experts, and guidelines. The choice of attributes and levels is where reviewers tend to focus. Therefore, it is important to be able to describe the rationale for your choice of attributes and levels. Usually DCEs have 4-7 attributes, and 3-4 levels for each attribute. Since you can test ideas that are not currently possible with DCEs, context is vital for respondents, and attributes must be described clearly.

Stage 2: Experimental design

This refers to the levels chosen for each attribute in constructing the alternatives, and how they’re paired together to make the choice questions. The total number of potential combinations can be very large. You must decide upon the number of choice questions, trading off the quantity of information received and respondent burden. Not all combinations are equally informative, therefore a trade-off must be created for respondents. Most researchers use experimental-design software to create a set of choices that optimises the amount of information that you will get from the respondent’s choices.

Stage 3: Survey development

Ethics must be considered in the development of the survey. This includes informed consent from respondents, data storage, and anonymity. The survey will include explanatory information at the beginning, using a variety of media, to establish context, attributes, levels and the different concepts. This can also include questions to help understand user choices, including socio-demographic questions and study-specific questions.

Stage 4: Sample recruitment and fieldwork

DCEs can use any of the common survey methods, including paper, face-to-face, and increasingly, online panels. They can deliver quotas for key characteristics, for example, a gender or age balance matching the population of interest. It is important to remember the groups of peoples who are underrepresented in the sample. DCEs don’t need large sample sizes because it uses multiple observations from the same respondents. For the minimal sample size, see the infographic below, where A is the number of attributes, and Q is the number of choice questions.

Stage 5: Analysis

The most commonly used analysis method in DCEs is conditional logistic regression. This models the probability of choosing option A, given that the alternative is option B, specifically taking into account the difference of the two alternatives presented. Another type of analysis used in DCEs is the mixed logit/random parameters logit. This model allows each respondent to have their own preferences, and the model outcome is a mean of the sample. Finally, there are the latent class models. This searches for respondents who make similar patterns of responses, and builds in effect a model for each of those groups.

An example of a statistical analysis along with a guide on how to interpret your statistical output can be accessed as a pdf below.

Whenever you are ready, click ‘next’ to proceed to the next step where we will explore the use of discrete choice experiments in AMS.