07.Health economics02.Discrete choice experiments - sporedata/researchdesigneR GitHub Wiki
Discrete choice experiments -- also known as conjoint analyses, multi-attribute compositional modelling, discrete choice modelling, or stated preference research -- evaluate how to assess different attributes associated with a given product (such as a medical device) or a service (such as Urgent Care service). Attributes can include specific features, functions, cost, ease of use, as well as other characteristics that patients or other stakeholders might see as benefits. -- see Location Isn't Everything: Proximity, Hospital Characteristics, Choice of Hospital, and Disparities for Breast Cancer Surgery Patients
- In-depth information on a product or service
- Current or potential users
The two main types of discrete choice experiments are choice-based conjoint and adaptive conjoint analysis. An example of a choice-based conjoint analysis question is "Would you rather schedule an Urgent Care appointment immediately for a sore throat for $300 or wait four days for a regular clinic appointment for $80?” In contrast, adaptive conjoint analysis tends to use Likert-scale type of question for its attribute-based questions. For example, one would describe the characteristics of an Urgent Care appointment (waiting time, price, etc) and then ask how likely the respondent would be to pay for that service.
From the perspective of academicians working within research consortia, the conjoint analysis will provide them with an estimate of the market value of their technologies, thus decreasing the information asymmetry between them and potential intellectual property buyers [1].
For starters, conjoint analysis means that you conduct a study to evaluate which characteristics of a product, services, etc might influence your choice to get it. AMCE is a way to compare the effect of each of the characteristics of that product/service on your choices to get it or not. More specifically, AMCE represents how much the probability of making a certain choice would change, on average, if one of the characteristics of the thing you're choosing (a product, a service, etc) were to be switched from one level to another. For example, if you are choosing whether to get vaccinated or not for COVID, a result saying that the vaccine is 50% vs 70% effective in preventing that you end up in an ICU might increase willingness to get vaccinated everything else being held constant [12]. The tricky part in the definition is the "on average" portion, since the traditional AMCE assumes that all choices are equally likely - new methods are now trying to make this assumption connected to the actual choices in the real world [13]. Regarding modeling, these are usually simple ordinary least square models.
- Using qualitative methods for attribute development for discrete choice experiments: issues and recommendations [2].
- Common packages for health economics
- Conjoint: An Implementation of Conjoint Analysis Método.
- Mlogit: Multinomial Logit Models.
-
Books
-
Articles combining theory and scripts
- Main Applications of Conjoint Analysis [5].
- Conjoint R package.
- [How to Use R for Conjoint Analysis [6].
- Survey software for discrete choice experiments.
- Conjoint analysis with QuestionPro.
Based on the analysis, one can conclude the users (patients/healthcare providers/participants) preferred choice from the available options for a given product (such as a medical device) or a service (such as a medical training program). Moreover, one can also conclude the most critical attribute or deciding criterion of the product or service from the user's perspective. This information could improve the quality of the product or service being offered to the users.
Below are the tables and plots for a discrete choice experiment or conjoint analysis evaluating the perception of surgeons participating in a advanced therapy training course.
Table: Estimated part worths (utilities) of each attribute of a medical training course based on multinomial logit model (McFadden pseudo R2 = 0.19901).
| Attribute | Levels | Utility score (SD) |
|---|---|---|
| Course duration | 2-day, 6-hour/day | -0.864536 (0.063) |
| 3-day, 4-hour/day | -0.322429 (0.059) | |
| Content | Module B | 0.443781 (0.0509) |
| Speaker experience | >2 to <=5 yrs | -0.712818 (0.0592) |
| >5 yrs | -1.383759 (0.0657) | |
| Fee | 350 USD | -0.869697 (0.0595) |
| 400 USD | -1.528377 (0.0658) |
The estimated part-worth coefficients for each level relative to each attribute's base levels [7] as well as the significance of test results indicate whether there is a detectible difference in preference for an attribute level relative to the base level [3]. In the example, the results indicate that the participants strongly liked 'Module B' over 'Module A' (base level) of the training program.
The pseudo R^2^ value is an additional criterion for the goodness of fit, reflecting the consistency of respondents' behavior [8]. A McFadden's pseudo R2 ranging from 0.2 to 0.4 indicates excellent model fit [11].
- Relative importance of training course attributes arranged in descending order.
The average or relative importance of attributes is based on differences between maximum and minimum part-worth utilities within each attribute [9] [10]. In the example, the training program's content was the essential deciding criterion among the participants.
- Subgroup analysis of relative importance of training course attributes by surgeons’ opinion of the compression therapy training on follow-up.
The subgroup analysis indicates that 'content' was the most important attribute for participants with a positive opinion of the training program on follow-up.
[1] Schwarz S, Voeth M, Herbst U. Information asymmetry in buyer-seller negotiations and its impact on effectiveness, efficiency and satisfaction. In26th IMP-conference in Budapest 2010.
[2] Coast J, Al‐Janabi H, Sutton EJ, Horrocks SA, Vosper AJ, Swancutt DR, Flynn TN. Using qualitative methods for attribute development for discrete choice experiments: issues and recommendations. Health economics. 2012 Jun;21(6):730-41.
[3] Chapman C, Feit EM. R for marketing research and analytics. New York, NY: Springer; 2015 Mar 9.
[4] Train KE. Discrete choice methods with simulation. Cambridge university press; 2009 Jun 30.
[5] Bock T. Main Applications of Conjoint Analysis. 27 February 2019.
[6] Telang A. How to Use R for Conjoint Analysis. Feb. 19, 19.
[7] Lee WC. and Joshi AV. and Woolford S. and Sumner M. and Brown M. and Hadker N. and Pashos CL. Physicians’ preferences towards coagulation factor concentrates in the treatment of Haemophilia with inhibitors: a discrete choice experiment. Wiley Online Library; 15 February 2008.
[8] Abdel-All, Marwa and Angell, Blake and Jan, Stephen and Howell, Martin and Howard, Kirsten and Abimbola, Seye and Joshi, Rohina What do community health workers want? findings of a discrete choice experiment among accredited social health activists (ASHAs) in India. BMJ global health; May 2019.
[9] Zimmermann, Thomas M and Clouth, Johannes and Elosge, Michael and Heurich, Matthias and Schneider, Edith and Wilhelm, Stefan and Wolfrath, Anette Patient preferences for outcomes of depression treatment in Germany: a choice-based conjoint analysis study. Journal of affective disorders; 2013 Jun;148(2-3):210-219.
[10] Bak, Andrzej and Bartlomowicz, Tomasz Conjoint analysis method and its implementation in conjoint R package. Wroclaw: 2011.
[11] Lee D. A comparison of choice-based landscape preference models between British and Korean visitors to national parks. Life Science Journal; 2013; 10(2).
[12] Kreps S, Prasad S, Brownstein JS, Hswen Y, Garibaldi BT, Zhang B, & Kriner DL. Factors associated with US adults’ likelihood of accepting COVID-19 vaccination. JAMA Netw Open. 2020 Oct; 3(10): e2025594.
[13] De la Cuesta B, Egami N, & Imai K. Improving the external validity of conjoint analysis: The essential role of profile distribution. Cambridge University Press: 14 January 2021.