How Chinese Restaurants Can Help With Robust Preference Insights: A Novel Dirichlet Mixture Model to Account for Complex Heterogeneity in Individual Preference Weights
Author(s)
Keila Meginnis, PhD1, Nicolas Krucien, PhD2, Christine Michaels-Igbokwe, BA, PhD3, Caitlin Thomas, BSc, MSc2, Sebastian Heidenreich, BSc, MSc, PhD4.
1Thermo Fisher Scientific, Glasgow, United Kingdom, 2Thermo Fisher Scientific, London, United Kingdom, 3Thermo Fisher Scientific, Calgary, AB, Canada, 4Executive Director, Patient-Centred Research, Thermo Fisher Scientific, London, United Kingdom.
1Thermo Fisher Scientific, Glasgow, United Kingdom, 2Thermo Fisher Scientific, London, United Kingdom, 3Thermo Fisher Scientific, Calgary, AB, Canada, 4Executive Director, Patient-Centred Research, Thermo Fisher Scientific, London, United Kingdom.
OBJECTIVES: Individual level preference elicitation has recently become an active area of research, with small sample sizes and shared decision-making being relevant applications. A key challenge for these methods is the characterisation of the sample level average. We introduce an analytical approach that endogenously accounts for multi-modal preferences.
METHODS: Individual level preferences are typically heterogenous. While some people may place the highest importance on avoiding risks, others might prioritise benefits. The analysis of individual level trade-off data using recently suggested Dirichlet models can be misleading if preferences reflect competing priorities. We used a simulation approach to quantify the bias of a standard Dirichlet regression under varying degrees of preference clustering and compared the findings to a model that uses the Chinese Restaurant Process to mix multiple Dirichlet distributions. Additional simulations were used to understand the implications for minimum sample sizes.
RESULTS: Assuming all preferences cluster around one treatment priority, the relative bias observed for the standard Dirichlet model was small (2.93%), which aligns with previous publications. However, the bias increased considerably (62.73%) even if preferences clustered only around two different priorities. In contrast, the continuous mixture of Dirichlet distributions resulted in a smaller bias (2.52%) for up to five different treatment priorities. The model identified the right number of priority clusters 99% of the time. While sample size requirements remained low compared to sample level preference elicitation methods, they increased with the complexity of the preference heterogeneity.
CONCLUSIONS: To avoid bias, accounting for preference heterogeneity in individual level preference data is important. This has real practical relevance. For instance, if some patients tend to prioritise efficacy while others prioritise avoiding risks, average estimates and resulting policy recommendations can lead to suboptimal treatment initiation or approval decisions. The proposed analytical approach can help overcome these challenges where sample sizes allow.
METHODS: Individual level preferences are typically heterogenous. While some people may place the highest importance on avoiding risks, others might prioritise benefits. The analysis of individual level trade-off data using recently suggested Dirichlet models can be misleading if preferences reflect competing priorities. We used a simulation approach to quantify the bias of a standard Dirichlet regression under varying degrees of preference clustering and compared the findings to a model that uses the Chinese Restaurant Process to mix multiple Dirichlet distributions. Additional simulations were used to understand the implications for minimum sample sizes.
RESULTS: Assuming all preferences cluster around one treatment priority, the relative bias observed for the standard Dirichlet model was small (2.93%), which aligns with previous publications. However, the bias increased considerably (62.73%) even if preferences clustered only around two different priorities. In contrast, the continuous mixture of Dirichlet distributions resulted in a smaller bias (2.52%) for up to five different treatment priorities. The model identified the right number of priority clusters 99% of the time. While sample size requirements remained low compared to sample level preference elicitation methods, they increased with the complexity of the preference heterogeneity.
CONCLUSIONS: To avoid bias, accounting for preference heterogeneity in individual level preference data is important. This has real practical relevance. For instance, if some patients tend to prioritise efficacy while others prioritise avoiding risks, average estimates and resulting policy recommendations can lead to suboptimal treatment initiation or approval decisions. The proposed analytical approach can help overcome these challenges where sample sizes allow.
Conference/Value in Health Info
2025-11, ISPOR Europe 2025, Glasgow, Scotland
Value in Health, Volume 28, Issue S2
Code
MSR120
Topic
Methodological & Statistical Research
Disease
No Additional Disease & Conditions/Specialized Treatment Areas