Bayesian D-Optimal Choice Designs for Mixtures

Aiste Ruseckaite, Peter Goos, Dennis FokSSRN 2014 · DOI

Introduction

Discrete choice experiments are a cornerstone of market research and economic analysis, used to quantify consumer preferences for various product attributes. However, a significant challenge arises when these attributes represent components of a mixture—such as the ingredients in a beverage or the allocation of time in a commute. In such cases, the attributes are not independent; they must sum to a constant total (e.g., 100%).

Standard experimental design tools often fail in this context because they assume independent variation. Furthermore, the optimal design of a choice experiment depends on the very preferences the researcher aims to measure—a classic "chicken-and-egg" problem. This paper introduces a systematic framework for creating Bayesian D-optimal experimental designs specifically tailored for mixture components, bridging the gap between industrial mixture experiments and discrete choice modeling.

Methodology

The proposed approach integrates several advanced statistical techniques to optimize the information gathered from choice experiments.

Model Integration

The authors utilize Scheffé mixture models (linear, quadratic, and special-cubic) within a Multinomial Logit (MNL) framework. This integration allows the model to account for the unique constraints of mixture data while capturing complex interactions, such as synergistic effects where the combination of specific ingredients yields a utility greater than the sum of their individual parts.

Bayesian D-Optimality

To address the dependency of the optimal design on unknown parameters, a Bayesian approach is employed. Rather than assuming a single set of "true" preferences, the design's performance is averaged over a probability distribution of potential preference parameters. This ensures the resulting design is robust across a range of likely consumer behaviors.

The Mixture Coordinate-Exchange Algorithm

A specialized algorithm was developed to search for optimal designs efficiently. Key features include:

  • Cox-effect Directions: When adjusting one ingredient's proportion, others are scaled proportionally to maintain the mixture constraint.
  • Minimum Support Designs: Using these as starting points significantly accelerates the computational search.
  • Quasi-Monte Carlo Integration: Halton sequences are used to evaluate the multidimensional integrals required for Bayesian criteria, providing a balance between speed and accuracy.

Experimental Validation

The methodology was validated using two realistic scenarios derived from existing literature:

  1. Sports Drink Sweeteners: A study involving a three-ingredient mixture (glycine, saccharin, and an enhancer) aimed at minimizing sweetness aftertaste.
  2. Cocktail Preferences: A three-ingredient mix (mango juice, blackcurrant syrup, and lemon juice) with specific lower-bound constraints on each component, requiring the use of L-pseudocomponents to transform the constrained space.

Discussion and Key Insights

A central finding of the research is the "Utility Balance" principle. As uncertainty regarding user preferences increases, the Bayesian optimal design naturally pairs alternatives that are more similar in utility. This prevents "dominant" alternatives from appearing, which would provide little information about the trade-offs consumers are willing to make.

Furthermore, the study highlights the pitfalls of "utility-neutral" designs, which assume indifference among all options. The authors demonstrate that even a coarse prior estimate of preferences within a Bayesian framework is significantly more efficient than assuming no preferences at all.

Results

The Bayesian designs demonstrated substantial efficiency gains over traditional benchmarks. In the sweetener case study, the optimal design was 20% to 33% more efficient than neutral benchmarks. Computationally, the algorithm performed on par with or better than commercial software like JMP for standard cases, while offering the unique capability to handle complex Bayesian scenarios that commercial tools currently cannot. Robustness tests involving 10,000 simulated draws confirmed that the Bayesian design outperformed locally optimal and neutral designs in the majority of cases (58% and 68%, respectively).

Conclusion and Future Work

While the current heuristic algorithm provides high-quality designs, it does not guarantee global optimality. Future research could extend this framework to Mixed Logit models to account for random taste variation across populations. Additionally, the algorithm could be adapted for other optimality criteria, such as A- or V-optimality, depending on whether the researcher's goal is parameter estimation or prediction of choice probabilities. This work provides a robust foundation for researchers to design more informative and efficient experiments in fields ranging from food science to transportation economics.

Summary updated

BibTeX
@misc{ruseckaite2014bayesian,
  author = {Aiste Ruseckaite and Peter Goos and Dennis Fok},
  title = {{Bayesian D-Optimal Choice Designs for Mixtures}},
  year = {2014},
  doi = {10.2139/ssrn.2435058},
  howpublished = {SSRN},
}
Figure