An economically-consistent discrete choice model with flexible utility specification based on artificial neural networks

Overview

• Random Utility Maximization (RUM) models are a core part of discrete choice modeling, but specifying the utility function is challenging and can impact interpretable outcomes.
• The paper proposes a new discrete choice model based on Artificial Neural Networks (ANNs), called Alternative-Specific and Shared weights Neural Network (ASS-NN).
• ASS-NN aims to balance flexible utility approximation with consistency with RUM theory and the fungibility of money.
• The model can derive economically-consistent outcomes like marginal utilities and willingness to pay without explicitly specifying the utility function.

Plain English Explanation

When people make choices, like choosing between different products or services, researchers often use discrete choice models to understand the factors that influence those decisions. One of the most common types of discrete choice models is the Random Utility Maximization (RUM) model.

RUM models assume that people choose the option that gives them the highest overall "utility" or satisfaction. However, defining this utility function can be tricky, and the way it's specified can significantly impact the results and insights you get from the model.

The paper introduces a new type of discrete choice model called the Alternative-Specific and Shared weights Neural Network (ASS-NN). This model uses artificial neural networks to flexibly approximate the utility function based on the data, rather than having to specify it upfront.

Importantly, the ASS-NN model also ensures that the results are consistent with two key economic principles:

1. The Random Utility Maximization (RUM) theory, which says people choose the option that gives them the highest overall utility.
2. The idea that money is "fungible" - that is, one euro is worth the same as any other euro.

By meeting these economic constraints, the ASS-NN model can generate insights like marginal utilities and willingness to pay without having to explicitly define the utility function. This can provide a more flexible and realistic way of modeling people's choices.

Technical Explanation

The paper proposes a new discrete choice model called the Alternative-Specific and Shared weights Neural Network (ASS-NN). This model uses artificial neural networks to approximate the utility function in a Random Utility Maximization (RUM) framework.

The key innovation is that the ASS-NN architecture includes both alternative-specific weights and shared weights across alternatives. The alternative-specific weights allow the model to capture heterogeneity in preferences across different options. The shared weights help ensure consistency with the RUM assumption that people choose the option that maximizes their overall utility.

Additionally, the ASS-NN is designed to be consistent with the principle of the fungibility of money - that is, one euro is worth the same as any other euro. This is achieved by constraining the neural network's output to satisfy certain economic properties.

Through a Monte Carlo experiment and empirical analysis on the Swissmetro dataset, the authors show that the ASS-NN outperforms conventional multinomial logit (MNL) models in terms of goodness of fit, under a variety of utility specifications. They also demonstrate how the ASS-NN can be used to derive economically-meaningful measures like marginal utilities and willingness to pay.

Critical Analysis

The paper presents an interesting and innovative approach to discrete choice modeling by leveraging neural networks while maintaining consistency with RUM theory and the fungibility of money. This is an important contribution, as flexibility in utility specification is a key challenge in discrete choice analysis.

One potential limitation is the computational complexity of the ASS-NN model, which may make it more challenging to implement and scale compared to simpler logit-based approaches. The authors do not provide a detailed analysis of the model's computational requirements or training times.

Additionally, the paper focuses on demonstrating the superior fit of the ASS-NN model, but does not extensively explore the interpretability and face validity of the derived marginal utilities and willingness to pay measures. Further research could assess how well these economically-consistent outputs align with real-world expectations and expert knowledge.

Incorporating uncertain or noisy data is another area that could be explored to further enhance the ASS-NN's applicability to real-world discrete choice problems.

Conclusion

The Alternative-Specific and Shared weights Neural Network (ASS-NN) proposed in this paper represents an innovative approach to discrete choice modeling that aims to balance flexible utility approximation with consistency with RUM theory and the fungibility of money.

By using a neural network architecture with both alternative-specific and shared weights, the ASS-NN can derive economically-meaningful insights like marginal utilities and willingness to pay without explicitly specifying the utility function. This could lead to more realistic and insightful models of human decision-making in a variety of domains.

The paper demonstrates the ASS-NN's superior performance compared to conventional multinomial logit models, but also raises questions about computational complexity and the interpretability of the derived economic measures. Further research exploring these aspects could help advance the state of the art in discrete choice modeling.