A Flexible Neural Network Model for Discrete Choice Analysis with Economically-Consistent Utility Specification

Core Concepts

The authors propose a new discrete choice model based on artificial neural networks, named "Alternative-Specific and Shared weights Neural Network (ASS-NN)", which provides a balance between flexible utility approximation from the data and consistency with random utility maximization theory and the fungibility of money assumption.

Abstract

The paper presents a new discrete choice model based on artificial neural networks (ANNs), called the "Alternative-Specific and Shared weights Neural Network (ASS-NN)". The key features of the ASS-NN are:

It provides a flexible utility approximation from the data, without the need to specify the utility functional form a priori.
It is consistent with random utility maximization (RUM) theory, as the utility functions are alternative-specific.
It satisfies the "fungibility of money" assumption, where the marginal utility of costs is the same across alternatives for the same individual and cost level.

The authors compare the performance of the ASS-NN with a conventional multinomial logit (MNL) model and another ANN-based model called the "Alternative-Specific Utility Deep Neural Network (ASU-DNN)".
The results of a Monte Carlo analysis show that the ASS-NN can accurately recover the true marginal utilities and willingness-to-pay measures, even when the data is generated from a non-linear utility function. In the empirical application using the Swissmetro dataset, the ASS-NN outperforms the MNL models in terms of goodness-of-fit, while providing economically-consistent outcomes such as marginal utilities and willingness-to-pay measures.

Stats

The average travel cost for train is 91.27 CHF, for Swissmetro is 110.19 CHF, and for car is 93.43 CHF.
The average travel time for train is 173.69 minutes, for Swissmetro is 91.38 minutes, and for car is 146.88 minutes.
The average headway for train is 70.04 minutes and for Swissmetro is 20.05 minutes.

Quotes

"Random utility maximisation (RUM) models are one of the cornerstones of discrete choice modelling. However, specifying the utility function of RUM models is not straightforward and has a considerable impact on the resulting interpretable outcomes and welfare measures."
"To balance these trade-offs, the analyst must provide a structure that provides enough flexibility to the ANN to approximate utility functions and, at the same time, satisfies consistency with RUM and economic assumptions that guarantee to derive meaningful interpretable outcomes and welfare measures."

Key Insights Distilled From

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

by Jose Ignacio... at arxiv.org 04-23-2024

https://arxiv.org/pdf/2404.13198.pdf

Deeper Inquiries

How could the ASS-NN framework be extended to incorporate additional decision-making factors beyond the traditional attributes, such as social influences, habit formation, or reference-dependent preferences?

Incorporating additional decision-making factors beyond traditional attributes into the ASS-NN framework can enhance the model's predictive power and provide more nuanced insights into individual choices. One way to extend the framework is by including social influences, such as peer behavior or social norms. This can be achieved by adding social network data or community-level variables to the model, allowing it to capture the impact of social interactions on decision-making.
Habit formation can also be integrated into the ASS-NN by including variables that reflect past behavior or inertia in decision-making. By incorporating lagged variables or indicators of past choices, the model can account for the influence of habits on current decisions. Additionally, reference-dependent preferences, where individuals' choices are influenced by comparison to a reference point, can be included by introducing reference levels or thresholds in the utility function.
By expanding the ASS-NN to incorporate these additional factors, the model can provide a more comprehensive understanding of decision-making processes and better capture the complexities of human behavior in choice scenarios.

What are the potential limitations of the fungibility of money assumption, and how could the ASS-NN model be adapted to relax this assumption in certain contexts?

While the fungibility of money assumption simplifies the modeling of cost-dependent utilities in the ASS-NN framework, it may not always hold true in real-world decision-making contexts. One limitation of this assumption is that it overlooks potential differences in the perceived value of money spent on different goods or services. In situations where individuals have specific budget allocations or mental accounting practices, the fungibility of money assumption may not accurately reflect their preferences.
To relax the fungibility of money assumption in certain contexts, the ASS-NN model can be adapted to allow for varying degrees of fungibility across alternatives. This can be achieved by introducing interaction terms or additional parameters that capture the differential impact of costs on utility for different goods or services. By incorporating these adjustments, the model can better account for individual variations in the valuation of money spent on different alternatives.
Furthermore, sensitivity analyses can be conducted to assess the robustness of the model to variations in the fungibility of money assumption. By testing the model under different scenarios and assumptions regarding the interchangeability of money across alternatives, researchers can gain insights into the potential biases introduced by the fungibility assumption and explore ways to mitigate them.

How could the insights from the ASS-NN model be used to inform the design of more effective transportation policies and infrastructure investments?

The insights derived from the ASS-NN model can offer valuable guidance for designing more effective transportation policies and infrastructure investments. By analyzing the marginal utilities and willingness to pay estimates generated by the model, policymakers can prioritize investments that align with travelers' preferences and maximize societal welfare.
For example, the model can identify which attributes, such as travel time or cost, have the greatest impact on individuals' choices and willingness to pay. This information can inform decisions on infrastructure improvements, service enhancements, or pricing strategies that are most likely to attract travelers and improve overall satisfaction.
Additionally, the ASS-NN can be used to simulate the effects of different policy scenarios, allowing policymakers to evaluate the potential outcomes of proposed interventions before implementation. By testing various policy options and assessing their impact on key metrics like mode choice or travel behavior, decision-makers can make more informed and evidence-based choices that lead to more efficient and sustainable transportation systems.
Overall, the insights from the ASS-NN model provide a data-driven approach to policy design, enabling policymakers to tailor interventions to meet the needs and preferences of travelers while optimizing the use of resources for transportation infrastructure and services.