A Multinomial and Rank-Ordered Logit Model With Inter- and Intra-Individual Heteroscedasticity

Anoek Castelein, Dennis Fok, Richard PaapSSRN 2020 · DOI

TL;DR

This paper introduces a novel framework for modeling individual choices that accounts for both within-individual (intra-individual) and between-individual (inter-individual) variations in decision-making randomness. By employing a hidden Markov model, the authors capture how individuals' decision-making "noise" can change over time and differ across people, leading to more accurate predictions and a deeper understanding of individual behavior in multinomial and rank-ordered choice scenarios.

What Problem Does This Paper Address?

Understanding how people make choices is crucial across many fields, from marketing to public policy. Traditional models like the logit model assume that the randomness in an individual's decision-making is constant. While extensions like the heteroscedastic logit model allow for this randomness to change over time (e.g., due to fatigue in surveys), they typically assume this change happens uniformly across all individuals. This is often unrealistic; some people might get fatigued faster, others might learn more slowly, and these dynamics can vary significantly from person to person. This paper addresses the limitation of existing models by developing a framework that explicitly models these inter- and intra-individual differences in the dynamics of choice randomness.

Methods

The core innovation is the development of a hidden Markov model (HMM) applied to both multinomial logit (MNL) and rank-ordered logit (ROL) models.

  • Heteroscedasticity: The standard approach models choice randomness using a scale parameter. This paper extends this by allowing this scale parameter to vary across individuals (inter-individual) and over time/ranks for each individual (intra-individual).
  • Hidden Markov Model: The HMM introduces the concept of "phases" that an individual can move through. Each phase is associated with a specific scale parameter, representing different levels of decision-making randomness. The transitions between these phases are governed by a Markov process, meaning the probability of moving to a new phase depends only on the current phase.
  • Multinomial Logit (MNL): For choices where only the top preference is observed, the HMM models how an individual's scale parameter evolves through different phases over time.
  • Rank-Ordered Logit (ROL): For rankings, the HMM models how an individual's scale parameter evolves across different ranks within a single ranking task.
  • Estimation: The models are estimated using maximum simulated likelihood estimation, which handles the complexity of unobserved individual states and preference heterogeneity.

Datasets

The paper uses two types of data for empirical illustration:

  • Discrete Choice Experiment (DCE) Data: Data from a food choice experiment in the Netherlands, collected from the LISS panel. This data is used to model learning and fatigue effects in choices.
  • Rank-Ordered Data: Survey data on political preferences from the Netherlands, where individuals ranked 16 political goals. This data is used to model differential capabilities in ranking.

Key Insights and Ideas

  • Individual Differences Matter: The most significant insight is that assuming uniform changes in decision-making randomness across individuals is often a poor assumption. Accounting for individual-specific dynamics leads to better model fit and more nuanced insights.
  • Phases of Decision-Making: The HMM framework effectively captures distinct "phases" of decision-making, such as a "learning" phase, a "most accurate" phase, and a "fatigue" phase, which can vary in duration and transition probabilities across individuals.
  • Bias Alleviation: The proposed models successfully alleviate biases found in standard MNL and heteroscedastic logit models, particularly those arising from unobserved heterogeneity and incorrect assumptions about scale parameter dynamics.
  • Distinguishing Heterogeneity: The models can distinguish between time-varying (intra-individual) and time-invariant (inter-individual) scale heterogeneity, providing a more granular understanding of individual behavior.

Key Results

  • Monte Carlo Studies: The proposed HM-MNL and HM-ROL models performed well in Monte Carlo simulations, showing unbiased parameter estimates and outperforming standard MNL and heteroscedastic models, especially when individual differences were present.
  • DCE Application (Learning & Fatigue): In the food choice experiment, the HM-MNL models provided a significantly better fit to the data compared to standard MNL and H-MNL, revealing distinct learning and fatigue patterns that varied across respondents. For instance, one HM-MNL specification was preferred over the H-MNL despite having fewer parameters.
  • Ranking Application (Differential Capabilities): For the political preferences ranking data, the HM-ROL models also showed superior fit. The results suggested that individuals might not assign middle ranks as accurately as top or bottom ranks, and these capabilities differ across individuals. The third HM-ROL specification, allowing for a decrease in accuracy for consecutive ranks, was found to be the most preferred.

Limitations and Future Work

  • Scale Parameter Limitation: A key limitation is that the current model scales all variables with the same factor. It cannot capture choice strategies where different subsets of attributes are used over time or where preferences change for specific attributes.
  • Future Work - Attribute-Specific Scales: The authors suggest extending the model to allow for different scale parameters per variable, each governed by its own Markov process. However, this could lead to overfitting and estimation uncertainty, especially with limited data per individual.
  • Future Work - Higher-Order Markov Processes: To impose restrictions on phase duration (e.g., minimum number of tasks in a phase), second- or higher-order Markov processes could be used.
  • Future Work - Combined Rank/Time Dynamics: For rank-ordered data, combining Markov processes for time and ranks could simultaneously model learning, fatigue, and differential ranking capabilities.
  • Future Work - Flexible Preference Heterogeneity: The authors recommend using more flexible forms of preference heterogeneity (beyond multivariate normal) to better capture individual differences, especially when scale parameters can range from one to infinity. This could help disentangle time-invariant scale heterogeneity from the Markov process dynamics.

Summary updated

BibTeX
@misc{castelein2020multinomial,
  author = {Anoek Castelein and Dennis Fok and Richard Paap},
  title = {{A Multinomial and Rank-Ordered Logit Model With Inter- and Intra-Individual Heteroscedasticity}},
  year = {2020},
  doi = {10.2139/ssrn.3701932},
  howpublished = {SSRN},
}

Loading related content...

Figure