Symposium In Honor of AAJ Marley
In the best-worst choice paradigm pioneered by Louviere (Finn and Louviere, 1992), the subjet is asked to select her best and her worst alternatives in any set of alternatives. Tony Marley conceived a random utility model providing a possible explanation for the choice frequencies (Marley and Louviere, 2005). He then asked for a characterization of the prediction range of the model, and since 2006 had been working from time to time on the problem with several collaborators (Samuel Fiorini, Mike Regenwetter, Reinhard Suck, and the speaker). The problem was quickly turned into the search of an affine description of the convex polytope formed by the model predictions. However, not much is known about the polytope. In the particular case of four alternatives, a description consists of 26 affine equalities and 144 affine inequalities (Doignon, 2023), and the Gale transform of the set of vertices fully reveals the polytope structure: the transform is a family of 24 vectors in a one-dimensional vector space. Thus the expert eye can read the full structure of the polytope from 24 vectors on a line. For more than four alternatives, Marley problem remains open. SYMPOSIUM IN HONOUR OF A.A.J. MARLEY Adele Diederich Jamal Amani Rad Jean-Paul Doignon Karim Kilani Konstantina Sokratous Marion Collewet Quentin Gronau Thomas Hancock Xinwei Li
This is an in-person presentation on July 21, 2023 (09:20 ~ 09:40 UTC).
Dr. Mohammad Hemami
Dr. Reza Khosrowabadi
Dr. Jamal Amani Rad
The prevalence of methamphetamine use disorder (MUD) as a major public health problem has increased dramatically over the last two decades, reaching epidemic levels, which pose high costs to healthcare systems worldwide and is commonly associated with experience-based decision-making (EDM) aberrant. However, precise mechanisms underlying such non-optimally in choice patterns still remain poorly understood. In this talk, to uncover the latent neurobiological and psychological meaningful processes of such impairment, we apply a reinforcement learning diffusion decision model (RL-DDM) while methamphetamine abuser participants (n = 18, all men; mean (±SD) age: 27.3±5) and age/sex-matched healthy controls (n = 25, all men; mean (±SD) age: 26.8.0±3.63) perform choices to resolve uncertainty within a simple probabilistic learning task with rewards and punishments. Preliminary behavior results indicated that addicts made maladaptive patterns of learning that mirrored in both choices and response times (RTs). Furthermore, modeling results revealed that such EDM impairment (maladaptive pattern in optimal selection) in addicts was more imputable to both increased learning rates (more sensitive to outcome fluctuations) and decreased drift rate (less reward sensitivity) compared to healthy. In addition, addicts also showed substantially longer non-decision times (attributed to slower RTs), as well as lower decision boundary criteria (reflection of impulsive choice). Taken together, our findings reveal precise mechanisms associated with EDM impairments in methamphetamine use disorder and confirm the debility of the options values assignment system as the main hub in learning-based decision-making.
This is an in-person presentation on July 21, 2023 (09:40 ~ 10:00 UTC).
Dr. Karim Kilani
This paper proposes new models for analyzing best, worst, and best&worst choice probabilities for logit and reverse logit models, building on the pioneering work of Tony Marley. We assume individuals have a stochastic underlying ranking of alternatives and posit a rationality assumption relating to the random utility model. We focus on models applicable to best and worst choice scaling experiments, utilizing an inclusion-exclusion identity to propose a variety of best-worst choice probability models that can be implemented in specialized software packages. We demonstrate the versatility and utility of logit and reverse logit models in capturing the underlying ranking of alternatives. Finally, we discuss the practical implications of these models and future research directions in this field. Our models can be implemented in popular software packages such as Apollo.
This is an in-person presentation on July 21, 2023 (10:00 ~ 10:20 UTC).
Prof. Stephane Hess
Prof. Charisma Choudhury
A. A. J. Marley
Decision field theory (DFT), although popular in mathematical psychology, has only recently been used in choice modelling for consumer and travel choices. A key difference that DFT has from standard choice models is that it has preference values for each alternative that update over the course of the decision-making process. This results in a different probability of picking each alternative depending on how long a decision-maker considers their alternatives. However, the computational complexities of DFT have resulted in failures to utilise its dynamic nature. Recent advances in the underlying computational methods for DFT have allowed for the calculation of the probability of choosing alternatives at any time point. Consequently, the number of preference accumulation steps can be linked to the choice response time. In the work in this paper, we develop an integrated choice and latent variable decision field theory model to predict choice responses and choice response time. We use these models to explore the confounding nature of choice response time. A key assumption within DFT and other accumulator models is that preference grows over time, contradicting a well-known result that a longer response time often indicates a less certain and hence less deterministic choice from a decision-maker. In line with DFT and preference accumulation, we find that across model results from three datasets, a longer mean response time indicates that a decision-maker appears more deterministic. However, within a decision-maker, our models suggest that fast decisions are typically more deterministic, demonstrating that a longer response time indicates a less certain decision. Whilst there is a weak correlation between choice response time and the estimated number of preference updating steps, results from multinomial logit (MNL) models suggest that DFT's time parameter performs a similar function to an MNL's scale parameter. This suggests that caution is required in interpreting the outputs from accumulator models. SYMPOSIUM IN HONOUR OF A.A.J. MARLEY Adele Diederich Jamal Amani Rad Jean-Paul Doignon Karim Kilani Konstantina Sokratous Marion Collewet Quentin Gronau Thomas Hancock Xinwei Li
This is an in-person presentation on July 21, 2023 (10:20 ~ 10:40 UTC).
Subjective value has long been measured using binary choice experiments to assess individual differences in intertemporal preferences. Dynamic, stochastic models of choice permit meaningful inferences about cognition from process-level data, explaining value in terms of underlying mechanisms in a way that simpler, static models cannot. However, the usability of complex generative models is severely limited by the technical difficulty of model fitting and model comparison steps, along with the computational power they require. In this talk, we develop and test an approach that uses deep neural networks to estimate the parameters of three behavioral models and perform model comparison between the three to assess their ability to better account for intertemporal choice. The models we explore differ in their complexity and the theoretical assumptions they make when it comes to the study of preference; the traditional and static hyperbolic discount and hyperboloid functions compared with a probabilistic attribute-wise model constructed by direct and relative differences in delay and payoff. Once trained, the neural networks allow for accurate and instantaneous parameter estimation and model comparison, as opposed to traditional methods that can take several hours and in some cases days. We compare different network architectures and show that they are able to accurately recover true intertemporal preferences related to each model's parameters, and then compare each model's performance in their ability to predict individual choice. The models were applied to a large data set of substance users in protracted abstinence from Sofia, Bulgaria who completed a short, 27-question choice task. The results illustrate the utility of machine-learning approaches for wider adoption and integration of cognitive and economic models, providing efficient methods for quantifying meaningful differences in intertemporal preferences from simple experiments.
This is an in-person presentation on July 21, 2023 (11:00 ~ 11:20 UTC).
Dr. Paul Koster
Point allocation experiments are widely used in the social sciences. In these experiments, survey respondents distribute a fixed total number of points across a fixed number of alternatives. This paper reviews the different perspectives in the literature about what respondents do when they distribute points across options. We find three main alternative interpretations in the literature, each having different implications for empirical work. We connect these interpretations to models of utility maximization that account for point and budget constraints and investigate the role of budget constraints in more detail. We show how these constraints impact the regression specifications for point allocation experiments that are commonly used in the literature. We also show how a formulation of a taste for variety as entropy that had been previously used to analyse market shares can fruitfully be applied to choice behaviour in point allocation experiments.
This is an in-person presentation on July 21, 2023 (11:20 ~ 11:40 UTC).
Dr. Scott Brown
Dr. Ami Eidels
Discrete choice (DCE) and rating scale experiments (RSE) are commonly applied procedures for eliciting preference judgments in a plethora of applied settings such as consumer choices, health care, and transport economics. An almost universal assumption underlying their use is that the two procedures elicit reports generated from a common internal preference state; that is, actual “ground truth” preferences are not dependent on which procedure is used to elicit them. It is usually not possible to test this assumption, because typical studies using DCE and RSE methods have response options for which there is no objectively correct response, and no ground truth. To facilitate a comparison of DCE and RSE, we conducted a perceptual discrimination experiment where response options varied on a single attribute -- stimulus saturation level -- with a known objectively correct response. We had the same participants complete both a DCE and RSE version of the experiment, allowing a direct examination of the assumption that a common representation underpins responses in both. For this purpose, we developed a cognitive model with a response mechanism for both DCE and RSE based on latent Gaussian stimulus representations. This enabled us to compare a model version that featured one shared latent stimulus representation across DCE and RSE versus a model version which featured a separate latent representation for DCE and RSE. Our results support the assumption that a single internal state supports both DCE and RSE responses, and also suggest that the DCE method might provide more sensitive measurement of internal states than the RSE method.
This is an in-person presentation on July 21, 2023 (11:40 ~ 12:00 UTC).
Dr. Prateek Bansal
Eye-tracking data such as the gaze patterns reveal important attention-related information about the evidence accumulation process in stated preference (SP) experiments but can only be collected in the lab with a relatively limited number of subjects owing to time and resource constraints. On the other hand, the online SP experiments offer a large sample size but at the expense of eye movement data. Extant literature uses online and lab-based eye-tracking data in isolation. This study develops an approach to elicit consumer preference by jointly leveraging both datasets and validatie it by collecting stated preferences of Singaporean ride-hailing drivers to rent electric vehicles in lab-based (N = 40) and online/street-intercept (N =300) choice experiments. This study is relevant in the local context due to the very high cost of vehicle ownership. To explore the general and accountable interactions between decoy effect strength, attention, and preference formation, an improved Multi-attribute Linear Ballistic Accumulator model concerning absolute attribute value with hierarchical structure (henceforth, HA-MLBA) is adapted and calibrated using both lab-based and online datasets. Specifically, the posterior distribution of HA-MLBA parameters estimated by lab-based data is considered as the prior distribution for the corresponding parameters (including process, alternative-specific parameters, and attribute-specific parameters) while estimating HA-MLBA using online data. To highlight the superiority of this data fusion method, in-sample and out-of-sample performances in fitting choice and response time distribution of the HA-MLBA model with non-informative prior (baseline) and data-fusion prior (informative) are compared. This research demonstrates the presence of the decoy effects (similarity effect and Attraction Effect particularly) in the vehicle rental market. With the increased online purchase of vehicle rentals, such context effects could be vital in nudging ride-hailing drivers to adopt electric vehicles.
This is an in-person presentation on July 21, 2023 (12:00 ~ 12:20 UTC).
The Cube model (Mallahi-Karai and Diederich, 2019) is a dynamic-stochastic approach for decision making situations including multiple alternatives. The underlying model is a multivariate Wiener process with drift, and its dimension is related to the number of alternatives in the choice set. Here we modify the model to account for Best-Worst setting. The choices are made in a number of episodes allowing the alternatives to be ranked from best to worst or from worst to best. The model makes predictions with respect to choice probabilities and (mean) choice response times. We show how the model can be implemented using Markov chains and test the model on data from (Hawkins et al., 2014b).
This is an in-person presentation on July 21, 2023 (12:20 ~ 12:40 UTC).