Journal of Choice Modelling最新文献

Hospital choice models often employ random utility theory and include waiting time as a choice determinant. When applied to evaluate health system improvement interventions, these models disregard that hospital choice in turn is a determinant of waiting time. We present a novel, general model capturing the endogeneous relationship between waiting time and hospital choice, including the choice to opt out, and characterize the unique equilibrium solution of the resulting convex problem. We apply the general model in a case study on the urban Chinese health system, specifying that patient choice follows a multinomial logit (MNL) model and waiting times are determined by M/M/1 queues. The results reveal that analyses which solely rely on MNL models overestimate the effectiveness of present policy interventions and that this effectiveness is limited. We explore alternative, more effective, improvement interventions.

Generating random draws from multivariate extreme value (MEV) distributions plays an important role in the microsimulation of travel behaviors, which can effectively avoid heavy computational burdens from simulation based on calculated probability values, particularly in simulations for a large population or choice behaviors from a large choice set. However, there are few practical and effective methods for drawing from MEV distributions. This paper proposes a simple and computationally efficient approach for drawing from MEV distributions in the nested logit (NL), cross-nested logit (CNL), and paired combinatorial logit (PCL) models. The proposed approach to draw from the MEV distribution for a CNL model provides a new perspective to understand the underlying choice mechanism of the CNL model. To our knowledge, this is the first study to draw from an MEV distribution in the PCL model. Random draws from the proposed approach approximately follow the standard Gumbel distribution, which is the marginal distribution of NL/CNL/PCL models, and approximate correlations among alternatives well. Simulation results of NL/CNL/PCL models show that the proposed approach provides high-level accuracy in recovering model parameters with the overall mean absolute percentage bias being less than 3%. The proposed approach is computationally more efficient than similar ones because it only needs to draw from Gumbel distributions. The proposed approach can be used to simulate NL/CNL/PCL models with a large choice set or a multiple discrete-continuous generalized extreme value model in various application settings such as joint destination-mode choices, time use allocations, etc.

This paper presents a micro-econometric framework to analyse choice data from participatory value evaluation (PVE) surveys. In a PVE survey respondents receive, similar to stated choice surveys, information on the social impacts of public sector projects before choosing the best policy portfolio according to their preferences. Respondents’ choices are limited by governmental and private budget constraints. The PVE data format is characterised by a mixture of discrete and continuous choice data. Building on recent literature of Kuhn–Tucker models, particularly the MDCEV model, a range of methodological and econometric contributions are provided facilitating model estimation and policy evaluation. We derive a set of closed form choice probabilities explaining the choice for the optimal portfolio with public projects, private consumption levels and whether to spend the public budget in full or not. The proposed policy evaluation framework is centred around the notion of social welfare maximisation. The parameter estimates are used to derive the optimal public sector budget and the corresponding portfolio maximising social welfare, but also to rank the set of feasible portfolios given a restricted budget, including sensitivity analyses. The proposed framework is illustrated using an empirical example on urban mobility investments in Amsterdam, The Netherlands.

The econometrics of the Linear Probability Model (LPM) cast as binary choice random utility model and where probabilities are constrained in the [0,1] interval is unexplored. The paper fills this gap. Assumptions are identified under which constrained maximum likelihood estimators exist and are unique, consistent and asymptotically normal. A consistent estimator of the covariance matrix is provided. Statistics that can be used to evaluate the prediction validity of binary choice models are reviewed. With income independent choices, the LPM has the merit of closed-form welfare change measure for the sub-population of consumers shifting from one alternative to the other. Two datasets illustrate the theoretical insights. One from the Swiss Mobility and Transport Microcensus related to choices between teleworking and commuting, one from the German Socio-Economic Panel related to add-on health insurance subscription. The signs and statistical significance at 5% level of the coefficients are concordant across LPM, Logit and Probit. Model prioritization based on prediction validity is data specific and dependent on the statistics used.

Decision field theory (DFT) is a model originally developed in cognitive psychology to explain behavioural phenomena such as context effects and decision-making under time pressure. Given this focus, the model has primarily been used to explain choices observed under controlled laboratory settings, with little attention paid to generalisability. Recent work has improved the mathematical foundations of DFT, making it a tractable model that is easier to apply to a wider variety of choice contexts. In particular, the inclusion of attribute importance parameters has led to successful applications to multi-alternative multi-attribute choice settings, notably with stated preference data in transport. However, thus far, implementations to real-life behaviour (i.e., revealed preference, RP, data) have been limited. The aim of this paper is to extend DFT for larger and more real-world applications, where data may be more ‘noisy’ and prone to larger variances of the error term. A theoretical extension for the model is presented, relaxing the assumption of independent normal error terms to capture heteroskedasticity. We apply the new model specification to two large-scale revealed preference datasets, also incorporating a range of sociodemographic variables. The new ‘heteroskedastic’ DFT model substantially outperforms the original version of DFT, as well as choice models based on econometric theory, in both estimation and validation subsets.

Previous work has identified attribute level overlap and level color coding as effective and attractive strategies to reduce task complexity and improve behavioral efficiency in discrete choice experiments (DCEs). However, the simultaneous and combined impact of level overlap and level color coding on attribute non-attendance and choice consistency has not yet been investigated. To address this limitation and to strengthen the available evidence base, this paper re-analyzed an existing randomized controlled DCE from the Netherlands (N = 2,731) and analyzed a new randomized controlled DCE conducted in the United Kingdom (N = 3,084) using heteroskedastic attribute non-attendance mixed logit models. Both randomized controlled experiments were based on a relatively complex instrument with 5 attributes with 5 levels each and the results from both experiments were remarkably similar. In the base-case study arms without level overlap and color coding, only about half of the attributes are attended to. Level color coding as a stand-alone strategy improves attribute attendance but reduces respondents' choice consistency. In contrast, level overlap as a stand-alone strategy improves attribute attendance while simultaneously increasing respondents' choice consistency. The combination of level overlap and color coding is even more effective: it results in approximately full attribute attendance and a 30% increase in respondents' choice consistency. Experimental designs with level overlap are therefore recommended as a default design strategy and level color coding recommended to further increase respondents’ behavioral efficiency in complex DCEs.

Spatial perceptions mediate human–environment interaction, and understanding spatial perceptions of humans can play a key role in the planning of activities. This study aims to analyze spatial multivariate binary choice data representing if an individual perceives a spatial unit to belong to a certain category (e.g., her neighborhood or set of potential activity places). To reasonably analyze such data, we present a spatial autoregressive mixed logit (SAR-MXL) model that accounts for both inter-individual heterogeneity and spatial dependence. We rely on the Bayesian approach for posterior inference of model parameters, where Pólya-Gamma data augmentation (PG-DA) is adopted to address the non-conjugacy of the logit kernel. The PG-DA technique eliminates the need for the Metropolis–Hastings step during the Markov Chain Monte Carlo process and allows for fast and efficient posterior inference. The high efficiency of the Bayesian SAR-MXL model is demonstrated through a numerical experiment. The proposed framework is applied to street-based neighborhood perception data, and we empirically analyzed the factors associated with the street perception probability of individuals. The result suggests a clear improvement of the model fit by incorporating spatial dependence and random parameters.

Whereas the number of alternatives per choice set in a labeled discrete choice experiment is often determined by the number of available labels, the choice set size in unlabeled choice experiments can be set more freely by the researcher. Determining the number of alternatives that will both yield enough information about the preferences and not overload the choice task for the respondents is, however, not an easy task. If the number of choice sets is restricted, the statistical efficiency of the designed experiment can be increased by increasing the number of alternatives per choice set. On the other hand, large choice sets are complex to deal with and could therefore lead to early fatigue and/or a plethora of screening heuristics that are hard to model. Moreover, although there is no compelling reason to keep the choice set size fixed in unlabeled discrete choice experiments, designs with varying choice set sizes have scarcely been studied. In this paper, we compute and investigate efficient designs with varying choice set sizes. We show that such designs can also be very efficient and we conjecture that such choice experiments are less monotonous for the respondents making it more likely that they will remain attentive. We report on two choice experiments that we conducted to check whether this assertion is correct. We compare designs with equal choice set sizes, with increasing choice set sizes and with random choice set sizes. The post-survey questions indicate that varying choice set sizes are indeed appreciated by the respondents while not reducing the statistical information obtained.

Existing literature has applied the sample selection modeling approach to disentangle the influence of the built environment (BE) and residential self-selection (RSS) on travel behavior. However, a limitation of the existing sample selection models is that they can handle only continuous or ordinal outcomes. The contribution of this study is twofold. First, we develop a sample selection model that can handle binary travel behavior outcomes in the RSS context. When the travel behavior outcome is binary, this approach's potential parameter identification issue can become serious. We employ a non-flat prior and Watanabe-Akaike information criterion in the Bayesian framework to address this issue. Second, we apply this proposed model to travel survey data in Kumamoto City, Japan, to disentangle the BE influence of a neighborhood type and RSS on car ownership. The neighborhood type is defined as the neighborhood being either less than 1,500 m (A) or greater than 1,500 m (B) from a station. We reveal that the true influence of the neighborhood type results in a mere 2.1 percentage point decrease in the car ownership probability. Additionally, we find that the share of the total BE influence (including the RSS influence) owing to RSS on the householder's car ownership is 45.7%. The proposed model is a new and useful tool for quantifying the influence of BE and the relative influence of RSS on binary travel behavior.