ABSTRACT. *** Please open the enclosed file to view figures and tables. ***

Introduction

Logit and probit are by far the most widely used discrete choice models. However, both models are constrained by strong parametric assumptions. Whilst the Gaussian distribution has a symmetric bell shape with light tails, the logistic distribution is also symmetric and exhibits slightly heavier tails than the Gaussian distribution. As a consequence, both logit and probit lack robustness to outliers in the response data. A robust model is insensitive to outlying observations (Gelman et al., 2013). Whereas an outlier in continuous data is an extreme data point, an outlier in discrete data is an observation that is unexpected through the lens of a non-robust model (Gelman and Hill, 2006). In discrete choice analysis, response data may be contaminated with outliers due to three reasons, namely i) aberrant behaviour, ii) misreporting, iii) misclassification. Ignoring outliers in logit and probit applications can result in biased and inconsistent parameter estimates (Hausman et al., 1998).

Method

In this study, we analyse two robust alternatives to the multinomial probit (MNP) model. Both alternatives belong to the family of robit models whose kernel error distributions are heavy-tailed t-distributions which moderate the influence of outlying observations. In the first model, the multinomial robit (MNR) model, a generic degrees of freedom (DOF) parameter controls the heavy-tailedness of the kernel error distribution. The second model, the generalised multinomial robit (Gen-MNR), is more flexible than MNR, as it allows for different marginal heavy-tailedness of the kernel error distribution. For both models, we derive efficient Gibbs sampling schemes, which also allow for a straightforward inclusion of random parameters.

Results

We first use simulated data to investigate the properties of the proposed models and their estimation methods in terms of parameter recovery and elasticity estimates. We are able illustrate the excellent finite sample properties of the proposed Bayes estimators and show that MNR and Gen-MNR produce more exact elasticity estimates when the choice data contain outliers through the lens of the non- robust MNP model.

We also compare MNP, MNR and Gen-MNR in a case study on transport mode choice in London, UK using a large revealed preference data set (Hillel et al., 2018). M-MNR and M-Gen-MNR markedly outperform M-MNP in terms of in-sample fit and also exhibit superior out- of-sample predictive accuracy. With just one additional parameter, M-MNR outperforms M-MNP by more than 100 log-likelihood points on the training data. M-Gen-MNR exhibits better in-sample fit and out-of-sample predictive accuracy than M-MNR due to its more flexible kernel error distribution. We also find that the differences in in-sample fit are reflected in different elasticity estimates.

Conclusion

On the whole, our analysis suggests that Gen-MNR is a useful addition to the choice modeller’s toolbox due to its robustness properties. In general, Gen-MNR should be preferred over MNR model because of its more flexible kernel error distribution.

References

Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., and Rubin, D. B. (2013). Bayesian data analysis. CRC press. Gelman, A. and Hill, J. (2006). Data analysis using regression and multilevel/hierarchical models. Cambridge university press. Hausman, J. A., Abrevaya, J., and Scott-Morton, F. M. (1998). Misclassification of the dependent variable in a discrete-response setting. Journal of econometrics, 87(2):239–269. Hillel, T., Elshafie, M. Z., and Jin, Y. (2018). Recreating passenger mode choice-sets for transport simulation: A case study of london, uk. Proceedings of the Institution of Civil Engineers-Smart Infrastructure and Construction, 171(1):29–42.

ABSTRACT. INTRODUCTION: Quantum probability, first developed in theoretical physics, has recently made the transition into cognitive psychology, where it is has been used to explain the impact of question order, fallacies in decision-making, and other effects that were previously difficult to explain using classical models (Pothos and Busemeyer, 2021). At the previous ICMC, we demonstrated that quantum probability theory could also make the transition into choice modelling and be used to explain moral choice behaviour (Hancock et al., 2020b). The results revealed that the quantum choice models can efficiently capture the effect of a change in choice context through ‘quantum rotations’ in stated preference choice settings. However, the applicability of the methods in real-world settings, where the context effects are more ‘fuzzy’ and difficult to capture in the data, remains unclear. In this paper, we aim to address the research gap by testing the quantum choice models on revealed preference (RP) data from two natural experiments involving behavioural interventions/nudges potentially leading to shifts in the choice context. The paper aims to test whether quantum choice models better capture behavioural changes arising from interventions, and whether they lead to similar or different behavioural insights in comparison to typical modelling approaches.

DATA: The first dataset tested is from the ‘Decisions’ survey (Calastri et al., 2020). In this survey, 273 participants completed a 2-week travel diary with the use of Rmove. At the end of the first week, some participants are given feedback as to whether they use more or less CO2 and burn more or less calories than other participants with similar demographics. Thus ‘quantum rotations’ will aim to capture the ‘nudge’ towards making greener choices. Initial tests using multinomial logit models imply that feedback reinforces behaviour, as opposed to changing behaviour. There is a shift towards choosing car if an individuals is told they use more CO2 than others, whereas there is a shift away from choosing car if an individual is told that they burn more calories than others. In the second dataset tested, 43 households (71 individuals) completed a 3-week travel diary where they `experience AV’ through having a chauffeur for the 2nd week (Harb et al., 2018). In this case, quantum rotations will be used to represent the shift in preference towards choosing car when the decision-maker has a chauffeur. Mixed logit models suggest that decision-makers have a substantially lower value of travel time (in car) during the chauffeur week (Harb et al., 2022).

MODEL FRAMEWORK: Under quantum choice models, each alternative is represented by an orthonormal vector in a ‘Hilbert space’ (similar to axes in Euclidean space, but with complex numbers and as many dimensions as there are alternatives). The preferences of the individual (e.g. to choose car, bus or train in the above case studies, if these are the available alternatives) are contained within a ‘state’ vector, which is a superposition of the alternative vectors. By ‘choosing’ an alternative, the decision-maker’s ‘state’ vector aligns with the vector representing the chosen alternative through a projection from the state vector. As the state vector is normalised to be of unit length and the vectors for the different alternatives are orthonormal, the set of squared projection lengths must sum to one, meaning that the probability for choosing the alternative can be set to its squared projection length (for a visual representation of this, see Hancock et al., 2020a). Consequently, we can build a quantum choice model simply by defining a normalised state vector, which, like a utility model, is based on the attributes of the alternatives. The state vector components (with respect to each alternative vector) can be based on utility function differences (e.g. regret functions work particularly well, Hancock et al., 2020a), as this results in larger projection lengths for alternatives with favourable attributes, which increases the probability of these alternatives being chosen. Under quantum choice models, the impact of a nudge can be captured by a ‘quantum rotation,’ which shifts the state vector. Preferences for two separate tasks/actions are represented by the same state vector (i.e. we would have the same state vector representing preferences/the propensity to choose which mode to travel to work, and to represent a decision to act more environmentally friendly). Vectors representing alternatives in the first task are not the same as vectors representing alternatives in the second, and are not orthonormal, meaning that making one decision (and thus `projecting’ onto the vector represented by the first chosen alternative), will shift the decision-maker’s state vector, but ultimately mean that they can still choose any of the alternatives in the second task. Consequently, the probabilities of choosing the different alternatives in the second task will change (i.e. trying to be more environmentally friendly will reduce the likelihood of choosing to drive). This concept of ‘entanglement’ is the key factor in driving the improvement found by adopting quantum models within case studies on ordering effects in cognitive psychology. In the current context, behavioural nudges are represented mathematically through the use of a quantum rotation to the state vector prior to the decision-maker choosing an alternative. For example, an individual may implicitly answer the question ‘Am I an environmentally friendly person?’ if they are nudged towards making greener choices. The more environmentally conscious the individual, the larger the rotation and consequential shift towards greener alternatives.

NEXT STEPS: Preliminary results for multinomial logit and mixed logit models imply significant behavioural change in the tested datasets, suggesting there is significant scope for quantum choice models also finding effects. Models will be compared on the basis of model fit and the different (or possibly similar) behavioural insights that they generate (i.e. who is more likely to change their behaviour).

ABSTRACT. Two broad approaches in the literature to model multiple discrete-count data include the multivariate count model approach and the joint single discrete choice/count model approach. Examples of the first multivariate count model approach include multivariate versions of the Poisson or negative binomial (NB) discrete distributions (such as the negative multinomial, but these only allow positive correlations in the counts), and those based directly on the multinomial distribution (such as the multinomial, Dirichlet-Multinomial, and random-clumped multinomial that allow only negative correlation in the counts). These multivariate count models have the advantage of a closed form, but become cumbersome as the level of multivariateness increases and/or as one superimposes additional mixing structures to relax the strict one-sided dependence between the counts. Further, these multivariate count approaches are not based on an underlying utility-maximizing framework, rendering them unsuitable for economic welfare analysis, an important consideration in many empirical contexts. The second approach, joint single discrete choice-count models, uses a combination of a total count model and a single discrete choice model for each event instance (the number of event instances constitutes the total count; also, by single discrete choice, we refer to the situation where, at each of the event instances, the individual is assumed to select one and only one alternative from the available set of event outcome categories). The single discrete choice is then linked to the total count, typically through the introduction of the expected value of the maximum utility from the single discrete choice model into the conditional expectation for the total count random variable. Noting that this structure is still inconsistent with utility theory, Bhat et al. (2015) proposed another linking function that considers the full distribution of the stochastic maximum utility from the single discrete choice model for introduction into the count model. But all these linking-based implementations, like those in the first multivariate count model approach, are based on the premise that the multivariate counts originate from the “vertical accumulation” of single discrete choice decisions at each of the many event instances making up the total count.

The current paper uses a fundamentally different perspective to multiple discrete-count modeling relative to the extant approaches in the literature. In particular, our new approach views the multivariate counts as arising from a combination of a total count decision and a fundamentally simultaneous “horizontal” choice among the multiple alternatives across all the counts, rather than as originating from the “vertical accumulation” of single discrete choice instances. For example, in the recreational literature, the interest may be on the total count of annual fishing trips and the split of this total count to different angler sites, or the total count of annual vacation trips and the split of this total count to different vacation destinations. In such choice situations, the alternatives are imperfect substitutes for one another, with the decision-making based on a simultaneous “horizontal” choice made over the total count. Equivalently, satiation effects set in as the investment in any single alternative over the total count increases. In other words, the reason that individuals partake in multiple activities is because of decreasing marginal utility at higher consumption levels of each alternative.

Based on the above discussion, we formulate a new econometric multiple discrete-count model. There are two components in our proposed model. The first is a multiple discrete-continuous extreme value (MDCEV) model where the discrete component corresponds to whether or not an individual chooses (consumes) a specific alternative over the total count, and the continuous component refers to the fractional split of counts in each of the consumed alternatives over the total count. The second component is a total count model, framed within a generalized ordered response (GOR) framework, with a linking function from the MDCEV model utility function appearing in the count model. Specifically, we start from Bhat’s (2008) MDCEV utility formulation for the case when only inside goods are present (with the budget of the fractions across all alternatives being equal to one), but adopt a reverse Gumbel distributional assumption for the stochastic terms in the baseline preferences of each of the alternatives. We also consider a reverse Gumbel distribution for the random error term within the GOR framework of the total count model. With these assumptions, the result is an incredibly simple, utility-theoretic, closed-form multiple discrete-count extreme value (MDCNTEV) model. Unobserved heterogeneity may be introduced in a straightforward way through mixing distributions.

An application of the MDCNTEV model is demonstrated in the context of individuals’ count of recreational episodes to each of multiple destination locations, using data drawn from the 2012 New Zealand Domestic Travel Survey (DTS). The DTS asked respondents to provide information on all long distance leisure trips over 40 kilometers one-way made four weeks prior to the survey date. For the analysis, New Zealand is partitioned into 16 aggregate destination regions, based on the classification used by the New Zealand Department of Tourism for its marketing campaigns. From the survey, the total count of long distance trips is obtained for each respondent, as is the count of trips to each of the 16 destination regions. Survey records are supplemented with a network level of service file and a disaggregate spatial land-cover database. The final data sample used in estimation includes 3508 individuals. The independent variables include the logarithm of the area of each region (a size variable), land-cover effects in each destination region, travel time and cost to each destination region, an overall destination-specific diversity index, socio-demographic variables, and a dummy variable for the presence of a ferry ride (needed if travel entails going from one island to the other). We expect the results to provide important information for effective targeting and strategic positioning to increase destination competitiveness. More generally, the MDCNTEV formulation should be a valuable methodology for marketing and positioning in the many consumer product/service markets that are characterized by multiple discrete-counts.

Reference Bhat, C.R., Paleti, R., and Castro, M. (2015). A New Utility-Consistent Econometric Approach to Multivariate Count Data Modeling. Journal of Applied Econometrics, 30(5), 806-825.

ABSTRACT. Many discrete choice models in the transportation literature have addressed issues that have occurred due to unobserved variables. In particular, many studies have attempted to overcome the two major methodological challenges to address (1) correlations between choice alternatives and (2) endogeneity. Two choice alternatives are termed “correlated” if they share unobserved variables. Additionally, if unobserved variables are correlated with an observed explanatory variable, the corresponding parameter is estimated inconsistently, which is referred to as the endogeneity problem (Guevara, 2015; Guevara and Polanco, 2016). On the contrary, ignoring these correlations and endogeneity may produce biased results in activity-travel behavior analysis and travel demand forecasting. Therefore, both the correlations and endogeneity caused by unobserved variables must be adequately addressed. In practice, however, the problems of correlations and endogeneity are rarely addressed simultaneously. Although they are caused by a common factor (i.e., unobserved variables), their solutions have been developed separately. For example, generalized extreme value models and a multinomial probit model, which can handle correlations between choice alternatives, cannot address the problem of endogeneity. Likewise, instrumental variable models (e.g., control-function method) and multiple indicator solutions, which are the typical solutions for endogeneity, generally assume independent choice alternatives for simplicity. Thus, to the best of our knowledge, there is no existing method to handle both problems. In this study, we propose a novel method that simultaneously handles correlations and endogeneity. Specifically, we propose a multinomial probit model that incorporates the instrumental variable method. More generally, the proposed model may be called an instrumental variable model for a multinomial outcome. It has the following practical characteristics: (1) it allows binary and/or continuous endogenous variables, (2) it allows any number of instrumental variables in each alternative, and (3) it allows positive and/or negative correlations between any choice alternatives. To the best of our knowledge, the proposed model is the first attempt at addressing the problem of endogeneity in a discrete choice model, while allowing correlations between choice alternatives. It should be noted that the probit models that incorporate the instrumental variable methods may have issues with the identification of a parameter that is essential for addressing the endogeneity problem. These issues may be severe when endogenous variables are binary. In such cases, Freedman and Sekhon (2010) pointed out that traditional maximum likelihood estimation methods cannot find the global maximum due to the existence of flat spots on the log-likelihood surface. Accordingly, in this study, we employ a Bayesian approach for parameter estimation to deal with potential parameter identification issues that are reported in related studies. We also propose a Markov chain Monte Carlo (MCMC) algorithm tailored to the proposed model. To check the properties of the proposed model, we conducted a simulation study. We generated 3,000 samples from a single setting and created two types of simulation data in which endogenous variables were continuous and binary. Then, we estimated the proposed model using the simulated data and examined whether the model can address endogeneity while allowing correlations between choice alternatives. Additionally, we conducted a sensitivity analysis to examine the effect of different prior settings on the parameter estimates in the proposed model. This simulation showed that the proposed model can adequately address endogeneity while allowing correlations between choice alternatives when endogenous variables are continuous. However, when endogenous variables are binary, the estimates are relatively far from the true values. Additionally, the prior sensitivity analysis indicated that the estimates are highly dependent on the prior settings when endogenous variables are binary. This may be due to a flat spot on the log-likelihood surface. Thus, it is necessary to pay special attention to the settings of the prior distributions when endogenous variables are binary. Based on the results of our analysis, we recommend that analysts conduct a prior sensitivity analysis and evaluate the prior settings based on posterior predictive checks, following a general Bayesian modeling workflow (van de Schoot et al., 2021). To conclude, we believe that the three characteristics of the proposed model mentioned above will be helpful in real-world applications. For instance, even though some existing methods only assume continuous endogenous variables, endogenous variables of interest are not necessarily continuous in practice. Moreover, strictly speaking, many (or all) explanatory variables can be endogenous variables in many empirical cases. Thus, it is generally recommended to include more instrumental variables in a model specification. To this end, the proposed model can be a useful tool because it can incorporate instrumental variables correlated with any number of binary and/or continuous endogenous explanatory variables. Additionally, in recent empirical analyses using choice models, allowing correlations between choice alternatives has become fairly common. Thus, it is desirable not to assume independent choice alternatives even if the research objective is to address endogeneity. Therefore, the proposed model will play an important role in disciplines in which both endogeneity and correlations between choice alternatives are major concerns.

ABSTRACT. Endogeneity arises when one or more of the explanatory variables of an econometric model are correlated with the model's error term. This may be caused by measurement/specification errors, omitted attributes and self-selection among other reasons (Guevara, 2015). The main problem of models affected by endogeneity is that their parameters may be inconsistent, leading to faulty policy analysis, forecasts, conclusions and/or behavioural assessments (Guevara and Ben-Akiva 2006). The correction of endogeneity has been studied in depth in the case of linear models, where the problem has been considered from several viewpoints (Stock and Yogo, 2005; Ebbes et al., 2011). In the case of discrete choice models (DCM), which are highly non-linear, some research gaps have been closed in recent years (Guevara and Polanco, 2016), and work by Guerrero et al. (2020), Wen and Chen (2017), Lurkin et al. (2017) and Mumbower et al. (2014), have successfully applied practical corrections for endogeneity in the transport field. Notwithstanding, several open questions still remain. There are several methods available to correct for endogeneity in DCM, such as Berry-Levinsohn-Pakes - BLP (Berry et al., 1995), Latent Variables - LV (Ben-Akiva et al., 2002), Maximum Likelihood (Park and Gupta, 2009), CF (Petrin and Train, 2010), Proxies (Guevara, 2015), and Multiple Indicator Solution - MIS (Guevara and Polanco, 2016). Louviere et al. (2005) provided an extensive compendium of advances in the treatment of endogeneity in DCM. The decision to use one or another method is based on what information is available, the assumptions the researcher is willing to make, and the associated computational costs (Guevara, 2015). We consider the MIS method for correcting endogeneity in DCM. The MIS method is based on the use of indicators, which often come from surveys designed for knowing respondents’ attitudes and/or perceptions about their decision making (Bahamonde-Birke et al., 2017). Indicators have been typically collected using Likert (1932) or verbal scales (Glerum et al., 2014), and they can be attitudinal because these represent the characteristic of the individuals toward life (Walker and Ben-Akiva, 2002; Daly et al., 2012; Bahamonde-Birke et al., 2017) or perceptual because they are exclusively related to the way certain alternatives are perceived, namely, they are intrinsically associated with an alternative (Bolduc and Daziano, 2009; Yáñez et al., 2010; Raveau et al., 2010). Guevara (2015) shows some advantages of the MIS method, such as the easiness of its applicability in practice and the fact that it is based on the use of indicators instead of instrumental variables, as in another popular method (CF). In several situations, indicators may be easier to obtain compared to instrumental variables. While indicators can be directly collected by the researcher, instrumental variables have to be gathered from existing information and need to satisfy two conflicting criteria: (i) to be exogenous (independent of the error term of the model) and (ii) to be relevant (strongly correlated with the endogenous variable). In practice, having proper instruments may be impossible and if the instruments are endogenous or weak (i.e., not sufficiently strongly correlated with the endogenous variable), their use can yield results as bad as when using an endogenous model (see, for example, Guevara, 2018; Guerrero et al., 2020). However, the relative easiness in the collection of indicators for the MIS comes with a caveat. In theory, to apply the MIS method to correct for endogeneity in DCM, the indicators must be continuous (Guevara, 2015), since this is a mathematical requirement for its derivation (Wooldridge, 2010), and this may be hard to fulfil. The problem is that, in practice, the indicators tend to be discrete since they are typically obtained through Likert scales. Although there is some preliminary empirical evidence suggesting that discrete indicators can be as good as continuous ones for correcting endogeneity with the MIS method in DCM (Guevara and Polanco, 2016; Fernández-Antolín et al., 2016), it should be clear that this is only an approximation (Guevara, 2015). So, this part of the research will focus on characterizing the impact of using discrete indicators to correct for endogeneity with the MIS method in DCM. For this, we will correct DCM, with discrete and continuous indicators, using both real and simulated data, and compare the results to determine the possible impacts and differences of each approach. This will allow identifying some of the conditions under which discrete indicators could work adequately, providing support or refuting existing empirical evidence. To do our tests we will use a purposely designed stated preference (SP) survey and Monte Carlo simulation. The SP survey was designed for the context of departure time choice, considering the main explanatory variables in this modelling context; that is, travel time, cost, variability of travel time and schedule delay (Arellana et al., 2012), where the last variable followed the scheduling model of Small (1982). The Monte Carlo experiments were designed to test the effect of: (i) different criteria for the discretization of latent continuous indicators, (ii) the sample size and (iii) the distribution of the indicator. Conclusions are based on the findings from the simulated and real data. From the Monte Carlo experiments, it was possible to show that the small sample sizes can lead to erroneous conclusions. Second, the most straightforward criteria to produce discrete indicators is rounding the continuous value to the nearest integer and assigning it to percentiles that worked properly. Similarly, we used four algorithms reported in the literature to discretize continuous indicators. We used these algorithms because they are typically used in computer science. Our findings show that the algorithm used to produce discrete indicators affects the endogeneity correction. Using real data, we showed that the correction with continuous indicators worked better than the correction with discrete indicators. In particular, we checked the parameter ratios, finding that the subjective value of time (SVT) for the model corrected with continuous indicators was closer to that of the Benchmark model than those computed from the model corrected with discrete indicators. Notwithstanding, the SVT achieved with the endogenous model was much worse, having a bias of approximately 250% concerning the benchmark model.

ABSTRACT. A major concern when designing discrete choice experiments (DCEs) to elicit consumer preferences is ensuring that respondents are presented with scenarios and choice sets that are both as realistic and relevant to the situation the respondent is likely to face in the real market as possible (or “verisimilitude”—see, e.g., Ben-Akiva, McFadden and Train (2019)). A specific example is development of vehicle choice models for forecasting consumer response to the introduction of new alternative fuel vehicles (AFVs, e.g., plug-in vehicles, hydrogen fuel cell vehicles, etc.). The challenges of using DCEs in this domain are the motivation for this work.

The vehicle market includes a very large number of offerings that vary widely along multiple dimensions, including basic functionality (passenger car, truck, van, SUV/crossover), size, prestige level, and most recently, fuel technology type. Moreover, projecting market evolution over time for purposes of policy analysis requires understanding demand for both new and used vehicles (varying by vintage). For these reasons, researchers have long relied on customization procedures to design vehicle choice DCEs. For example, if a respondent expects their next vehicle purchase to almost certainly be some type of new pickup truck, offering them lots of choices involving used sedans and SUVs would be both unrealistic (from their perspective) and statistically inefficient. Researchers typically elicit this type of information from the respondent and use it to customize choice sets that ensure verisimilitude. See, e.g., Bunch et al. (1993) and citations.

However, this customization implies that choice sets are being generated as a function of preference-related information, i.e., the choice sets are endogenously determined. This potentially compromises the statistical properties of choice model estimates. Our experience suggests that this has been an issue of concern for a long time, yet we are unaware of any theory-based efforts to systematically investigate and address it. (Some researchers have used so-called inertia variables as an attempted corrective measure. However, this approach is ad hoc with unknown properties.)

Another important transportation related DCE application area with endogeneity issues is mode/route choice. In contrast to vehicle choice, this has been investigated by, e.g., Train and Wilson (2008) [hereafter, TW2008] and more recently, Guevara and Hess (2019). Our approach to addressing endogeneity for the vehicle choice case is based on applying and extending results for the “SP-off-RP” case of Train and Wilson (2008).

SP-off-RP is used for the case of mode choice, where DCE choice sets are based on details from a recent real-world trip made by the respondent to ensure verisimilitude (and which induces endogeneity). A key similarity between SP-off-RP and vehicle choice is that there is generally a “label structure” used to identify specific alternatives, but also a profile of attributes that can be manipulated. However, there are important differences. Mode choice uses a small number of labels (e.g., private car, train, bus), and the same labels from the RP trip are typically used in the DCE. Moreover, because attributes for the RP trip are provided by the respondent, they will vary from respondent to respondent. Finally, investigating response to totally new modes of travel is not typically the purpose of SP-off-RP, which often focuses on obtaining measurements of value of travel time savings important to policy makers. (Although there are studies to evaluate new modes, e.g., high-speed rail.)

In contrast, vehicle markets have a very large number of labeled alternatives that are complex and multidimensional (e.g., a “new battery-electric small SUV”). Moreover the “RP choice context” is usually not a recent purchase, but the next intended purchase. Despite notable differences, we apply and extend the TW2008 approach to this case, taking advantage of, e.g., ongoing improvements in model estimation software. This study uses data from the series of California Vehicle Surveys (CVSs) administered by the California Energy Commission (CEC). Information is elicited from respondents about their next intended vehicle purchase and used to create 8 customized choice sets (with four vehicle attribute profiles each). Following usual practice, CEC uses logit-based choice models estimated on DCE data only, which are vulnerable to endogeneity.

The approach performs a joint RP-SP estimation that captures endogeneity effects by directly implementing the random utility maximization error structure (in this case, using a logit kernel). The likelihood for each respondent is an integral that is evaluated by simulating vectors of RUM disturbance terms conditional on the RP choice. These are then carried over and used as (otherwise unobservable) explanatory variables in the DCE choice probability expressions. So, even a simple logit model requires simulated likelihoods. TW2008 addresses this case, but also the random coefficients (mixed logit) case. Because of the multidimensional nature of vehicle labels, an error-components approach of this type is virtually a requirement for our analysis. Taken together, the overall approach simultaneously addresses multiple issues beyond endogeneity, including: correcting the relatively scaling of DCE responses, capturing correlated preference patterns among vehicle types, and capturing unobserved heterogeneity across respondents (which also addresses the lack of independence across multiple DCE tasks). By addressing these, this study addresses a number of gaps in the literature for this application area.

References

M. Ben-Akiva, D. McFadden and K. Train (2019), “Foundations of Stated Preference Elicitation: Consumer Behavior and Choice-based Conjoint Analysis”, Foundations and Trends in Econometrics: Vol. 10, No. 1-2, pp 1–144. DOI: 10.1561/0800000036.

Bunch, D. S., M. Bradley, T. F. Golob, R. Kitamura, G. P. Occhiuzzo (1993), "Demand for Clean-Fuel Vehicles in California: A Discrete-Choice Stated Preference Survey" Transportation Research A, 27A(3): 237-253.

Guevara, C. A. and S. Hess (2019), “A control-function approach to correct for endogeneity in discrete choice models estimated on SP-off-RP data and contrasts with an earlier FIML approach by Train & Wilson,” Transportation Research Part B 123, 224–239, https://doi.org/10.1016/j.trb.2019.03.022.

Train, K., W.W. Wilson (2008), “Estimation on stated-preference experiments constructed from revealed-preference choices,” Transportation Research Part B 42, 191–203, doi:10.1016/j.trb.2007.04.012.

ABSTRACT. One of several factors shown to potentially affect the validity and reliability of stated preferences is the type and amount of information presented to survey respondents before preference elicitation. The influence of variations in information about the valuation scenario on stated preferences has been researched for many years (e.g., Munro and Hanley 2001), but the mechanisms underlying these information effects are less understood. This study contributes to this line of research by proposing and empirically analyzing two potential, unstudied pathways explaining how the provision of information about the valuation scenario may affect stated preferences.

The first examined pathway is perceived survey consequentiality. Changes in the information provision may change respondents’ perceptions of the survey being consequential (i.e., potentially affecting future policy decisions), which, in turn, can matter for stated preferences. Numerous studies have provided evidence on the influence of perceived consequentiality on stated preferences (e.g., Vossler and Watson 2013), but none have researched the effects of information about the valuation scenario on perceived consequentiality and hence on stated preferences. The second investigated pathway is perceived scenario credibility. Providing information about the valuation scenario may affect respondents’ perceptions of its credibility (i.e., perceived feasibility to implement the considered scenario), and this in turn can influence stated preferences (the latter effect has been demonstrated by Kataria et al. (2012)). If information provision affects perceptions of consequentiality and/or credibility, and if these shifts in the perceptions affect stated preferences, these pathways offer new insights into how information effects emerge in stated preference surveys.

To examine these potential pathways of information effects, this study utilizes data from a discrete choice experiment survey on urban climate adaptation. The survey was conducted in April and May 2019 with 1,276 residents of the city of Bremen in Germany. It elicits respondents’ preferences towards extending urban green as a climate change adaptation measure. We design and randomly assign two survey versions that differ only with respect to the information about the valuation scenario displayed before eliciting preferences. While both versions provide necessary information for understanding the scenario, one version presents additional information. The additional information describes an existing policy document of the city of Bremen containing potential climate adaptation measures and explains that the valuation scenario consists of a selection of these measures. This information emphasizes the link between the survey responses and actual policies and makes the scenario appear more feasible to be implemented, which may strengthen consequentiality and credibility perceptions, respectively. To study these possible effects, we estimate mixed logit models and hybrid choice mixed logit models. The latter include the perceptions as latent variables to simultaneously estimate the effects of the additional information on perceived consequentiality and credibility as well as on stated preferences. We analyze how the additional information affects stated preferences, how it shifts the perceptions, and how these perceptions matter for stated preferences.

Our results confirm that the information effects frequently found for a range of goods also emerge in the context of urban green and climate change adaptation. Being provided with additional information about the climate adaptation context of the valuation scenario increases willingness to pay for the urban green measures. To investigate whether information-induced shifts in consequentiality and credibility perceptions may explain this effect, we look closely into two parts of these pathways. First, our data suggests that the additional information strengthens the credibility perceptions, while its effect on the consequentiality perceptions is not statistically significant. Second, we observe that stronger perceptions of both consequentiality and credibility correspond to larger willingness to pay. These results indicate that the information-induced shift in the perceived credibility may explain part of the information effect. In contrast, because the effect of the information script on the consequentiality perceptions is small and insignificant, this pathway appears unlikely to explain a meaningful portion of the information effect.

The study provides a novel explanation of information effects: part of the effects may emerge as a result of information-induced shifts in the perceptions of credibility of the valuation scenario. Our application also demonstrates the use of the hybrid choice framework for studying mechanisms underlying information effects. Understanding the mechanisms that lead information effects to emerge is essential for correctly designing stated preference questionnaires and for obtaining valid value estimates to support public decision-making. The pathways proposed here appear particularly important, as both consequentiality and credibility are acknowledged as desired characteristics of stated preference surveys. The perceived credibility of the scenario has been little studied, and there is little advice on how it can be reinforced. Our findings show that additional information on the policy context and mechanism of the policy change in the valuation scenario can strengthen the credibility perceptions. Further, our results indicate that strengthening credibility perceptions with information scripts may affect stated preferences. This suggests that when including additional information designed to improve credibility, changes in value estimates can be a desirable consequence rather than a sign of bias.

References:

Kataria, M., I. Bateman, T. Christensen, A. Dubgaard, B. Hasler, S. Hime, J. Ladenburg, G. Levin, L. Martinsen, C. Nissen. 2012. “Scenario Realism and Welfare Estimates in Choice Experiments – A Non-Market Valuation Study on the European Water Framework Directive.” Journal of Environmental Management 94(1): 25–33.

Munro, A., N.D. Hanley. 2001. “Information, Uncertainty, and Contingent Valuation.” In: Valuing Environmental Preferences, 258–79. Oxford University Press.

Vossler, C., S.B. Watson. 2013. “Understanding the Consequences of Consequentiality: Testing the Validity of Stated Preferences in the Field.” Journal of Economic Behavior & Organization 86: 137–47.

ABSTRACT. Stated preference (SP) surveys should ideally present scenarios and choice settings that stimulate respondents to answer the choice tasks truthfully. Failure to ensure such incentive compatibility may lead to a range of undesirable response behaviors causing hypothetical bias in Willingness-To-Pay (WTP) estimates (Carson et al., 2014). An important survey design feature in this regard is so-called consequentiality, which refers to a situation where respondents perceive that 1) their answers to the survey are likely to affect the policy decision in question; 2) the described policy will lead to the described environmental changes, and 3) they will have to pay the stated costs if the policy is realized (Carson et al., 2014; Herriges et al., 2010; Vossler et al., 2012).

Several papers have investigated the role of consequentiality in SP surveys. One empirical approach concerns controlling for variations in respondents’ perceived consequentiality elicited through follow-up questions, which are subsequently incorporated into econometric models to test for impacts on WTP estimates. For instance, studying the relationships between respondents’ perceived level of consequentiality and their WTP, Herriges et al. (2010) found that the WTP distributions were not equal between those respondents who perceived their responses to be consequential and those who did not. Similar results are echoed by Groothuis et al. (2017). However, other authors did not find impacts of perceived consequentiality on WTP estimates (e.g. Lloyd-Smith et al., 2019; Oehlmann & Meyerhoff, 2017). Another approach focuses on ways to induce consequentiality, e.g. through various forms of consequentiality scripts (e.g. Czajkowski et al., 2017; Herriges et al., 2010; Lloyd-Smith et al., 2019; Oehlmann & Meyerhoff, 2017). Findings are ambiguous, though with most studies finding limited or no impact on WTP.

We contribute to the literature on consequentiality by testing three different consequentiality scripts, designed to induce different degrees of consequentiality. We use data obtained from an online stated choice experiment (CE) aimed at eliciting Danish citizens’ WTP for reducing the negative effects of new motorways on nature and outdoor recreational activities. A treatment-control design with three sample splits is used. In the first sample split, no particular efforts are made to induce consequentiality (Control). Respondents are essentially only instructed to be realistic and answer the questions carefully and honestly in order to ensure that results can be used for research purposes, whereas any linkage to actual decision-making is not mentioned. In the second sample split, respondents are additionally instructed that the researchers will present findings to the Danish Road Directorate, and, thus, it could potentially affect future planning of motorways in Denmark (Standard Consequentiality Treatment). This treatment is largely in line with current guidelines for SP studies (Johnston et al., 2017; Mariel et al., 2021). In the third sample split, along with the invitation to the online survey, respondents also receive a letter from the Danish Road Directorate stating that the results of the research will be instrumental for future planning of motorways in Denmark (Enhanced Consequentiality Treatment).

Respondents were presented with a case-specific scenario describing construction of 180 km of new motorway through nature areas, which the respondents use for outdoor recreational activities. The CE entailed five attributes associated with impacts of the new motorway in nature areas: levels of noise annoyance, amount of nature area converted to motorway, whether negative impacts on rare and endangered animal species will occur, whether mitigation measures to reduce roadkill numbers are introduced, and costs of adjusting the standard motorway layout. The payment vehicle was additional annual income tax, ensuring a high degree of payment consequentiality. Each respondent faced eight choice sets consisting of three alternative routes for a new motorway. A zero-cost status quo alternative was constant across all choice sets. This was described as a previously proposed route by the Danish Road Directorate with standard considerations for avoiding negative impacts on nature. The two experimentally designed alternative routes entailed further considerations for avoiding negative impacts on nature and outdoor recreational activities at the cost of an increase in annual household income tax. Figure 1 shows an example of a choice set. Respondents were randomly selected from a national registry and questionnaires were sent out to their digital postbox via a digital platform so-called “e-boks”. Data was collected from around 3,616 respondents in total, each respondent randomly allocated to one of the three sample splits. Each sample split thus has about 1,200 respondents, providing a very solid basis for comparisons between sample splits.

The data was analyzed using random parameter logit model based on WTP space specification allowing for correlation between the random parameters. All quality attributes were specified to follow a normal distribution, while cost was specified as lognormal. Model fit is generally quite satisfactory, and all attributes are significant and of the expected sign. More importantly, our preliminary results suggest that the WTP estimates obtained from the two treatments do not differ significantly from the control. Hence, varying the induced consequentiality appears to have had no effect in our empirical case. This is somewhat surprising given that we vary the degree of consequentiality from what may be considered a very low level of consequentiality to what may be considered a very high level of consequentiality compared to previous environmental CE studies. While we of course cannot generalize based on a single study, this finding is in line with, and arguably expands on, the studies mentioned above finding no effect of measures to induce consequentiality. We refrain from speculating whether this is because hypothetical bias has not been an issue in these empirical studies in the first place – in which case we would not expect an effect of such scripts – or it rather implies that consequentiality scripts simply do not work.

ABSTRACT. Choice experiments (CE) and other stated preference methods are known for eliciting willingness to pay (WTP) estimates that are larger than payments actually made, a problem known as the hypothetical bias (HB). HB is a crucial problem in CE, because it limits their external validity and potentially undermines reliability of WTP estimates from CE for their use in policy analysis. One way to decease HP in CE is to increase its payment consequentiality by making the CE real. Such ‘real CE’ approach (also termed non-hypothetical, incentivized or consequential CE) makes one randomly selected choice in the CE setting binding - i.e. the selected choice situation is realised and the payment/contribution is actually made. This approach was shown in the existing literature (e.g. Liebe et al., 2019; Moser, Raffaelli and Notaro, 2014) to decrease the WTP in contrast to hypothetical settings and hence decrease the HB. While real CE show promising results, their limitation is their significant cost and difficulty of applying real CE in non-market situations valuing environmental public goods where it is more difficult to devise a realistic payment mechanism.

In this study, we aim to adress some of these problems by exploring to what degree can CE with a probabilistic payment consequentiality (CE with probability 0<p<1 of the situation becoming binding) have similar effects on HB as real CE and how the two methods compare to conventional hypothetical CE. Existing theories on risk perception and previous empirical literature have shown that people are generally insufficiently sensitive to order risk well, often leading people overestimating low probabilities. This suggests that there is space to achieve (some of) the effects of real CE with just probabilistic CE, for significantly lower costs. In addition, we demonstrate how real CE and their probabilistic derivatives can be conducted in valuations of non-market environmental public goods where no realistic payment mechanism generally exists. Our study is novel in its testing of the probabilistic payment consequentiality with significant financial incentives for the respondents, implemented in a environmental public good setting online, and the study extends previous applications which were predominantly either implemented in a lab setting, focused on private market goods (e.g. food products) and/or were incentivised with low financial rewards (e.g. few dollars).

We adopt a between-subject randomized experiment study design which compares effects of various types of consequentiality on WTP. An online CE survey concerning ecosystem restoration in the Czech Republic will be administered online to a representative sample of Czech respondents in early 2022. Respondents will be informed that they would receive, with different probability, an additional reward of 800 CZK (approx. 31 EUR) for the participation in the survey and that they can divide this amount between donation to the government agency nature restoration fund and what they would receive at the end of the survey. Four treatments were designed with probability of the reward (i.e. the DCE being ‘real’): (i) approx. 400 respondents with p=0 (i.e. hypothetical DCE setting); (ii) approx. 200 respondents with p=0.1 (10% to win the lottery for potential donation); (iii) approx. 200 respondents with p=0.4 (40% to win the lottery for potential donation); (iv) approx. 200 respondents with p=1 (i.e. the real DCE setting).

The study will be pre-registered in order to adhere to the best possible standards in data analysis and scientific conduct in general. The analysis of the choices will employ standard multinomial logit, as a base model, and mixed logit models estimated in willingness to pay space in order to facilitate comparisons across treatment groups. Alternatively, a pooled mixed logit model estimated in willingness to pay space would be also estimated with a priori assumptions about the effects of the treatments on marginal WTPs (similar to e.g. Johnston et al. 2013 Land Econ).

In terms of anticipated results (the data will be collected in Jan/Feb 2022), we will test the following hypotheses. First, we are interested whether the Real DCE will indeed decrease the WTP in contrast to hypothetical DCE, in line with the majority of similar applications. We anticipate this as very likely, potentially confirming the usefulness of real DCE for decreasing the HB in DCE. Second, we aim to test the relationship between the estimated WTP and probability of the DCE being binding (i.e. the probability of winning the additional award and hence possibility to donate to ecosystem restoration) and whether this relationship is linear. We are particularly interested to see whether even low probability of DCE being binding produces significant decrease in hypothetical bias – this would suggest that there is potentially a cost-effective strategy to improve the reliability of DCE for policy. Finally, the results of the present study will provide relatively rare evidence about the values (to different degree free of hypothetical bias) that the public holds for ecosystem restoration which can be of particular use to inform national policy in light of EU biodiversity strategy 2030 and UN’s decade on ecosystem restoration and new EU Biodiversity Strategy 2030.

ABSTRACT. 1 Introduction: The Covid-19 pandemic has had an unprecedented impact on human life, and among other things, it has led to drastic changes in mobility patterns, employment statuses and occupation types of people around the world. In the United Kingdom, despite government measures such as the furlough scheme, the unemployment rate raised from 3.8% in the fourth quarter of the year 2019, to a peak of 5.1% in the fourth quarter of the year 2020 [1]. More than 11 million jobs on furlough in the UK by the end of 2020 [2] better shows the enormous impact of the pandemic on the labour market. Aside from workers with regular employment conditions, workers in flexible employment statuses, such as temporary subcontracted workers, gig workers, etc. were affected during the pandemic [3], as they often cannot rely on stable incomes labour protections policies offered to regular employees [4]. Some evidence suggests that the economic impact of the pandemic has been uneven among different ethnic groups [5]. While the unemployment rate among people from White ethnic groups raised from 3.6% in the first quarter of 2020, to 4.0% in the second quarter of 2021, the change in the unemployment rate among minority ethnic groups was much higher, from 6.2% to 8.0% in the same time span [5]. This study aims at inferring the occupational shifts due to the Covid-19 pandemic and investigating the heterogeneity in these shifts. This can include spatial heterogeneity (i.e., small vs large city) and sociodemographic heterogeneity (e.g., ethnicity, income, etc.). To do so, we utilize passively collected mobility datasets. While most studies and reports on the topic involve survey data, mobility patterns derived from mobile phone data can provide valuable information from larger sections of the population and help reduce the biases involved in traditional surveys [6]. 2 Data: To infer mobility patterns of the population, we use data from Cuebiq Inc. [7], which provides access to aggregated, anonymous, privacy-compliant mobility data for research. Using a Software Development Kit (SDK) included in smartphone applications, location data from users who opted-in is collected from various sensors of the smartphone. Mobility indicators, including trip numbers, distance, destinations, and motifs are derived from the Cuebiq data. At an aggregated level, counts of the number of households/individuals in the neighbourhood level, the average number of locations visiting other than household, number of courier delivery drivers/cyclists, and other occupation categories that are included in the dataset are retrieved. By estimating travel patterns and observing their changes during the pandemic, an aggregated index of a change in travel pattern and occupation choice can be estimated. Several data sources will be used as complementary data sources, to cross-validate the data and enrich it with ethnicity and social attributes. The complementary data sources include: • Labour market data [8]: includes monthly and seasonally updated data on employment in the region and postal code levels. • Modelled Ethnicity Proportions [9]: include estimates on the proportion of different ethnic groups at neighbourhood level from 1997 to 2020. • Census data [10]: includes data on the population of ethnicity groups in postal code level, conducted in 2011. 3 Methodology and expected results: We hypothesize that observing changes in mobility patterns help determine the changes in occupation status, including starting to work from home, shifting to gig-economy, unemployment and increase or decrease in the working hours. For example, a change in the number of daily trips, trip destination or trip distance can be an indicator of unemployment or change of occupation. On the other hand, change in the number of visits to certain points of interest, or change in trip motifs (e.g., from Home-Work-Home to Home-Visit-Visit-Home-Visit-Home) can indicate an increase in working hours in different locations, or employment in a higher risk job (e.g., a gig worker). In general, by analyzing the changed structure of visit point locations and their frequencies (i.e. the shape of the distribution to visit settings), we might detect deviations from the dominant Home-Work pair during the pandemic, which will be an indication of a change in the work status. In an earlier study using the data, changes in household visitations are investigated during the pandemic and the national lockdowns in the UK [11]. Furthermore, an additional step of the analysis would be to analyze the secondary and tertiary visit points, which could have been related to workplaces in the pre-COVID time, but during the pandemic, have been replaced by other points of interest, e.g., grocery stores.

References [1] Office for National Statistics. Labour market statistics time series. https://www.ons.gov.uk/employmentandlabourmarket/peopleinwork/ employmentandemployeetypes/datasets/labourmarketstatistics. Accessed: 2021-11-25. [2] HM Revenue and Customs. Coronavirus job retention scheme statistics. https://www.gov.uk/government/statistics/ coronavirus-job-retention-scheme-statistics-7-october-2021/ coronavirus-job-retention-scheme-statistics-7-october-2021. Accessed: 2021-11-25. [3] Daniel Spurk and Caroline Straub. Flexible employment relationships and careers in times of the covid-19 pandemic, 2020. [4] Bénédicte Apouey, Alexandra Roulet, Isabelle Solal, and Mark Stabile. Gig workers during the covid-19 crisis in france: financial precarity and mental well-being. Journal of urban health, 97(6):776–795, 2020. [5] Andrew Powell and Brigid Francis-Devine. Coronavirus: Impact on the labour market, 2021. [6] Li Shen and Peter R Stopher. Review of gps travel survey and gps dataprocessing methods. Transport Reviews, 34(3):316–334, 2014. [7] Cuebiq inc. https://www.cuebiq.com/. Accessed: 2021-11-25. [8] Nomis. official labour market statistics. https://www.nomisweb.co.uk. Accessed: 2021-11-25. [9] Consumer Data Research Center. Modelled ethnicity proportions (lsoa geography). https://data.cdrc.ac.uk/dataset/ cdrc-modelled-ethnicity-proportions-lsoa-geography. Accessed: 2021-11-25. [10] office for National Statistics. 2011 census. https://www.ons.gov.uk/census/ 2011census. Accessed: 2021-11-25. [11] Ross, Stuart, George Breckenridge, Mengdie Zhuang, and Ed Manley. "Household visitation during the COVID-19 pandemic." Scientific reports 11, no. 1 (2021): 1-11.

ABSTRACT. COVID-19 has severely affected travel behaviour globally. People travelled less often and less far, partly due to the dangers of the virus itself and partly due to measures intended to restrict people’s movement. These substantial changes to travel behaviour may also affect the value of travel time (VTT). This might happen as a result of two possible effects:

1. An intrinsic effect, resulting from people experiencing travelling differently. For example, people might feel less comfortable when traveling due to the risk of getting infected. This reduced comfort might increase the VTT.

2. A composition effect, caused by the fact that different people make different changes to their traveling behaviour as a result of COVID-19. This changes the composition of travellers using a specific mode. The change in composition could have an effect on the (average) VTT of public transport users during (and possibly after) the pandemic compared to before the pandemic.

The VTT is mainly used for infrastructure appraisal. Since infrastructure investments are typically long-term investments, policymakers need to be reasonably confident that the VTT is stable in the future. As mentioned above, COVID-19 can potentially change the VTT. To ensure that the measured VTT can be assumed to be stable in the future, we need to know whether or not COVID-19 has actually affected the VTT and whether or not this effect will be structural in a post-pandemic future. Our research objective is to measure the effect of COVID-19 on the VTT, separating the intrinsic and composition effects. The separation of these effects is important, as whether or not they will be structural depend on different circumstances.

We used a survey with stated preference (SP) choice tasks, which we held at two points in time with a panel design. Both waves consist of SP route choice tasks, where respondents are given the option of choosing between two routes based on travel time and travel costs. The first wave was collected in February and early March 2020. The original goal of this wave was to test SP designs for a larger study. During this time there were no movement-restricting COVID-19 measurements in the Netherlands. This February wave consisted of 184 fully participating respondents, recruited using an existing internet panel. The second wave was collected in September and October 2020, when COVID-19 incidence was rising rapidly in the Netherlands and the government was reintroducing strict measures (after relaxing the initial set of measures introduced in March 2020). The goal of this wave was to determine the effect of COVID-19 on the VTT. First, we invited the 184 February respondents again; 128 of them participated in the second wave as well. These respondents were asked whether they were still making the same trip they described in the first wave. If so, they were asked to consider their most recent version of this trip when doing the route-choice SP-experiment. If not, they were asked to make the choice tasks for the hypothetical situation where they had to make the same trip in October as they made in February. In total, we distinguish four categories: people who used the same travel mode as before, people who stayed at home, people who switched to bicycle, and people who switched to car (the latter three of which were asked to make the hypothetical choice tasks). Next, we screened other members of the internet panel and we recruited 309 additional respondents based on their travel behaviour during and before Covid, so as to fill each category with a sufficient number of respondents for reliable estimations.

The two waves allow us to estimate the difference in VTT between February and October within-individuals. We are able to link the differences in VTT to the changes in travel behaviour in the same period, which allows us to estimate an intrinsic and a composition effect. The intrinsic effect is calculated as the difference in VTT between February and October for the set of respondents that participated in both waves. The composition effect is calculated as the difference in VTT caused by people moving away from public transport, where we multiply the VTT-change for each segment (for example the change for people who stay home now instead of travelling) by how often the behavioural change happens in the population (for example the share of people who travelled with PT before pandemic, but stayed home during the pandemic). The SP choice data is analysed using Mixed Logit choice models. The choice models use functions based on log random-valuation. The VTT parameters are estimated as random variables following a log-normal distribution. We estimate separate choice models for the train and bus, tram and metro.

The parameters of the estimated models are interpretable and effects are in the expected direction and we attain realistic VTT’s. We find only a fairly small intrinsic effect, especially for the train. Perhaps this is the result of the sharp decrease in the number of travellers with PT: as the carriages were relatively empty at the time of measurement, travellers were possibly not so worried about infection risk. We find some evidence of potential composition effects, as people who switched to the car have a higher VTT and people who switched to the bicycle have a lower VTT than people who stayed in PT. However, people who stopped travelling have a roughly similar VTT to people who stayed in PT. Since this last change is by far the most common in the population, the overall composition effect is small as well. In conclusion, our study finds that the VTT of public transport travellers seems to be relatively stable in the face of COVID-19. This would mean that VTT’s attained pre-COVID are roughly similar to expected VTT’s post-COVID and that VTT’s measured during or slight ‘after’ the pandemic can be valid as well. However, our measurement was done during a relatively low-incidence and low-measurement time. Results might change when using measurements taken during lockdowns or other times with strong PT-related COVID measures and a high COVID incidence.

ABSTRACT. Background Reducing the burden of COVID-19 is on people already facing health disparities was among the main national priorities for the COVID-19 vaccine rollout in the United States. Early reports from U.S. states releasing vaccination data by race showed that White residents were being vaccinated at significantly higher rates than Black residents. Public health efforts were targeted to address vaccine hesitancy among Blacks and other minority populations. However, healthcare interventions intended to reduce health disparities that do not reflect the underlying values of individuals in underrepresented populations are unlikely to be successful. The objective of this study was to use choice models identify key factors underlying the disparities in COVID-19 vaccination uptake rates.

The reason for the differing rates of vaccinations is unclear. The disparity could be caused by structural or systematic racism; it could be also be caused by higher vaccine hesitancy among Black and Hispanic communities. It has been recommended that health care providers engage with these communities to overcome vaccine hesitancy and provide appropriate public health information. But these collective efforts do not acknowledge that vaccine hesitancy can be caused by a myriad of underlying differences among subgroups – or that the difference could be due to factors other than hesitancy. Other potential factors range from easily observable attributes, such as lack of income or education, to attributes that are harder to observe such as effects of structural or institutional racism.

Methods We gathered longitudinal data asking respondents from the four largest states (New York; California; Texas; Florida) to imagine a situation where a number of vaccines for COVID-19 had been developed. These vaccines would have undergone all required testing and received regulatory approval for use in humans. They were then faced with six scenarios, or choice tasks, where in each task, two possible vaccines were described with seven attributes, including risk of infection, risk of serious illness, estimated protection duration, risk of mild side effects, risk of severe side effects, waiting time, and the fee: how much people need to pay to obtain the vaccine immediately. The vaccines also varied by two key population attributes: the share of the population that have already been vaccinated, and exemption from international travel restrictions.

For this study, we estimated five sets of choice models and compared model fit, including a simple multinomial logit, a nested logit (NL) grouping together all vaccine options, an NL with socio-demographic effects, a latent class (LC) with purely random heterogeneity and a LC with the same socio-demographic effects as the NL models. Our final model specification was a latent class model with three classes, where in each class, we estimated a NL model, nesting together the four vaccine options given in the survey. Respondents were sampled from an online Qualtrics panel from August 10 through September 3rd, 2020, and were representative with respect to the state and U.S. population in terms of age, gender and race.

Results Overall, 15.7 percent of respondents indicated they would not accept a COVID-19 vaccine, either because of attribute levels or regardless of its attributes. Of these, 14.7 percent were White and 2.4 percent were non-White (1.3 percent Blacks, 0.5 percent mixed race and 0.5 percent other). Of Black residents in the sample, 10.9 percent were in the group completely unwilling to consider a vaccine, regardless of any other factor. Of White respondents, 16.8 percent would not consider a vaccine, a higher proportion than among Blacks.

The latent class model probabilistically segmented respondents into a class with an overall preference for the paid vaccine options (Class 1); a class dominated by respondents who were most likely to choose the no vaccine option (Class 2) and a class that highly valued the free vaccine options (Class 3). In Class 1 (“Anxious”), 62 percent of respondents preferred a vaccine option that would require them to pay a fee, but not wait. In Class 2 (“Evaluative”), 52 percent of respondents chose the “no vaccine” option most often, meaning that they did not like the other options given to them. In Class 3 (“Cost-Conscious”), 89 percent of respondents preferred vaccine options that were given for free, but with a wait time. We found that there are 8 percent more Blacks in Class 1 (“Anxious”), compared to the sample mean, but 2 percent fewer in Class 2 (“Evaluative”) and 5 percent fewer in Class 3 (“Cost Conscious”). In other words: relatively more Blacks prefer vaccine options with a fee but no wait time, and relatively fewer Blacks prefer free vaccine options with a wait time or no vaccine options.

Overall, we found that individuals who identify as Black had lower rates of stated vaccine hesitancy than those who identify as White. This was true overall, by latent class, and within latent class. This suggests that contrary to what is currently being reported, Blacks are not universally more vaccine hesitant than other groups. Across classes, adjusting for other sociodemographic factors, we found that individuals with lower income had lower rates of vaccine uptake, as did individuals with low to no formal education. Combining the respondents who would not consider a vaccine (17%) with those who would consider one but ultimately choose not to vaccinate (11%), our findings indicate that more than 1 in 4 (28%) persons will not be willing to vaccinate. The no-vaccine rate is highest in Whites and lowest in Blacks.

Lower rates of vaccination among Black Americans do not reflect lower rates of racially motivated vaccine hesitancy. Instead, these lower rates reflect a higher proportion of Blacks among groups with vaccine hesitancy – namely lower income and lower educated individuals . To reduce racial disparities in vaccination rates, it will be necessary to address vaccine hesitancy more broadly in disadvantaged populations. Public health efforts are currently being targeted to address vaccine hesitancy among Blacks and other minority populations. Our model results help point the way to more effective differentiated policies.

ABSTRACT. While many people used to fly to distant holiday destinations, during the COVID19 pandemic many people have spent their holidays at a destination closer to their homes. One could hear people say that they did not mind because others did the same. This raises the question (van Wee, 2021) whether holiday destination is a positional good (Hirsch 1976). Positional goods are goods or services of which the individual’s utility is influenced by the level of consumption of other people. Examples involve cars, clothes, luxury watches and luxury houses, which may provide status when individuals consider their own goods to be ‘better’ than that of others.

A couple of studies have examined the positionality of goods by conducting stated choice experiments. Hoen and Geurs (2011) found evidence for positionality of passenger cars. Respondents had to imagine that they lived either in a country in which cars on average are either a class lower or a class higher than the cars they currently experience. Carlsson et al. (2007) constructed choice sets for positional goods such as cars, income, leisure time and car safety, in which individuals chose among hypothetical futures in which attributes represent a value for the individual and the average for the population. Both studies allowed examining the impact of the distribution of a good over the population on the utility a person derived from the good.

The objective of this paper is to examine whether holiday destination is to some extent a positional good. To that effect, we developed a context-dependent stated choice experiment. We first constructed choice sets of two alternatives each based on an orthogonal fractional factorial design. The attributes involve holiday destination rank, travel costs (1000, 1500 and 2000 euro) and travel time (6, 9 and 12 hours). The levels of the holiday destination rank involve the customized preference rank numbers of five destinations. Each respondent therefore first rank ordered five given holiday destinations. In the analyses, a higher number represents a more attractive destination. The choice sets were then combined with a context variable, which represented the destination of their friends (relevant ‘others’, which also may involve family and colleagues). Hence, while making a choice, respondents have to assume that their friends travel to the presented destination. Since the levels of the own destination are varied in the rank levels 1, 3 and 5 and the friends’ destination in the levels 2, 3 and 4, rank differences between own and friends destination range from -3 to +3, where a positive value means that the own destination is more attractive in the view of the respondent than the friends’ destination. __________________________________________________________________________________ Your friends travel to: Eastern Asia. Which vacation scenario would your choose? • Eastern Asia; cost = €1500 and travel time = 12 hrs. • South America; cost = €2000 and travel time = 9 hrs. __________________________________________________________________________________

The utility function included main effects for destination rank, travel cost and travel time and an interaction effect for destination rank and rank difference. Both an MNL model and a four class Latent Class model were estimated from choices made by a convenience sample of 159 respondents recruited from the authors’ networks.

On average, the parameter for destination rank has a positive value, which means that as expected, higher ranked destinations yield on average a higher utility. The interaction between destination rank and rank difference has a negative value, which means that the same destination becomes more attractive to an individual when their friends travel to more attractive destinations in their view, and likewise, becomes less attractive when their friends travel to less attractive destinations. The interaction effect denoting positionality is statistically significant in three of the four latent classes (class 1, 2 and 4), suggesting that positionality plays a role for about 85% of the respondents. Thus, these results suggest that holiday destination choice for most people depends on the destination choice of relevant others, which provides evidence for the hypothesis that holiday destination is a positional good.

Remarkable is that for class 3 both parameters for destination and the positionality effect are not statistically significant, while it has a significant but positive parameter for travel cost. Typically, such a result is considered implausible. Alternatively, one may speculate that this is also an expression of positionality: this class does not care where it is flying to, but it derives a positive utility from a more expensive ticket, suggesting that a more expensive ticket provides status.

That holiday destination is a positional good, is consistent with the observation that over the years increasing numbers of people flew to distant holiday destinations. Based on the same mechanism, we may speculate that currently a reverse trend is developing. At least part of the people who have spent their holidays closer to home during the Covid19 pandemic, may have enjoyed this and may continue doing this more frequently in the future. Hence, if friends travel to less attractive destinations, the positionality mechanism predicts that individuals derive less utility from their most attractive (often: distant) destinations. Due to this, they may be more inclined to also choose holiday destinations closer to their homes with the likely result that the total number of flights will decrease.

Moreover, this has an important implication for flying reducing policies: the welfare loss of such policies will be smaller than currently expected. After all, the results suggest that if also relevant others will no longer visit the most attractive destinations, the utility individuals derive from distant holiday destinations will be lower than currently believed.

ABSTRACT. Conservation of cultural heritage often depends on public funding. Determining the socially-optimal extent of this support and designing economically efficient policies requires the measurement of economic values associated with specific heritage items. Although heritage valuation studies form an increasingly important field of applied research, there is a lack of formalized approach. The shift in the political debate towards an emphasis on the benefits for a society has made cultural policymakers increasingly interested in monetary evidence of this value. As citizens became an important source of knowledge about optimal ways to govern resources, the number of contingent valuation method (CVM) applications to cultural goods started growing rapidly. In recent years, more and more of these studies utilize discrete choice experiments (DCE), however, the proliferation of this method to the cultural economics has been slower than in other fields and hence its potential has not been fully exploited yet.

CVM and DCE evoke controversies and ethical objections among those professionally involved in culture. DCE studies devoted to heritage present an overly simplified approach in which they have focused on the quality of provision or experience instead of the quality of assets (Wiśniewska 2020). Many other special concerns arising while applying DCE in the cultural arena (i.e. general lack of quality measures, poorly defined and ambiguous attributes) have been neglected. Researchers have been trying to replicate environmental valuation studies outright while not devoting enough attention to culture-specific methodological developments. Hence, most of the applications available in the empirical literature are not on-par with the current state-of-the-art non-market valuation guidelines (e.g., see the report by Lawton et al. 2021).

In this paper, we present a carefully designed study that leverages the above-mentioned limitations and elicits consumer preferences towards heritage conservation strategies. We estimate the monetary value of heritage goods, as expressed in general population WTP. While previous valuations have been carried out for individual sites in the context of a specific plan, our study includes a novel approach to evaluate programs for various categories of cultural heritage. To the best of our knowledge, there is only one study of this kind (Throsby, Zednik, and Araña 2021).

We implement DCE with three different valuation tasks – one for each category of potential cultural heritage conservation: Buildings, Historic sites, and Landscapes. Moreover, we divide each category by the type which is the only category-specific attribute. Other attributes are generalized across categories. We carefully selected them to address the needs of those involved in heritage protection. Some of the attributes are directly related to a given asset and its features, while others describe protection amendments. The attribute matrix has been verified with the large sample of existing tangible assets included in the register. In total, we investigate respondents' preferences towards 33 different types of immovable heritage with various characteristics. Together they form a valuation scenario describing the implementation of a hypothetical protection plan for a hypothetical asset.

The case study deal with various items included in the Victoria Heritage Register (Australia). The data is collected from a representative sample of n=1613 inhabitants of the southeastern state of Victoria. To convey the valuation task for respondents and reduce their fatigue, typical images of given assets were added. To control for the associated aesthetic bias, we implement three experimental treatments: two alternative images and no image. The design of the study included the necessary components for incentive compatibility of the study, such as consequentiality, and coercive payment, and a binary choice setting (including the status quo alternative), ensuring that respondents reveal their true WTP.

We built three different models using the mixed logit approach to interpret assets in separated categories. We estimated our models in WTP-space which means that the coefficients can be readily interpreted in monetary values by lay readers. Overall, the results show how much respondents are willing to pay for the conservation of various types cultural heritage and the types of assets that are the most preferred. Using imagery increased respondents' engagement in valuation tasks and does not produce any significant biases towards certain heritage types. "No image" treatment resulted in reduced WTP for all types except industrial and mining landscapes. Alternate images do not influence the results.

Overall, our study significantly contributes on the methodological level, demonstrating a new universal tool that enables stakeholders to easily evaluate protection plans for multiple sites with different profiles simultaneously. It is designed, conducted and documented in a manner that supports extrapolating findings to new policy scenarios. This creates the potential for future use of the results in benefit transfer. Providing respondents with credible choice scenarios with well-defined and unambiguous goods allows us better to understand their choices, including social and cultural motivations. As there is a need to raise the quality of SP methods applied to cultural goods, our research can help draw some systematic conclusions. We believe that results will help make cultural policies better matched to contemporary issues, public preferences, and the idea of sustainable development, providing useful implications and leading to more justifiable fund allocation.   BIBLIOGRAPHY

Lawton, R., D. Fujiwara, M. Arber, H. Maguire, J. Malde, P. O’Donovan, A. Lyons, and G. Atkinson. 2021. “DCMS Rapid Evidence Assessment: Culture and Heritage Valuation Studies - Technical Report.” Simetrica. Throsby, D., A. Zednik, and J. E. Araña. 2021. “Public Preferences for Heritage Conservation Strategies: A Choice Modelling Approach.” Journal of Cultural Economics, February. https://doi.org/10.1007/s10824-021-09406-7. Wiśniewska, A. 2020. “Quality Attributes in the Non-Market Stated-Preference Based Valuation of Cultural Goods.” Central European Economic Journal 6 (53): 132–50. https://doi.org/10.2478/ceej-2019-0012.

ABSTRACT. Aquaculture is the fastest growing food production industry globally, but it often conflicts with competing uses like recreation and nature-based tourism, and its impacts on sensitive coastal ecosystems are substantial. Public perception and social acceptability are key factors in a more sustainable growth of finfish aquaculture. However, the empirical evidence considering these factors is scant, which makes the planned growth of this industry uncertain. In Norway, there are plans for a three-fold increase in the production of farmed Atlantic salmon by 2030 (to the level of 3 million tons). Using the current plans as the Business-as-Usual (BAU) scenario, we designed and conducted a Discrete Choice Experiment (DCE) to investigate the general public’s support for this plan. Table 1 shows the attributes and attribute levels used (see attachment for table 1). The Meals produced indicate the daily number of meals worldwide based on Norwegian farmed salmon, and jobs are the number of employees directly occupied in the aquaculture sector. Plastic indicates the number of plastic items originating from aquaculture and being released into the marine environment per year. Sea lice indicate the share of young wild salmon that can die due to sea lice infestation when migrating from a river to the ocean; the intensity of aquaculture production in fjords, where wild salmon migrate, has a significant influence on sea lice infestation rates. An example of a choice card is given in Figure 1 (see attachment for figure). The survey was administered to 11,588 online panel participants, of which 5669 opened the invitation. Of these, 604 withdrew from the survey due to self-screening, which resulted in a contact rate of 47.7%, and a response rate (of those who opened the survey) of 89.3%. The sample characteristics are shown in table 2 (see attachment for table 2). The preliminary analysis of the data shows that overall, the Norwegian population prefer plans that are lower in scale and more sensitive to environmental issues (plastic and sea-lice) rather than economic (the number of salmon-meals and jobs) attributes. However, there is a large heterogeneity, as evidenced by relatively high and statistically significant coefficients of standard deviations of the estimated WTP distributions. Table 3 presents the results of the MXL model (see attachment for table 3). The study was designed to allow for several methodological tests. The pilot demonstrated a very high acceptance rate of the costly alternatives, which led us to including a split-sample experimental treatment varying the levels of the costs presented to respondents. In addition, we tested if presenting the choice situation as a BAU plus 1 or 2 alternatives influenced the observed results. We included forced vs non-forced choices in the cards (the latter means that respondents could proceed in the survey without making choices in the choice cards). Finally, our analysis includes the hybrid-choice model based investigation of the relative importance of various socio-demographic factors in explaining the preference heterogeneity.

ABSTRACT. Extensive empirical evidence suggests that including new alternatives into a choice-set may affect the way in which the original alternatives are evaluated and the relative probabilities among them (Huber et al.,1982; Doyle et al., 1999; Shafir et al., 2002, etc.). Hence, the inclusion of irrelevant alternatives carries the potential of affecting the decision-making process and, consequentially, of affecting the individuals’ choices. A corollary of the former is that it is possible to include irrelevant alternatives into a choice-set with the only purpose of favoring the choice probability of another alternative. The later phenomenon is known as decoy effect (Huber et al.,1982) and the irrelevant alternative is known as decoy. Customary behavioral theories such as random utility theory (Thurstone, 1927; McFadden, 1974) are not capable of explaining this phenomenon, as they assume that the utility ascribed to a given alternative is only dependent on the characteristics of the decision-maker and on the attributes of the alternative themselves; hence, the utility is independent from the choice-set. Consequentially, behavioral model departing from the compensatory mechanisms associated with the homo œconomicus s are required to represent the phenomenon. In the past, different behavioral explanations have been proposed for the phenomenon. The weight change theory (Ariely and Wallsten,1995; Huber et al., 1982) postulates that the existence of a decoy modifies the weights assigned to the different attributes of the alternatives. Instead, the value shift paradigm (Wedell 1991) consider instead that the inclusion of the decoy somehow modifies the way in which the attributes of the alternatives are perceived. Common to both behavioral explanations is the fact that the inclusion of the decoy changes the way in which alternatives and attributes are evaluated. Consequentially, it is not possible to operationalize the approaches into an integrated behavioral model (that preserves consistency regardless of the decoy being considered) without further assumptions that are required to be able to assess the impact of a potential decoy alternatives based on their characteristics. Hence, approaches aimed at integrating behavioral rules, are normally based on alternative behavioral theories. First, models based on regret theory (Loomes and Sugden, 1982), such as regret minimization models (RRM; Chorus, 2010), have been proposed to model the decoy effect. Under these assumptions, individuals aim at minimizing the regret caused by the independent comparison of all attributes of the chosen alternatives with all remaining alternatives in the choice-set. This way, including additional irrelevant alternatives (i.e. the decoy) should affect the level of regret, given how the remaining alternatives compare with the decoy. Guevara and Fukushi (2016) proposed a different approach to represent the phenomenon, on the basis of the emergent value approach (EV; Wedell and Pettibone, 1996). It aims at identifying how the inclusion of a given alternative directly influences the utility of competing alternatives. Under these assumptions, utility (i.e., the emergent value) is directly derived for the alternative that is being favored by the decoy. Guevara and Fukushi (2016) establish that, for their dataset, the EV outperforms the RRM to represent the decoy effect, but the EV model has the limitation that it requires defining which alternative act as decoy, which is take out of the choice-set, while RRM is agnostic to that respect. A different possible explanation for the decoy effect is given by prospect theory (Kahneman and Tversky, 1979). Here it is assumed that the utility of a given alternative is measured relative to a reference level, which is unknown to the modeler, and that losses and gains (relative to this reference) are valued differently. Then, including a new alternative into the choice-set has the potential of affecting the reference level, relative to which all losses and gains are measured. Bahamonde-Birke (2018) proposed a model that allows capturing prospect theory and estimating the reference level. This model can be used to establish the level of decoy caused by an irrelevant alternative. In this paper we explore different ways to characterize the decoy effect and offer an in-depth discussion on the theoretical and empirical implications of the different alternatives. Along these lines, we also consider a stated preference experiment, including fully dominated decoy alternatives that was developed by Fukushi et al. (2021). We model the phenomenon relaying upon different specifications of the RRM, EV and prospect theory model (PT), which, in turn, considers different specifications for the reference level. The empirical results are considered based on the level of adjustment to the dataset, the capacity of the models to behave in accordance with the underlying theory behind the decoy effect and the impact of decoy alternatives on the outcomes. The results show that, at least for this dataset, models based on prospect theory exhibit the best performance for our dataset, followed by emergent value models (which is not surprising, as we show that emergent values models can reasonably approximate PT models). A closer look at the estimated parameters reveals that PT and RRM models exhibit a congruent behavior in the sense that in the PT models, gains measured from the reference level do not seem to affect the choices and gains are, per definition, irrelevant in RRM models as they do not produce regret. However, both kinds of models differ regarding the reference levels. PT models reveals that the actual reference level seem to be considerable below the level suggested by the choice-set, which is the base to determine regret in RRM models, and that the decoy alternative affects the reference to a lesser extent than the other alternatives. Finally, a diagrammatic analysis of the probabilities associates with the alternatives being favored by the decoy reveals that the response to the decoy associated with PT models aligned much better with the theoretical expectations than other approaches. (see attachment for full list of references)

ABSTRACT. Short abstract

We propose an indirect measurement approach to approximate altruistic motivations underlying preferences for public health policies. The approach relies on associations between on the one hand decision makers’ perceived health risk for themselves and for close relatives, and on the other hand their observed preferences for health policies that reduce such risks. Our approach allows to draw a distinction between selfishness (i.e., self-protection), local altruism (i.e., protecting relatives) and global altruism (i.e., protecting the general public) as motives behind preferences for health policies. We illustrate the proposed approach using data obtained from a discrete choice experiment (DCE) in the context of policies to relax (or not) coronavirus-related lockdown measures in the Netherlands. Our results show that responses to policies are driven, besides global altruism, by considerable degrees of selfishness and local altruism, the influence of the latter being stronger than that of the former.

Extended abstract

Many public health-related decisions that people make involve combinations of, and trade-offs between, individual self-interest and collective wellbeing; the wearing (or not) of facemasks and the (un)willingness to receive a covid-19 vaccine being two highly salient recent examples. To motivate people to make decisions that promote long-term collective interests than just self-interest, public health interventions routinely emphasize altruistic (moral) values. Studies show that alluding to altruistic values, for example, increases support for a healthcare financing plan that promotes welfare of others and encourages people to accept vaccination against influenza. Quite recently, local and national governments across the world have started to use public messages that underline the need to protect others, in order to encourage people to adopt desired coronavirus safety behaviors. In particular, the use of public messages that highlight protecting close relatives or family members has been suggested, to increase compliance with mandated coronavirus protective measures.

The evidence suggesting the important role of altruistic motivations in encouraging actions that advance public health, is based on direct elicitation mechanisms such as asking people directly if such motivations played a role and to what extent. Such direct elicitation approaches to measure the relative importance of altruistic compared to selfish motivations is problematic at least for the following reasons. First, to avoid judgement from others or obtain social approval, people tend to suppress their own self-interest and respond to questions about moral motivations in a way they consider to be morally correct; alternatively, they may attempt to obfuscate their true moral intentions. As a result, choices between public health-related options (e.g., in a stated choice experiment) that are perceived by the participant to reveal their concern to others may not reflect actual behavior when such preferences are elicited directly. Second, people may consider the sacrifice that is required to not behave selfishly to be negligible if asked directly, leading to an overstated weight for altruistic motivations in their decisions, mainly when the choice environment is hypothetical which is often the case in public health studies. Finally, many of the moral judgements and choices people make occur spontaneously, without a strong conscious awareness of the underlying decision processes. Therefore, the use of explicit questioning approaches to uncover morally sensitive behavioral phenomena is likely to produce inaccurate responses.

To avoid these pitfalls associated with the direct measurement of moral motivations behind people’s (stated) responses to public health policies, we propose an indirect inquiry approach to measure the degree of altruistic motivations in a public health context. Our approach relies on the notion that an individual’s perceived risk of being negatively affected by some negative health event, could increase their acceptance of health policies aimed at reducing such risks for society at large. This would signal a degree of self-protection-related motivation behind their support for public health policies. Similarly, the extent of an individual’s local altruism can be approximated by associating their perceived risk of close relatives being negatively affected by the health hazard, with their support for public health policies. In case the individual does not perceive themselves and neither their relatives or friends to be at any health risk, then any support for the health policy would presumably be driven by global altruism for society at large. At the sample level, this implies that strong statistical associations between perceived risks for self and for close relatives and policy preferences, respectively are indicative of a significant degree of selfishness and local altruism-related motivations.

The contribution of this paper lies in putting forward this indirect approach to approximate moral motivations and in providing an empirical proof of concept related to measuring public support for health policies. Our data come from a stated choice experiment (and associated risk-perception survey) concerning public preferences for Dutch covid-19 policies.

Our main results are as follows. First, we observe that health risk perception is significantly and positively associated with willingness to accept societal sacrifices per avoided corona-fatality. This shows that people with a strong perception of health risk are more willing to accept a societal (e.g., an increase in the number of educationally disadvantaged children) or personal (e.g., corona tax) sacrifice per avoided fatality; presumably, because they believe that they themselves or one of their relatives could be one of the hypothetical fatalities.

Our second result indicates that perceived health risks for close relatives have a higher influence on people’s willingness to accept sacrifices, than perceived health risks for themselves. This suggests that responses to coronavirus-related health policies are driven more by local altruism than by selfishness. Third, we find that when the severity level of perceived health risk increases, willingness to accept societal sacrifices in order to avoid coronavirus-related fatalities increases. This finding holds mainly when the health risk is perceived to be a threat to oneself.

We derive and discuss policy implications from our results, and outline future avenues of research into proxy measures for moral motivations underlying decision making.

ABSTRACT. Discrete choice models (DCMs) have become a popular technique to uncover the preferences of economic agents for different alternatives across a range of applied economic fields. Applied either to revealed preference or stated preference data, DCMs assume economic agents trade between the attributes that describe the various alternatives on offer and select the alternative that provides the highest level of utility. Whilst most models consider only the attributes of the alternatives influencing utility, many behavioural economists recognise that psychological, cognitive, emotional and social factors can also contribute to the determination of economic individual decisions (see e.g., Thaler, 1980, 1985). To accommodate this view within the DCM structure, McFadden (1986) extended the framework to incorporate these latent psychological constructs, and hence allow the analyst to model how they can influence the decision-making process in the context of discrete competing alternatives. This framework, referred to hybrid choice models (HCM), is now extensively used to incorporate attitudinal and perceptual data, particularly within the transportation literature.

The HCM typically consists of multiple linked models. Attitudes or perceptions are modelled via a structural equation model (SEM), the outputs of which are latent psychological constructs that then enter the utility functions of a classical DCM. In the SEM, a structural equation is used to link the latent factor to covariates (usually socio-demographics), whilst measurement equations are defined to represent the relationship between psychometric indicators and the latent construct. The assumption under the measurement model is that the psychometric indicators are a manifestation of the (latent) psychological factor and therefore they act as dependent variable in the measurement equations. A common example within the transport literature relates to the attitudinal construct environmental preferences. Environmental preference is a multifaceted psychological construct, making any attempt to directly measure it very difficult. Instead, the construct is measured using specific indicators that measure what the researcher believes are related to the underlying psychological construct being considered. E.g., , Kim et al. (2012) use instrumental variables such as “I’m likely to choose an alternative mode in order to protect the environment”, “I’m willing to pay more for environmentally friendly items” and “I believe that using public transit is more helpful to the environment than driving a car” to measure environmental inclinations when exploring the preferences towards a new ecofriendly water transit system.

Outside of the transportation literature, it is common knowledge that not all indicator variables are the same in terms of how they relate to the underlying attitudinal constructs being measured. Indeed, there exist two types of constructs, these being reflective and formative, yet despite differences between the two, they are often treated as being the same when applied to the HCM framework. The two differ in terms of the direction of causation between the instrumental variables and the latent construct. Whereas reflective measures are seen to be caused by the underlying latent construct, the opposite is true for formative measures, which are thought to cause the latent construct (Bollen and Lennon, 1991).

The HCM framework as currently defined is suitable for handling reflective constructs as the measurement equations define the causal indicators on the basis of the latent variable operationalised in the structural equation component of the model. Although indicators can be manipulated (i.e., reworded) to a certain degree to adapt a formative measure to a reflective measure, some latent variables are formative by nature and should be included in an econometric setting that mirrors a different theoretical framework to the HCM as it currently exists. Despite this, it is common for transport researchers to treat formative measures as reflective measures and apply them incorrectly using the HCM framework. For example, Daziano (2012) defines appreciation of car features using items measuring eight aspects, these being purchase price, vehicle type, fuel economy, horsepower, safety, seating capacity, reliability, and styling. Such constructs are formative in nature insofar as they represent aspects of the vehicle rather than attempt to measure underlying attitudes towards vehicles. In a more recent study, Guzman et al. (2021) measure the satisfaction of public transport system considering the specific satisfaction for fare, comfort, security, as well as general overall satisfaction towards the public transport system. As with the measurement items used by Daziano (2012), these indicator variables appear to measure constructs that contribute to the satisfaction of a public transport system, rather than measure latent satisfaction. That is to say, levels of comfort, security and fare influence satisfaction, rather than being measures of underlying satisfaction.

In this research, we present the results of two separate modelling exercises, both dealing with electric vehicle choice. Firstly, we extend the existing HCM framework using reflective attitude questions, however in doing so we randomly expose respondents to different experimental treatments involving unique information states. Respondents undertaking this experiment are exposed to either a positive information state about the impact of electric vehicles on the environment, a negative information state, or no information (control group). In this manner, the information state assigned to a respondent is used within the SEM model to determine how information can influence attitudes towards electric vehicles, thus extending the framework beyond the simple inclusion of socio-demographic variables as is typically the case with such models estimated within the transportation literature. The second modelling exercise involves respondents being exposed to formative rather than reflective attitudinal questions. As with the first experiment, these respondents are exposed to either a positive, negative or no information state, however given the formative nature of the instrumental variables used, we adopt a variation of the HCM that specifically accounts for the nature of the questions being asked and their relationship to the latent constructs being measured. As such, this paper seeks to extend the literature on the inclusion of attitudes into DCMs in two directions, firstly by demonstrating how existing DCMs incorporating attitudes can include more than just socio-demographic differences, and secondly, by demonstrating how the framework can be adapted to handle formative attitudinal constructs.

ABSTRACT. This study examines the effect and relative importance of various factors shaping the preference to attempt to migrate again (“re-migration”) among Sudanese migrants who have returned to Sudan using a discrete choice experiment (DCE). East Africa is a region characterized by high mobility, including internal and international migration and displacement as a result of conflict, persecution, lack of economic opportunities, and draughts, with several countries hosting also large refugee populations. In recent years, many migrants from Sudan, Eritrea, Ethiopia and Somalia travelled northwards towards Libya and Egypt in search of opportunities there or in an attempt to reach Europe. Alternative migration routes are the so-called Eastern corridor towards Gulf countries from Djiboutian or Somalian shores, and the Southern route towards the Southern Africa region.

Some migrants return to their origin countries following failed migration attempts, sometimes assisted by “humanitarian evacuation” and “assisted voluntary return” programmes implemented by the International Organization for Migration (IOM) and others. Despite reintegration assistance offered to some returnees and awareness of the risks involved, some intend to re-migrate abroad via irregular means. While this is not necessarily inconsistent with what constitutes “sustainable” reintegration according to IOM, irregular remigration may well lead to the worsening of pre-existing vulnerabilities. A better understanding of the decision process leading to remigration is needed to inform policies that protect migrants without discouraging remigration per se.

This study examines the factors driving the re-migration decision of Sudanese migrants who have returned from Libya (84.5%), Egypt (7.6%), Niger (5.8%), Algeria (1.2%), or Chad (0.9%). Many potential factors shaping migrants’ decision-making process have been proposed in the literature including expected incomes, risk perceptions, costs of migration, debt, stigma, and community or peer norms. Much of the evidence available is based on qualitative studies and grey literature. Previous approaches have not been able to provide insights on which factors drive migrants to remigrate and which factors are more important than others. Quantitative evidence on the role of individual factors is needed to work towards a coherent theoretical model of remigration decisions in the context of irregular migration and to inform how policy actors address return to and reintegration in low-income countries.

To address this gap, we employed a DCE to study the re-migration preferences of returnees in Sudan. The DCE can derive and quantify the relative importance of certain drivers of remigration against alternative drivers. At the same time, DCEs can reduce the risk of social desirability bias – a key concern especially regarding sensitive issues such as preferences regarding potentially dangerous (irregular) migration journeys.

In our study we employed a labelled choice experiment and respondents were asked to make four choices, each between two labelled alternatives: 1) stay home and 2) re-migrating. Pre-tests revealed concerns regarding cognitive load and complexity of the questionnaire. As a result, the choice tasks were reduced to the key theoretical determinants referring to costs, risks, and benefits of (re)migration choices. The attribute levels were set following an iterative process in consultation with the local IOM office who assists migrants and returnees and has expert knowledge on the local context, as well as several interviews with returnees. In the final set-up, the option to “stay home” contained information on the attribute “potential income at home” (0, 30, 60, 90, 120, 240 USD per month). The option to “re-migrate” consisted of three attributes and featured information on “potential income abroad” (150, 300, 600, 900, 1200, 1900 USD per month), the “risk of failure” (high vs. low), and the “migration costs” (500, 800, 1100, 1400 USD one-off payment).

To construct choice sets, we used a D-efficient design for a conditional logit model with priors for all parameters. The priors mainly refer to the expected sign of the effect and were small in magnitude. This resulted in 24 choice sets that were blocked in six blocks with four sets each. Thus, each respondent was randomly assigned one block and answered four choice sets. To capture additional theoretical determinants, additional survey questions were included regarding perceived stigma, debt, peer preferences, social networks, etc.

In July 2021, we administered a CAPI phone survey including the DCE and surveyed returnees in Sudan that have been assisted by IOM. The net sample for the analysis is based on 327 respondents yielding 1308 individual choices between remigration and staying home alternatives.

Overall, the results show that any policy change which directly or indirectly affects remigration preferences may only apply to a fraction of returnees. In our sample, we found that 28% of returnees made trade-offs between different remigration choices that were presented to them; 48% had consistent preferences to stay home and 24% had consistent preferences to remigrate. This means that, in our sample, approximately one in four returnees has strong preferences towards remigration which are likely not going to be malleable by policy changes, at least as far as they relate to earning potentials, costs, and risks.

For those who made trade-offs between income, costs and risks involved in migration, conditional logit models indicate that income at home and risks were the most important factors driving remigration preferences. For example, holding all other attributes at their mean values, an expected income at home of 30 USD per month results in a remigration probability of 87%, while the highest income home of 240 USD per month decreases this probability to 20%, a 67-percentage points difference. For low migration risks, the predicted probability of remigration is 72%, compared with a probability of 37% for high risks. Additional models show that the strong consistent “stay vs. leave” preferences were associated with expected theoretical determinants such as owing debt, feeling pressure through social norms, or community stigma.

Despite limitations referring for example to sample size, our study provides a unique test of theoretical determinants of (re)migration preferences in the context of irregular migration, using a unique sample of vulnerable returnees – a population which is rarely captured in survey research.

ABSTRACT. In Japan, people’s shopping destinations are shifting from local stores to large-scale retail stores as a result of motorization; consequently, the number of local stores is declining. The establishment of large-scale retail stores can have a negative impact on the regional economy: For example, in areas that are dependent on large-scale retail stores for shopping, local profits may be outflowed to outside areas, place attachment may be weakened. Given this situation in Japan, revitalization of local stores could make an important contribution to social welfare. Local currency is one of the options for revitalization of local regions. A local currency is one that circulates only inside a specific region. In general, the aims of introducing local currency are (1) to revitalize the local economy by promoting local consumption, and (2) to facilitate mutual assistance among residents of the community. Although it is not very difficult to start a local currency project, it is often difficult to operate and maintain it. To maintain active transactions with it, both consumer choice behavior in using the currency and merchant choice behavior in accepting the currency as a means of payment are crucial. We have been working on exploring the consumer behavior and the merchant behavior, where we viewed local currency transactions as a two-sided market between consumers and merchants via the operator (Yoshioka et al., 2021) [1]. They incorporated social interaction aspects into the models and showed that 1) there is an interdependency between consumer and merchant behavior, 2) there is a significant conformity effect among consumers and strong place attachment increase the local currency users. However, Yoshioka et al. (2021)[1] don’t incorporate variables related to the spatial distribution of consumers and merchants. Given that the decline of local stores is attributed to the decrease in transportation costs due to motorization, people's shopping destinations are affected by their transportation costs. In addition, since the use of local currency is closely related to people's shopping destinations (shopping in the local shopping district vs shopping in the large-scale retail store in the suburbs), their willingness to use local currency would also be influenced by their transportation costs. If local currency is positioned as a part of urban and transportation planning, it is essential to explicitly introduce spatial information such as consumers' residential areas and locations of stores in the choice model for reflecting the aspects of transportation costs. Integrating spatial information into the choice model would clarify how the spatial distance between consumers and merchants affects consumers’ preference to use local currency. For example, consumers who live within a certain radius of a local shopping district would be more likely to use the local currency. In addition, this study has the potential to lead to new directions in urban and transportation planning, by examining how the urban form will change with the introduction of a local currency system. Therefore, we first model consumer and merchant choice behavior, where spatial distribution of consumers and merchants are considered. For the consumer side, we model whether they use local currency and if so, how much money they exchange utilizing the Type Ⅱ Tobit model. To reflect the spatial distribution in the model, we first develop a shopping destination choice model by extracting only the shopping trips observed in the target area from a travel diary data. Then, using the estimation results, we calculate the log-sum based accessibility index and introduce it as an explanatory variable in the consumer model. For the merchant side, we model whether they accept local currency payment utilizing a panel random-effects binary choice model. As in the consumer model, the accessibility index calculated from the shopping destination choice model is reflected in the merchant choice model. The destination choice model also depends on the consumer and merchant behavior. We use stated preference survey data collected from consumers and merchants in four cities of Hiroshima Prefecture (Higashi-Hiroshima, Kure, Takehara, Akitakata). As per our knowledge, no research has utilized the two-sided market framework on the local currency system that explicitly takes into account spatial characteristics and accessibility index. In SP questions for consumers, assuming a situation in which a local currency was introduced, four different scenarios were displayed for each consumer. The respondents were asked to choose whether they would exchange money for local currency, and if they would, they were also asked how much they would anticipate exchanging. Specifically, we examined the extent to which consumers consider the following six attributes in their decision: cashback rate, means of payment, share of stores that accept local currency, expiration date, local currency rewards for volunteer activities, and diffusion rate of local currency. In the SP questions for merchants, we assumed a situation in which a local currency was introduced and presented four different scenarios with the six attributes and asked the respondents to choose whether they would accept the local currency. The following six attributes are displayed: merchant fees, the ratio of government subsidies in the merchant fees, the type of local currency operator, the ratio of local currency use to the total consumption of residents, the ratio of merchants who accept local currency among all stores in the region, and the type of local currency. After modelling consumer choice behavior and merchant choice behavior, we conduct sensitivity analysis to discuss the efficiency of the local currency system by utilizing a methodological framework for estimating the transaction volume developed by Yoshioka et al. (2021)[2] with our proposed choice model. We believe that this is the first study utilizing the two-sided market framework that explicitly takes into account land use and transportation systems.

Reference: 1. Yoshioka, T., Chikaraishi, M., and Fujiwara, A. (2021) Empirical Models of Consumer and Merchant Behavior in the Two-Sided Market of Local Currency, Accepted for Proceedings of the Eastern Asia Society for Transportation Studies. 2. Yoshioka, T., Chikaraishi, M., and Fujiwara, A. (2021) An Empirical Analysis of Interaction in Two-Sided Market of Local Currency, Accepted for Proceedings of the Japan Society of Civil Engineers.

ABSTRACT. In the analysis of housing demand, discrete choice models implicitly or explicitly assume that each household can choose both freely and independently from other households in a given set of alternatives. However, limitations and interactions can occur when the number of available slots (supply) of one type is smaller than the demand for this type, as in some very attractive neighborhoods with an inelastic supply. The existence of long queues as in Paris or New York (Sieg and Yoon; 2020) reveals that capacity constraints are particularly severe in the social housing sector and could generate some feedback to the private sectors (Chapelle, Wasmer, Bono; 2020). Most analyses of the housing demand neglect the fact that some households are often denied their preferred alternative, and the observed choices do not fully reflect preferences. In the presence of capacity constraints, observed choices are poor predictors of household behavioral responses to changes in the supply of social or Private housing, or for public policy evaluation. Differential capacity constraints in these two submarkets can play a major role in explaining residential segregation in many megalopolis over the world, especially in Europe, which has a long tradition of funding social housing. Our analysis builds on de Palma, Picard and Waddell (2007), who developed a methodology to address such constraints, both from the theoretical and empirical points of view, with an application in Paris region. We generalize their analysis in several directions. First, we recognize that the dwelling renting market is split in two submarkets (social housing and Private sector) interacting with each other, but affected by dramatically different capacity constraints. Second, we analyze the differential role of real estate prices between these two submarkets. Third, we consider an overflow from social housing towards the Private renting sector, but no overflow in the opposite direction. This unilateral overflow adds to, and interact with, the overflow within each sector. It is consistent with the observation of real estate market in Paris region. We extend their axiomatic approach, theoretical developments and iterative procedure to estimate and predict the respective demands for social housing and for renting in the Private sector, given the discrepancy of capacity constraints in these two interacting submarkets. We assume that the HLM sector is globally constrained, in the sense that total regional supply of HLM is lower than the total regional demand for HLM. By contrast, Private renting sector and the system are not globally constrained, in the sense that the supply in the Private renting sector is sufficient to absorb the global demand addressed ex ante to the Private sector plus the global overflow from the social housing sector. We extend the main assumptions made by de Palma et al (2007) to the case of joint choice of renting sector and residential location. Our algorithm relies on two axioms: “No priority rule”: the probability that a household actually lives in a location constrained ex post (taking into account both ex ante demand and overflow) is proportional to the probability that it prefers this location (same ratio for all households in this location). “Extended IIA”: the probability that a household actually lives in a location unconstrained ex post is proportional to the inflated probability that it prefers this location (same ratio for all unconstrained locations for this specific household). The main difference with de Palma et al (2007) is that our notion of inflated preference probability integrates the probability that the household is denied access to the HLM sector, in relation to the global overflow from the HLM sector to the Private sector. In the simplest version of the model, we further assume that the households’ preferences for social housing are independent from their preferences for locations. However, this assumption is not consistent with the observation that rents in a given location are dramatically different between social housing and Private sector. This price differential induces an increase in the complexity of the model and of the algorithms used to disentangle the roles of preferences and capacity constraints, but we show that the model is still tractable in this more realistic case. The model is first illustrated on pedagogical examples, which highlight the role of capacity constraints and analyze in detail the mechanism governing the overflow within the social housing sector. We then use novel datasets in order to estimate the demand accounting for capacity constraints: the exhaustive data on housing in Ile-de-France (FIDELI) allows to track households’ characteristics and their movements between housing units between 2015 and 2017, and to identify the vacant dwellings. We combine this first dataset with a large dataset describing the supply dwellings for rent including a large sample of rental units on the market and their characteristics (price, surface ) to create rental price index (see Chapelle and Eyméoud; 2018). This paper complements the literature estimating the demand for subsidized dwellings (Geyer and Sieg ; 2013, Sieg and Yoon; 2020). As in these two papers, we allow for the existence of excess demand in the social housing sector, but we also allow for the existence of capacity constraint in the private sector and jointly study the tenure choice and the neighborhood choice while these two papers ignore the spatial dimension. We thus also contribute to the vast literature estimating the demand for neighborhoods and housing that follows Rosen (1974) and use discrete choice approach following Berry, Levinsohn, and Pakes (1995). This stream of literature relies on the implicit assumption that prices clear the market, whereas our model allows for excess demand. More broadly, our research question is related to the literature on rent control (Diamond et al; 2019), subsidized dwellings (Olsen and Zabell; 2018 ) and tenure choice (Sommers and Sullivan; 2018).

ABSTRACT. Motivation

The latent class model (LCM) is a well-established nonparametric model that captures inter-individual heterogeneity [1]. The LCM assumes that the population can be divided into latent classes that show homogeneity within the same class but heterogeneity among different classes, widely used in the choice modeling field. However, the log-likelihood function of the LCM is non-convex, and the commonly used Expectation-Maximization (EM) algorithm for parameter estimation has a problem of falling into local maxima. Therefore, a multi-start local search heuristic is widely used, where the searches are started from a large number of different initial values. However, as the number of classes increases, the number of local maxima increases, and there is no guarantee that a reliable solution can be obtained. Therefore, we propose to re-derive the Deterministic Annealing EM (DAEM) algorithm [2] and use it to estimate the LCM. This algorithm solves the maximization problem of the log-likelihood function by reformulating it as a minimization problem of the free energy depending on the temperature parameter, which is a natural extension of the EM algorithm. Taking advantage of the fact that the free energy function is smoothed when the newly introduced temperature parameter is high, the algorithm is expected to reach convergence without falling into poor local maxima by gradually decreasing the temperature parameter.

Uniqueness of this work

The originality of this study is twofold. First, we used the DAEM algorithm after re-deriving it in a theoretically consistent form. Ueda and Nakano (1998) state that the algorithm can be derived by using the principle of maximum entropy and statistical mechanics analogy. However, in their derivation, the correspondence with the concepts in statistical mechanics is unclear. As a result, the E-step and M-step were derived based on two different arguments, the principle of maximum entropy and free energy minimization. Hence, we show that the algorithm can be derived as the problem of minimizing the free energy, which is expressed as the antagonistic state of entropy and internal energy, by assuming that each individual belonging to each latent class in the LCM is equivalent to each subsystem taking each energy level in statistical mechanics. As a result, the E-step and M-step can be derived as the maximization of entropy and the minimization of internal energy, respectively. Second, we empirically compared the convergence properties of the two algorithms by estimating the LCM. The estimation of LCM with a high number of classes will frequently fail to converge or experience other problems [3]. However, to our knowledge, the reason for this was not examined. We have shown that this is caused by the EM algorithm falling into poor local maxima and that the DAEM algorithm can avoid such local maxima and converges to relatively good local maxima.

Data and analysis methods

We use RP data of the evacuation choice behavior from the tsunami. It is obtained from a survey conducted by the Japanese Ministry of Land, Infrastructure, and Transport (MLIT) among survivors of the Great East Japan Earthquake in 2011. The survey was undertaken via interviews with 10,603 people across 49 cities. Additionally, we use the situation data at the time of the disaster, such as the time of the issuance of the evacuation order, which was collected and constructed through various primary sources and inquiries to various administrative and research institutions. To analyze the heterogeneity in evacuation choice behavior, we estimated the LCM with the Sequential Logit model [4], which represents sequential decision making. To evaluate the performance of the two algorithms, we randomly set the initial values of the parameters in the range of [-1, +1], and conducted 20 trials for each number of classes and each algorithm.

Application potential and policy relevance

First, the proposed DAEM algorithm and the comparative evaluation of the algorithms will allow us to obtain more reliable parameters for disaster mitigation measures that directly affect human lives. Second, by estimating the LCM, we can comprehensively evaluate disaster mitigation measures, considering that the effectiveness of short-term measures such as evacuation orders may vary depending on long-term measures such as risk education. Thirdly, we re-derive the algorithm by clarifying the correspondence with statistical mechanics, which will enable us to develop the theory. For example, although we assume that each individual makes decisions independently, it is possible to consider the interaction among individuals. As Brock and Durlauf (2001) propose the social interaction model using the Mean-Field Approximation of the Ising model, further theoretical development is possible based on the discussion of phase transitions in statistical mechanics.

Obtained and anticipated results

The estimation results (Fig. 1) show that the DAEM algorithm does not necessarily converge to the global optimal maximum. However, as the number of classes increases, the multimodality of the likelihood function certainly increases, and the EM algorithm easily falls into local maxima. In such cases, we found that the DAEM algorithm tends to return a relatively good final likelihood. Additionally, the estimation results (Fig. 2) confirm the existence of heterogeneity in evacuation behavior. For example, people who live near the coast, at a lower elevation, or who have higher risk knowledge tend to evacuate earlier. Validation results have not been obtained yet. We expect that once validation results are obtained, a more detailed comparative analysis of the properties of the solutions obtained by the EM algorithm and the DAEM algorithm, such as the variation in prediction accuracy, will be possible.

Reference

[1] Bhat, C. R. (1997). An endogenous segmentation mode choice model with an application to intercity travel. Transportation science, 31(1), 34-48. [2] Ueda, N., & Nakano, R. (1998). Deterministic annealing EM algorithm. Neural networks, 11(2), 271-282. [3] Yuan, Y., You, W., & Boyle, K. J. (2015). A guide to heterogeneity features captured by parametric and nonparametric mixing distributions for the mixed logit model (No. 330-2016-13990). [4] Fu, H., & Wilmot, C. G. (2004). Sequential logit dynamic travel demand model for hurricane evacuation. Transportation Research Record, 1882(1), 19-26. [5] Brock, W. A., & Durlauf, S. N. (2001). Discrete choice with social interactions. The Review of Economic Studies, 68(2), 235-260.

ABSTRACT. The revealed preference (RP) data for mode choice modelling often suffers from measurement errors. Measurement errors in the level of service variables, which include travel time and travel cost, are associated with the inaccuracies in the perceived travel times and travel costs of the respondents. The literature have highlighted significant differences between the reported travel times and the actual travel times of the public transport. The existence of measurement errors cause a correlation between the reported travel time and the error term of the utility function (endogeneity), and subsequently results in a biased estimate for the parameters in a mode choice model. These biased estimates may result in an incorrect Value of Time (VoT), and further, an inaccurate prediction of the market responses. The issue of endogeneity in discrete choice models is addressed by employing various procedures, including the control function method, the method of proxies, the multiple indicator solution, and the integration of latent variables. Among these, the latent variable method with a Hybrid Choice Modelling (HCM) architecture has been accepted as an effective technique, although at a high computational cost. In literature, the analysis of measurement errors in travel time and travel cost was performed using a HCM framework, where in both the structural and measurement part had either an additive (normal distribution) or multiplicative (lognormal distribution) error specification. Until now, none of the studies have tested a set-up in which the structural part has a different error specification compared with the measurement part. Moreover, there is an absence of studies that explore the modelling of heteroskedastic measurement error, and further the incorporation of multiple variables with the measurement errors. In the context presented above, the current study aims to contribute to the existing literature by testing: 1. The effect of incorporating multiple variables with measurement errors on the parameter estimates and 2. The effect of size of measurement error variance on the parameter estimates. These tests are performed by a systematic analysis of the different error term specifications, corresponding to the measurement and structural models. Four different specifications of error terms are used in the current study: 1. An additive structural equation and an additive measurement equation (S1), 2. An additive structural equation and a multiplicative measurement equation (S2), 3. A multiplicative structural equation and an additive measurement equation (S3) and 4. A multiplicative structural equation and a multiplicative measurement equation (S4). For testing, synthetic data generation is done using a pre-defined decision process with three alternatives (car, bus and metro), two independent variables (travel time and travel cost) and fixed (true) utility parameters. In the data generated to test the first objective, both the travel time of bus and travel cost of car are assumed to suffer from measurement error. In the dataset, the true value of travel time is drawn from a mixture of gaussian and log-normal distributions with each distribution contributing towards 50% of the sample. Further, the observed travel time of bus is assumed to be distributed around the true travel time, with two components of the measurement error: an additive component drawn from a normal distribution and a multiplicative component drawn from a log normal distribution. A total of 100 datasets, with 2000 observations per dataset, are generated for each specification under each objective to check the capability of the HCM to derive the best parameters in the presence of measurement error. Parallel to these HCM specifications, estimation is also performed on all the datasets using a multinomial logit model (L1), to understand the bias in parameters, when measurement error in data is not considered in estimation. As hybrid choice models do not have well specified statistical goodness of fit measures, the best model is chosen from three different approaches: 1) proximity of parameters to their true value, 2) Quantile-Quantile plots of the residual of measurement equation and 3) cross validation using 70-30 data split. The current study further proposes a novel approach to incorporate the heteroscedastic measurement error specification in the hybrid choice model. For testing this model, the data is generated in a similar manner as mentioned above, with an added assumption that the measurement error is proportional to the true travel time of the individuals. The proportionality is introduced both in the additive and the multiplicative parts of the measurement error. This data generation mimics a real world scenario in which the variance around true travel time increased with an increase in the true travel time. The estimation on this data is performed using HCM frameworks with both homoscedastic, and heteroscedastic specifications. A comparison between these two models highlighted the importance of a heteroscedastic measurement error specification. Finally, the best specification among the tested models is utilized to estimate the parameters in a real-life dataset. The estimated model is checked using Quantile-Quantile plots and the prediction accuracy using cross validation procedures. Further, estimated models are compared with standard models to understand the advantages of complex models that are better specified. The study concludes by presenting the future directions of the work. The results from the currents study are expected to elicit the best performing model among the four different specifications in the homoscedastic and heteroskedastic set-ups. Further, the study also determines the implications of using an HCM framework when there are multiple variables suffering from measurement error. Lastly, the study aims to understand the efficacy of aforementioned specifications at different variances of the measurement error.

ABSTRACT. When modelling repeated discrete choice occasions, two dominant model families are multinomial logit (MNL) and multinomial probit (MNP). Between these two, the MNP offers more modelling flexibility while suffering from larger computational costs. Moderately sized MNP models were estimated using maximum simulated likelihood (MSL) methods, which show computational disadvantages with a growing number of choice alternatives and choice occasions. An alternative approach is to use a composite marginal likelihood (CML) consisting of the product of the choice probabilities for pairs of choices instead of using the joint probability for all choices of an individual. This reduces the dimension of the required integrals and reduces the computational burden. CML models and advancement of these, such as maximum approximate CML (MACML), have thus become popular in the discrete choice modelling field ((Bhat, Transportation Research Part B: Methodological, 2011), (Varin, AStA, 2008), (Varin, Reid, and Firth, Statistica Sinica, 2011)). The CML formulation allows the assignment of different power weights to each bivariate margin. The chosen weights affect the statistical properties of the CML estimator, and by setting a portion of the weights to zero significantly reduces the number of required bivariate probabilities, further reducing the computational burden. There has been a fair bit of research into weighting schemes in the past (Bhat, Foundations and Trends in Econometrics, 2014), (Varin and Vidoni, Biometrika, 2005), (Pedeli and Varin, Statistical Methods in Medical Research, 2020), (Bevilacqua et al., JASA, 2012). This paper aims to contribute to the literature by looking at the impact of weighting strategies in two special cases (1) unbalanced panel data, (2) unaccounted autocorrelation of the errors. Unbalanced panel data: When using full-pairwise weighting strategy every of the T_i observations of individual i is part of (T_i−1) CML margins and there is a total of (T_i−1)T_i/2 pairs of choice occasions for this person. So, each observation of an individual with more observations has a higher total weight in the CML function than the observations of an individual with fewer observations. Individuals with more observations, hence, have a disproportionately higher influence on the CML function. Asymptotically this will only affect the efficiency of the estimator but will not affect its other properties (Bhat, Foundations and Trends in Econometrics, 2014). (Cessie and Houwelingen, Applied Statistics, 1994), (Kuk and Nott, Statistics & Probability Letters, 2000), (Joe and Lee, JMVA, 2009) have investigated this in binary choice occasions and came up with different weighting strategies to account for this imbalance in weights. Following (Joe and Lee, JMVA, 2009), the asymptotic variance of the estimator can be calculated as a mixture of variances for deciders with an identical number of choice occasions. Focusing on the trace of the variance matrix (or a particular parameter component) the analytical expressions can be used to derive the optimal weights specific to deciders with a particular number of choice occasions. We develop explicit formulas that can be used in a two-step procedure to provide estimators using optimal weighting schemes. A demonstration example shows that the gain in terms of asymptotic variance can be as much as 10%. To investigate the effects of the individual weights on the CML estimator in the MACML setting we simulate unbalanced panel data sets with an underlying MNP model with mixed effects for the alternative specific constants (ASCs) and estimate the model using MACML with different weighting strategies for individuals with different numbers of observations. In a first simulation study with 105 replications of a simulation of a population with 2500 individuals with T=3 and 2500 with T=7 observations, we weighted the margins of the individuals with 7 observations with c and the margins of individuals with 3 observations with (1-c), calculated the robust covariance matrix at the true parameter value. In these simulations c = 0.161 minimised the trace of the robust covariance matrix, where (Cessie and Houwelingen, 1994) would suggest c=0.25 and one strategy by (Joe and Lee, 2009) would suggest c=0.143. Unaccounted autocorrelation of the errors: Correlations between repeated choices are classically modelled by introducing random effects. This accounts for taste heterogeneity and creates correlations across the observations for one individual. Accounting for autocorrelation of the error terms, however, is tedious to implement in multinomial discrete choice models, and includes the estimation of many additional parameters, if no restrictive assumptions on the autocorrelation pattern are made. In the case of positive autocorrelation, introducing mixed effects for the alternative specific constants (ASCs) captures these effects to a certain extent. The taste perseverance would then be interpreted as taste heterogeneity, implying that the misspecified model is robust to a certain degree. This effect is enhanced when margins with close by observations are assigned higher weights. In the case of negative autocorrelation of the error terms, however, including mixed effects would not be able to capture these effects. In both cases, the misspecification of the model can lead to biased estimates and a loss in the efficiency of the estimator. When assuming stationarity for the error terms, the correlation of the errors will decrease with increasing temporal distance between the observations. This implies that for distant pairs of observations the misspecification of the autocorrelation of the error terms is of less importance. Our suggestion is to give more weight to margins with more distant observations. In these cases, the correlation of the errors due to autocorrelation is reduced, while the correlation due to the mixed effects can still be estimated. This strategy would go in the opposite direction of the weighting strategies favoured in the current literature which mostly favour higher weights for margins with closer observations (Crastes et al., working paper, 2020). We investigate the effects of unaccounted autocorrelation of the errors by simulating data from a MNP model with mixed effects for the ASCs and various structures of autocorrelated errors. We then estimate the model using the MACML approach with different weighting strategies to mitigate the effects of the unaccounted autocorrelation. We evaluate the different weighting strategies by measuring: -the finite sample bias, -the trace of the robust variance-covariance matrix.

ABSTRACT. Motivation The Value of Travel Time (VTT) is a pivotal input for Cost-Benefit Analyses (CBA) for new national transport infrastructure projects and policies. Travel time improvements are typically a substantial economic benefit in CBA. Hence, accurate estimation of the VTT is of high societal, financial and practical relevance. Given the importance of VTT metrics for transport projects and policies, national VTT studies are regularly conducted.

In a number of Western-European countries these national VTT studies use two-attribute two-alternative SP-experiments where respondents are asked to choose between a slow/cheap and a quick/expensive route. As such, each choice task has an implicit price of time determined by the ratio of the differences in travel time and travel cost between the two alternatives. This ratio is called the Boundary Value of Travel Time (BVTT). Despite ongoing debates on whether or not better VTT values are obtained through SP-experiments with more attributes (Hess et al. 2020), for reasons of consistency with previous surveys alone these SP-experiments are still widely used.

The objective of this paper is to consolidate the knowledge on SP designs for VTT studies. In particular, we aim to draw lessons for the SP design of the upcoming Dutch VTT, scheduled for 2022. To do so, we review the VTT literature and conduct Monte Carlo studies to test properties of designs.

Results for the literature review Our review revealed there has been quite some development regarding what is considered to be the best design strategy over the years. In fact, the debate is ongoing: even the most recent studies express strongly different views. Examples of different design strategies: - The Dutch 1988 and 1997-studies (Gunn 2001) used a fixed set of SP-questions pivoted around current travel time / cost - The UK 1994-study and the Dutch 2009/2011-study (Kouwenhoven et al. 2014) used a so-called “Bradley”-design, i.e. a kind of orthogonal design while preventing dominant questions. - The Scandinavian studies, including the Danish 2004 (Fosgerau et al. 2007), Swedish 2008 (Börjession & Eliasson 2014) and Norwegian 2009 and 2019 studies (Ramjerdi et al. 2010), used a semi-random design that ensured that each respondent was offered a set of choice pairs in which each BVTT was in a different pre-defined range. - The latest UK 2014-study (Hess et al. 2017; Batley et al. 2017) used a D-efficient design in which the BVTT-range was not included as a design parameter.

From our review it is clear that both the Scandinavian and the D-efficient design strategies have their merits. D-efficient designs have the edge over Scandinavian designs when it comes to statistical efficiency, but Scandinavian designs offer much wider ranges of BVTTs. The latter is important since especially the Danish VTT study showed that the full VTT distribution should be captured for a reliable estimate of the mean (“catching the tail”).

Results from the Monte Carlo studies We carefully generated synthetic choice data in four steps: (1) trip characteristics were sampled from observed trips in order to get realistic combinations of travel mode, travel time and travel cost; (2) each synthetic decision maker was randomly assigned a VTT from a given distribution (e.g. uniform, lognormal, loguniform); (3) choice tasks were created based on the design rules of one of the design strategies mentioned above; and (4) synthetic choices were generated assuming a decision-making model (e.g. random utility with additive or multiplicative errors). After this, we estimated mixed logit models to recover the parameters of the VTT-distribution.

In line with expectations, the simulations shows that if the shape of the VTT distribution is the same in step (2) above and in the mixed logit estimation; and if the decision making model is the same in step (4) above and in the model estimation, all design strategies are capable to recover the true mean VTT. That is, no bias is observed, even if the BVTT-range covered only a (small) part of the full distribution. Hence, under these conditions, there is no need for a design that covers the full domain of the VTT distribution by including relatively high BVTT.

However, if the shape of the VTT-distribution used for the choice data generation is not fully known to the analyst, or if the data generating process is not exactly random utility with additive (or multiplicative) errors, all designs struggled to recover the mean VTT with the exception of the Scandinavian design which is more robust against misspecification.

Design of the upcoming Dutch VTT study Based on these insights from the literature the design and from the Monte Carlo analysis the new Dutch survey combines the best of both worlds. That is, each respondent will be offered four choice pairs based on a D-efficient design, with BVTTs that cover the central part of the expected VTT-distribution, and four choice pairs based on the Scandinavian design strategy covering the more extreme parts of the VTT-distribution.

This paper furthers the ongoing research and debates into the design of two-attribute SP VTT studies. However, the implications apply to any two-attribute experiment (and maybe even multiple-attribute experiments), also outside choice modelling in transport.

References Batley et al. (2019) New appraisal values of travel time saving and reliability in Great Britain. Transportation 46, 583-621. Börjesson & Eliasson (2014) Experiences from the Swedish value of time study. Transp.Res.PartA 59, 144–158. Hess et al. (2017) A framework for capturing heterogeneity, heteroskedasticity, non-linearity, reference dependence and design artefacts in value of time research. Transp.Res.PartB 96, 126–149. Hess et al. (2020) A critical appraisal of the use of simple time-money trade-offs for appraisal value of travel time measures. Transportation 47, 1541–1570. Fosgerau et al. (2007) The Danish value of time study, DTU Report Gunn, H. (2001) Spatial and temporal transferability of relationships between travel demand, trip cost and travel time Transp.Res.Part.E 37, 163–189. Ramjerdi et al. (2010) Value of time, safety and environment in passenger transport–Time, TØI report 1053-B/2010. Kouwenhoven et al. (2014) New values of time and reliability in passenger transport in the Netherlands. Res.Transp.Econ. 47,37–49.

ABSTRACT. Introduction: There is a lack of guidance in current literature regarding the suitability of different patient preference elicitation methods for different preference-sensitive contexts. Limited evidence exists for how patient preference elicitation methods compare directly. Rating methods, such as swing-weighting (SW), are often regarded as a simpler approach to eliciting patient preferences since they do not force simultaneous trade-offs between multiple attributes. However, other health economists state that direct pairwise comparisons in a discrete choice experiment (DCE) are easier for patients than a direct numerical assessment of relative value present in SW. Therefore, the aim of this study is to compare the performance and results of DCE and SW through empirical research, namely by eliciting preferences for glucose-monitoring devices in a population of diabetes patients.

Methods: A sample of Dutch adults with type 1 or 2 diabetes (n=459) completed an online survey assessing their preferences for glucose-monitoring devices, consisting of both a DCE and a SW exercise presented in random order. Each exercise was followed by debriefing Likert-scale questions related to the ease of understanding and ease of completing the exercise.

For the DCE, NGene software was used to develop a Bayesian D-efficient design, consisting of three blocks of 12 choice tasks. Each contained two hypothetical glucose-monitoring with seven attributes of varying levels in a ‘best-best’ or ‘dual response’ DCE, with a follow-up task asking participants if they would prefer the hypothetical device chosen or a standard finger-prick test. Observations were analysed in NLOGIT by a latent-class model and a mixed-logit model. Based on model fit, the mixed-logit was the model best suited to the data. For the SW, participants ranked the seven, randomly-listed attributes based on how they would prioritise improving an attribute-level from its worst to its best state. Participants allocated points, from 0 to 100, to each of the swings relative to their first choice which was automatically allocated 100 points. The SW analysis was conducted by examining each participant’s point allocation for each attribute-level improvement relative to the total number of points allocated. The weighted average of each attribute was calculated across the entire participant sample.

A comparison of how important each attribute was reported through the methods was examined by through the proportion of preference for one attribute compared to the summed preferences for all attributes. For the DCE, this involved examining the absolute difference between the best level coefficients and the worst level divided by the sum of all these differences across the attributes. For the SW, one attribute’s weight was calculated as a percentage of the total summed attributes’ weights. The relative weights for both methods, reflected as a proportional percentage, were then directly compared. In order to determine whether there were significant differences in attribute rankings between the methods, the respondent-level ranking of the attributes in the DCE and the SW were compared using a (generalised) ordered logit model. Self-reported feedback from participants indicating how easy the method was to understand and answer was used to compare the methods, stratified by order of the exercises and health literacy and numeracy. Drop-out rates during the completion of the exercises were also compared as a proxy for participant burden.

Results: The two attributes with the highest importance were cost and precision, respectively, for both the DCE and SW, but the relative weight of costs was much higher in the DCE. For the DCE, the following order of attributes based on their relative importance weight was: skin irritation, fingerpricks, effort, alarms, and glucose information, respectively. For the SW, these were fingerpricks, glucose information, effort, skin irritation, and alarms, respectively. The relative weights of all these attributes differed significantly between the two methods. The weights derived from the SW were almost evenly distributed across all attributes. All attributes in the SW received between 12-17% of the designated importance. The SW point allocation had a 1.4-fold difference between the most and least important attribute, while the DCE had a 14.9-fold difference. The (generalised) ordered logit model also indicated that there were significant differences in the respondent-level rankings of the attributes between the DCE and SW. Fingerpick frequency was more likely to be ranked as the highest attribute in the DCE rather than in the SW and alarms and precision were more likely to be ranked among the bottom-ranked attributes in DCE rather than in the SW. The DCE was better received by participants than the SW, regardless of the order completed, or the level of health literacy or health numeracy reported by the patient.

Conclusions: Incorporating point allocation into SW is often praised for being a simple way to elicit the relative valuation of the attributes by allowing respondents to directly report this valuation for each attribute. However, there was little difference between the mean attribute weights in the SW, and poor discriminatory power resulting from insufficient variability in the point allocation. Over 55% of participants allocated the same number of points to at least two attributes.

This study supports the theory that direct pairwise comparisons in a DCE are easier for patients than a direct numerical assessment of relative value presented in SW, and the complexity of applying it against multiple attributes. Essentially, it is easier to say which of two attributes is more important, rather than trying to quantify how much more important it is. From the researcher’s perspective, the DCE may appear more complex compared to the SW in terms of design and analysis, but respondents view the DCE as the simpler method to understand and complete. An additional benefit of DCEs is the ability to assess preference heterogeneity using mixed logit models, as evidenced by this study. A SW is still a viable alternative in cases when the number of attributes cannot be feasibly integrated into a DCE, or when a sample size is too small for a DCE.

This method comparison provides further evidence of the degree of method suitability and trustworthiness of these methods for measuring preferences for decision-making. Further research should compare these methods in different disease areas and decision-contexts.

ABSTRACT. Three modeling approaches are generally used to quantify the value of travel time reliability: mean-variance, scheduling, and integrated. The mean-variance approach directly adds a variable that represents travel time variability, such as the standard deviation of travel times, into the user’s cost function. The drawback of the mean-variance approach is that there is no theoretical foundation for its underlying behavioral principle. This is obvious because the mean-variance approach does not distinguish schedule delay late (SDL) from schedule delay early (SDE) because of travel time variability. The scheduling approach explicitly analyzes the departure time choices of travelers with stochastic travel time. As demonstrated by Noland and Small (1995), this approach models departure time choice behavior based on Vickrey’s (1969) scheduling costs, including the expected cost of SDL and the expected cost of SDE both of which are timing disparities from the preferred arrival time (PAT) of each traveler. The scheduling approach is preferable from the perspective of its theoretical foundation; however, a major drawback is that the traveler’s PAT needs to be observed for the model to be identified empirically. The integrated approach was proposed by Fosgerau and Karlström (2010) to incorporate the benefits of the mean-variance and scheduling approaches. A mean-variance type cost function of travelers was derived as a maximal utility from the scheduling model under several relatively weak assumptions. The virtue of integrated models is that, although they are theoretically more rigorous, it is not necessary to obtain information about the PAT in the empirical model estimation.

Most previous empirical studies that estimated the value of travel time reliability focused on a single behavioral dimension, that is, either of route choice, mode choice, or departure time choice, and few researchers have paid attention to consistent parameter estimation across these different behavioral dimensions. In the present study, we attempt to fill this gap by developing a unified survey and estimation framework for valuing travel time reliability. The behavioral dimension analyzed in the study is not independent of the modeling approach. Specifically, the scheduling approach is, in principle, used to model departure time choice behavior but not for mode and route choice behavior because it essentially handles preferences for the timing of activities. It is also obvious that optimal departure times differ across modes/routes; that is, each mode or route has its own optimal departure time. Thus, in empirical modeling applications, the travel time and reliability of each mode/route need to be specified under mode or route-specific optimal departure time. To summarize, route/mode choice behavior should be modeled using the indirect utility of the underlying departure time choice behavior with stochastic travel times. The integrated approach of Fosgerau and Karlström (2010) was motivated by this notion; thus, it exhibits a theoretical link between departure time choice and mode/route choice. However, to the best of our knowledge, no empirical study has been conducted to consistently estimate preference parameters across departure time and route choices.

Empirically estimating a simultaneous model for departure time and route choices has two major challenges. The first is in data collection. Constructing a realistic choice context is essential to reduce hypothetical bias in stated preference (SP) surveys. Recently, using the revealed preference (RP) to construct SP questions has been a major approach used to alleviate the bias. In this study, we use RP information obtained from the loop detector data of target road segments to improve the representation of travel time reliability. A similar approach was used by Small et al. (2005); however, the novelty of our survey is that information about the travel time distribution obtained from the RP data is used in multiple choice tasks, which allows for the simultaneous estimation of the departure time and route choice behavior. The second challenge is in the model estimation framework. As we discuss in the next paragraph, the expected indirect utility for each route/mode under the condition that the traveler leaves at the optimal departure time contains the so-called mean lateness factor, which is characterized by the standardized travel time distribution, and SDL and SDE parameters, which requires a special model estimation procedure. In this study, we propose a two-step estimation approach: in the first step, the expected value of the mean lateness factor is computed given all parameter values, and in the second step, scheduling preference parameters are updated using a conventional maximum likelihood approach given the expected value of the mean lateness factor obtained in the first step.

To empirically apply the proposed framework, we conducted an SP survey for urban expressway drivers in the Kinki region of Japan in July 2020. In the survey, each respondent was asked to answer SP questions in one of the following choice contexts: leisure, business, and logistic delivery. The choice context was determined based on the respondent’s daily use of the expressway. Each SP question contained three choice tasks: route A’s departure time choice, route B’s departure time, and route choice between A and B on the condition that the departure time choice was made. The travel time and its variability measured as a standard deviation of travel times could vary across the SP questions (we assigned six attribute levels to each SP question). We generated the probability distribution of travel times from the loop detector data of a particular expressway route segment to appropriately mimic real-world travel time variations. The use of loop detector data was also essential for the model estimation task, allowing for the construction of the mean lateness factor from the empirical data.

Using the collected SP data, we applied the proposed two-step estimation approach under both multinomial logit (MNL) and mixed logit (ML) model specifications. All the estimated parameter signs were logically understandable for all the models. Finally, we found that the estimated reliability ratios were 0.954 (MNL) and 1.037 (ML) for leisure trips, 0.945 (MNL) and 1.671 (ML) for business trips, and 1.417 (MNL) and 1.572 (ML) for logistic delivery trips. These results indicate that the reliability ratio varied depending on trip purpose, in addition to model specification.

ABSTRACT. Kindly see the PDF version for the correct display of mathematical equations and figures.

Word count: 974 Figures: 1 Tables: 1

ABSTRACT. Over the past few decades, there has been a trend for privacy preservation in wake of the proliferation of big data, reflected in public sentiments and legislation (e.g. the General Data Protection Regulation in Europe and the Data Protection Act in the UK). In accordance with these regulations, data owners/providers tend to perform suitable aggregation on the data so as to prevent individuals being re-identified from the datasets. One noticeable impact of the growing privacy concerns on the applications of big datasets involves the increasing difficulties to obtain from such datasets the socio-demographic information of respondents. This type of information is, however, essential for various analyses such as the modelling of behaviour patterns. In light of the prevailing benefits of big data sources, extensive attempts have been made to enrich the socio-demographic attributes of anonymous big datasets with different aggregation levels.

The transferability of most existing socio-demographic enrichment methods are found to be limited as they are unable to forecast their performance prior to the enrichment. Nor are they capable of interpreting the variation in their enrichment performance with respect to the change in the underlying data conditions (e.g. quality, variable types, sample sizes, correlation structure and contexts). As anonymous big datasets often retain various data qualities and aggregation levels due to their origin purpose of collection and the underlying privacy protection requirements, the ability to forecast and interpret the enrichment performance is particularly vital to ascertain the generalisability and transferability of an enrichment approach under different data conditions or across application contexts.

Amongst previous efforts, the Inverse Discrete Choice Modelling (IDCM) framework proposed by Zhao et al. is, to our best knowledge, the first attempt for enabling the predictability and interpretability of socio-demographic enrichment performance. This is achieved by systematically linking the IDCM enrichment performance with known information about the underlying data distributions. Specifically, Zhao et al. established an algorithm (named the IDCM performance theory) for the algebraical estimation of the socio-demographic enrichment performance using the IDCM approach in a doubly binary enrichment context.

Zhao et al. has verified that the aforementioned IDCM theory is capable to forecast the enrichment performance of an IDCM model estimated based on the estimation samples having similar data distributions with the anonymous big data to be enriched. In the real-world applications, however, the distributions of big datasets are less likely to be similar with the corresponding estimation samples because of the variation in the qualities and aggregation levels of the anonymous data, as well as the scarce availability of estimation samples. This hence inspires us to investigate the reliability of the IDCM performance theory for samples with different data qualities and aggregation levels, which is crucial as determining the transferability of the IDCM framework in more practical application contexts.

In the current contribution, we implement the IDCM performance theory in the enrichment for an anonymised dataset collected by a source different from the corresponding estimation sample. In particular, we employ the 2015/16 London Travel Demand Survey as the estimation sample, which informs sufficient socio-demographic attributes of the respondents in London and their stated travel choice behaviour patterns. On the other hand, we enrich the gender and age attributes of an anonymised mobile network dataset collected from London in October and November 2016 given their departure time-of-day choice behaviour extracted from the data. The aforementioned two datasets were collected from different sources and aggregated at different spatial levels, hence having different covariance structures in terms of the input behaviour features and the enriched attributes. Although different, these two sources still cover the same geographical area during a relatively close time period, thus avoiding, at least for the current study, issues associated with either spatial or temporal transferability. We explore the impact of different data distributions between the estimation sample and the anonymous dataset in the doubly binary context from the following two aspects: 1) Distribution of the enriched socio-demographic attribute(s) – this is revealed by the difference in the marginal probability distributions of the enriched attribute(s) in the two datasets; hence to understand the impact of data qualities, i.e. the quality of the prior information about the enriched attribute provided by the estimation sample. 2) Distribution of the choice behaviour variable – this is captured by the different marginal probability distribution of the choice(s) in the two dataset; hence to obtain the impact of data aggregation levels of the input choice behaviour.

In addition, we also compare the enrichment performance of the IDCM approach with representatives (logistic regression and support vector machines) of machine learning methods as they play the mainstream role in previous socio-demographic enrichment attempts. This would not only reveal the relative enrichment performance of the IDCM approach as compared to other methods, but also provide a comprehensive assessment for the transferability of the IDCM framework, i.e. whether it can be applied for predictable socio-demographic enrichment while preserving the privacy.

The presented findings are crucial in a number of ways. By assessing the performance of the IDCM approach in the enrichment for samples with different data distributions, we reveal whether the socio-demographic information of anonymous datasets can be enriched in light of different data qualities and aggregations due to privacy regulations. Furthermore, we explore the reliability of the IDCM theory for such enrichment, which, once verified, acknowledges the possibility of predictable and interpretable socio-demographic enrichment under the strict legislation. Additionally, the verified capability of the IDCM theory can provide a guideline for selecting the input choice behaviour patterns or an insight into which socio-demographic attribute(s) can be effectively enriched from specific input choice patterns. This is a considerably desirable property as the IDCM theory can be used to expect the enrichment performance based on different choice features without performing the enrichment. On the other hand, this can help in the design of an appropriate data collection protocol to minimise the cost. The enabled capability mentioned above would improve the usability of the extensively available big data sources without infringing privacy considerations.

ABSTRACT. 1 Introduction Revealing transport route choice behaviour and preferences among passengers in public transport systems is an important base for evaluating strategies to improve attractiveness of public transport. Such analyses require detailed information on the full trip performed by travellers from their point of origin to their final destination. Traditionally, such analyses have been performed using detailed travel survey data, e.g. [1]–[4]. However, such datasets are costly and limited wrt. sample size. Recently, more focus has therefore been on using automatically collected data from smart card-based automated fare collection systems (AFC). Such data are increasingly used within public transport planning and modelling [5], also for route choice analyses, e.g. [6]–[11]. However, one important drawback of using AFC data for route choice behaviour persists in all these studies, namely the lack of knowledge on the full journey since AFC data only includes trip segments within the public transport system. Hence, the data lacks information on the actual origin and destination, i.e. the access and egress segments of the journey. Not explicitly considering this might introduce bias in the estimation of route choice preferences. An example is when passengers have two options for their first (or last) segment of a trip and can choose to either i) take a bus to the train station, or ii) walk/cycle/drive to the station. If choosing i) then both options will be available in the choice set for estimating the route choice, whereas if choosing ii) then the choice set will not include option i), thus introducing bias to the model estimation. More specifically, this will introduce biases to the estimates of the in-vehicle times for the different public transport modes where the value of in-vehicle time for bus might be under-estimated. This study contributes to existing literature within transport route choice in two important aspects. First, by highlighting the potential biases obtained when using AFC data for estimating route choice models in multimodal public transport systems. Second, by proposing a method for estimating route choice models based on AFC data, which reduces estimation bias. 2 Methods and data For estimating behavioural preferences from transport route choice a traditional two-stage estimation process is applied. This involves i) choice set generation of relevant alternatives to the actual observed choices, and ii) route choice model estimation. The difference between using AFC data and traditional travel survey data lies in the first step. Most previous studies using AFC for route choice analysis have simply generated alternative choice sets based on the revealed stop pairs chosen by the traveller, i.e. points of tap-in and tap-out of public transport, and thus neglected potential access and egress travel between the chosen stops and the actual origin/destination, as well as possible alternatives at nearby stops in the vicinity of the origin and destination stops. This study instead suggests to explicitly consider that the point of origin (and destination) is not the chosen stop, but rather a point some distance away from the chosen stop. As this point is not known when using AFC data, we propose to simulate random points in the vicinity of the chosen stop, e.g. within a 1,000 meter radius, cf. Figure 1. These points are used for calculating (pseudo) access and egress distances (and travel times) to stops in the choice set. The model estimation was hereafter performed using Monte Carlo simulation in PandasBiogeme [12]. The methodology was tested on case study data from the Greater Copenhagen area consisting of 4,810 revealed preference multimodal public transport trips from the Danish travel survey, as was originally also used in [1], [2], [13]. For each observed route choice a number of alternative routes in the multimodal public transport network was generated, which i) used either the chosen stop or a stop in close proximity, and ii) at or close to the chosen departure time. For more details on the choice set generation we refer to [14]. The comparison to estimation based on full knowledge of the entire trip was done by gradually restricting the choice set of alternative routes for each model estimation. In the simplest estimation replicating raw AFC data routes were restricted to be between the actually chosen stops only, whereafter multiple model estimations were done with gradually allowing stops located within certain distances from the observed stop in the choice set. 3 Results The main results show the accuracy of the parameter estimates, i.e. the ratio between the parameter estimate and the true parameter estimate, with the latter being defined as the parameter estimate when estimating the model based on the full choice set, and knowledge on the door-to-door travel rather than tap-in to tap-out, cf. Figure 2. Hence, the ratios representing full data are always equal to one. The results show that all parameter estimates were highly biased when simply excluding access/egress from the model formulation. Adding parameters for (pseudo) access and egress, based on the random origin and destination points, the parameter estimates are still notably off compared to the true estimates. Accuracy improves by including stops in close proximity, and becomes stable when including stops within 500-1,000 metes. Thus, the results suggest that it is very important to include not only routes between the observed stops, but also routes between stops in the vicinity of the chosen stops when estimating route choice models based on AFC data. Even if parameter estimates for the pseudo access/egress are off, they are important to include to improve accuracy of remaining parameters. However, even though the biases are reduced for in-vehicle times, there are still very large biases related to transfer walking time and hidden waiting time. Next, models will be formulated in which route choice is a conditional probability of choosing the observed OD-stops, thus incorporating differences in stop choice attractiveness – and in which stop choice is dependent on service levels between OD-pairs. These will be expressed through a nested logit model which include feedback mechanisms from the lower levels (the route choice) to the top level (stop choice).

ABSTRACT. Introduction

The purpose of this paper is to build and investigate a Discrete Choice Model using neural networks with non-linear activations. The final model is built in three steps from three sub-models, each learning one of activity, destination or mode. Tests show how additional non-linearities increase performance and can simulate activity-based travel demand on a very detailed level.

Traditionally, discrete choice has successfully been modeled by the Multinomial Logit Model (MNL). While interpretation of variables are convenient, non-linear functions are challenging. If, for example, non-linearities were to be explored, countless non-linear combinations would have to be tested such as $x_1*x_2$ or ${x_1}^2$. The impracticality of this limits the modeler to only test a subset of possible combinations. These would presumably be based on educated guesses or by using some pre-existing knowledge of their forms, thus introducing specification bias.

With neural networks receiving increased attention, \textcite{MNLvsNN_2000}, compares the two methods with respect to Discrete Choice modeling and covers both differences and similarities. In fact, the MNL has a neural representation if the network uses a log-linear output node, also known as a softmax, and is removed of any hidden nodes.

Neural networks have, however, a harder time explaining the meaning of every parameter, but when this is acceptable, it can provide multiple beneficial additions. Particularly, neural networks achieve non-linear mappings between its inputs and outputs. Besides this, the network is also constrained to find the most descriptive characteristics in the data given the nodes and parameters available. This automatic search for the very best combinations of non-linear characteristics prevents the specification bias present in the classical MNL. This feature also acts as a lever that gives the modeler control. If for example data size increases, more nodes and parameters can very easily be added to the network to allow for increased complexity, while still performing equally as well on unseen data.

Data

Each participant/ID has recorded every trip taken during a day between 05:00 am and 11:00 pm. A trip is here a combination of one of the following attributes in each category:

Activities: Home, Work, Social, Recreation, Other and Shopping. Zones: 1240. Modes: Walk, Bike, Public Transport (PT), Car.

The following land use data was also used:

Land use: Total Population, Total Employment, Employment Retail, Employment Restaurants, Employment Recreation.

Time tables data for each mode was also available for every trip between any of the 1240 zones.

The Model

The model is separated into the three following steps; chosen activity, chosen location and chosen Mode, modelled by Sub-Model 1, 2 and 3 respectively. While being very different in how it operates, the model share several features with the dynamic MNL model Scaper, developed in ref1.

Sub-Model 1

Sub-Model 1, shown in figure Sub_Model1, learns whether the ID in the next time-step choose Stay and keep on doing the activity in the same location, or selects any of the other six activities, meaning a trip will happen where the selected activity will be performed at the destination.

Sub-Model 2

Sub-model 2, shown in figure Sub_Model2, learns the distribution over destinations given the activity to be performed and zone specific information containing Mode times distribution and land use data. Walk and bike times were both a function of distance, which made their normalized input to sub-model 2 identical and only one input node was here necessary for these two modes. Noteworthy is the correlation layer, shown in figure Sub_Model2, which is not only reducing attraction to a zone linearly based on each mode's travel time, but is forced to learn substitution willingness between modes in order to output the correct attraction to each zone. Another important addition is the Reduced time from Zone x to mandatory module, which inputs how the distance to the mandatory zone improves by visiting a zone. The mandatory zone is defined as work when not been to work, and home when been to work.

Sub-model 2 was tested with three different architectures, each adding additional non-linearities. These three architectures are shown in figure Sub_Model2_arch1, figure Sub_Model2_arch2 and figure Sub_Model2_arch3 in the Appendix. Table arch_results shows their results.

Finally, Sub-model 3 shown in figure Sub_Model3, learns the choice of mode to the selected zone.

Simulation

After training, the entire model with all sub-models are evaluated, by sampling day paths for every ID in the left out validation data set. Each ID begin in their respective home ID zone performing the home activity. A flow-chart of the simulation procedure for each time-step is shown in figure model_flow_chart. A spatial visualization of activity-destination probabilities is shown in figure visual_id1 in the Appendix.

Evaluation Metrics

Daily Activities and trips

For 1000 sampled paths, activities per day and minutes travelled per day are summarized from both ground truth data and model's sampled paths. These results are shown in table per_day}.

Activity Distribution

Figure GT_act_dist and figure model_act_dist show how activities are distributed between 05:00 and 11:00 pm in the ground truth data and simulated model paths respectively.

Modes trip lengths in minutes

Trips lengths were gathered from each sampled trajectory and the distribution of these are visualized in figure mode_times_summary.

Applications

By letting the model simulate people's travel patterns, it can be used to predict how much time people need to spend travelling and give information on traffic load.

The model also has promising properties to simulate how people change their daily travel pattern as a reaction to a change in the travel network's mode times from for example a new sub-way line or changed road speed limits.

Since no unique output bias exist for any zone, its probability is entirely based on land use data and mode times. This makes the model able to predict well on cities even after land use data is changed.

The utility values before the softmax output layer reveal the overall travel utility in the city. These utilities can be analyzed from different input features and could therefor be used to evaluate whether travel utility increase or decrease from an intervention in the city.

ABSTRACT. Introduction Proportional Odds Models (POM) or Ordered Logit Models is the most traditional and popular approach for modeling ordinal problems where an unobserved continuous variable relates to ordinal responses through thresholds. In recent years, the promising performance of machine learning algorithms, especially, Deep Neural Networks (DNNs) provide an opportunity for modelers to achieve more flexible models structure with more accuracy of prediction. Despite achieving highly accurate prediction tools using neural networks, there has been a widespread debate over the restrictions of DNN algorithms among modelers and researchers, such as the lack of model interpretability, statistical analysis, extracting behavioural indicators, etc. Recently, Wong and Farooq proposed a novel deep learning version of the Multinomial Logit model (MNL) employing Residual Neural Networks or ResNets architecture. The resulting RestLogit model captured the underlying unobserved heterogeneity, where the residual layers of the deep network, represented the influential unobserved behaviour, estimated independently by means of skipped connection. To evaluate ordinal problems, some researchers exactly focused on the structure of threshold models and some others tried to use classification algorithms in the literature. Broadly, in some classification problems, a popular reduction framework was used in which the ordinal regression problem is transformed into multiple binary classification sub-problems. However, this framework depends on misclassifications cost matrix which should be priory defined by modelers and also satisfy the convex condition to achieve a rank-monotonic threshold model. Inconsistency among binary classifiers tasks is the main problem of some theoretical algorithms proposed for ordinal data. To address this problem, Cao et al. recently developed a deep learning ordinal regression approach called COnsistent RAnk Logits (CORAL) framework. Our study presents a new version of ResLogit specifically designed for ordinal data. We define this model structure as an Ordinal-ResLogit model. Precisely, the Ordinal-ResLogit model combines the CORAL framework with the ResLogit so as to propose a deep learning version of the threshold model.

Ordinal-ResLogit Model Structure The proposed learning-based ordinal regression model inherits the theoretical concept of both threshold models and binary classification algorithms. The deep neural network is applied only at the residual layers, instead of the whole structure of the model by using ResNet architecture with the intention of keeping the deterministic term of the utility function of traditional discrete choice models and also capturing random heterogeneity from the data. Therefore, in the Ordinal-ResLogit model, the unobserved utility of ordinal categories for individual n, consists of three components: 1) deterministic component consisting of explanatory variables, 2) residual component capturing random heterogeneity and 3) error term capturing remaining unobserved errors. The final layer of our proposed model is fed into the CORAL framework, which consists of K-1 binary classifiers. The Ordinal-ResLogit is illustrated in Figure 1. The main purpose of CORAL structure is to guarantee consistency among multiple binary classifiers using the same weight parameters (Wk), for K-1 binary classifiers. In general, the Ordinal-ResLogit model provides an opportunity to capture heterogeneity from the data and exploit the interpretability of the ResLogit structure, which is rooted in the operation of residual layers. In fact, although the deterministic part of utility function is constant among ordinal categories, the residual matrices related to residual layers consider the correlation between the categories. It is noticeable that although the CORAL framework can be classified as a binary classification method, its structure does not require a misclassification cost matrix. As in ordinal problems the cost values are not the same for different prediction errors and the formulation of the cost matrix strongly influences the performance of the model; being free from constraints of the reduction framework is the main advantage of CORAL framework leading to a more reliable approach for evaluating ordinal problems.

Case Study and Result To evaluate the performance of the Ordinal-ResLogit model, the ordinal categories of pedestrian wait time in the presence of automated vehicles on the sidewalk are investigated by using a virtual reality dataset. The data give us an opportunity to analyze futuristic scenarios and comprise of crossing information related to street characteristics, traffic condition and environment situation as well as socio-demographic and travel patterns. Our results demonstrate that the Ordinal-ResLogit model outperforms the conventional ordered logit model in terms of both accuracy and considering significant variables. Consistency among binary classifiers is another advantage of our proposed model. To the best of our knowledge, the Ordinal-ResLogit model is the first deep learning-based ordinal regression model that maintains model interpretability and allows us to statically analyze the deterministic part of the utility function. Regarding explanatory variables, our proposed model is able to consider the highly significant impact of more variables for the prediction of pedestrian waiting time in comparison to conventional ordered logit model. Results of the Ordinal-ResLogit model further confirm that mixed traffic conditions increase the pedestrian propensity to choose longer waiting times. However, a converse result is obtained for fully automated conditions. In addition, walking habit, gender, age group, lane width, road density and road type are estimated as contributing factors in the prediction of pedestrian waiting time. The Ordinal-ResLogit model as an interpretable deep learning-based behavioural model provides the feasibility of analyzing ordinal problems with more accuracy as well as enjoying the structure of traditional discrete choice models. Similar to previous studies, our proposed model highlights that deep learning algorithms and discrete choice modelling can be complementary methods, despite inherently having different performance.

ABSTRACT. 1 Introduction Including richer data in Discrete Choice Models (DCM) has been identified as a crucial task for promoting future transport research [1]. The default method of encoding categorical variables, one-hot encoding, poses constrains to the number of explanatory variables that can be included. As it adds an extra variable per category, the model’s complexity is increased proportionally to the cardinality of the categorical variables considered. This can pose severe challenges in statistical modeling with an exponentially increased sample size requirement to avoid problems such as overfitting or poor parameter estimation [2,3,4,5]. Given that travel data collection is considered the most expensive and time-consuming part of the transportation model development process [6], rather than increasing the sample size, a more efficient treatment would be to find alternative methods for encoding categorical variables, that would allow for more compact yet informative representations of the categorical data. We address that aspect in the current study by incorporating a data-driven neural network (NN) approach for encoding categorical variables into continuous vector representations, called Embeddings. Compared with previous work, we (i) estimate jointly the embedding weights and the parameters of the DCM (ii) enforce interpretability to the representations by formally associating each of their dimensions to a choice alternative. The proposed model not only improves thepredictive performance compared to benchmark models but also provides rich, behaviorally meaningful outputs that can be used in travel demand analysis and policy decisions. (PLEASE SEE THE ATTACHED PDF FILE FOR CONTINUING READING THE EXTENDED ABSTRACT)

ABSTRACT. see pdf

ABSTRACT. The extended abstract has been submitted via a pdf file. Please find the file attached.

ABSTRACT. [Please see attached pdf] Introduction The transport sector is a major emitter of greenhouse gases, accounting for 15% of all human emissions in the world during 2016 (Ritchie & Roser 2020). Focusing on passenger travel, car is the most polluting mode, and reducing its use would not only diminish emissions but also increase the quality of life in urban areas (Rojas-Rueda et al. 2012). As emissions are roughly proportional to the distance travelled in each mode, efficient environmental policies should not only aim to reduce car’s modal share, but also to reduce the total amount of vehicle-kilometre travelled (VKT) by car. The Multiple Discrete Continuous (MDC) modelling approach has been used in the literature to simultaneously predict what modes individuals use, and how much they travel on each of them (Bhat & Sen 2006, Fang 2008). Within this family of models, Bhat’s MDCEV (Bhat 2008) is among the most popular. Yet it has limitations: (i) it assume the total distance travelled (i.e the travel budget) to be an exogenous variable, and (ii) it ignores complementarity and substitution effects. The first limitation means that out-of-sample forecasting requires two stages: first an auxiliary model to predict each individual’s travel budget, and a second stage distributing that travel budget among different modes using the estimated MDCEV. This process is inefficient, and makes it difficult to calculate confidence intervals for the forecast. The second limitation implies that if one mode becomes less (more) attractive, the distance travelled by all other modes increases (decreases) proportionally, without any particular mode benefiting (reducing) more than the others. This is not realistic, as public transport modes are probably closer substitutes of each other than car. To avoid these limitations we use the extended MDC model (eMDC, Palma & Hess 2021), for which the total distance travelled is endogenous (i.e. does not need to be defined a priori), and incorporates complementarity and substitution effects. We compare our results to those of a traditional MDCEV model. Data sources We studied the first two phases of the MOBIS dataset (Molloy, 2021; Hintermann et al. 2021). Each phase records the mobility of 3680 individuals in Switzerland during four weeks, collected through GPS tracking between September 2019 and January 2020, before the Covid-19 pandemic in Switzerland. While the data collection included splitting the sample into control and treatment groups, none of the treatments were found to have a significant effect in this study using a simple modelling approach, and are therefore ignored. We only considered individuals with full socio-demographic information in our analysis. We aggregated the distance travelled at the week level, keeping only full weeks. We considered six travel modes: walking, bicycle, bus, tram, train, and car. Table 1 summarises engagement (i.e. percentage of sample using each mode) and average distance travelled when those modes were used. Methodology MDC models are based on consumers’ classical problem of utility maximisation, as stated in equations (1) and (2). Where xnk represents the distance travelled by individual n on mode k; while uk and ukl represent additively separable, and non-additively separable utility functions, respectively. Bn is the maximum distance individual n could have travelled during the week under analysis (their travel budget). The formulation includes an outside good (indexed as zero) representing the additional distance the individual could have travelled, but decided not to. The MDCEV model assumes ukl = 0, preventing any complementarity or substitution effects. We assume Bn to be equal to the observed total distance travelled, making xn0 = 0 and dropping out of the formulation. Bn must be predicted with an auxiliary model when forecasting out of sample. The eMDC model assumes〖 u〗_kl=δ_kl (1-e^(-x_nk ) )(1-e^(-x_nl ) ), where δkl > 0 implies complementarity, δkl < 0 substitution, and δkl = 0 no significant interaction. It also assumes u0 to be a linear function, causing Bn to drop out of the model likelihood and forecasting algorithm, meaning that no auxiliary model is needed when forecasting out of sample. Results and conclusions Models were estimated on 70% the available data, the rest being used to measure out-of-sample fit. Only socio-demographics were used as explanatory variables. The same formulation of uk was used in both models, but the eMDC incorporates additional parameters inside u0 and ukl. Full estimation results are available online. We used an auxiliary linear regression with all available socio-demographics as explanatory variables to predict Bn when forecasting with the MDCEV (R2=0.04). Table 2 presents the RMSE of the out-of-sample forecast of each model for each mode, as well as the total distance travelled. The eMDC achieves consistently smaller errors, except for bus. This is to be expected, as the eMDC has more parameters, nevertheless, both models use exactly the same amount of information. This reveals that complementarity and substitution patterns have a significant effect in individual’s behaviour. Indeed, the model captures complementarity or a non-significant interaction between all active and public transport modes. Car, on the other hand, present substitution with all other modes, except for walking. This makes sense as people who don’t drive tend to use a combination of modes depending on the trip, while drivers tend to rely on their car for all their mobility needs, except for very short trips. Future work We will compare the performance of the eMDC against newly developed MDCEV formulations with unobserved budgets, and investigate preference heterogeneity across individuals in the sample.

ABSTRACT. The Covid-19 pandemic has influenced how individuals engage in activities, including work-related activities. During the peak of the pandemic commute trips were often replaced by telecommuting; and, in particular, a notable shift from shared modes (public transit, ridehailing etc.) to private cars was observed. Two main mechanisms leading to strong shifts in individuals’ behavior were: a) restrictions imposed by governments, closure of wor and public places, and b) individuals’ fear of contracting infections. In either case, individuals may have gained new experiences with activities and travel modes that differ from their pre-pandemic patterns and habits. As the U.S. (the geographic scope of this study) emerges from the pandemic, external restrictions on activities and travel can be gradually lifted, and personal fears of contracting infections can be somewhat mitigated by the mass availability of vaccines. An important question for the future: to what degree will pandemic-induced changes in individuals’ activities and mode choices persist in the post-pandemic “new normal”?

In the US, Covid-19 was declared a pandemic in March 2020. The period of March 2020 to January 2021 saw spells of curfews being imposed, closure of schools, workplaces and restaurants at different times and parts of the country (depending on the local spikes in cases). Vaccines were made available to the general public by February 2021, and by July 2021 a significant proportion of the country was vaccinated. The middle of 2021 also saw several external restrictions being eased up.

Shortly after the start of the pandemic, researchers at UC Davis launched a panel survey to explore these issues. The UC Davis collected data regarding individuals’ engagement in activities, attitudes, home- and work-locations, and socio-demographics using three waves of surveys. This data set will be augmented with additional variables from external sources on the timeline of Covid cases and lockdown measures imposed in geographical locations of respondents taking the surveys. Wave 1 was conducted in May-June 2020; Wave 2 was conducted in December 2020 – January 2021; and the last survey, Wave 3, was conducted September – October 2021. In each wave, the respondents were asked to report their behaviors (engagement in activities, mode use) both currently and retrospectively. To preserve clarity, respondents were asked to report their current behavior at a specific point in time, plus retrospective data for one year earlier. For example, in Wave 2 of the survey the respondents were asked to report their current behaviors (Nov-Dec 2020) and retrospective behaviors from one year ago (Nov-Dec 2019). In this way, we collected behaviors at four points in time: November-December 2019 (retrospective from November-December 2020)), March-April 2020, November-December 2020, and June-July 2021.

Surveys were conducted online and took 30-40 minutes to completed. Respondents in Wave 1 were recruited using commercial online opinion panels and through social media platforms (email chains, Facebook etc.). Recruitment through commercial online opinion panels was focused on recruiting respondents from metropolitan areas in the U.S. (rather than focusing on entire country), so our sample is skewed toward city dwellers. The same respondents were invited again for Waves 2 and 3. In total, 1,398 individuals completed all three waves of the survey.

Our analysis focuses on a discrete-continuous outcome measure for mode choice to work, where working from home via telecommuting is treated as a potentially available mode in addition to physical commuting using private cars, active modes and public transit/other shared modes (e.g., ridehailing, shuttles and shared scooters), using individuals that remained employed across all four Waves (n=692). The subsample is therefore likely to be skewed toward individuals in certain job categories. However, our main concern is on identifying dynamic effects related to behavioral change.

Table 1 reports summary statistics on mode choice and usage for the four time periods. The mode choice measure is the percentage of individuals adopting the mode at least once during the period, and an individual’s intensity of mode use is measured as the number of times the mode is used per month. There is a clear pattern in changes of mode usage in response to the pandemic timeline. There is a large increase in telecommuting, and a concomitant drop-off in other modes in response to the start of the pandemic in March 2020. As the pandemic evolved, telecommuting dropped marginally, and use of active and private modes recovered closer to pre-pandemic levels. However, recovery of transit/shared mode usage was more limited. Our objective in this study is to use the individual-level panel data to develop dynamic behavioral models to identify factors that explain the changes in these patterns.

Our approach uses latent class segmentation to capture unobserved heterogeneity mode choice and usage via class-specific multiple discrete continuous extreme value (MDCEV) models, as in Sobhani et al. 2013. Models are extended to address panel data as in El Zarwi et al. (2017), so the latent class structure suggests a Hidden Markov Model (HMM) framework. The focus is on estimating models for an individual’s transition probabilities among classes when going from period t to t+1.

Finally, in contrast to most current models, we explore model extensions that identify feedback effects, e.g., if mode choice/usage experiences due to pandemic-related factors in time t have an effect on latent class transition probabilities between periods t and t+1, and the degree to which these effects might persist post-pandemic.

Table 1: Distribution in the changes in telecommuting and use of modes for physical commute at four points of observation for 692 employed respondents in the dataset

Nov-Dec March-April Nov-Dec June-July 2019 2020 2020 2021 Work from home % adopters 25.69% 77.75% 74.31% 75.14% mean of days per month 13.24 18.77 17.96 15.36

Car commute % adopters 82.97% 35.98% 44.44% 61.42% mean of days per month 8.96 7.85 8.23 8.23

Active commute % adopters 39.39% 12.43% 14.14% 21.10% mean of days per month 6.06 4.61 5.95 5.15

Transit commute % adopters 41.27% 10.26% 13.28% 17.77% mean of days per month 8.09 9.49 9.04 7.06

ABSTRACT. Single-leader car-following models, where the subject vehicle is assumed to follow a single lead vehicle ahead of it, have been extensively used to model driving behaviour in both homogeneous and heterogeneous disordered (HD) traffic conditions. Further, most literature in this area focuses on the longitudinal movement of vehicles; albeit lane-changing models consider lateral movements separately from longitudinal movements. However, emerging literature (as well as some early research) argues that assuming a single leader’s influence (Bexelius, 1968; Hoogendoorn and Ossen, 2006) and/or separating longitudinal movements from lateral movements (Kanagaraj and Treiber, 2018; Sarkar et al., 2020) may not adequately represent driver behaviour. Particularly, in the context of HD traffic conditions, drivers’ manoeuvres are not only restricted to vehicle-following behaviour but also exhibit characteristics such as staggered following, passing behaviour, gap-filling, and following between two vehicles (Asaithambi et al., 2016). Additionally, current traffic flow models lack due consideration of human factors such as multi-vehicle anticipation. Therefore, many of these behaviours in HD traffic conditions can possibly be better represented by considering two-dimensional movements while also considering influences (or stimuli) from multiple vehicles in the vicinity of the subject vehicle. Further, the extent of acceleration and two-dimensional movement executed by drivers while travelling on the road are usually a consequence of drivers’ intentions to accelerate, decelerate, or maintain constant speed and to steer left, right, or straight along the longitudinal access. However, small acceleration/deceleration values and slight angular deviations from the direction of traffic flow are common even if the driver intends to maintain a constant speed state and to move in the straight path, respectively. This is due to the difficulty in maintaining a constant speed and avoiding any lateral movement (i.e., angular movement). However, the actual intentions of drivers are unobserved or latent to the analyst, and only the outcome of driver’s actions (such as the extent of acceleration and angular deviation from the direction of traffic flow) can only be observed. This makes it important to take drivers latent intentions into account while analyzing driving behaviour to improve the realism of driver behaviour models. Motivated by the above discussion, this study proposes a novel, two-dimensional, multi-vehicle anticipation, and multi-stimuli based latent class framework for modelling microscopic driving behaviour in HD traffic conditions. Specifically, the following non-trivial extensions are proposed to a typical stimulus-response based driving behaviour framework: 1. two-dimensional movements of the subject vehicle are modelled, as opposed to only longitudinal movements, represented as a combination of the angle of movement with respect to the longitudinal access and the extent of acceleration or deceleration along the angle, 2. a latent class framework is used to incorporate driver’s latent intentions along two dimensions – (a) the intent to accelerate, decelerate, or remain in constant speed and (b) the intent to keep in the left, right, or straight (longitudinal) direction, 3. multi-vehicle anticipation is accommodated, wherein the driver is assumed to consider the movements of multiple vehicles in its influence zone to make higher-level driving decisions (intents), 4. a multi-stimuli acceleration\deceleration model is formulated based on the theory that drivers choose an angle of movement that offers them the highest possible longitudinal acceleration (if they intend to accelerate) or allows them to move with the least possible deceleration (if they intend to decelerate), 5. and drivers’ execution errors are modelled as the difference between their intended acceleration or deceleration extents and the observed values. To do so, subject vehicles’ latent intentions and two-dimensional movement are represented in the form of three components: (a) latent intent model (class membership model) (b) latent intent-specific choice model (class-specific choice model), and (c) extent of acceleration and deceleration model. The latent intent model formulates the probability that a driver intends to carry out a particular manoeuvre latent to an analyst. The latent intents considered in this study comprise a combination of whether to accelerate, decelerate or maintain a constant speed and the steering direction (i.e., steer in the straight (longitudinal) direction, steer left to longitudinal axis, or steer right to longitudinal axis) – a total of nine possible intent combinations. To probabilistically relate each of the three intended steering directions for the subject vehicle with observed angle (with respect to longitudinal axis) of movement, the two-dimensional roadway space ahead of the vehicle is viewed as several radial cones. For a given combination of accelerate/decelerate and steering direction intentions, it is assumed that the driver would select a specific radial cone that allows the driver to accelerate (decelerate) as much (low) as possible in the longitudinal direction. The driver attempts to manoeuvre in a particular direction with the calculated maximum (minimum) longitudinal acceleration (deceleration). However, due to the driver’s vehicle capability, driving skill, machine error, and other unobserved factors, the observed acceleration might be different from the calculated acceleration. This difference is labelled the execution error. Therefore, the proposed model considers execution error while modelling the extent of acceleration and deceleration. Moreover, multi-vehicle anticipation is included in the modelling framework using the concept of an influence zone around a subject vehicle (Nirmale et al., 2021) where it is assumed that vehicles within the influence zone influence the subject vehicle’s driving behaviour. Properties of the multivariate normal distribution are employed to derive the joint log-likelihood function. An empirical application of the proposed framework is presented on an HD traffic conditions trajectory dataset from Chennai, India, for two-wheeler motorcycles. Proposed modelling is employed to simulate HD traffic conditions. Then, simulated HD traffic streams is evaluated using macroscopic properties traffic streams, i.e., speed-flow-density relationships. The empirical results underscore the importance of incorporating multi-vehicle anticipation and latent intents while modelling driver’s two-dimensional movement in HD traffic conditions. The results also suggest that drivers’ higher-level intents are strongly influenced by the microscopic traffic environment variables than their microscopic decisions of which specific angle to move along and how much to accelerate or decelerate. The proposed framework and the insights from this study might inform improvements to existing approaches to simulate two-wheeler movements non-lane-based HD traffic streams.

ABSTRACT. 1 Introduction The presence of Autonomous Vehicles (AVs) on urban roads will lead to drastic changes in transport systems. To plan, design, develop and prepare for the transport systems of the future, understanding the public’s opinions, attitudes and perceptions towards these new modes is of vital importance. Traditionally, stated preferences (SP) surveys have been used to collect data on attitudes and perceptions towards innovations and new technologies. SP surveys, however, can be limiting in terms of sample size and response rate, time, cost, social desirably and hypothetical bias. The growth of social media platforms in the past decade, on the other hand, has allowed billions of users around the world to express their opinions and emotions about various events or topics, or log their daily activities or feelings. This has provided a massive source of data on different behavioural aspects of users, which can be tracked over time to understand the attitudes and perceptions of users toward certain events, topics, or concepts. Furthermore, the global usage of social media platforms has made it possible to access data from users of different countries, which is of particular interest when the heterogeneity among users might play a role in their attitudes and preferences. 2 Objective In the context of attitudes towards new modes of transportation, global heterogeneity in public perception and preferences can lead to different behaviour towards the new unseen modes in different parts of the world. This study utilizes data from social media, in particular Twitter, to explore the global heterogeneity in the preferences and perceptions of potential AV users. By extracting tweets related to AVs from English-speaking users located in Australia, Canada, the United Kingdom and the United States, we attempt to analyze the general sentiments of the public in these countries towards AVs and provide insights into their similarities and differences. 3 Background Several studies have used traditional surveys and have shown significant differences in the attitudes towards AVs among people from different countries (Continental (2018); Hudson et al. (2019); Potoglou et al. (2020)). However, using social media data to understand the attitudes towards AVs is gaining popularity in recent years. Sentiment analysis methods have been applied to tweets (Ding et al. (2021); Bakalos et al. (2020); Wang et al. (2021); Trivyza (2021)), Reddit comments (Bakalos et al. (2020); Chen et al. (2021)), and Facebook comments (Chen et al. (2021)) to estimate the public opinion towards AVs. Bakalos et al. (2020) developed their sentiment analysis using Reddit and Twitter data to assess public perception of AVs. By defining a set of relevant keywords, the authors extracted around 8000 tweets and 500 Reddit posts. The extracted posts were then used to categorize tweets and comments to positive and negative public opinions and provide confidence levels in AVs based on the sentiments in the texts. Wang et al. (2021) expanded the number of tweets analyzed to around 1 million tweets. Taking sentiment score as the dependant variable, the authors developed a linear mixed model to explore the factors impacting the public opinions. In another study, Ding et al. (2021) extracted and scored around 600,000 tweets over a three-month period, and using time series analysis, they defined the critical points that shaped sentiment changes. Chen et al. (2021) used Facebook, Twitter, and Reddit data to analyze the sentiments about AVs after the Uber and Tesla crashes, as well as the Covid-19 pandemic. In their study, they manually extracted sociodemographic of users, including their age, gender, occupation, and location. Despite the recent growing interest in using social media data to assess public opinions towards AVs, the comparison of attitudes towards AVs in different countries has yet to be explored in the literature. Availability of location information in a portion of tweets makes the cross-country comparing possible, which provides insights into the future market shares for AVs and adoption of them. 4 Methodology and expected results By defining a set of words relevant to AVs, and by specifying the desired location, language, and time frame of the tweets, tweets of interest are extracted using web scraping methods. After pre-processing and cleaning the retrieved data, sentiment analysis techniques are used to quantify feelings, emotions or opinions and rank the tweets retrieved from each country from mostly negative to mostly positive (Kharde et al. (2016)). The sentiment towards AVs is compared among the understudied countries, and the spatial and temporal variations are highlighted. Comparison of the findings with the results based on analyses of survey data is expected to provide interesting insights regarding AV penetration rates in different parts of the world.

References Bakalos, N., Papadakis, N., Litke, A., 2020. Public perception of autonomous mobility using ml-based sentiment analysis over social media data. Logistics 4, 12.

Chen, X., Zeng, H., Xu, H., Di, X., 2021. Sentiment analysis of autonomous vehicles after extreme events using social media data, in: 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), IEEE. pp. 1211–1216.

Continental, 2018. Where are we heading? paths to mobility of tomorrow: The 2018 continental mobility study. https://cdn.continental.com/fileadmin/ __imported/sites/corporate/_international/english/hubpages/10_ 20press/03_initiatives_surveys/mobility_20studys/2018/mobistud2018_ 20studie_20pdf_20_28en_29.pdf. Accessed: 2021-11-20.

Ding, Y., Korolov, R., Wallace, W.A., Wang, X.C., 2021. How are sentiments on autonomous vehicles influenced? an analysis using twitter feeds. Transportation research part C: emerging technologies 131, 103356.

Hudson, J., Orviska, M., Hunady, J., 2019. People’s attitudes to autonomous vehicles. Transportation research part A: policy and practice 121, 164–176.

Kharde, V., Sonawane, P., et al., 2016. Sentiment analysis of twitter data: a survey of techniques. arXiv preprint arXiv:1601.06971 .

Potoglou, D., Whittle, C., Tsouros, I., Whitmarsh, L., 2020. Consumer intentions for alternative fuelled and autonomous vehicles: A segmentation analysis across six countries. Transportation Research Part D: Transport and Environment 79, 102243.

Trivyza, M.F., 2021. Autonomous vehicles: Multi-class twitter sentiment analysis .

Wang, S., Li, M., Yu, B., Bao, S., Chen, Y., 2021. Investigating the impacting factors on the public’s attitudes towards autonomous vehicles using sentiment analysis from social media data. arXiv preprint arXiv:2108.03206 .

ABSTRACT. Please see the pdf attached.

ABSTRACT. 1. Introduction Autonomous vehicles (AVs) are expected to drastically affect travel behaviour and travel systems [1-3]. The rate and the extend of the effect depend on the adoption rate of AVs in the market. Previous studies have shown the adoption process of AVs depends on the operational features of this innovation (e.g. price) [4], customers demographics [5, 6], and the perceived benefits and concerns [7-10]. While the significant role of psychological constructs in the adoption decision is acknowledged [8, 11, 12], little is known about how these psychological constructs are formed or altered. Ghasri and Vij [13] conducted a stated-preference survey to understand the potential role of informational cues on attitude formation towards AVs. This study will continue the study by Ghasri and Vij [13] in three directions. First, this study will use SP data from Ghasri and Vij [13] to develop a model of the influence of social media sentiment on willingness-to-pay (WTP) for automation. Second, it will scrape tweets that include keywords about autonomous vehicles, to understand how social media sentiment on autonomous vehicles has changed over time. Third, this study will use these sentiments as inputs to the behaviour model developed in the first step to examine how willingness to pay has changed over time. Consistent with previous studies, we assume the decision of purchasing AVs depends on vehicles attributes and consumers’ attitudes. The novelty of this study is including attitude formation in the behavioural model and combining it with the scraped sentiments from social media, which enables us to calculate temporal variations in WTP following the changes in sentiment on this technology. This approach can estimate the dollar value for positive or negative informational cues. This study contributes to the literature by developing a framework to model correlation between social media sentiment and willingness to pay with regards to any good or service. Subsequently, we show how this framework can be used in practice to track social media sentiment over time to understand how willingness to pay has changed in the past, and to predict how it may change in the future. 2. Methodology The methodology of this study consists of two parts: (i) formulating WTP for automation as a function of received informational cues, and (ii) quantifying social media sentiments. 2.1. Willingness-to-pay and informational cues This study assumes the utility of selecting AVs is formulated as a linear function of vehicle attributes, and decision makers’ attitudes towards AVs. Attitudes towards AVs are assumed to be formed based on the received information. We assume a linear construct between received informational cues and attitudes towards AVs WTP for automation is defined as the ratio of marginal utility of automation over the marginal utility of price. The marginal utility of price will be the estimated coefficient for price. The marginal utility of automation is derived from two components of attitudes towards AVs and the alternative-specific-constant of automation. Therefore, the WTP for automation will be a function of informational cues. 2.2. Social media sentiment analysis Twitter as one of the most widely used and popular social media platforms has been used in this research for data collection. Sentiment of each tweet text will be obtained using the Syuzhet method [14]. This method is developed in the Nebraska Literary Lab. The default dictionary of Syuzhet comprises 10748 words with associated sentiments spanning between -1 and 1. This dictionary is extracted from a collection of 165,000 human coded sentences taken from a corpus of contemporary novels.

3. Data 3.1. SP survey The SP survey conducted by Ghasri and Vij [13] is used to calibrate the behavioural framework. The survey collects respondents’ media consumption habits, and it includes a discrete choice experiment (DCE) designed to elicit preferences towards AVs. The DCE in this survey is a labelled vehicle-choice task, where four hypothetical vehicles are presented, and respondents are asked to choose the alternative they most prefer. A preliminary analysis of respondents’ media consumption suggests a diverse range of consumption behaviours. Investigating the frequency of level of engagement with the most frequently checked online outlets shows individuals receive information from different media outlets, and they have different level of engagement with their source. 3.2. Sentiment analysis Almost all the main media outlets have a Twitter account. Twitter posts, known as tweets, including AV related keywords are extracted using Twitter Search API. The list of keywords includes Autonomous vehicles, Autonomous cars, AV, AVs, Connected and autonomous vehicles, CAV, CAVs, Self-driving cars, Self-driving vehicles, Driverless car, Robo-car, Robotic car, Google car, Waymo, Tesla Autopilot, Cruise automation, Cruise LLC, GM Cruise, Argo AI, and Argo AI LLC. Those tweets including any of the keywords is extracted. This exercise started in March 2021 and is still ongoing. So far, 639,867 are collected in the dataset. The Syuzhet Package [15] is used to assess the sentiment of each tweet text, providing a value as a proxy for the received informational cues by potential consumers about the technological advancements or AV related events. Investigating the temporal trend in the sentiments based on a preliminary analysis on tweets posted by ABC, BBC and all other media outlets shows the sentiment of messages broadcasted from different outlets is different, which in turn can have different effects on audience WTP for automation. 4. Conclusion This study proposes a behavioural framework to link attitudes towards autonomous vehicles to received informational cues from media outlets. The proposed framework is then used to monitor temporal variations in willingness-to-pay for automation as a response to the media broadcasts about AVs. Such a tool can inform policy makers about the psychological effect of policies and regulations in the society, and it can inform the industry about consumers’ reaction to technological advancements or AV related events.

ABSTRACT. The concept of social capital describes how individuals acquire beneficial assets and services through social interactions. Among various definitions of social capital, Lin’s formulation of social capital as embedded social resources has a strong methodological synergy in activity-travel behavior – aligning with the individual-level basis of most research. Specifically, Lin (2001) proposes three primary elements of social capital: (1) resource embeddedness in social networks, (2) resource accessibility, and (3) resource use for action-oriented aspects. Lin (2001) further defines three processes involved in the creation and use of social capital: “(1) investment in social capital, (2) access to and mobilization of social capital, and (3) returns of social capital” (p. 19). Häuberer’s (2011) schema (Figure 1) clarifies Lin’s theory of the three processes and thus provides causal relationships between preconditions, social capital, and outcomes. Individuals are preconditioned in a societal context and have access to individually owned resources and assets. Access to social resources is mobilized through social networks and their structural properties. Smaller, denser networks help maintain social connections and promote continued access to group resources through trust and reciprocation. This leads to more resources for expressive actions and subsequently, capitalization of expressive outcomes. In contrast, larger, wider social networks enable new contacts but results in less intimate social support. These diverse, lower-maintenance connections can broaden access to new resources for profit or influential gain and lead to more resources for instrumental action.

This study proposes to create a latent variable framework to explicitly test hypotheses related to social capital outcomes as theorized by Lin and Häuberer. Two case studies in leisure activity behavior will be explored related to leisure activity variety and value of leisure time. This framework is applicable across many disciplines including health and sociology – two fields that use this resource conception of social capital. The first case study tests a proposed theory that leisure activity variety is an instrumental outcome and thus mostly affected by instrumental social capital. The theory underlines two hypotheses that (1) social capital is an integral determinant of leisure activity participation, and (2) having access to instrumental support promotes instrumental outcomes demonstrated by increase in more leisure activity variety. This theory was extensively tested on the number of different unique leisure activities collected from 1,297 survey respondents. An integrated count and latent variable (ICuLV) model structure was developed. Generally in count models, the expected value of the count takes the form: λ ̇_n=E(y_n )=exp⁡{βx_n } – where x_n is a vector of explanatory/observable variables for individual n. Using the Poisson model as an example, this results in the likelihood for a particular count being: p(y_n│x_n;β)=(λ_n ) ̇^y/y! e^(-λ ̇ )=(e^(y_n x_n β) e^(-x_n β))/y! To create an integrated count and latent variable structure, the expected value of the count is modified to incorporate the latent variable: λ_n=E(y_n )=exp⁡{βx_n+δ_s ln⁡(w_n^* ) } – where w_n^* is a vector of latent variables. For an integrated Poisson and latent variable model, the resulting likelihood of a particular count is thus an integral of this count probability integrated across the indicator distribution and latent variable distribution. This integrated count and latent variable structure requires a system of three models: a count model for the count variable of interest, a structural model for each latent variable, and measurement models for the indicators of each latent variable. For studying the impacts of social capital on leisure activity variety, the model structure would take the following form: Count Equation: "Activity Variety"=λ_n=E(y_n )=exp⁡{βx_n+δ_s ln⁡〖(s_n^* )+ε_n 〗 } Latent Variable Structural Equations: "Instrumental Social Capital"=w_(1,n)^*=α_1 s_n+v_(1,n) "Expressive Social Capital"=w_(2,n)^*=α_2 s_n+v_(2,n) Latent Variable Measurement Equations: "Indicators of Instrumental Social Capital" 〖=i〗_(1,n)=Γw_(1,n)^*+η_(1,n) "Indicators of Expressive Social Capital" 〖=i〗_(2,n)=Γw_(2,n)^*+η_(2,n)

The instrumental social capital indicators came from a position generator and the count of instrumental resources from a resource generator. The expressive social capital indicators came from a resource generator and strong ties count from a name generator. Results from a negative binomial regression models with latent social capital variables demonstrated that instrumental support indeed had the largest influence on predicting activity variety outcome. Between an individual with the lowest instrumental social capital and one with the most, there was an expected increase in leisure activity variety of approximately 75%. In the second case study, a stated preference study is designed to explore variations in leisure time valuations across people with varying access to social capital. The study will present respondents with three different activity options that varying in the expected companion, activity type, exposure to new people, activity length and cost, and travel time and cost. It is hypothesized that individuals with more instrumental resources have: 1. higher value of time for leisure activities with weak ties, 2. prefer leisure activities that help broaden contacts, 3. spend less time traveling to social activities, and 4. spend more money on social activities. It is expected that these hypotheses would be reversed for individuals with higher levels of expressive resources. An example of a scenario is presented in Figure 2.

To model this choice and individuals’ preferences, an integrated choice and latent variable model formulation will be formulated. Like the formulation above but with an utility maximization choice model, the measurement equation will introduce social network structural properties – as described in the access to social capital section of Figure 1. This will allow the authors to test if social network characteristics have the expected effect direction on instrumental and expressive resources. This presents a prime opportunity to explore if social resources act as an intermediary in leisure activity behavior thus helping to unify previous studies that only used social network characteristics as direct descriptors of leisure activity. The survey and stated choice design has been developed and data collection is expected in Winter 2022 with analysis expected in Spring 2022.

Lin, N., 2001. Building a network theory of social capital. In: N. Lin, K. Cook, R. Burt, eds. Social capital: Theory and Research. Transaction Publishers. pp. 3-29.

Häuberer, J., 2011. Social capital theory. Towards a methodological foundation. Wiesbaden: VS Verlag.

ABSTRACT. see pdf

ABSTRACT. Stanisław Łaniewski, Mikołaj Czajkowski, Marek Giergiczny, Maciej Sobolewski

Valuation of public goods is commonly done using Stated Preference (SP) methods, such as the Discrete Choice Experiment (DCE). In DCE respondents are asked to choose between presented alternatives, which may represent possible policies that differ with respect to attribute levels, including the cost. This approach allows for using observed choices to model respondents’ utility functions and eventually their Willingness To Pay (WTP) for new policies and their attributes.

However, the hypothetical nature of choices made by respondents makes them prone to various behavioural phenomena (Johnston et al., 2017). One of them, observed in a variety of contexts and settings is anchoring (Ariely et al., 2003) – a cognitive bias whereby an individual’s decisions are influenced by a particular reference point. If anchoring to the cost levels used in experimental design occurs then the resulting value estimates may be biased. For example, using lower or higher cost levels of the presented alternatives could change the resulting WTP estimates.

In this study, we test whether the cost vector used in experimental design is anchoring respondents’ WTP and investigate its potential drivers. We use experimental treatments to test, if such an anchoring occurs and investigate whether it is driven by the mean or the range of the cost vector used in a DCE study. In a 2-by-2 experimental treatment we vary (1) mean levels of the cost vector and (2) range of the cost vectors, around these means. This allows us to independently evaluate the effects of these design dimensions.

The empirical application concerns valuation of Galileo – the European geolocation system, which is developed as an alternative to the American Global Positioning System (GPS). The data comes from an CAPI-based survey of the population of Poland which run in waves from November 2018 to August 2019. The representative sample of 1,862 respondents were selected based on social-demographic quota and only active users of GPS were invited. Each participant faced 12 treatments (in a random order) with 2 or 3 alternatives, with the first one being ‘opt-out’. In each treatment, a respondent would see 3 variables: cost, sharing anonymous and sharing personal location data (obligatory, optional or no sharing at all). We model the data with the state-of-the-art mixed logit model while controlling for the treatment-specific scale.

We find that changing the mean cost levels appears to influence the estimated WTP associated with the ‘opt-out’ constant and do not significantly influence other attributes. The effect for the opt-out constant, however, significantly changes the WTP associated with a hypothetical policy. The effect is the strongest for the low range of cost levels and is mitigated by using the highest range of bids. On the other hand, we find that the range of bids (for the same mean level of costs) does not significantly influence WTP.

In addition, we observe the impacts of the cost vector on the share of ‘opt-out’ choices answers and using systematic strategies, supporting findings from (Glenk et al., 2019). These results are in line with theoretical expectations, including higher probability of reaching ‘choke price’ in high mean or high range treatments (Mørkbak et al., 2010). Higher range of bids decrease number of opt outs and number of people using systematic strategies. However, they do not disappear, and as we usually have lower boundary for cost (0), the upper boundary can be arbitrarily high and still be accepted by yeah-sayers (Kragt, 2013). This finding lends support to the fat-tails problem (Parsons & Myers, 2016): extremely high bids which will be unconditionally accepted can be abused to introduce arbitrarily high bias in WTP estimates.

The paper concludes with providing guidance for future studies, such as using larger rather than smaller range of cost levels. The anchoring, however, likely remains a concern for SP-based valuation studies.

Ariely, D., Loewenstein, G., & Prelec, D. (2003). "Coherent Arbitrariness": Stable Demand Curves Without Stable Preferences. The Quarterly Journal of Economics, 118(1), 73-106. https://doi.org/10.1162/00335530360535153

Glenk, K., Meyerhoff, J., Akaichi, F., & Martin-Ortega, J. (2019). Revisiting cost vector effects in discrete choice experiments. Resource and Energy Economics, 57, 135-155. https://doi.org/10.1016/j.reseneeco.2019.05.001

Johnston, R. J., Boyle, K. J., Adamowicz, W., Bennett, J., Brouwer, R., Cameron, T. A., Hanemann, W. M., Hanley, N., Ryan, M., Scarpa, R., Tourangeau, R., & Vossler, C. A. (2017). Contemporary Guidance for Stated Preference Studies. Journal of the Association of Environmental and Resource Economists, 4(2), 319-405. https://doi.org/10.1086/691697

Kragt, M. E. (2013, 2013/02/01). The Effects of Changing Cost Vectors on Choices and Scale Heterogeneity. Environmental and Resource Economics, 54(2), 201-221. https://doi.org/10.1007/s10640-012-9587-x

Mørkbak, M. R., Christensen, T., & Gyrd-Hansen, D. (2010, 2010/04/01). Choke Price Bias in Choice Experiments. Environmental and Resource Economics, 45(4), 537-551. https://doi.org/10.1007/s10640-009-9327-z

Parsons, G. R., & Myers, K. (2016). Fat tails and truncated bids in contingent valuation: An application to an endangered shorebird species. Ecological Economics, 129, 210-219. https://doi.org/10.1016/j.ecolecon.2016.06.010

ABSTRACT. Anchoring, similarly referred to as reference-dependence or starting point bias, is a well-known phenomenon resulting from a heuristic suggested by Tversky and Kahneman (1974, Science, 185(4157)), where behavior, such as survey responding, is guided by information faced earlier. This issue has been broadly documented in plentiful contexts, including stated-preference surveys used for non-market valuation when series of questions are employed for eliciting preferences (e.g., Herriges & Shogren, 1996, Journal of Environmental Economics and Management, 30(1); Ladenburg & Olsen, Journal of Environmental Economics and Management, 2008, 56(3); Meyerhoff & Glenk, Resource and Energy Economics, 2014, 41). The dependence of stated-preference survey responses on prior information causes a challenge to modelling the preferences and deriving valid welfare measures, credibly representing true values of public goods to society. With this study, we aim to contribute to the understanding of the anchoring effect in discrete choice experiments along two dimensions. First, we investigate how anchoring varies across willingness-to-pay and willingness-to-accept welfare measures. This way, we explore the interaction of anchoring and endowment effects within a single study and a consistent decision setting. Second, we examine the occurrence of anchoring despite the advanced disclosure of the full range of attribute levels, which should mitigate the problem in theory.

While anchoring has been broadly studied, there is a paucity of systematic investigations of the effect across willingness-to-pay and willingness-to-accept measures. We analyze the relative importance of anchoring and a related magnitude of the bias in value estimates across these two types of welfare measures to understand which one is less and which one is more susceptible to the problem. To this end, we focus on two different versions of the same discrete choice experiment. In one, we ask respondents whether they are willing to pay for specific increases in the attributes’ values, and in another, we ask whether they are willing to accept a monetary compensation for decreases in the attributes’ values. We examine the anchoring especially with respect to the cost attribute value displayed in the first task in the discrete choice experiment. We hypothesize that facing an extreme (high or low) value of the cost attribute in the first task can affect consumers’ stated-preference behavior. In particular, the research design helps us assess whether the anchoring effect emerges despite respondents being disclosed in advance the full range of the possible cost amounts.

Data to this study comes from a discrete choice experiment survey administered online in Poland between May and December 2018 on a nationwide, representative sample. The valuation scenario concerns the theater offer in Poland – specifically, the number of performances and show premieres provided by public (state-subsidized) theaters. Two separate (split-sample) versions of the questionnaire are developed and randomly distributed in the sample: one starts with the discrete choice experiment eliciting willingness to pay for proposed extensions of the theater offer (1,437 respondents) and another begins with a similar discrete choice experiment but concerning willingness to accept for considered reductions in the offer (1,426 respondents). Each version consists of 8 choice tasks with 2 alternatives: the status quo and a proposed program (i.e., a change in the theater offer). The choice alternatives are described by 6 attributes: (i-iv) four attributes regarding changes in the numbers of theater shows of a different type, including entertainment, drama, children’s and experimental performances (i.e., a separate attribute for each type), (v) a geographical scope of the theater offer change (i.e., a selected province or the whole Poland), and (vi) a change in annual expenditures to an individual (an obligatory cost or compensation to all).

Preliminary results from our discrete choice models (including random parameter logit and latent class models) demonstrate pronounced anchoring effects. In particular, in the willingness-to-pay version, respondents who face a low cost amount in the first choice task are less sensitive to changes in the expenditures for the theater offer and those who are displayed a high cost in the task are more sensitive to the cost, when compared to respondents shown a non-extreme cost level. Correspondingly, facing a low cost amount increases willingness to pay for the proposed changes in the theater offer, while facing a high cost leads to a decrease of that measure. Stated-preference behavior of respondents who face the willingness-to-accept version is different. For them, we observe increased sensitivity to the cost attribute when faced the extreme (highest or lowest) value of the cost in the first task compared to respondents shown a non-extreme cost level. This translates into lower values of the willingness-to-accept measure. Summing up, these results show that facing the lowest and highest values of the cost attribute in the first task of a discrete choice experiment has an impact on subsequently reported stated preferences, and the direction and magnitude of this impact varies across willingness-to-pay and willingness-to-accept measures.

These results have important practical implications for stated-preference design and modelling. First, the advanced disclosure of the full range of attribute levels does not appear to make stated-preferences free from anchoring. Second, while we acknowledge that the usual randomization of choice tasks across the entire sequence in the discrete choice experiment might counterbalance the anchoring on the low and high cost amounts, the issue still remains for respondents who face the extreme (lowest and highest) levels of the cost attribute. This raises a question of a potential benefit from omitting the extreme cost levels from the initial choice tasks to reduce the anchoring. This is further tackled in the empirical analysis and theoretical discussion in the paper.

In conclusion, our research shows that anchoring affects willingness-to-pay and willingness-to-accept measures. The results lead to a question whether controlling for possible anchoring on the first choice task might improve the validity of the estimates. We further examine this issue by taking into account attitudinal questions that help enhance the validity of the estimates (such as, consequentiality and credibility perceptions). Overall, the study suggests that taking into consideration anchoring may improve the accuracy of the value estimates by providing a better representation of respondents’ preferences that excludes some cognitive biases.

ABSTRACT. In environmental economics, the contingent valuation method is frequently applied to elicit willingness to pay for non-market goods and services. Considering the hypothetical nature of contingent valuation, the validity of the willingness to pay estimates it produces impinges on incentive compatibility of the choice settings (i.e., respondents have an incentive to provide answers that reflect their true preferences). There are several elicitation formats commonly used for contingent valuation, each having its advantages and disadvantages. For an incentive-compatible elicitation of stated preferences, however, the theoretical literature clearly points to the use of a single binary choice (SBC). In a recent contribution, Vossler and Holladay (2018) provide a theoretical framework, which – under certain assumptions – can make open ended (OE) and payment card (PC) contingent valuation questions incentive-compatible. OE and PC formats have several advantages over SBC, including increased statistical efficiency, and absence of “yea saying”, which may lead to overstated welfare measures. If incentive-compatible OE and PC formats provide valid results, they may become the preferred elicitation format, allowing for valid welfare measures at lower costs.

In their empirical application, Vossler and Holladay did not find convergent validity between incentive compatible OE, SBC and PC (i.e., willingness to pay estimates based on different elicitation formats significantly differ). They discuss that there might be "a behavioral effect of providing respondents with a set of possible payment amounts” (p.144). The presence of this effect has been shown in empirical studies for PC (e.g., Covey et al., 2007), and, to a lesser extent, for SBC (e.g., Boyle et al., 1997). Thus, convergent validity may be compromised by the choice of the bid vector (i.e., the range of costs or prices respondents are presented with). At the same time, a lack of incentive compatibility may have only a minor effect on the estimated WTP.

In this study, we test the hypothesis that behavioral effects induced by the bid vector are lower in SBC than in incentive-compatible PC and OE formats. Using a sample with more than 10,000 adults from six European Union member states, we estimate willingness to pay for a hypothetical program to increase agricultural biodiversity measures. The preference elicitation is accompanied by several follow-up questions on understanding and consequentiality.

Using a split-sample approach, we elicit willingness to pay with different contingent valuation elicitation formats and using different bid vectors. Our experiment is based on a 2×2+1 between-subject design with five treatments. Treatments 1 and 2 use the PC format, treatments 3 and 4 the SBC format, and treatment 5 the OE format. In treatments 2 and 4, we use a higher bid vector than in treatments 1 and 3. This design allows us to identify the interaction between bid vectors and elicitation format.

We estimate discrete choice models to calculate willingness to pay from PC and SBC questions, and use other regression techniques for the OE question. We investigate our hypotheses using formal parametric tests of differences in willingness to pay and two-sample tests of proportions for differences in frequencies of yes and no votes. In addition to our main investigation, we test for convergent validity of all five treatments and estimate the effects of the elicitation formats on understanding and consequentiality perceptions. Overall, we can assess the three elicitation formats with respect to their robustness and validity.

Our results are relevant for the ongoing debate on the validity of stated preference methods in general and contribute to the literature on convergent validity, incentive-compatibility, and bid vector effects. We show to what extent incentive-compatible valuation formats are robust to changes in bid vectors. We also provide new evidence of convergent validity of different incentive-compatible elicitation formats comparing incentive-compatible OE, PC and SBC and the effects on consequentiality perceptions.

The results support applied researchers in selecting an appropriate method for value elicitation with stated preferences methods. If it turns out that bid vector effects are similar between OE, PC and SBC methods, a clear argument in favor of the incentive-compatible PC formats can be made, given the statistical efficiency of these methods. If bid vector effects are significantly larger in PC, a SBC may remain the preferred format.

References:

Boyle, Kevin J., F. Reed Johnson, and Daniel W. McCollum. "Anchoring and adjustment in single-bounded, contingent-valuation questions." American Journal of Agricultural Economics 79.5 (1997): 1495-1500.

Covey, Judith, Graham Loomes, and Ian J. Bateman. "Valuing risk reductions: Testing for range biases in payment card and random card sorting methods." Journal of Environmental Planning and Management 50.4 (2007): 467-482.

Vossler, Christian A., and J. Scott Holladay. "Alternative value elicitation formats in contingent valuation: mechanism design and convergent validity." Journal of Public Economics 165 (2018): 133-145.

ABSTRACT. Over the past two years, the COVID-19 pandemic has upended the routines and lifestyles of almost every person across the world, and work is no exception. While a small fraction of employees had flexible work arrangements before the pandemic, COVID-19 forced millions to shift into a virtual environment. With the initial stay-at-home orders in place, working from home became the norm. But as society has begun opening up again, individuals have sought to revisit pre-COVID commute patterns, and are exploring a portfolio of work arrangements and work locations. In terms of work location, three possibilities (and their combinations) have become quite popular: the regular work place (the pre-COVID norm), the home (the COVID era norm), and a third workplace location (such as coffee shops, designated co-working locations, hotels and restaurants). Each of these work locations has positive aspects as well as not-so-positive aspects in terms of social, personal and professional considerations (of course, what is positive and what is not-so-positive may itself vary across individuals). Thus, while working from home may be associated with a better work-family life balance for many individuals, it also may be viewed as precluding the ability to socialize at the work place or reducing productive discussions as a unified team through face-to-face in-person conversations. Further, as in pre-COVID conditions, many employees may correlate extended absences from the work place with a lack of visibility to management and potential career stagnation implications. On the other hand, commuting every day to the work place (as was the case for a large fraction of workers pre-COVID) may promote socialization opportunities and provide more visibility, but can also be draining in terms of financial considerations (for example, investment in clothing/attire and formal day care facilities for children) and time/emotional considerations (for instance, long and tiresome commutes). The third work place may appeal to some individuals because of fewer distractions and enriching small-scale socialization possibilities relative to working from home, while also eliminating a long commute and avoiding larger-scale gatherings at the usual work place. But this third workplace location arrangement may be expensive because of the need, for example, to rent a hotel space; besides it is still characterized by the decreased visibility associated with working away from the regular work place.

It is clear from above that the three work location arrangements are not perfect substitutes of each other, but are better viewed as imperfect substitutes. Each location arrangement may satisfy specific functional, social, productivity, emotional, privacy, professional, visibility, networking, and financial objectives to different extents. Indeed, as an employee invests more and more in one work arrangement, the employee may feel a need to switch to another work arrangement to feel a sense of “optimal” social and professional balance. In other words, the true decision process may be better characterized as “horizontal” variety-seeking, where the consumer selects an assortment of alternatives due to diminishing marginal returns for each alternative. Accordingly, we characterize the workplace location choice situation as a horizontal choice structure, where the individual decides on an optimal combination of the three work location arrangements over a certain time period (in this analysis, we use the month as the time period of analysis). We then apply a multiple discrete-continuous model to identify the effects of sociodemographic characteristics, work-related and occupation characteristics, attitudes toward the pandemic and pre-COVID commute patterns, built environment variables of both the residence and out-of-home workplace, as well as COVID risk measures. The discrete component here refers to whether an individual partakes in a specific work location arrangement at all or not, and the continuous component refers to the fraction of the work day instances in a month of each work location arrangement. The “budget” here is exogenous and effectively takes a value of one for all individuals (basically, the number of days of work per month is considered pre-determined, and the fractional split among the three work arrangements is investigated). This multivariate fractional split approach is different from traditional multivariate fractional split models, which assume that the fractional observations arise from multiple “vertical” choice occasions, at each of which there is a single discrete choice.

The primary data for the current study will be obtained through a workplace location choice survey, scheduled to be deployed across Texas, US in January, 2022. This survey is part of a Texas Department of Transportation-funded study, and will collect information on pre-COVID, during the worst of COVID, and the current work location patterns of Texas residents. The survey, which is currently at the pilot stage, includes a stated preference experiment -- each respondent is presented two scenarios and asked to allocate their number of working days per month across the three different workplace alternatives. The attributes used in each scenario include commute time to the usual work place, time to a third workplace location, measures of distraction and crowding level for different work locations, flexibility of work hours, and COVID-19 risk intensities associated with each possible location. In addition, socio-demographics and the regular work/home location information are obtained. The work/home location information will be subsequently associated with built environment information already available from secondary data sources.

This study on work place location arrangement choices can provide important insights on how roadways, office space, virtual communication technology, and other work-related infrastructure may evolve in the future. What will the workplace environment look like across the world as employees strive to retain some of the work location flexibility brought about by the pandemic? How will employers decide on their real estate needs, and how might this affect land-use patterns and commute patterns? How might employers respond to changing work pattern demands of employees, and what policies might they adopt to best harness and balance their employees’ productivity, motivation and mental/emotional health? Can the paradigm change in work location perceptions brought about by the pandemic be harnessed to promote transportation equity across population groups? All these questions start from the more fundamental question of how employees will choose their portfolio of work location arrangements.

ABSTRACT. Introduction The unprecedented challenges brought by the COVID-19 pandemic have introduced several new considerations into people’s daily activity-travel routines including their shopping behavior. State governments’ stay-at-home orders together with consumers' desire to avoid crowds and public places, as well as the continued need for groceries, were among the most significant contributors to the exponential growth in online shopping. As such new considerations have urged people to rethink the way they conduct their daily activities, it is also essential that transportation planners and modelers update their predictive tools in light of such implications. In recognition of this critical need, the current study develops a modeling framework to understand the dynamics of people’s online shopping behavior in the era of the COVID-19 pandemic. In particular, we aim to understand how people have adjusted their shopping behavior due to the pandemic crisis, and whether these adjustments will persist afterward or will bounce back to the pre-pandemic situation. In doing so, we focus on the dynamics of individuals’ online shopping behavior in three distinct stages of pre-pandemic, during the pandemic, and post-pandemic situations. Scrutinizing such dynamics engendered by the spread of the Covid-19 pandemic plays a crucial role in updating and recalibrating transportation models and simulations so they can be reliable tools in the post-pandemic world. Further, as consumers’ shopping behaviors for various retail sectors are quite different, we only focus on online grocery shopping to keep the scope of our study manageable.

Data To provide the required information for our analysis, a multidimensional travel behavior survey was designed and implemented in the Chicago metropolitan area from April 25th to June 2nd, 2020. The survey collected a rich set of data regarding Chicagoans’ socio-demographic details, their health-related background, as well as an extensive set of information about their daily activity-travel behavior, including their shopping habits. Also, the survey employed Google Maps API to collect respondents’ residential locations in a high resolution (i.e., a block-group level). The final dataset comprises 915 respondents after rejecting observations with either missing, invalid, or inaccurate information. The dataset consists of 54% male and 45% female participants who live in the Chicago region. As for the age, 6% are less than 24 years old, 18% are between 25 and 34, 15% are between 35 and 44, and the rest are older than 45 years old. We have also incorporated the data from the survey with built-environment information from the Smart Locations database provided by the Environmental Protection Agency (EPA). This data which is available at the census block group level provides insights into the effect of built-environment settings on online shopping behavior. In this study, the dependent variables are derived from three questions in the survey which asked the respondents to indicate the frequency of their online grocery shopping before and during the pandemic as well as their expectations about their future shopping behavior. These variables were collected in an ordinal format.

Methodology Given the ordinal nature of the three target variables, they can be modeled using the well-established ordered probit model. It is worth noting that the three dependent variables can be correlated due to some shared unobserved factors affecting them. Furthermore, there might exist causal effects of people’s shopping habits on their future shopping behavior. To consider these potential interrelations and capture the complex interactions of the three ordered variables (i.e., (1) frequency of ordering groceries online before the pandemic; (2) frequency of ordering groceries online during the pandemic; (3) likelihood of e-shopping groceries more frequently in the future) they are modeled through a joint framework while considering potential endogenous effects of past and current experiences on future behavior. Our modeling framework accounts for the correlation among the three dependent variables by imposing a multivariate normal distribution on the error terms. Deriving the likelihood function for our utilized joint structure yields a complex formulation involving computation of a three-dimensional integral which, as can be expected, is computationally extensive and can lead to convergence issues. To overcome this issue, the current study applies the pairwise marginal likelihood estimation method, which forms a composite margin as a combination of bivariate marginal probabilities.

Results The results indicate that a wide range of explanatory variables including demographic variables such as education level, household income, and working status, as well as built-environment characteristics, and travel habits affect people’s online shopping preferences and behavior. However, many of the variables that are found to be significant in different periods are dissimilar. Even the common variables have different impacts. Focusing on the endogenous variables which represent the impact of people’s online shopping habits before and during the pandemic on their future behavior, the results show that those who had the experience of online grocery shopping before the pandemic (at any non-zero level) and mostly/always are using online shopping services during the pandemic are likely to stick to online shopping more frequent than the before-pandemic case. Overall, our findings reveal major changes due to the COVID-19 in underlying factors behind people’s online shopping preferences.

ABSTRACT. 1 INTRODUCTION The world is still reeling from the COVID-19 pandemic that has upturned the global economy and individual lifestyles. To prevent virus transmission, countries around the world have witnessed a dramatic reduction of travel by air, mass transit and other shared modes. The use of private vehicles initially declined during the early stages of the pandemic but has later rebounded (1) (2). New and used car sales have taken off during the pandemic both in the US and other countries (3)(4)(5). In light of the long-lasting problem of traffic congestion and air pollution associated with the use of private vehicles, it is extremely important to understand what factors have been affecting consumers’ change in vehicle ownership during the pandemic, and how car dependence can be contained in the post-pandemic future.

Previous studies have reported the substitution with travel modes that are perceived to be safer following major incidents. For instance, it was found that Americans increased car travel to substitute air travel after the 9/11 terrorist attacks (6, 7). This phenomenon is referred as the “dread hypothesis” in the psychology literature (8). In this study, we hypothesize that the COVID-19 pandemic might have also caused such a psychological dread risk effect that led some individuals to increase their vehicle ownership. In addition, we assume that people have different personal characteristics and motivations for changing vehicle ownership. For instance, a new vehicle for a previously zero-car household may mostly be the substitution of modes that became unavailable during the pandemic (“mobility utility” ), whereas an additional vehicle for a car-owner might also for the preparation of the unexpected mobility needs in the face of uncertainty (“psychological utility”).

By estimating an integrated choice and latent variable (ICLV) model, this study will explore individuals’ socio-demographic characteristics, latent perceptions/attitudes, residential built environment attributes, other life events during the pandemic (e.g., change of employment status) on their vehicle ownership change, where respondents are further segmented based on current vehicle ownership status resulting from the change.

2 DATA & METHOD The dataset used in this study is derived from a large COVID-19 Mobility Study, administered by a research team at the University of California, Davis to better understand the impacts of the pandemic on travel behaviors and mobility decisions. It is a multi-wave survey that was launched in spring 2020, fall 2020 and summer 2021 in 17 metropolitan regions in the US and Canada. More details on the sampling method and survey tools used for this study are available at https://postcovid19mobility.ucdavis.edu.

This study focuses on 3,560 respondents from the fall 2020 survey in the greater Los Angeles area. The survey asked about respondents’ current vehicle ownership status in fall 2020 (i.e., zero-car/ car-deficient/ car-sufficient) and the vehicle ownership change in comparison to fall 2019 (i.e., increase/ decrease/ replace/ no change). The survey also contains information on respondents’ socio-demographics, residential location, employment status, and a series of attitudinal statements to measure latent constructs of the participants, such as attitudes towards car ownership, environmental issues, home location choice and so forth.

One logit-kernel based ICLV model will be constructed to model the effects of individual characteristics, latent perceptions/attitudes, change of residential location and commuting status during the pandemic on respondents’ vehicle ownership (i.e., increase/ decrease/ replace/ no change), where respondents are further segmented based on current vehicle ownership status (i.e., zero-car/ car-deficient/ car-sufficient).

Figure 1 ICLV framework for household vehicle ownership model (please see PDF version)

3 ANTICIPATED RESULTS & IMPLICATION The measurement component of the ICLV model will suggest the relationship between each latent variable and their corresponding indicators, while the structural component will test our hypothesis that exogenous socio-demographic attributes significantly influence people’s perceptions and attitudes. These findings will help build the clusters of population and to predict values of the unobserved latent factors. An important policy implication of understanding the heterogeneity among population is to understand how individuals will react to a policy which can be used to support the design of efficient policies tailed for certain group of population. Finally, the vehicle ownership change model will reveal the factors impacting the change in vehicle ownership of individuals who are from non-car, car-deficient and car-sufficient households, respectively. Findings of this study will help better understand the factors associated with vehicle ownership change by population segments during the pandemic. Policymakers and other stakeholders can thus predict future trends, design efficient policy provisions and marketing efforts in the latter phase of the pandemic as well as in the post-pandemic era.

4 REFERENCES

1. Fatmi, M. R. COVID-19 Impact on Urban Mobility. Journal of Urban Management, Vol. 9, No. 3, 2020, pp. 270–275. https://doi.org/10.1016/j.jum.2020.08.002. 2. Abdullah, M., C. Dias, D. Muley, and M. Shahin. Exploring the Impacts of COVID-19 on Travel Behavior and Mode Preferences. Transportation Research Interdisciplinary Perspectives, Vol. 8, 2020. https://doi.org/10.1016/j.trip.2020.100255. 3. SCMP. Coronavirus : China Car Sales Mark First Gain in Almost Two Years after Restrictions Are Eased. South China Morning Post. https://www.scmp.com/economy/china-economy/article/3083797/coronavirus-china-car-sales-mark-first-gain-almost-two-years?module=perpetual_scroll_0&pgtype=article&campaign=3083797. Accessed Dec. 12, 2021. 4. Rosenbaum, E. The Used Car Boom Is One of the Hottest, and Trickiest, Coronavirus Markets for Consumers. CNBC. https://www.cnbc.com/2020/10/15/used-car-boom-is-one-of-hottest-coronavirus-markets-for-consumers.html. Accessed Dec. 12, 2021. 5. Park, K. Car Sales Are Surging Nationally , Stressing Vehicle Manufacturers and Delighting Salespeople. The Philadelphia Inquirer. https://www.inquirer.com/. Accessed Dec. 12, 2021. 6. Gigerenzer, G. Dread Risk, September 11, and Fatal Traffic Accidents. Psychological Science, Vol. 15, No. 4, 2004, pp. 286–287. https://doi.org/10.1111/j.0956-7976.2004.00668.x. 7. Gigerenzer, G. Out of the Frying Pan into the Fire: Behavioral Reactions to Terrorist Attacks. Risk Analysis, Vol. 26, No. 2, 2006, pp. 347–351. https://doi.org/10.1111/j.1539-6924.2006.00753.x. 8. Schweitzer, L. Planning and Social Media: A Case Study of Public Transit and Stigma on Twitter. Journal of the American Planning Association, Vol. 80, No. 3, 2014, pp. 218–238. https://doi.org/10.1080/01944363.2014.980439.

ABSTRACT. It is widely accepted that the COVID-19 pandemic has dramatically changed travel patterns in the last two years, largely due to restrictions to people’s movement and work-from-home practice. A wealth of studies have been conducted to understand such changes from the traveller perspective, collecting different types of mobility data in different parts of the world. Some studies adopted a longitudinal perspective in a given country, allowing the observation of changes over the different phases of the pandemic (e.g. Molloy et al., 2021 for Switzerland or de Haas et al. 2020 for the Netherlands) while others have compared different countries at a point in time (e.g. Barbieri et al, 2021).

The proposed study incorporates both perspectives through the use survey panel data. The authors collected an extensive amount of information about travel behaviour in the United Kingdom, Australia, Colombia and South Africa using identical surveys (with minor adaptations to conform to the different cultural settings). Respondents are asked to provide information about their pre-covid travel behaviour and to then complete questionnaires at different points in time during the pandemic, characterised by different levels of infection and different restrictions. So far, 3 waves are available for each country except Australia (2 waves), and the collection of a final wave will take place shortly. While the data is not collected to be representative of the population, we believe it can still be used to produce key insights, and the final results can be weighted to produce a more realistic picture of the relevant contexts.

The key strength of the data is not only that it allows international comparison over time, but that it allows to investigate a wide range of research questions due to the large number of questions. For this particular study, we focus on data about current and expected behaviour, attitudes as well as hypothetical future scenarios. In particular: • we analyse how attitudes (towards fear of the virus, appropriateness of measures imposed by governments and acceptability of masks/tracking) change over time and aim to establish a causal link between attitudes and behaviour in different countries; • we compare what people expect to be their behaviour in the future with what they actually report in following waves; • we analyse hypothetical choices in specific scenarios about anticipated use of different modes of transport, in particular public transport, taxi and shared ride services. The scenarios are characterised by different levels of restriction on travel so that they can be compared with data collected at subsequent points;

Different kinds of analysis will be conducted. Latent variables have been identified from the attitudinal statements and a hybrid choice model has been estimated to analyse their link with travel behaviour. We have also modelled the answers to the hypothetical questions in the different scenarios and the link between attitudes and such choices, obtaining intuitively meaningful results. For example, we observe that people who are more concerned about catching the virus and who are supporting the enforcement of travel restrictions are more likely to expect that they will decrease their level of travel by any mode under any scenario. We also unveil the effect of different socio-demographic characteristics as well as country-specific effects on the evolution of attitudes and behaviour. For example, we see that for all countries, the level of concern decreases over time, while only in the UK it is higher among women and only in the UK and Australia it increases with age, while the opposite is true for South Africa.

While these first results are promising, further modelling will allow us to also understand whether changes in attitudes precede, follow or happen simultaneously with changes in behaviour as well as how attitudes and behaviour relate to future expectations and their realisation. The work to be presented at the conference will also include the final wave of data and comprehensive analysis.

References

Barbieri, D. M., Lou, B., Passavanti, M., Hui, C., Hoff, I., Lessa, D. A., ... & Rashidi, T. H. (2021). Impact of COVID-19 pandemic on mobility in ten countries and associated perceived risk for all transport modes. PloS one, 16(2), e0245886.

Molloy, J., Schatzmann, T., Schoeman, B., Tchervenkov, C., Hintermann, B., & Axhausen, K. W. (2021). Observed impacts of the Covid-19 first wave on travel behaviour in Switzerland based on a large GPS panel. Transport Policy, 104, 43-51.

de Haas, M., Faber, R., & Hamersma, M. (2020). How COVID-19 and the Dutch ‘intelligent lockdown’change activities, work and travel behaviour: Evidence from longitudinal data in the Netherlands. Transportation Research Interdisciplinary Perspectives, 6, 100150. Chicago

ABSTRACT. Introduction In health economics, discrete choice models almost exclusively assume individuals adopt random utility maximisation (RUM) as a decision rule when making choices. A recent trend in health economics, following the broader choice modelling literature, has been to challenge this assumption in a variety of ways. One such way has been to incorporate behavioural choice heuristics, such as attribute non-attendance. Other ways test alternative decision rules that individuals adopt when making decisions. In health, an alternative decision rule to RUM, random regret minimisation (RRM), has been tested. There is theoretical support for the role of regret in health behaviours, and the empirical results suggest that some individuals do indeed adopt regret minimization when making choices. The RRM literature in other fields is more developed, and in a range of settings, RRM was found to improve on RUM as a depiction of choice behaviour. Recently, in transport economics, researchers have developed another alternative decision rule: Decision field theory (DFT), a theory of decision-making with psychological foundations (rather than econometric). Under DFT models, the decision-making process is explicitly modelled through dynamic, stochastic preference updating of the different alternatives as the decision-maker considers the alternatives. Like RRM, early empirical comparisons with RUM show promise. Moreover, there is a theoretical basis to support DFT in health behaviours. Thus, this study introduced DFT to health economics, and empirically compared DFT to RUM and RRM. This study also introduced new forms of DFT models using latent classes, larger-scale DFT models with substantially more parameters for deterministic heterogeneity, and novel testing of DFT models. Methods This was a secondary analysis of a DCE of cigarettes and e-cigarettes conducted in 2017 in the United States. Attributes were flavours, nicotine, health harms, and prices. 2,031 participants were recruited online, and were matched to the population on age, gender, and region using data from the 2014 Behavioural Risk Factor Surveillance System, a nationally representative survey collecting data on behavioural risk factors. Each responded to 12 choice tasks generated from a Bayesian D-optimal design, with priors taken from a pilot study on 100 smokers. Attribute-only (base) models were estimated and tested for each of the decision rules. Next, models with deterministic heterogeneity were estimated and tested. In a further step, latent class models were estimated to accommodate unobserved heterogeneity. Lastly, decision rule heterogeneous models, using latent classes, were developed accommodating multiple decision rules in the sample. Model fit was compared between RUM, RRM and DFT using non-nested likelihood-based Vuong tests for differences in model fit. Parameter ratios, predicted choice shares, and pseudo-elasticities were computed. Bootstrapped standard errors and hypothesis tests for model differences were derived. The presence and impacts of decision rule heterogeneity on model outputs was investigated. Results Model fit of DFT was significantly higher than RUM and RRM in base models (Vuong tests: p<0.001, for both comparisons). In models with preference heterogeneity, DFT again outperformed RUM for latent class models (p=0.0143), but this was not significant in models with deterministic interactions (p=0.2819). Parameter ratios, predicted choice shares and pseudo-elasticities significantly differed for both DFT and RRM, compared to RUM. In RUM models, for life years lost, the ratio implied that a change from 2 to 10 years of life lost may be offset by a 6.701 USD decrease in price to keep utility constant. The relative importance of life years lost to cost was significantly higher for RRM (7.064 USD) but significantly lower for DFT (5.024 USD). For some attributes, significant differences were found for pseudo-elasticities. Between RUM and DFT, elasticity differences were of a larger magnitude than between RRM and RUM. The strongest (significant) differences were obtained for price. DFT appeared significantly more price elastic (cigarettes: -0.1414 vs. -0.1158, and e-cigarettes: -0.1914 vs. -0.1634), and cross price elasticities varied strongly for the opt-out choice (cigarettes: 0.4498 vs. 0.1843). Choice shares, however, despite being statistically significantly different, were very similar in magnitude. The presence of decision rule heterogeneity within the sample was shown, especially between RRM and DFT, which outperformed a latent class DFT model (p=0.0125). When incorporating multiple decision rules, point estimates appeared to better reflect behaviour of both decision rules. Conclusions Tobacco choices were better depicted using DFT in this DCE, compared to RRM and RUM. Models that incorporated decision rule heterogeneity outperformed models that imposed a single decision rule. These models were preferred to latent class models that included only preference heterogeneity, suggesting the presence of decision rule heterogeneity as well as preference heterogeneity. Contributions were made to the literature by deriving standard errors and test-statistics for model differences using bootstrap methods, and by deriving latent class, decision rule heterogeneous, DFT models. The significant differences demonstrate that care should be taken when choosing a decision rule in health-based choice models, but further evidence is needed for generalisability.

ABSTRACT. When facing a conflictual decision - like choosing between two desirable options - people tend to use simplifying strategies (e.g., preferring status quo) in order to deal with the complexity of the task and reduce cognitive load. This phenomenon should impact the decision-making process, from the way information is integrated to the weight given to considered evidence. However, a great majority of studies investigating this effect have been focused on nonmoral decisions, whereas we could expect different mechanisms in a moral realm. Moral decisions tend to be more extreme than nonmoral decisions, as morality often involves strong and clear-cut opinions. However, in the specific case of conflictual decision, extreme positions may be less psychologically comfortable to endorse and induce more nuanced responses. In order to investigate this effect, we built a computational model based on differential equations (dynamic neural fields coupled with sensorimotor control, extending classical drift diffusion models) in which a 1D population of neuronal units maps a continuous decision space. Neural fields usually operate on continuous spaces (e.g., sensorimotor), but allow the emergence of spatially localized attractors (which determine final decisions). In the decision space, activations of the corresponding accumulators account for the preference in the decision option continuum. Spatiotemporally coherent activity across the neural field reflects convergence in the decision space, and are able to model the non-linear dynamics of decision-making (e.g. measured in humans using mouse tracking techniques). We propose that conflict emerges when two regions of activation appear at qualitatively different places in the decision space (preventing the fusion of evidence into a consensual decision), and that activations are stronger for moral elements (leading to their prevalence in final decision). Simulated data with a priori defined parameters successfully reproduced empirical patterns, giving insights about what differs between conflictual moral and nonmoral decision processes and how different pieces of information (moral and nonmoral) are integrated during decision-making.

ABSTRACT. The trend for choice modellers to adapt ideas about behavioural processes from other fields shows no signs of diminishing. In particular, there has been an increase in the adoption of ideas from behavioural economics and mathematical psychology. This has led to the investigation of behavioural concepts such as reference dependence, gains-loss asymmetry and zero price effects, the development of new models such as regret minimisation and the adoption/refinement of models from elsewhere, such as decision field theory (DFT) or the multi-attribute linear ballistic Accumulator model. Choice modellers are also increasingly engaging with the field of neuroscience, and increasingly attempt to capture additional process data, mainly through neuroimaging techniques such as EEG or FMRI. However, what is lacking thus far is an engagement with the theories specifically developed in neuroscience, in relation to how the human brain perceives stimuli and how those stimuli are acted upon. The majority of choice modellers would readily agree that any model used thus far in choice modelling, even if based on behavioural theories, only provides an abstract representation of the way decisions are actually made. This is of course entirely reasonable and thus does not imply that the predictions made by these models are in any way inferior by definition. However, it raises the question whether the implementation of theories grounded in neuroscience would allow for deeper insights and a potentially more accurate representation of behaviour. In recent years, one topic in particular has received extensive attention and generated much excitement in neuroscience. The free energy principle, developed by Karl Friston, has become the leading unifying account of brain function and accounts for much of the empirical data on decision making and most grandly, our very existence. It has been lauded as the key to true Artificial Intelligence and has seen Friston ranked as one of the 22 most likely people to receive a Nobel prize. At the same time, the sheer complexity of the Free Energy Principle has so far seen very few attempts to actually implement it, and no efforts to use it in the context of choice modelling. The key notion behind the free energy principle is that the brain seeks to minimise surprise. We do this by decreasing the gap between our expectations of the world and our perceptions of the world, either by adapting our expectations or by taking actions to bring our surroundings (and hence perceptions) in line with our expectations. The free energy principle applies to any biological system that minimises disorder (entropy) [2]. The “system” is divided into internal and external states, and the agent continually acts to minimise surprise. The internal and external states are separated by a “Markov blanket” (they do not directly influence each other). External states inform internal states through sensory states and internal states inform external states through active states. As a consequence of its top-bottom approach, the free energy principle can be used to explain multiple phenomena, including (but not limited to) memory, attention, value, reinforcement, salience [2], emotion [3] and exploration behaviour [4]. Whilst it has not yet been used in the context of choice modelling, the free energy principle has been demonstrated to reduce to expected utility theory as a special case [1]. This is as a result of the fact that an agent’s posterior distribution Q for the probability of external states that they will finish in can be decomposed into the entropy of the distribution over final states (exploration bonus) plus the expected utility of the final state, both given the current state and policy [1]. In short, choices are based on beliefs about the possible alternatives, or strategies, and the free energy principle aims to additionally maximise the precision of beliefs about the consequences of choosing an alternative. In this paper, we consider not only the incorporation of ideas from free energy minimisation into typical choice models, but also ‘pure’ operationalisations of the free energy principle for modelling choice behaviour. For example, typical choice models only include error terms to represent uncertainty from the viewpoint of the analyst. Our model looks at the incorporation of parameters for the uncertainty from the perspective of the decision-maker. Additionally, costs of deliberating (exploring) can be taken into account through examination of the response times of a decision-maker. We will make use of data already collected in Leeds where, across 3 experiments undertaken in a controlled virtual reality environment, 126 participants were asked to make several hundred decisions on a 2-arm bandit task in which reward probabilities varied between options and the optimal choices needed to be learned through experience. While our work is still preliminary, our recent successes with providing the first operationalisation of DFT and MLBA [5] to both stated and revealed preference gives us confidence that we will provide a fully flexed implementation in time for ICMC. This work will form an important bridge between choice modelling and the neuroscience of decision making.

References [1] Friston, K., Schwartenbeck, P., FitzGerald, T., Moutoussis, M., Behrens, T., & Dolan, R. J. (2014). The anatomy of choice: dopamine and decision-making. Phil. Trans. R. Soc. B, 369(1655), 20130481. [2] Friston, K. (2009). The free-energy principle: a rough guide to the brain?. Trends in cognitive sciences, 13(7), 293-301. [3] Joffily, M., & Coricelli, G. (2013). Emotional valence and the free-energy principle. PLoS computational biology, 9(6), e1003094. [4] Schwartenbeck, P., FitzGerald, T., Dolan, R., & Friston, K. (2013). Exploration, novelty, surprise, and free energy minimization. Frontiers in psychology, 4, 710. [5] Hancock, T. O., Hess, S., Marley, A. A. J., & Choudhury, C. F. (2021). An accumulation of preference: two alternative dynamic models for understanding transport choices. Transportation Research Part B: Methodological, 149, 250-282.

ABSTRACT. Background Consider the following widely used additive random utility specification:

[EQUATION 1 - please see the attachment]

In this specification, the alternative attributes in x_qi are typically assumed as deterministic. However, this assumption can be contested for the following reasons: (1) the alternative attributes may be inherently stochastic; for example, stochasticity in travel times owing to day-to-day or intra-day variability in travel conditions on the network (Biswas et al., 2019), (2) errors made by the analyst in measuring the attributes (Bhatta and Larsen, 2011) and/or (3) difference between travellers’ perceptions and the analyst’s measurements of the attributes (Daly and Ortuzar, 1990). In the context of incorporating variability in explanatory variables describing alternative attributes in x_qi, some discrete choice modelling studies use the errors-in-variables approach to represent erroneous data on such variables (Sanko et al., 2014; Steimetz and Brownstone, 2005; Nirmale and Pinjari, 2020). Other studies use the Integrated Choice and Latent Variable (ICLV) approach (Ben Akiva et al., 2002) to incorporate measurement errors by the analyst (Walker et al., 2010) and perception errors by individuals (Varotto et al., 2017). Recently, Biswas et al. (2019) used the ICLV approach to accommodate inherent stochasticity in travel time variables in choice models. In a well-established and large stream of the literature, the parameters in \beta corresponding to the alternative attributes in x_qi have been specified as random to capture unobserved taste heterogeneity among decision-makers (Mc Fadden and Train, 2000; Bhat, 2000; Hensher and Greene, 2003; Revelt and Train, 1998; Hess and Polak, 2005). Most studies that do so use random coefficients on alternative attributes such as travel time through frameworks such as the mixed multinomial logit and the multinomial probit. While each of the two distinct directions discussed so far has witnessed substantial contributions in the literature, little effort has been made in terms of identifying both sources of variability simultaneously. As discussed in Biswas et al. (2021), this is likely because mixed logit or mixed probit and ICLV models on typical choice data do not allow the simultaneous identifiability of both sources of variability. This is because two different sources of data – one that helps in identifying variability in x_qi and the other that helps in identifying variability in the corresponding parameters in \beta – are needed for the simultaneous identifiability of both the sources of variability. At the same time, specifying either only x_qi or only \beta as random while keeping the other fixed can potentially lead to biased parameter estimates, thereby leading to distorted model predictions and policy analyses. Simultaneous identification of these two separate sources of variability is important for understanding travel behaviour in complex environments characterized by variable travel conditions. A recent effort in this regard, and the first of its kind as per the authors’ knowledge, is the work by Biswas et al. (2021), which proposes an ICLV framework to disentangle variability due to stochastic travel times from random coefficients on travel time. Current research In this research, we conjecture that in situations where stated preference (SP) and revealed preference (RP) data are pooled, one can disentangle the variability in alternative attributes from random coefficients on those attributes. This is because SP data is typically free of variability in x_qi (because the attribute values are presented to the decision-maker), which makes heterogeneity in \beta an important source of variability in stated choices. In RP data, however, various sources of variability in x_qi will likely have a bearing since RP data is collected from the “field” where it is not possible to control the different sources of variability. Therefore, combining both these data sources can potentially help in simultaneously identifying variability in x_qi and that in the corresponding parameters. To verify this conjecture, we formulate a choice modelling framework for pooled SP-RP datasets that allows the analyst to disentangle both the sources of stochasticity. The utility structure of such a framework is briefly outlined here for the analysis of urban mode choice considering mode-specific travel times as stochastic. Define the utility function associated with an alternative i, if it is an RP choice alternative as:

[EQUATION 2 - please see the attachment]

Alternatively, if is an SP choice alternative, the utility function is defined as:

[EQUATION 3 - please see the attachment]

In the above equations, TT_qit is the travel time for individual q for an SP alternative i which is assumed to be free of variability due to measurement errors, perception errors and inherent variability, is the corresponding attribute for an RP alternative i (which is considered stochastic due to various sources of uncertainty discussed earlier), \gamma_q is the random coefficient on travel time and does not vary across choice occasions or alternatives, x_qit is a vector of other alternative-specific and individual-specific attributes such as travel cost of the mode, income and gender, \varphi is the corresponding vector of parameters, \delta_qt,RP is a dummy variable which assumes the value 1 if the t^th choice occasion of individual q corresponds to his RP choice and 0 otherwise, \delta_qt,RP = 1 - \delta_qt,SP. \theta_q is the individual-specific effect which captures the difference across the RP and SP utilities for an alternative i. To account for the scale difference between SP and RP choice occasions (Ben Akiva et al., 1994), the scale of the additive random error term is fixed to 1 for the RP choice situation for a particular q, while it is specified as \lamda for an SP choice occasion. To do so, the following function (\upsilon_qt,RP =1 for an RP observation and 0 for an SP) is used for the scale parameter:

[EQUATION 4 - please see the attachment]

The complete model formulation will be discussed in the full paper. Next, the model will be applied to an empirical context of a joint SP-RP data collected from Bengaluru, India. In addition, simulation experiments will be presented to verify the conjectures made in the paper.

ABSTRACT. From a policy perspective, it is crucial to allocate the public resources for environmental interventions in an effective and socially acceptable way. However, most of economic evaluation approaches which are used to help decision makers tend to maximize the efficiency of proposed changes, disregarding fairness aspects (Richardson and Schlander, 2019). Despite the increasing expenditure on environmental programs, there has been relatively little attention paid to the issue what constitutes a fair distribution of new spending. Fairness concerns regarding the distribution of costs across socioeconomic groups may constitute a key determinant of the public acceptance of new policy measures.

In this study, we investigate individuals’ preferences for the air quality improvement in four Polish cities. The vast majority of Polish cities violate the European Union’s air quality target. Excessive airborne particulate matter (PM) pollution (often referred to by the mass media as PM smog phenomenon) has been occurring in Poland for decades. Almost all winter seasons in Poland are characterised by very high concentrations of PMs. The main source of this phenomenon is increased emission of particulate pollutants generated by combustion processes in stationary sources, predominantly the municipal and household sectors. Ambient air pollution, especially PMs constitutes a major environmental risk to human health (Landrigan et al., 2018). Both the WHO and the EU Environment Agency estimate that air pollution causes tens of thousands of premature deaths in Poland each year.

The main objective of this study is to examine the potential impact of concerns regarding distributive justice in payments on willingness to pay (WTP) for the health risk reductions resulting from air quality improvement. Instead of using single survey items, we applied a Factorial Survey Experiment (FSE) to elicit fairness concerns. FSEs are a type of multifactorial survey experiment which has become an important method in sociology for the study of justice concerns and social norms, among others (Liebe et al., 2020). In a FSE, respondents face one or more descriptions of a situation that differ from each other in a discrete number of factors. The respondents are then asked to evaluate those situations according to criteria such as support, agreement, or perceived fairness. Due to the systematic variation of the factors or situational attributes presented in the situations, a FSE is an experimental setup that can separate effects of single situational dimensions. Thus the causal influence of relevant situational factors can be determined.

As far as we know this is the first study which considers FSE and Discrete Choice Experiment (DCE) in a unified framework. To do that we employ a hybrid choice model in which respondents’ answers in FSE are used as measurement equations to identify latent factors representing their beliefs regarding general fairness of the air quality improvement programs as well as their distributive justice concerns. These factors are then used as explanatory variables in the choice model. Note that in our study FSE and DCE were completed by the respondents in two separate surveys with at least a week-long break between them. Such design increases the chances of identifying a casual relationship between the two, rather than simply identifying correlations. To avoid “spillover effects” and experimenter demand effects, in the FSE and DCE we focus on different aspects of the environmental programs.

The FSE was designed to elicit distributive fairness concerns regarding the payments required from various income groups for the investments in houses and communal flats that are aimed to increase the air quality in cities. Each vignette situation was described by five attributes. With respect to distributive justice the investments differ in terms of the share of subsidy across income groups. For creating vignettes, we used an orthogonal design with two-way interactions in which the attributes vary independently of each other within and across vignettes. This resulted in 72 vignettes. Each respondent faces six randomly drawn vignettes from this set. Respondent are then asked to assess the perceived fairness of the proposed program on each vignette using a 11-point Likert scale. The DCE follows the study by Jin et. al (2020). Respondents are asked to choose between two programs that prevent non-fatal illnesses and premature deaths caused by the PMs pollution and the status quo option. The programs differ across choice sets on the size of mortality and morbidity risk reductions, time needed for an implementation, and cost. The 36 choice sets with two alternatives were created using the D-efficiency design in Ngene. Each respondent faced eight randomly drawn choice sets from the full design.

Both surveys are currently conducted on the same group of respondents, with a time gap between each survey from one up to two weeks (i.e., two-waved panel data). The sample consists of 800 respondents, 200 from each of four Polish cities. Two cities are characterized by the lower air pollution level than the average in Polish cities, two others are with higher emission levels. The respondents from each city will not differ in terms of socioeconomic characteristics. The field time for the survey is December 2021 to January 2022. Until the conference, we will have a full working paper on this study including a hybrid model that combines both the FSE data and the DCE data.

References:

Jin, Y., Andersson, H. and Zhang, S., 2020. Do preferences to reduce health risks related to air pollution depend on illness type? Evidence from a choice experiment in Beijing, China. Journal of Environmental Economics and Management, 103, p.102355. Landrigan, P.J., Fuller, R., Acosta, N.J., Adeyi, O., Arnold, R., Baldé, A.B., Bertollini, R., Bose-O'Reilly, S., Boufford, J.I., Breysse, P.N. and Chiles, T., 2018. The Lancet Commission on pollution and health. The lancet, 391(10119), pp.462-512. Liebe, U., Moumouni, I.M., Bigler, C., Ingabire, C. and Bieri, S., 2020. Using factorial survey experiments to measure attitudes, social norms, and fairness concerns in developing countries. Sociological methods & research, 49(1), pp.161-192. Richardson, J. and Schlander, M., 2019. Health technology assessment (HTA) and economic evaluation: efficiency or fairness first. Journal of market access & health policy, 7(1), p.1557981.

ABSTRACT. Most real-world (revealed preference) data sources used for discrete choice modelling are samples from a larger statistical population. Deviating from this practice, in this paper we present a new car ownership model for the Netherlands that contains several submodels, most of which are estimated not on a sample, but on the population as a whole. For this, we use data from various person/household registers and the motor vehicle register, which are administered by Statistics Netherlands (CBS) and Netherlands Vehicle Authority (RDW), concerning (practically) all the 17 mln persons, 8 mln households and 8.5 mln passenger cars in The Netherlands. The Netherlands has a strong history of car ownership modelling, and at the moment there are several models available for the short, intermediate and long term. There is, however, a demand for a new car ownership and use model that combines the functionalities of the existing ones, while also being able to incorporate emerging trends such as the rapid uptake of electric vehicles or private lease. The new model, called SPARK, is developed by Significance and Demis for the Dutch Ministry of Infrastructure and Water Management and the PBL Netherlands Environmental Assessment Agency. It aims to provide national level forecasts under various scenarios, both for a few years ahead and for the very long term (up to 2060), give impacts of fiscal and other policies (especially on the penetration of electric vehicles) and yield car ownership inputs for the national transport model (LMS). SPARK contains several innovative features: • The implemented model works through micro-simulation of households and cars, both for the market for new cars and the second hand car market; • SPARK contains several dynamic discrete choice models for car market transactions of households (similar to those in Mohammadian and Miller, 2003); • Car use is estimated using negative binomial regressions, which provide a distribution that fits the data well; • The transaction and car use models are estimated on the population as a whole, not a sample; • For car type choice, the discrete choice models use a combination of revealed preference data, stated preference data on electric cars and diffusion curves for the penetration of electric cars over time (as in Norton and Bass, 1987), all in a single choice modelling framework (building on Jensen et al., 2016). The new model contains several submodules: • A dynamic household-simulator to forecast the future input variables at the household and person level (such as income, age, occupation) from year to year; • A car ownership module containing dynamic transaction models to forecast changes in the number of passenger cars per household in a year; • A car usage module containing negative binomial regression models to forecast the annual number of vehicle kilometres driven and the corresponding emissions. • A car type choice module containing discrete choice models to forecast car type choice: e.g. petrol/diesel/LPG/fully electric/hybrid, market segment (size), brand group, vintage and import/domestic purchase; Within the latter three types of submodules we distinguish between models for private cars (including private lease), for business cars in households (including company lease) and for business fleets of passenger cars (which are not used by households). The car ownership module for private cars contains three transaction models and one static model: • A dynamic model for households that start the simulation year without private cars, forecasting whether they will purchase a private car in the course of the simulation year or not, estimated on 2018-data for 2.1 mln households; • A dynamic model for households that start the simulation year with one private car, forecasting whether they will add, replace or dispose of a private car, or will do nothing, estimated on 2018-data for 3.6 mln households; • A dynamic model for households that start the simulation year with two or more private cars, forecasting whether they will add, replace or dispose of a private car or do nothing, estimated 2018-data on 1.5 mln households. • A static model for new households. These are households that either result from a break-up of an existing household or result from immigration. The model forecasts whether this household will get 0, 1 or 2 cars and was estimated on 2018-data on 0,5 mln households. The car ownership module for business cars uses a similar structure. Just as for private cars, these models are estimated on the population data from the registers. This is made possible by estimating the models on the CBS server using strict security and confidentiality rules. The transaction models use as explanatory variables characteristics of the households and persons and data on earlier transactions of the households since 2010. The models for annual car use are also based on the population data from the registers. The vehicle type choice decisions in the model depend on person/household attributes, characteristics of the available car types (e.g. cost variables) and attributes of the cars owned at the beginning of the year (showing for instance brand group or fuel type (dis)loyalty). These models have no less than 1800 choice alternatives, and the model estimation took place on a sample of the population data available at CBS, to save computing time. Moreover, these models also incorporate diffusion curves to obtain a better picture of the inflow of electric cars over time, and are also use based on new stated preference data focusing on electric cars. The estimation of all submodules of SPARK has been completed and the project has moved to the implementation, testing and validation phase. The model consists of more than 30 modules in total, with around 1800 estimated coefficients. The paper/presentation will therefore discuss the aims of the model, the model structure and estimation and first application results for the submodels for the private and business market. It will also comment on the implications of model estimation on large sets of register data (e.g. run time in estimation, t-ratios that can get as high as 3000, abundance of significant effects).

ABSTRACT. Travel demand, a derived demand, ensues from the demand to pursue certain activities. The success of any travel demand management strategy depends on how well that strategy fits the activity demand of decision-makers. Therefore, to better understand people’s travel behaviour, it is important to explore the role of different activity dimensions (e.g., activity type, duration etc.) in people’s mobility (Calastri et al., 2017, Palma et al., 2021). Following the activity-based approach, the linkage among activities and travel demand allows the researcher to understand complex household and individual travel choices by focusing on the constraints of time and space (Bhat, 2008, Enam et al., 2018). Previous studies attempted to explore the heterogeneity associated with activity choice and its different dimensions considering sociodemographic factors, built-environmental conditions, etc. Contemporary research attempted to capture the heterogeneity in the activity choice behaviour by classifying the activity demand into similar activity patterns within a class but different between classes (Liu et al., 2021, Rafiq and McNally, 2021, Victoriano et al., 2020). However, the heterogeneity may be due to the underlying differences in the time budget for different activities and a latent class allocation based on the time budget (unobserved in revealed preference data) may yield better results. In this context, this research aims to examine the heterogeneity in activity participation behaviour and the relationships between socio-demographic characteristics considering the time budget and heterogeneity in satiation associated with the activity duration. To do so, this study used a detailed travel diary survey data accumulated for 7 consecutive days (both working days and weekends) from 175 residents of Greater Concepción Area, Chile, between 2015 and 2016. We estimated a latent class model of activity type and duration choice in a single model following multiple discrete-continuous extreme value (MDCEV) frameworks. The proposed MDCEV based latent segmentation is then compared with the classical latent segmentation approach of the multinomial logit framework. The MDCEV based segmentation considers the time budget of 24 hours in a day while MNL based segmentation highlights the importance of participation rather than activity duration. In other words, in MNL based classification participation in leisure activity for one hour is equivalent to participation in the same activity for 4 hours as the classical MNL base segmentation ignores the effects of satiation which is an important dimension in activity decision. Unlike the other choice decision, getting closer to the end of the day decreases the satiation rate for other types of activities. Therefore, we contributed to the literature in two ways: using the MDCEV framework we highlighted the importance of satiation which can be used as a new dimension for segmenting activity demand, and we provide a better understanding and description of the differences in activity type and duration choice among different population segments. The proposed MDCEV based classification framework uncovered three distinct activity styles in the sample population that differ in terms of their socioeconomic characteristics, their predisposition towards different activity choice, and sensitivity to different activity duration.

ABSTRACT. Transport studies on cable cars have usually focused on different aspects of travel time savings (Garsous et al., 2019), accessibility improvements (Heinrichs and Bernet, 2014), or social benefits (Biberos-Bendezú and Vázquez-Rowe, 2020; Sarmiento et al., 2020). However, studying time allocation as a structural benefit of a transport project is scarce in the literature, particularly in Latin American cities. In-depth analysis of how public investments affect time use is crucial. Current evidence shows that travel time savings obtained from transport projects are likely to perform better in a cost-benefit analysis than comparable projects serving excluded and disadvantaged population groups (Martens and Di Ciommo, 2017), which is our case study. Travel time savings could be spent in more pleasurable activities such as leisure or more sleep time improving wellbeing. Therefore, it is necessary to understand individual time use decisions since transport-related choices are directly linked (Jara-Díaz and Rosales-Salas, 2017). We analysed the time use before and after implementing a new cable car in Bogotá (Colombia) in an underserved and peripheral community with low accessibility levels. We aim to understand the impact of a transport intervention on the time use and whether the project made residents more likely to perform some activities using a quasi-experimental prospective natural experiment. The project is located in the southwest of Bogotá in a district named Ciudad Bolívar, where 88.8% of the population does not meet their minimum consumption needs. In addition to the cable car, the project includes recreation facilities, community centers, markets, citizen service offices, and a housing physical improvement program (Sarmiento et al., 2020). The studied area comprises households within an 800-m buffer around each of the current cable car stations. We gathered socioeconomic information, travel patterns, and activity diaries using a household panel survey. A first wave collected baseline data before implementing the project between August and October 2018. Then, we collected follow-up data six months after the project inauguration, between August and November of 2019. The database includes 1,031 individuals for the baseline, 391 reporting activity diaries. The follow-up consists of 824 respondents, 191 activity diaries. After exhaustive analysis and data cleaning, the final sample used for this study considered 112 individuals who reported valid information in both stages of their activity diaries. We classified activities into work, study, stay at home, travel, leisure, errands, exercise, carry/collect things/people, and others, following activity time allocation (Bhat and Misra, 1999) and activity type and duration literature (Calastri et al., 2017). The last category was included, as most activity diaries did not completed 24 hours a day. We estimated two multiple discrete-continuous extreme value models (MDCEV) to represent time use decisions (Bhat, 2008, 2005). The first model refers to the before project situation, while the second one, represents the scenario six months after implementation. We considered systematic taste variations in the discrete and continuous model components to explore differences in activity participation and duration. Mainly, we explored observed heterogeneity regarding gender and cable car use in both models. The results show that participation and duration of all previously described activities changed after the project implementation. The cable car increased the probability of participating in work, study, carry/collect activities, and travel. It also decreased the duration of study activities and travel. Moreover, staying at home, errands, working, exercising, and leisure activities increased in duration. Those activities where participation increased (e.g., work, study, carry/collect and travel) suggest an accessibility improvement to perform productive opportunities after the cable car implementation. Reduced travel time allows more time for leisure and recreational activities. Results also allowed to identify gender and cable car users gaps. Women participated less in exercise activities than men before the cable car, and their leisure activity participation was also shorter. After the cable car implementation, women's probability of participating in activities is similar to men's. The cable car also led to increases in the activity duration of women in the study and paid work. Cable car users increased the probability of performing work, study, and leisure activities. Therefore, the cable car and the urban renovation in the neighborhood positively influenced the community by helping close time use gender gaps of residents. It also impacted the low-income community by providing better accessibility to opportunities and more free time.

ABSTRACT. Motivation and objectives

Many studies have investigated how people divide their time across activities throughout the day. However, little is known about the timing of these activities, i.e., when exactly the activities are scheduled, and the impact of external factors such as a pandemic on these behavioral patterns. The COVID-19 pandemic provides a unique opportunity to investigate its effect on activity time use and timing behavior as the risk imposed by the virus as well as the measures put into force to contain the pandemic had a significant impact on people and their everyday mobility patterns.

In the past, multiple discrete-continuous extreme value (MDCEV) models have proven to be a promising approach for estimating activity time use (Bhat, 2005) and timing behavior (Pinjari and Bhat, 2010). The application of a MDCEV model to time use data for a deeper analysis of behavioral changes induced by the COVID-19 pandemic allows us to:

(1) Improve policies related to time use and transport planning. (2) Identify first indicators of what the “new normal” may look like. (3) Predict behavioral changes based on external factors and hence analyze the reaction of the population to various policy measures, eventually supporting policy makers in their decisions to manage and find a way out of a crisis.

Based on tracking data from Switzerland and building upon our initial work (Mesaric et al., 2021; Winkler et al., 2021), this paper aims to provide insights into the drivers of activity scheduling and the consequences of the different phases of a pandemic on these schedules. The extension of the previously used dataset and the comparison with simpler statistical models (cf. Hammer, 2012) allow for testing the following two questions:

(Q1) The impact of the various phases of the pandemic can be clearly seen in the activity patterns for all socio-demographic groups. However, individual groups are affected differently in terms of their activity-travel repercussions of the policy measures based on profession and risk of infection due to the pre-existing health status.

(Q2) The use of a complex MDCEV model adds value compared to simpler statistical models, i.e., results in a better model fit and better estimates / forecasts for time use and timing behavior.

Data

This paper uses extensive Swiss GPS tracking data of the MOBIS and MOBIS:COVID-19 studies from September 2019 until September 2021 (Molloy et al., 2021) as well as of the LINK panel (different socio-economic sample characteristics) from November 2020 until September 2021. After the first round of data cleaning, the sample includes 4,542 respondents who reported 8 million trips and activities over a total of around 600,000 days. The data is scheduled to undergo a second round of data cleaning to further improve the data quality (filling of gaps, re-imputation of trip purposes (Gao et al., 2021) and new imputation of work from home activities). The tracking data is complemented with detailed surveys which provide information about the participants’ socio-demographics and work situation.

Model formulation

The following three models will be presented:

(1) First, the panel mixed multiple discrete-continuous extreme value (MMDCEV) model approach used by Mesaric et al. (2021) is applied to the extended dataset, including unemployed people and weekends. Due to the longer time period under investigation, there will be more pandemic phases compared to the first model version. The data will be further enriched with measures of perceived risk during the pandemic (e.g., infection data and Oxford stringency indices (Hale et al., 2021)).

(2) In a second step, the model is estimated separately for the different phases of the pandemic. The pre-pandemic model serves as the baseline for fixed parameters which are assumed to remain constant throughout the pandemic. The model is then compared to the more complex model in the first step to assess advantages and disadvantages.

(3) The two previously described MDCEV models are compared with simpler statistical models which are also used to estimate time use and timing behavior (e.g., Tobit models or (generalized) linear models; cf. Hammer, 2012) to find out if and to what extent the MDCEV model formulation improves the estimation results and – more importantly – the findings.

The MDCEV models include 5-7 activity purposes (e.g., home, work, shop, leisure and travel), 3-5 time periods throughout the day (e.g., peak-hours, off-peak hours, night hours), a budget (the 24 hours of a day) and an outside good (home activities at night). The utility function reflects the utility an individual accrues spending time on different activities on a given day. The model includes translation parameters to allow for corner solutions (i.e., zero consumption) for all activities except for the outside good. The panel structure of the data is accounted for by including a vector that generates correlations for each individual’s choices. The introduction of the pandemic phases as time segments allows us to examine the differential impact of the variables in each phase as independent variables in the model.

Preliminary results

Our first MMDCEV model analyzed time use and activity timing of working participants on weekdays during the first 6 months of the pandemic compared to the pre-pandemic baseline (Mesaric et al., 2021). The preliminary results of this data subset reveal that socio-economic variables such as age, gender, education and income have a distinct impact on activity participation and time use.

Compared to their activity participation rates before the pandemic, women perform more out-of-home activities, higher income individuals spend more time at home and older individuals go shopping more often during peak hours. People with higher workloads show a higher propensity to travel. There is a visible effect of the mask obligation on public transport and daily temperature. Overall, the total number of trips decreases, but the duration of the trips tends to increase.

Therefore, we can conclude that the preliminary results answer the first question that the identified activity patterns reflect the individual phases of the pandemic and the main influences vary across pandemic phases, activities, time of day and socio-economic groups. It remains to be seen if the continuation of the pandemic attenuates or weakens this effect.

ABSTRACT. Autonomous vehicles (AVs) have been lauded as the next stage in transportation for decades, although they suffer from several shortcomings, including interactions with pedestrians (Tabone et al. 2021). Previous studies have focused on investigating communication techniques between autonomous vehicles and pedestrians or on the interpretation of pedestrian actions (de Clercq et al. 2019; Stanciu et al. 2018; Habibovic et al. 2018). A recent informative review suggested several priorities for research on AV-pedestrian interactions, including the requirement for universal algorithms and large-scale datasets, covering a range of road types, car designs and pedestrian demographics (Rasouli and Tsotsos 2020). This study contributes to these requirements, by investigating perceived safety and road crossing decisions in response to traditional and autonomous vehicles, at a range of crossing points. Investigating the pedestrian-AV interactions can be difficult, given the relative expense and danger of physical experiments including an unpredictable machine and vulnerable participants. This has led to previous studies on pedestrians and AVs performing stated-preference (SP) approaches, rather than physical experiments, to investigate relevant details surrounding situational outcomes and AV design. There is a wealth of knowledge surrounding the design of SP surveys, however there are questions remaining surrounding the best medium in which to present them, and how much of the relevant information is interpreted by the participant. Visualization techniques such as images and videos are often used as a dense, easily interpretable medium to provide participants with exact replicas of a situation. However, such data is usually presented in 2-dimensional formats (e.g. a computer), which necessarily reduces some of the information available, and can also reduce the level of immersion experienced by the participant. Virtual reality (VR) is an established research tool that provides an immersive environment while allowing user interaction. However, there remain numerous questions surrounding the validity of using VR and the extent of its benefits as a data generating tool (Arellana et al. 2020; Mokas et al. 2021; Rossetti and Hurtubia 2020). As a result of these limitations, it is still unknown how much the level of immersion can influence a participant’s SP responses. This study investigates participant SP responses to scenarios surrounding autonomous vehicles and road-crossing behaviour using both computer screens and VR headsets as experiment paradigm. This will investigate the relationship between perceived safety and road crossing choices as a function of car type (autonomous or traditional vehicles) and of nearby infrastructure (pedestrian crossings, signalised crossing, and basic road crossing). This builds on the previous studies that have performed SP investigations into pedestrian-AV interactions using VR (Farooq, Cherchi, and Sobhani 2018; Nuñez Velasco et al. 2019) but extend it by performing identical experiments on computer screens in addition to within VR headsets. Thus, we will detect any differences in the level of participant immersion, and crucially, any differences in participant responses. In this study, participants were shown a series of 360° videos, filmed from the perspective of a pedestrian walking towards a crossing point on the pavement in London, UK. Within these video clips, there are either traditional cars, or AVs. The AVs are easily identifiable with numerous sensors (e.g. LIDAR) on the roof of the vehicle. The videos stop just as the pedestrian has the option to cross or wait for the vehicle to drive past. The videos show scenarios for zebra crossings, signalised crossings, and simple road crossings, with each scenario being repeated (in identical videos, as far as practicable) for traditional human-driven vehicles and AVs. The participant was then asked to respond to several SP questions, detailing their impressions and their hypothetical responses. Finally, after all six videos had been viewed, the participants were asked to rate their level of immersion, using the IPQ immersion questionnaire. The study is ongoing, but at the moment of writing there have been 150 participants (we will continue to recruit up to 350) for computer screens, 100 (200) participants for VR headsets, of which 80 (180) participants took part in both. These participants were sampled from members of the public and from staff/student populations. The members of the public were recruited using a social media advertising campaign, as well as in-person engagement events run at major institutions around the United Kingdom. The staff/student population were recruited over a month-long period, where they were invited to perform both the VR and the desktop survey. This study will analyse the difference between SP surveys performed on computer screens and within VR environments, using simple difference testing based on the levels of immersion (as measured by the IPQ score), and on the responses. It will then continue to assess any differences in perceived safety based on crossing type, vehicle type and experimental paradigm. Finally, this study produces structural equations relating the choice to cross a road with the perceived safety, participant demographic, crossing type, vehicle type, and further predictors. Provisional data shows being more immersed in VR does not translate into a difference in perceived safety but distinguishes between the respondents' interactions with autonomous vehicles and traditional vehicles.

ABSTRACT. Introduction To better understand the role of shared e-scooters in urban mobility, it is critical for policymakers and planners to answer the questions "who" are the first potential users and "how frequently" do they use e-scooters? To this end, this study explores these questions by jointly studying the e-scooter's potential adoption and frequency of use. In this sense, this study first differentiates potential e-scooter users and investigates the underlying socio-demographic characteristics and built-environment factors. Then, it jointly predicts the frequency of e-scooter use and explores the underlying mode choice behavior as well as socio-demographic and built-environment factors. In the context of intent to use micro-mobility modes, previous studies have mainly focused on bike-sharing (1–3). Dockless electric bikes, as a micro-mobility mode, have similar operations to shared e-scooters. However, e-scooters and electric bikes have different usage patterns (4), and the bike sharing adoption pattern cannot be extended to e-scooter adoption. Moreover, a main body of literature investigates the first adoption of e-scooters (5–7). However, the integration between first adoption (i.e., potential use) and usage frequency has been overlooked. In this sense, this study first contributes to the previous studies by distinguishing the potential users from the frequency of use. A line of literature shows that behavioral and individual-level factors affect the first use differently than subsequent frequency of use (8). Second, it advances the earlier studies by estimating a joint model for e-scooter potential users and frequency of use, as having the potential to use plays an important role in subsequent frequency of use. Data A dataset with approximately 500 respondents from the city of Chicago in June 2021 will be used in this study. The data consists of: (i) the person and household socio-demographics, (ii) the person's daily travel mode choice, and (iii) residential location. In addition, using residential locations, the built-environment characteristics from EPA smart location, crime rate, crash rate, and average daily e-scooter trips were added to the main dataset. In the context of first potential users, respondents indicated whether they would consider e-scooters as a transportation mode option while also deciding whether to use an e-scooter and indicating how frequently they use e-scooters. The distribution of e-scooter usage frequency in the dataset implies an excessive number of potential users that do not use e-scooters. Methodology The excessive number of e-scooter non-users implies that many potential users do not consider e-scooters as a viable transportation mode. To better understand the behavioral patterns among all potential users, we employ the Zero Inflated Ordered Probit (ZIOP) model (9) with random parameters, which divides the respondents' decisions into two stages, each with a different set of explanatory factors. This approach can be interpreted as whether an individual considers e-scooters as an option, and if so, how frequently they use e-scooters, which also has a zero occurrence. Moreover, the random effect of the model facilitates taking the unobserved heterogeneity across observations into the consideration. Expected results It is expected that the socio-demographic characteristics and built environment variables associated with residential locations, such as population density, distance from transit stations, and crime rate, have a significant impact on differentiating potential e-scooter users. Transportation mode choice-related behaviors, on the other hand, are expected to have a greater impact on predicting the frequency of use than socio-demographic and built environment variables. The resulting models will enable planners and policymakers to predict the e-scooter's potential users and frequency of use, thereby providing insights into mobility, environment, and other outcomes. Acknowledgment This submission has been created by UChicago Argonne, LLC, Operator of Argonne National Laboratory (“Argonne"). Argonne, a U.S. Department of Energy Office of Science laboratory, is operated under Contract No. DE-AC02-06CH11357. The U.S. Government retains for itself, and others acting on its behalf, a paid-up, nonexclusive, irrevocable worldwide license in said article to reproduce, prepare derivative works, distribute copies to the public, and perform publicly and display publicly, by or on behalf of the Government. Argonne National Laboratory's work was supported by the U.S. Department of Energy, Office of Vehicle Technologies, under contract DE-AC02-06CH11357. This material is based upon work supported by the U.S. Department of Energy, Vehicle Technologies Office, under the Systems and Modeling for Accelerated Research in Transportation Mobility Laboratory Consortium, an initiative of the Energy Efficient Mobility Systems Program. References 1. Aguilera-García, Á., J. Gomez, and N. Sobrino. Exploring the Adoption of Moped Scooter-Sharing Systems in Spanish Urban Areas. Cities, Vol. 96, No. December 2018, 2020, p. 102424. https://doi.org/10.1016/j.cities.2019.102424. 2. Bi, H., Z. Ye, and Y. Zhang. Analysis of the Integration Usage Patterns of Multiple Shared Mobility Modes and Metro System. Transportation Research Record: Journal of the Transportation Research Board, 2021. https://doi.org/10.1177/03611981211013351. 3. Kim, B., and D. Kim. Exploring the Key Antecedents Influencing Consumer’s Continuance Intention toward Bike-Sharing Services: Focus on China. International Journal of Environmental Research and Public Health, Vol. 17, No. 12, 2020, pp. 1–14. https://doi.org/10.3390/ijerph17124556. 4. Blazanin, G., A. Mondal, K. E. Asmussen, and C. R. Bhat. E-Scooter Sharing and Bikesharing Systems: An Individual-Level Analysis of Factors Affecting First Use and Use Frequency. 5. Eccarius, T., and C. C. Lu. Adoption Intentions for Micro-Mobility – Insights from Electric Scooter Sharing in Taiwan. Transportation Research Part D: Transport and Environment, Vol. 84, No. April, 2020, p. 102327. https://doi.org/10.1016/j.trd.2020.102327. 6. Cao, Z., X. Zhang, K. Chua, H. Yu, and J. Zhao. E-Scooter Sharing to Serve Short-Distance Transit Trips: A Singapore Case. Transportation Research Part A: Policy and Practice, Vol. 147, 2021, pp. 177–196. https://doi.org/10.1016/J.TRA.2021.03.004. 7. McKenzie, G. Spatiotemporal Comparative Analysis of Scooter-Share and Bike-Share Usage Patterns in Washington, D.C. Journal of Transport Geography, Vol. 78, 2019. https://doi.org/10.1016/j.jtrangeo.2019.05.007. 8. Biehl, A., A. Ermagun, and A. Stathopoulos. Utilizing Multi-Stage Behavior Change Theory to Model the Process of Bike Share Adoption. Transport Policy, Vol. 77, 2019, pp. 30–45. https://doi.org/10.1016/J.TRANPOL.2019.02.001. 9. Harris, M. N., and X. Zhao. A Zero-Inflated Ordered Probit Model, with an Application to Modelling Tobacco Consumption. Journal of Econometrics, Vol. 141, No. 2, 2007, pp. 1073–1099. https://doi.org/10.1016/J.JECONOM.2007.01.002.

ABSTRACT. Research motivation Infrastructure-to-vehicle [V2I] and vehicle-to-vehicle [V2V] communication technologies are expected to increase safety. Many researchers have investigated people’s subjective evaluation of the new transport services supported such technologies. Such services should be well supported by road infrastructure to form a safer transport environment. However, existing studies only focus on subjective aspects of the services, but not on the transport environment, which is usually formed by transportation mode, infrastructure, and people, etc. It is crucial to properly present the transport environment under study to people for a better understanding of their subjective evaluation. In this context, conventionally, texts and images are used to design the popular stated preference (SP) surveys; however, texts and images may not be sufficient for people to understand SP scenarios of the environment. It is therefore necessary to explore the roles of non-text/image information by accumulating more scientific evidence.

Research purpose and methodology The purpose of this study is to clarify what types of non-text/image information improve people’s subjective responses to a new transport environment with advanced technologies by jointly addressing the influence of people’s attitudes towards the changed environment on the responses. Subjective responses are measured in terms of perceived safety under the new environment, while attitudes are used to control the potential self-selection biases related to voluntary survey participation. In this study, a demonstration experiment (i.e., real experience) and showing a video clip of the experiment (i.e., virtual experience) are used to compare with the provision of text/image-based information. An SP survey was implemented. A hybrid choice modeling (HCM) framework is built to jointly represent perceived safety and attitude variables. Here, the new transport environment attributes used to form SP scenarios refer to the deployment of new transport services (i.e., connected bus–tram, autonomous vehicle [AV], and autonomous tram) and road infrastructure attributes (i.e., people-friendly tram tracks allowing people to walk freely across tracks and safety fences along tracks for keeping people away from tracks).

Experiment and survey The SP survey was conducted online in January 2021 in Hiroshima City, Japan, which has a tram system. The respondents were divided into three groups: (1) participating in the demonstration experiment implemented in December 2020 (n=34, “demonstration-group”), (2) watching a video showing/explaining the experiment (n=480, “video-group”), and (3) being presented with only text/image information about the new transport environment (n=474, “base-group”). The adopted new technologies under study refer to (1) V2V (tram–bus) communications for assisting bus drivers to smoothly enter and exit from tram tracks, and (2) V2I communications for controlling traffic signal timings that prioritize the operation of buses and trams. An orthogonal fractional factorial design was made based on the above-explained attributes, resulting in 18 representative SP scenarios. These scenarios were randomly grouped into 2 blocks, each of which was assigned to a respondent who answered 9 SP questions. Perceived safety is measured with respect to each scenario, by using the statement “do you feel that the transport environment under study is safe?”, based on a five-point Likert scale (1: strongly disagree, 5: strongly agree). Attitudes are measured by using the following two types of questions “how do you think about the following statements” (1: fully disagree, 5: fully agree): (1) Questions about risk-taking attitudes: - when leaving from home, it is necessary to sufficiently confirm the source of fire (e.g., gas stove) and closure of doors/windows - even during a short leave from home, it is necessary to lock the home door (2) Questions about the ELSI (Ethical, Legal and Social Implications) attitudes towards AI adoption: the necessity of, - clarifying the responsibility in case of accidents/failures caused by AI - making a new insurance system for compensation in case of accidents caused by AI - amending the parts of existing laws/regulations that are not compatible to the AI environment - confirming how often AI-equipped robots fail to work and what kinds of failures may be observed

Modeling The HCM is employed to jointly represent the aforementioned perceived safety and attitudes. An ordered logit model structure is adopted to describe the perceived safety, while a modeling structure with multiple indicators multiple causes (MIMIC) is used to describe the attitudes. Considering that each respondent answered multiple scenarios, a random-effect component is added in the above modeling framework to accommodate the individual heterogeneity. To better quantify the influences of real and virtual experiences, interaction terms between the dummy variables of experiences and SP attributes are introduced. Variables capturing people’s knowledge about the new transport system were included to measure the usefulness of the knowledge to enhance perceived safety.

Results and findings Both real and virtual experiences increase the perceived safety, while the influence of real experience is three times greater than that of virtual experience. For the demonstration-group, installing safety fences along tracks and people-friendly tram tracks increases perceived safety. For the video-group, having AVs and connected bus–tram system increases perceived safety. Both of the two attitude variables significantly affect the perceived safety. Risk-taking respondents are more likely to feel safer under the new transport environment. Respondents concerning about the ELSI in adopting AI technologies tend to perceive a higher level of safety. It is found that risk-taking behavior, real experience, virtual experience, and caring about ELSI of AI affect the perceived safety in that order. The larger influence of real experience than virtual experience is understandable because of richer information obtained from real experience. This research is among the first studies to understand the different impacts of virtual and real experiences on perceived safety under the new transport environment with advanced communication technologies. The details will be reported at the conference.

ABSTRACT. Introduction and Background In the transport literature, discrete choice models are usual tools to analyze travel behaviour. The Random Utility Maximization (RUM) is the most recurrent approach to derive discrete choice models in the literature (Mcfadden, 1974). However, different behavioral approaches have been proposed alternatively to the RUM, such as the Random Regret Minimization (RRM) in which an individual aims to minimize the regret derived from the bilateral comparison of the attributes of the available alternatives (Chorus, 2010; Chorus et al., 2008). The advances of the RRM approach comprise new specifications of the regret function with the extension of the Classical RRM (CRRM) model (Chorus, 2010) to the GRRM model (Chorus, 2014), and the P-RRM and μRRM models (van Cranenburgh et al., 2015). Though it has been recognized the importance of accounting for spatial dependency in the transport phenomena (Bhat & Zhao, 2002), discrete choice models have been less explored in this context due to the autoregressive spatial dependency effects that incur in heteroscedasticity and inversion of the spatial weight matrix (SWM) in maximum likelihood estimators. The SAR (Spatial Lag) and SEM (Spatial Error) autoregressive specifications of spatial dependency are the most explored from the theoretical and applied perspective. Although the SLX (Spatial Lag of X) model is simpler to estimate since it does not requires the inversion of the SWM and neither implies in heteroscedasticity, empirical applications of such models are less common (Vega & Elhorst, 2015). The paper analyses spatial dependency in RRM models by means of SLX specifications in the variables of interest from the transport planning perspective. We explore these effects under the RUM and RRM approaches extending the applications of spatial dependency under the RUM approach found in the literature and incorporating such features into the RRM approach. We derive measures used in transport project appraisal (Value of Time) to analyze the impacts of ignoring spatial dependency.

Data and Method We used revealed preference data of travel mode choices from an Origin and Destination Household Survey conducted in the Metropolitan Area of São Paulo (Metrô-SP, 2017). The choice set considered is composed by Car, Bus and Rail (suburban trains and subway systems). Travel times by Bus and Rail were defined using the OpenTripPlanner using GTFS data from 2017. Travel times by Car were determined using the TomTom API that considers congestion throughout the day. Travel costs by Bus were calculated based on the public transport charging value, and by Rail were calculated according to the distance to access train/metro stations. Travel costs by Car were calculated considering a fixed value and a variable component as a function of the distance. The attributes of the alternatives were travel time and travel cost, while the sociodemographic attributes were age, sex, number of individual daily trips, number of household members, social class, trips made in morning peaks (between 7 am and 9 am) and level of education. We defined two spatial weight matrices: distance (individuals who are up to a certain threshold distance receive a weight equal to one, otherwise zero); and k-nearest neighbor (knn, individuals considered neighbors receive weight equal to one, zero otherwise). We varied the thresholds between 1,000 meters to 2,000 meters by 100 meters for the distance criterion, and from 10 to 100 neighbors by 10 for the knn criterion.

Model specification We specified non-spatial RUM models with additive and linear-in-parameters utility functions, and non-spatial RRM models with the CRRM regret function specification. SLX models are specified by including lagged variables, specified by the product between the spatial weight matrix and the respective attribute as shown in Eq. (1). The specification of the spatial RUM and the proposed spatial RRM model are shown in Eq. (2) and Eq. (3), respectively. All Logit models are estimated via maximum likelihood in the apollo package in R (Hess & Palma, 2021). (1) Uploaded file (2) Uploaded file (3) Uploaded file where z_ikn is the k-th spatial lagged attribute of the k-th attribute x of alternative i of individual n; w_pq is the 〈p,q〉-th element of the spatial weight matrix of dimension NxN; U_in is the utility of alternative i for individual n; R_in is the regret of the alternative i for individual n;

Results Preliminary results show that non-spatial RRM spatial models outperforms their RUM counterparts. The results also indicate that incorporating spatial dependency in the independent variables improves their performance in terms of log-likelihood. Differences in VTT measures are found between spatial and non-spatial RUM models of around 35% for the Car alternative where spatial parameters are significant, showing that neglecting spatial dependency biases the estimations. The opposite was found when comparing the RRM models, where no significant differences on the VTT among spatial and non-spatial specifications were found.

References Bhat, C. R., & Zhao, H. (2002). The spatial analysis of activity stop generation. Transportation Research Part B: Methodological, 36(6), 557–575. https://doi.org/10.1016/S0191-2615(01)00019-4 Chorus, C. G. (2010). A New Model of Random Regret Minimization. EJTIR, 10(10), 181–196. https://doi.org/10.18757/ejtir.2010.10.2.2881 Chorus, C. G. (2014). A Generalized Random Regret Minimization model. Transportation Research Part B: Methodological, 68, 224–238. https://doi.org/10.1016/j.trb.2014.06.009 Chorus, C. G., Arentze, T. A., & Timmermans, H. J. P. (2008). A Random Regret-Minimization model of travel choice. Transportation Research Part B: Methodological, 42(1), 1–18. https://doi.org/10.1016/J.TRB.2007.05.004 Hess, S., & Palma, D. (2021). Apollo: a flexible, powerful and customisable freeware package for choice model estimation and application version 0.2.5 User manual. www.ApolloChoiceModelling.com Mcfadden, D. P. (1974). Conditional logit analysis of qualitative choice behavior. In Analysis of Qualitative Choice Behavior (1st ed.). Metrô-SP. (2017). Pesquisa Origem e Destino 2017. http://www.metro.sp.gov.br/pesquisa-od/resultado-das-pesquisas.aspx van Cranenburgh, S., Guevara, C. A., & Chorus, C. G. (2015). New insights on random regret minimization models. Transportation Research Part A: Policy and Practice, 74, 91–109. https://doi.org/10.1016/j.tra.2015.01.008 Vega, S. H., & Elhorst, J. P. (2015). The slx model. Journal of Regional Science, 55(3), 339–363. https://doi.org/10.1111/jors.12188

ABSTRACT. How do preferences for biodiversity vary over space and time? This research uses a choice experiment (CE) to investigate the Spatio-temporal stability of preferences, expressed in Willingness-To-Pay (WTP) for different attributes of woodland biodiversity. This is, to the best of our knowledge, one of the few studies that have investigated spatial and temporal heterogeneity in environmental preferences. Furthermore, we use unique CE attributes that ask respondents about their sensory experiences of woodlands. With this design, we elicit preferences for changes in the range of sounds, smells, colours, and ecological processes experienced by respondents in woodlands.

To date, most work into the temporal stability of CE results has been typically as part of a test-retest approach with two periods. Test-retest designs have been used to evaluate how preferences change over time by conducting multiple waves of a survey (Mariel et al., 2021). Some studies, (Elsassre, Englert, and Hamilton, 2010., Fedrigotti et al, 2020), have also tested seasonal differences in WTP albeit with few seasons. However, these studies focus on the stability of preferences themselves without examining how exogenous factors influence preferences. One such factor important to the choice modelling literature is seasonality. If preferences and WTP are shown to vary seasonally, this has implications to practitioners for the appropriate timing for survey-based research. A second motivation to study the effect of time-variant factors on WTP is the appropriate location of new woodlands. For instance, if levels of local estimates for willingness-to-pay for experiences in nature can be used to help site new woodlands or invest in restoring and conserving existing woodlands. Finally, spatio-temporal methods also suggest the possibility for predicting the direction of WTP in the future.

Data: The data has four seasonal components (Jan-March, April-May, June-August, November-December) each with ~1500 completed usable respondents for 6000 total respondents. Quota sampling was used to ensure that respondents were representative of the general population in the study area (Great Britain). Each respondent answered the questionnaire in a single season, undertook nine choice tasks, and identified the location of their household and nearest woodland. Other data collected measured socioeconomic and wellbeing factors to control for observed preference heterogeneity. In this research, we present choices models with the seasonal and pooled data and then spatio-temporal statistics using estimated WTP. However, we also comment on further lines of inquiry facilitated by this data structure.

Glenk et al (2020) described an emerging trend of CEs integrating spatial aspects, for instance spatially explicit designs (a location attribute) or sampling (comparing between areas). This research asked respondents from across Great Britain (England, Scotland, and Wales) about preferences for their local woodlands. Both respondents and their local woodlands are geocoded to facilitate investigation into global and local spatial autocorrelation.

Analysis: Following the approach pioneered by Cambell et al (2009) and later used by Czajkowski et al (2017) and Toledo-Galegos et al (2020), we use a two-stage process of estimating a model, we then estimate spatio-temporal statistics using the individual-level WTP estimated from the conditional posterior distributions. We estimate mixed logit models for each season and the pooled all-seasons data. The model is estimated in WTP-space, with an Alternative-Specific-Coefficient (ASC) for the status quo alternative and features one coefficient for each non-zero level of the non-monetary attributes.

Latent-Class Models where segmentation is driven by spatial (visits to and distance from woodland, country of respondent) and temporal (season of response) factors are also estimated but perform relatively poorly to mixed logits in terms of goodness-of-fit and plausibility of WTP.

In the second stage, WTP is then aggregated at the NUTS3 sub-regional level leading to 168 areas in Great Britain. The average population of for each area was 360,000 and corresponds to English County, Scottish Districts, and Welsh unitary authorities. This level of spatial resolution is used to balance the number of respondents per area and permit the identification of local differences in WTP. The spatial weights are calculated using the common K-Nearest-Neighbours approach where inverse distance weighting is used to link respondents. The calculation of spatial weights allows the estimation of Global and Local measures of spatial autocorrelation (Moran’s I and Getis-Ord statistics) and spatial lag models. Again, each measure is calculated for each season and then with the pooled data.

Results: The initial results can be described in terms of WTP and spatial clusters. With regards to the WTP, respondents were willing to pay between -£10 and +£20 in additional local taxation per year for a marginal change from a low to a high level of colours, sounds, smells, and ecological processes in woodlands. Colour was the most preferred attribute, then smells. However, there were non-linear preferences for sounds and ecological processes whereby medium levels were preferred to higher levels. This result suggests that increasing the range of colours and smells is important to garner support for the management of new woodlands. The research then examines spatial and temporal autocorrelation. No global spatial correlation was detected for any attribute in any season. Furthermore, there was limited evidence for local spatial correlation. The existence of spatial clusters in each season are tested using both Local Moran’s I statistic and Getis-Ord statistic to validate although only a few hot (cold) spots of WTP were detected. An important result for the choice modelling literature is that those spatial clusters of WTP that do exist are time-sensitive and rarely persist for more than a season. Indeed, no seasonality in WTP was detected and there was very limited temporal autocorrelation in WTP. The absence of persistent clusters indicates that WTP for woodland biodiversity was stable across seasons.

The result of no spatial correlation in WTP values means that policymakers may locate new woodlands without welfare losses. Moreover, policymakers may count on sustained support for woodland management given the temporal stability of preferences. Finally, this research shows that preferences for biodiversity are relatively stable over spatial and temporal dimensions.

ABSTRACT. Stated preference surveys are frequently used for the valuation of environmental goods, and increasingly so for the valuation of urban green. To derive valid and reliable value estimates for policy impacts on spatially distinct environmental goods, it is important to understand how individuals’ values depend on their distance to the goods and on the distance to and quantity of their endowment and potential substitutes (Glenk et al., 2020). We aim to investigate how residents’ valuation of an extension of green spaces relates to the distance from their place of residence and their current endowment. For this goal, we compare model fit and the precision of estimated welfare measures of different approaches proposed in the distance decay literature, as well as new model specifications which we hypothesize to be more suitable for area based changes in the quantity of an environmental good.

Distance decay in the marginal utility for additional green spaces can be expected based on findings that individuals mostly visit green spaces in close proximity. Urban green may also provide value beyond recreational use, such as the amenity value during other activities like commuting. Distance decay is plausible also for this value component, because activities tend to be more frequent closer to the place of residence. Further, economic theory predicts marginal utility for additional green spaces to decrease with the quantity of the endowment in green spaces. This matters for valid value estimates, because endowments can vary greatly between different neighborhoods.

The most common approach for investigating distance decay is to model utility as a continuous function of distance between the place of residence and the environmental good (Glenk et al, 2020). Holland and Johnston (2017) propose a quantity-within-distance approach as an alternative: Instead of the distance to the good, they include the quantity of the good within a certain distance in the utility function. We hypothesize that this a more suitable framework if the policy change is rather area than point based and not homogeneously distributed. In such cases, the point to use for the calculation of the distance between residence and the policy change is not obvious. Also, this area-based approach may better capture amenity value during activities in motion. Quantity-within-distance further enables an integrated analysis of diminishing marginal utility with distance and with the quantity of the endowment in the environmental good. These factors may make this approach more suitable for the context of urban green.

Our study investigates how suitable a quantity-within-distance approach is to jointly model distance decay and diminishing marginal utility for urban green spaces. It contributes to the literature in two ways. First, it applies the quantity-within-distance model as proposed by Holland and Johnston to the new context of urban green. Holland and Johnston suggest a specification targeted at investigating quality changes. However, many valuation studies, as also our case study, investigate the value of quantity rather than quality changes. Thus, second, we develop and test alternative model specifications that may be more suitable for quantity changes.

We use data from choice experiments that were conducted in August and September 2021 in the German cities of Berlin and Leipzig with 759 and 966 respondents, respectively. The choice experiments considered the extension of green spaces within a predefined policy area with a size of 1.4 and 2.3 km², respectively. Respondents were recruited within and outside of the policy area by sending invitation letters to a random sample from the official registries of the two cities. As part of the questionnaire, respondents marked the place of their residence on a map. With this and further data on current land use, the distance and quantity-within-distance variables are calculated.

We estimate six random parameter logit specifications that model distance decay in different ways. First, we include the distance between the place of residence and the center of the policy area as an additive interaction term, as is the most common approach in the literature. Second, we apply Holland and Johnston’s model to our data. Third, we propose and estimate a new quantity-within-distance model based on the quantity of policy change. Fourth, we combine the proposed model with Holland and Johnston’s model, enabling us to simultaneously investigate distance decay of the policy change and diminishing marginal utility with increasing quantity of endowment. The fifth and sixth models are specifications with less restrictive assumptions about distance decay, allowing non-zero marginal utility beyond the boundary of the first quantity-within-distance variable.

We compare the models based on their fit measured in LogLikelihood, AIC, and BIC, as well as on the size and confidence intervals of the estimated total economic value of the extension in green space. Results on model fit show how well the modelling approaches explain the stated preferences in the case study. The differences in total economic value between the approaches serve as indications of how important the choice of the right specification is in similar valuation studies. Deriving confidence intervals for the total economic value allows conclusions about which approach is most precise in calculating estimates of the economic value.

Our preliminary model results suggest that we will be able to derive recommendations for model specification decisions in choice experiment studies where the value of an environmental good is subject to distance decay. These recommendations will be based on our results on model fit and precision of welfare estimates, the ease of calculation and interpretation of the models, as well as a theoretical discussion of how the size and homogeneity of the policy change, the relevance of amenity values, and the relevance of substitutes mediate the relative advantage of quantity-within-distance over more common modelling approaches.

References:

Glenk, K., Johnston, R.J., Meyerhoff, J. and Sagebiel, J., 2020. Spatial dimensions of stated preference valuation in environmental and resource economics: methods, trends and challenges. Environmental and Resource Economics, 75(2), pp.215-242.

Holland, B.M. and Johnston, R.J., 2017. Optimized quantity-within-distance models of spatial welfare heterogeneity. Journal of Environmental Economics and Management, 85, pp.110-129.