ABSTRACT. Please see attached PDF.

ABSTRACT. (Note that there are 41 authors in total for this paper, only the lead author is shown in the submission)

Despite unprecedented progress in developing COVID-19 vaccines, global vaccination levels needed to reach herd immunity remain a distant target, while new variants keep emerging. Obtaining near universal vaccine uptake relies on understanding and addressing vaccine resistance. Information on uptake amongst those people offered a vaccine so far (as opposed to absolute numbers of vaccines administered) is difficult to obtain, with the potential of upwards bias due to many (but not all) countries prioritising vulnerable groups for early vaccination, where uptake might be higher. In addition, comparing vaccination rates across countries leads to another potential source of bias as access to vaccination varies substantially, especially in developing countries. This then motivates a focus on potential vaccine uptake in the overall population.

Simple questions about vaccine acceptance however ignore that the vaccines being offered vary across countries and even population subgroups, and differ in terms of efficacy and side effects. At the same time, there is extensive scope for heterogeneity in preferences across individuals, across countries, and over time. An increasing number of studies have looked at this issue using stated choice (SC) surveys. The present paper adds to this, but is different along a number of key dimensions:

1. Data was collected in 18 countries/territories covering all 6 inhabited continents, using a consistent survey design, including the core stated choice component. 2. The survey was longitudinal in nature, with respondents sampled three times, at regular intervals. 3. Alongside indicating their own preferences, respondents were also asked how they would choose if they were to make the decision for other members in their household, including children.

Participants were faced with six hypothetical vaccination choice scenarios. In each choice scenario, they were presented with two different vaccines. These were described on the basis of key vaccine characteristics such as efficacy, risk of side effects, and expected protection duration. Respondents were told that, given the need to vaccinate very large parts of the population, and limits on supply, there would be a wait before they could receive a vaccine. However, they could also obtain vaccination immediately by paying a one-off fee.

We use latent class nested logit models on the resulting data, and show a substantial influence of vaccine characteristics, but with substantial heterogeneity within and across countries. Uptake increases if more efficacious vaccines (95% vs 60%) are offered (mean across study areas=3.9%, range of 0.6% to 8.1%) or if vaccines offer at least $12$ months of protection (mean across study areas=2.4%, range of 0.2% to 5.8%), while an increase in severe side effects (from 0.001% to 0.01%) leads to reduced uptake (mean=-1.3%, range of -0.2% to -3.9%). Additionally, a large share of individuals (mean=55.2%, range of 28% to 75.8%) would delay vaccination by 3 months to obtain a more efficacious (95% vs 60%) vaccine, where this increases further if the low efficacy vaccine has a higher risk (0.01% instead of 0.001%) of severe side effects (mean=65.9%, range of 41.4% to 86.5%). Our work highlights that careful consideration of which vaccines to offer can be beneficial. In support of this, we provide an interactive tool to predict uptake in a country as a function of the vaccines being deployed, and also depending on the levels of infectiousness and severity of circulating variants of COVID-19 (https://stephanehess.shinyapps.io/COVID19_Shiny/).

In addition to the analysis of vaccination preferences by individual respondents at a single point in time, we also study how preferences change over time. Here, we find an astounding level of stability in preferences across the three waves, suggesting that the status of the pandemic and personal experiences play less of a role in vaccine uptake than the characteristics of the vaccines that are offered.

Finally, respondents were first asked to make choices for themselves, before they were then presented with scenarios that also involved the other members of their household. Here, we can make interesting behavioural findings, with individuals changing the choices for themselves in the presence of needing to also choose for others (notably a lower uptake of the paid rapid access options) while also choosing different vaccines for different members of their household, or choosing to only vaccinate some people.

ABSTRACT. INTRODUCTION

The new SARS‐CoV‐2 coronavirus that created the COVID-19 pandemic has led to an avalanche of research in this virus as well as the impact it has on our lives (Haghani and Bliemer, 2020). Choice modellers in applied economics have also taken an interest in COVID-19 related research or pandemics in general. For example, choice experiments have been conducted regarding vaccination uptake and distribution during pandemics (e.g., Determann et al., 2016; Luyten, 2020) or to assess trade-offs between health impacts and economic impacts (e.g., Chorus et al., 2020; Manipis et al., 2021). Such trade-offs are typically considered morally problematic, diabolical, or taboo (Chorus et al., 2018).

In our study we investigate preferences of citizens regarding restrictions imposed during future pandemics. Restrictions affect (in opposite directions) the number of lives lost as well as the number of jobs lost, which means that governments are faced with a moral conundrum. While governments in some countries have imposed extended lockdowns to save lives, governments in other countries have prioritised the economy. We developed a stated choice experiment in which we asked residents of 8 countries across 6 continents to choose between two options that varied in the level of restrictions and the impact on deaths and job losses. Each participant in the choice experiment was given both a local and a national context. In the local scenario, respondents were asked to consider restrictions and impacts in their local community of 100,000 people, while in the national scenario the same restrictions were considered to affect the whole nation (with the same rates of death and job losses applied to the country population, the impact was represented by much larger absolute numbers of deaths and job losses).

There were some unique challenges in both the design of the choice experiment as well as the estimation of the model as described below.

STATED CHOICE EXPERIMENT

Each alternative is represented by eight attributes, namely the level of restrictions with 13 levels, the number of deaths in each age category (child, adolescent, young adult, middle aged, senior) and job losses in two categories (short term, long-term). To reduce choice task complexity, we tested several graphical representations for each attribute and its corresponding level, ensuring that they would be understandable across all countries.

For each of the 13 restriction levels, where level 1 means no restrictions and level 13 is lockdown, we determined allowable activities based on expert advice regarding the risk level of each activity. Activities ranged from visiting gyms and churches to visiting restaurants, cinemas, and dance clubs. Each activity was depicted with a pictogram, with increasing restrictions meaning more activities crossed out.

A complication with the 11 levels for the number of deaths and job losses was that they were exponentially increasing to account for a large variation in death rates across countries (as was also observed in the COVID-19 pandemic). We assumed the following levels for the number of deaths per 1,000,000 inhabitants and job losses per 10,000 inhabitants: 1, 5, 10, 20, 50, 100, 200, 500, and 1000. Graphical representations with grids or bars do not work well with a large number of exponentially increasing levels, therefore we designed a novel graphical representation where the colour brightness of the corresponding pictogram was used to accentuate its level.

In generating the profiles, we assumed that a future virus could either have a higher mortality rate in younger people (e.g., Spanish flu) or in older people (e.g., COVID-19). An efficient experimental design with 1000 choice tasks based on noninformative priors was generated using the modified Federov algorithm with a candidate set of 10,000 choice tasks where constraints were imposed on attribute level combinations to ensure realism and avoid dominance. The choice experiment was part of an online survey conducted in December 2020 and will be repeated in December 2021 across the 8 countries.

PRELIMINARY MODEL ESTIMATION RESULTS

Since the attribute levels for deaths and job losses were exponential, it is not possible to estimate a traditional discrete choice model with linear additive terms. While a logarithmic transformation of the levels leads to an estimable model, trade-offs between deaths and job losses for the various categories cannot be easily interpreted. Therefore, we resorted to a multiplicative discrete choice model proposed by Fosgerau and Bierlaire (2009). Via a novel representation of the parameters via orthogonal polynomial coding we estimated linear, quadratic, and cubic effects in the aversion towards deaths for different age categories.

In the local scenario we found that there was a willingness to make 450-500 workers unemployed to save 1 life, whereas in the national scenario (where the number of deaths shown is much larger) we found this to increase to 900-950 workers. On average, our respondents valued a life of a child 3x more than a life of a senior.

Further analysis will be conducted to uncover differences between countries and to compare shifts in preferences from a pre-vaccination world in December 2020 to a mostly post-vaccination world in December 2021.

REFERENCES

Chorus, Pudane, Mouter & Campbell (2018) Taboo trade-off aversion: a discrete choice model and empirical analysis. Journal of Choice Modelling.

Chorus, Sandorf & Mouter (2020) Diabolical dilemmas of COVID-19: An empirical study into Dutch society’s trade-offs between health impacts and other effects of the lockdown. PLoS ONE.

Determann, Korfage, Fagerlin, Steyerberg, Bliemer, Richardus, Lambooij & de Bekker-Grob (2016) Public preferences for vaccination programmes during pandemics caused by pathogens transmitted through respiratory droplets – a discrete choice experiment in four European countries. Eurosurveillance.

Fosgerau & Bierlaire (2009) Discrete choice models with multiplicative error terms. Transportation Research Part B.

Haghani & Bliemer (2020) Covid-19 pandemic and the unprecedented mobilisation of scholarly efforts prompted by a health crisis: Scientometric comparisons across SARS, MERS and 2019-nCoV literature. Scientometrics.

Luyten, Tubeuf & Kessels (2020) Who should get it first? Public preferences for distributing a COVID-19 vaccine. Covid Economics.

Manipis, Street, Cronin, Viney & Goudall (2021) Exploring the Trade-Off Between Economic and Health Outcomes During a Pandemic: A Discrete Choice Experiment of Lockdown Policies in Australia. The Patient.

ABSTRACT. In many different contexts deciders choose between a finite set of alternatives repeatedly. Prime examples are mode choice in transportation, product choices in marketing and surveys in general. A prominent model for such data is the multinomial probit (MNP) model. Often in this context unobserved heterogeneity in a panel data context is modelled using mixing of parameters leading to the mixed MNP (MMNP) model.

The estimation of a MMNP model is hampered by computational complexity involved in the evaluation of a high dimensional Gaussian cumulative distribution function. As a potential remedy (Bhat and Sidharthan, “A Simulation Evaluation of the Maximum Approximate Composite Marginal Likelihood (MACML) Estimator for Mixed Multinomial Probit Models”, Transportation Research B, 2011) proposed to use the composite marginal likelihood (CML) approach (see (Varin, Reid, and Firth, “An overview of composite likelihood methods”, Statistica Sinica, 2011)), in situations where the number of repeated choices per decider is large. This proposal reduces the computational load by replacing the full likelihood by a CML formed by including the probabilities of all pairs of choices. This approach is called full pairwise (FP) in the following.

To further reduce the numerical load, a number of different CMLs beside the FP approach have been suggested, including the restriction to adjacent pairs (only including consecutive pairs of decisions; in the following called AP). (Cox and Reid, “A note on pseudolikelihood constructed from marginal densities”, Biometrika, 2004) use a weighted sum of univariate and bivariate marginals. We add the univariate marginals to the AP approach (APU in the following). These proposals reduce the computational costs, but they also have consequences with respect to the statistical efficiency of the corresponding estimates. Using the CML in place of the full likelihood, we incur a penalty in the form of a larger asymptotic variance of the estimators of the various parameters.

The penalty is not necessarily the same for all parameters. While in the literature it is suggested that the loss in efficiency is small (Bhat and Sidharthan, “A Simulation Evaluation of the Maximum Approximate Composite Marginal Likelihood (MACML) Estimator for Mixed Multinomial Probit Models”, Transportation Research B, 2011) lists a worst case of 20% efficiency loss compared to maximum likelihood estimation for the probit models in the cases investigated; (Katsikatsou et al., “Pairwise likelihood estimation for factor analysis models with ordinal data”, Computational Statistics & Data Analysis, 2012) indicate only modest efficiency losses for ordered probit), it is currently unknown to the best of our knowledge which factors contribute to a larger relative loss in efficiency and hence a larger statistical cost of using the CML approach.

This is where the contribution of this paper lies: Based on recently implemented analytical expressions for the asymptotic variance (conditional on the regressors) we investigate using a simple model the main driving factors of the efficiency losses. The simple model here consists in a MMNP with two alternative varying regressors, both chosen independent and identically distributed from a continuous uniform [−1,1] distribution. The corresponding coefficient of the first regressor contains a normally distributed random effect with expectation β_1 and variance ω^2. The coefficient for the second regressor is fixed to 0.5 for identification. There are T choice decisions per decider to choose between J alternatives. The random error in the random utility is normally distributed with zero mean and variance σ^2I_J.

We investigate the influence of all five characteristics, the expectation β_1 and the variance ω^2 of the mixing distribution, the variance σ^2 of the error term, the number T of choice situations per decider, the number J of choices within each situation.

We obtain the following results: •In all cases the full likelihood results in the best accuracy followed by FP. •AP and APU consistently provide less accurate results than FP, while the ranking between AP and APU differs from experiment to experiment. The estimation of the parameter β_1 is improved in some situations by adding the univariate marginals. •The efficiency loss for FP in almost all cases is small (ranging up to 10%, say) for the estimation of β1. •For σ efficiency losses are largest if it is close to zero. The maximal observed efficiency loss in our experiments amounts to 25% for FP. •For ω the efficiency loss is largest for small values close to zero. Again for FP only modest efficiency losses are observed up to 12% while the losses for AP range close to 50% already for T = 3. •For the number of choice alternatives we do not get a clear picture in our calculations. •Predictably the strongest influence is the number of choice situations. In this respect for the estimation of β_1 and σ the efficiency loss is bounded and modest, while for ω we witness a roughly linear increase ranging up to 76% for FP and T = 6 (for AP we obtain a loss of 136%).

Summing up we conclude that the efficiency loss effects the various parameters differently. The FP for most parameters only leads to modest efficiency losses up to roughly a magnitude of 10% in our experiments. Only the parameter ω corresponding to the random effects show a bigger loss of efficiency which can be quite substantial for a large number of repeated choice situations. The losses are most pronounced if the variances of the random effects are small. Intuitively this makes sense: Estimation efficiency is most valuable if the information in the data is small. The same happens for the parameters for σ while the parameter estimators for β_1 appear to be less affected.

Therefore, we find the largest loss of estimation accuracy in cases with many choice situations per person and only a modest unobserved heterogeneity. In such cases the larger relative loss of efficiency is countered by a generally greater accuracy due to larger samples. Additionally our experiments illustrate that the efficiency loss for the parameters not pertaining to the random effects is limited. Both results provide some support for the usage of the CML approach for large panel data sets and reconfirm intuition.

ABSTRACT. Discrete choice models based on random utility maximization theory typically treat the choice alternatives as independent and the choice made is absolute in the sense the utility of the single chosen alternative must be larger than the utility of others. These are without any theoretical problems when a single choice is made among the multiple alternatives, e.g., someone chooses car in a transportation mode choice. However, in some situations where an individual may want to choose more than one option from the provided alternatives, the application of utility maximization-based models may have to be simplified, e.g., by assuming one alternative overrides the others. Such a choice situation is not rare and is perhaps getting more popular with the progress of ICT and people’s increasing awareness of the various products in the market. For instance, in the case of a stated choice experiment which deals with the transportation modes choice involving car, bus and shared bike, the three alternatives are commonly assumed independent. However, in a fully general sense, people may prefer both bus and bike because he/she may have a subscription of MaaS. Conventional choice experiment may just allow one choice or simply offer bus+bike as one independent alternative. Relative to the single choice among the three, we could allow people to indicate their scaled preference for each alternative, based on which the real preference on more than one alternative can be captured.

In another example of energy equipment choice where we allow company stakeholders in tertiary section to choose one or more innovative energy facilities among the three offered alternatives, including solar panel, heat pump and ventilation. The idea is to understand to what extent stakeholders of small companies would like to purchase the innovative energy facilities, such that the energy efficiency can be dramatically improved for the tertiary sector. Therefore, we designed a stated choice experiment considering the three alternatives and a number of alternative attributes and generated a fractional factorial design. Here, one of the interesting issues is that the purchase of innovative energy facilities is relatively a long-term investment, thus one could assume companies may purchase one more equipment at once such that the benefits can be maximized over time. Statistically data already indicate the intention people to purchase/install combined/integrated innovative energy equipment, e.g., ventilation with heating function, the electricity consumed by ventilation may be compensated by solar panel. Therefore, this seems to be an ideal case to investigate the choices of more than one alternative.

We implemented the design based on the principle of stated adaptation, considering the various attributes of different alternatives. For instance, the price of each energy equipment is set as relative to the area of offices. For solar panels, we consider municipal subsidy, the reduction in energy charges, CO2 reduction and payback years. For heat pump, we additionally included the functionality (cooling + heating or heating only). For ventilation, instead of the inclusion of energy cost reduction, payback year, CO2 and functionality, we considered binary attributes of having demand control, can heat recovery and filtering out pollutants. In total, we have twenty 4-level attributes and three 2-level attributes, a full factorial design results in 420×23 combinations. Therefore, we use an orthogonal fractional factorial design which includes 64 combined profiles. The choice sets were then randomly assigned to the respondents. We distributed the questionnaires online and obtained the data from 126 companies located mainly in the Netherlands. The manager in each company was invited to finish eight choice sets which were randomly selected from the experiment, which leads to a total of 1008 choice observations.

The final choice scenarios we presented to the company stakeholders involves the three alternatives, while for each alternative, we allow people to indicate a scale of the preference (i.e., from 1 to 5 to indicates ‘definitely not choose’ to ‘definitely choose’), without any constraints on their decisions. In an ideal situation, people who prefer one of the three alternatives should indicate 5 for the chosen alternative and 1 for the other two alternatives, while people who prefer two out of the three should indicate a same score for the chosen (two) alternative which is higher than the score of the left alternative. Because we do not control the choice made by the respondents, it could be also possible that people indicate an equal preference among the three. This case may be approximated as no choice or a random choice among the three alternatives.

What is interesting is also related to the scales of the alternatives which are not the highest, because any indication larger than 1 could indicate a preference of that alternative, which is different from ‘not choosing’, and this may violate the utility maximization for single choice. The possible challenge is regarding the modeling approach because respondents compare the different alternatives, and also indicate the scale of preference for each alternative. Because the scales between different alternatives is comparable, the chosen alternatives become correlated. Therefore, traditional modeling approach which simply merge the possible alternative combinations is considered impropriate.

Therefore, in this paper, we propose a copula-based model which can flexibly incorporate different types of margins to describe the joint distributions. More specifically, the copula models integrate the ordered Probit models and prospect theory models. For each alternative, an order Probit model will be specified according to the attributes of alternatives. Because it is important to select a proper copular function, we intend to examine the main types of copula functions, including Gaussian copular, the Farlie-Gumbel-Morgenstern copular, Student’s t-copula and the Archimedean copulas. In addition, random error components will be also included to deal with the potential correlation between alternatives. The integrated model is expected to interpret the choice preference of each alternative and the correlations between the multiple alternatives.

ABSTRACT. Various random utility models that have been proposed in the past for modelling ordered choices, such as the ordered logit, ordered probit, ordered GEV or the generalized ordered Logit model (Small, 1987). These models are based on the cumulative probability distribution of the response variable, and the basic idea is to assume an underlying continuous latent process and discretizing it into finite number of thresholds on a real line. Although the thresholds are assumed to be in an ascending order, the posterior probabilities are not guaranteed to be unimodal. Furthermore, the ordered logit approach causes the model to lack the flexibility of the multinomial logit (MNL) or probit (Small, 1987).

The objective of this study is to improve on current methods of modelling ordered choices by capturing the unimodal characteristics of sequential ordered choices. We introduce a new choice modelling framework, the Unimodal Ordered Logit model, which incorporates a utility correction term into a multinomial logit model and applies a unimodal probability mass function on the ordered choice responses. This allows the model to have the flexibility of the MNL model, while also capturing the correlations of sequential ordered responses. The probability mass function can be in the form of a Poisson, zero-truncated Poisson, or negative binomial distribution. Our assumption is that for any set of choices which are ranked or ordered in a qualitative manner, e.g., from best to worst, the probability of an individual ranking a set of ordered choices follow a unimodal distribution.

While Ordered Logit models account for the ordering of the choices through the estimation of a set of thresholds, unimodality of the resulting choice probability distribution is not explicitly taken into account. This means that the choice probabilities of an individual may not follow a logical sequence and may generate illogical probability mass function. For instance, using the presidential questionnaire example given in Train (2003) with five possible ranked choices (from "very poor job" to "very good job") it would not make sense if a respondent's opinion probabilities from highest to lowest is ranked as "good job", followed by "very poor job", then "neutral", and so on. A graphical illustration of this problem is presented in (Wong et al., 2021).

Unimodality implies that the overall preference distribution is unimodal and, for any given alternative that is the preferred choice of an individual, the non-selected alternatives adjacent to the preferred choice have a monotonically decreasing order of preference, in both directions (Da Costa et al., 2008). We mathematically define that the model captures the ordering of choices if there exists an integer $c \in J$ such that: $p(y_{i}|X) \geq p(y_{i+1}|X)$, for all $i \geq c$, and $p(y_{i-1}|X) \leq p(y_{i}|X)$, for all $i \leq c$, in a set of $J$ alternatives. This approach imposes a unimodal distribution constraint on the utility such that the choice probability distribution is unimodal and consistent with the ordering of the choices.

A lack of unimodality may result in model misspecification. In a previous study (Wong et al., 2021), we analyzed a crash severity dataset and found that unimodality improves model fit and crash severity prediction over an ordered logit or MNL model. However, this study was only limited to discrete count models.

In this conference, we shall present the theoretical background of the Unimodal Logit Model and improve the utility formulation of our proposed model on choice preference surveys. We note that this is an on-going work, and a continuation from Wong et al. (2021). To illustrate the benefit of the Unimodal Logit model, we apply our method on the COVID Impact Survey Open Dataset (Wozniak et al., 2020), COVID Impact Survey dataset: https://www.covid-impact.org/. This study applies the Unimodal Ordered Logit model to analyze the effects of different socio-economic variables and perceptions on the willingness to participate in testing and tracking during the pandemic. The dataset is a U.S. nationwide survey of the U.S. adult household population, conducted by NORC at the University of Chicago during the pandemic in April, May, and June of 2020. The survey includes likert-scale questions on social perception, physical and mental health, economic security and employment, and basic demographic information such as age, gender, household size, income, race, and ethnicity. We define the dependent variables as the respondents' willingness to participate in testing and tracking in order to slow the spread of the virus (from "extremely likely" to "not likely at all"). We evaluate the applicability of the Unimodal Ordered Logit approach on the COVID survey dataset and understand how it differentiates from an Ordered Logit model in terms of model estimates, goodness of fit indicators, and out-of-sample predictions. The out-of-sample predictions are assessed using three evaluation metrics: discrete classification accuracy (DCA), geometric mean probability of correct assignment (GMPCA) and quadratic weighted kappa (QWK).

References

Small, K., 1987. A discrete choice model for ordered alternatives, Econometrica: Journal of the Econometric Society, 409–424.

Train, K. E. (2003) Discrete Choice Methods with Simulation, Cambridge university press.

Da Costa, J., Alonso, H. and Cardoso, J., 2008. The unimodal model for the classification of ordinal data, Neural Networks, 21 (1) 78–91.

Wozniak, A., Willey, J., Benz, J. and Hart, N., 2020. COVID Impact Survey. Chicago, IL: National Opinion Research Center.

Wong, M., Martín-Baos, J.Á. and Bierlaire, M., 2021. A Unimodal Ordered Logit model for ranked choices. In: 21st Swiss Transport Research Conference}, Sept. 12--14, Monte Verita, Ascona, Switzerland.

ABSTRACT. Research question

This study examines how individuals’ preferences for transport safety depends on the mode of transportation using a stated preference (SP) survey on a Swedish web panel. Whereas there is a rich literature, both in Sweden and internationally, on the value of road safety, less evidence exists for the value of safety for other transport modes. Moreover, many studies have focused on estimating the value of reducing fatality risk, which means that there is less knowledge concerning the value of reduced injury risk. The aim of this study is to elicited monetary values for traffic safety in Sweden to be considered for policy purpose, and two specific objectives are to: (i) examine if preferences differ between mode of transport, and (ii) to estimate both the value of a reduced fatality risk, often defined as the value of a statistical life (VSL), and the value of a reduced injury risk, defined as the value of a statistical injury (VSI). Another objective is to examine to what extent these values may be influenced by the valuation method. To examine these three objectives an SP study was conducted in Sweden in the spring of 2021 using both discrete choice experiments (DCE) and the contingent valuation method (CVM). Moreover, a forth objective was to examine scope sensitivity in line with Andersson et al. (2016). Andersson et al. found that DCE can be expected to be plagued by the same lack of strong scope sensitivity as has been found in CVM studies. However, the risk scenario in that study can be considered less familiar to respondents than traffic risk, and hence this study will contribute to the literature by examining how robust their findings are.

Methodology

This study uses a Swedish web panel to examine the research questions describe above. The survey is divided into a DCE and a CV sample, and within each sample different treatments are run. Respondents are randomly allocated to the different samples and treatments. In total the survey consist of 14 treatments (8 DCE and 6 CV) with a total of 7500 respondents. The sampling, data collection, and survey design follow state-of-the-art, including pre-testing.

The mode-specific WTP for safety is elicited in a DCE with one fatality, one injury, one cost attribute, and one attributes being the modes studied (road, railway, and air-traffic). We argue that including mode as an attribute is superior to between sample analysis since respondents are faced with choice situations where they need to decide to what extend the transport mode is important for their choices. However, to examine what effect the mode attribute has on the marginal WTP, i.e. VSL and/or VSI, for road risk separate samples are run excluding the mode attribute. To further examine the robustness of the findings the DCE is replicated with a CVM scenario, where the later involves choice situations only considering one risk attribute and the cost attribute.

The between sample scope analysis is conducted by replicating one of the DCE subsamples and one of the CVM subsamples, but dividing the changes in the risk attributes by 2. Hence, expect for the size of the risk attributes everything else is identical.

Results

Preliminary results suggest that respondents in the DCE sample react as expected to the choice situations, i.e. they prefer policies that reduce the risk of death or serious injury more, and that cost less. Moreover, respondents are also more likely to choose a policy for road safety compared with a rail or air-traffic safety policy. This results in measures of the marginal WTP to reduce the fatality risk, i.e. the VSL, that are higher for reducing road fatality risk, compared with railway and air-traffic risks. We find only a small effect from including the mode attribute on the estimates of the VSL for road. Moreover, our analysis also suggests that the VSL and the corresponding VSI are higher when estimated using the CVM compared with the DCE approach.

Regarding between sample scope sensitivity. As indicated above, we find scope sensitivity when analyzing data separately for the different subsamples, i.e. respondents are more likely to choose options where the risk reduction is larger (ceteris paribus). However, we do not find any scope sensitivity between samples. That is, when the risk reductions are halved, the VSL and VSI are doubled, which suggest that WTP is not sensitive to the size of the risk reduction. We find that these results are robust for the DCE and CVM, and is troubling for the SP approach since it suggest that VSL and VSI can be influences by the analysist when deciding on what risk reduction levels he/she chooses to use.

We would like to stress that the results are preliminary at this stage, but seems robust.

Reference

Andersson, H., Hole, A.R., and Svensson, M., 2016, “Valuation of small and multiple health risks: A critical analysis of SP data applied to food and water safety”, Journal of Environmental Economics and Management, 75, 41-53.

ABSTRACT. Welfare measurement in a discrete choice framework has a long record of theoretical investigations and applications and is a key ingredient in cost-benefit analysis. Since McFadden (1999), the compensating variation has been used to evaluate welfare changes. The Compensating Variation (CV) corresponds to the income variation an agent would need ex-post, after changes in price and/or quality attributes of the alternatives, in order to reach the initial or ex-ante utility level.

The literature has focused its attention on the expected CV, which has been widely used in theoretical and applied work in transportation and industrial organization. In the case of linear-in-income logit, the expected CV equals the monetized difference of logsums. This canonical formula plays a key role in the appraisal of transportation projects and policies (see, e.g., De Jong et al., 2007), but is unable to characterize the individuals in terms of winners and losers from policies (see Delle Site, de Palma and Kilani, 2021).

For a given policy, some individuals may stick to their choice before a price or a quality attribute changes, while others may alter their choices after the change. In the spirit of discrete choice models, compensations are individual specific and described by the modeler as a random variable. In our approach the agents keep their observable idiosyncratic terms, before and after the changes, and these factors are part of the agent’s welfare.

The distribution of the CV for additive random utility models has been derived by de Palma and Kilani (2011). Based on the distribution of the compensating variation, the analysis of inequality as a consequence of welfare change across the population, measure by the Gini coefficient and the Lorenz curve is tractable, in principle. This is the scope of this paper, although it requires to solve some technical issues in the definition of the inequality measurement.

In the literature on public economics and on inequality, methodologies for income inequality analysis have been extensively investigated. Different inequality measures have been proposed. The Lorenz curve relates the cumulative proportion of individuals to the cumulative proportion of income when agents are reordered by ascending order of their income. The Gini coefficient is the ratio of the area between the Lorenz curve and the unit slope line through the origin (the area of concentration) to the area between the unit slope line through the origin and the axes. The unit slope line through the origin represents complete equality (all have the same income).

The use of the CV per se for the analysis of welfare change is, however, problematic, since welfare change can be either positive or negative. By contrast, in their original definition, the Gini coefficient and the Lorenz function refer to positive income. The few contributions that have relaxed the positive income assumption have been restricted to distributions with some negative income values but with positive expectation.

The welfare impact of policies is measured in this paper according to the Hicksian equivalent income, i.e. the income to be provided to the agent's in the (ex-ante) state without the change to bring her to the level of utility in the (ex-post) state with the change. We derive new theoretical and exact formulas to characterize the distribution of the equivalent income. We then derive two theorems related to the Lorenz function and the Gini index for any discrete choice model. For the linear-in-income logit, it turns out that both the Lorenz function and the Gini index have a closed form. For the disaggregated version of the CES model, we also derive practical inequality measures.

Our theory is applied to the data on automobile characteristics and demand at the household level in the United States (based on Berry et. al.,1995). The data contains mainly aggregate demand (number of models of cars purchased) and characteristics of 997 distinct models observed over a 20 years period, with a total of 2217 observations. We are evaluating numerically how the computation time of the Lorenz function and the Gini index changes when exact formulas are used instead of complete simulation techniques.

Finally, we discuss the Gini index and the Pigou-Dalton principle for non-infinitesimal transfers, which however preserve the wealth ranking.

References BERRY, Steven, LEVINSOHN, James, et PAKES, Ariel. Automobile prices in market equilibrium. Econometrica: Journal of the Econometric Society, 1995, 841-890.

DELLE SITE, Paolo, DE PALMA, André. and KILANI Kilani (2021) Consumers’ welfare and compensating variation: survey and mode choice application, THEMA Working Paper n°2021-11, CY Cergy Paris Université, France.

DE PALMA, André et KILANI, Karim. Transition choice probabilities and welfare analysis in additive random utility models. Economic Theory, 2011, vol. 46(3) 427-454.

DE JONG, Gerard, DALY, Andrew, PIETERS, Marits, et al. The logsum as an evaluation measure: Review of the literature and new results. Transportation Research Part A: Policy and Practice, 2007, vol. 41(9), 874-889.

MCFADDEN, Daniel. Willingness-to-pay in random utility models. Trade, Theory, and Econometrics: Essays in Honour of John S. Chipman, 1999, 15, 253-274.

ABSTRACT. Choice models are extensively used to analyse consumer behaviour in a variety of application contexts. Key objectives of such analyses include i) obtaining a better understanding of the drivers of choices, ii) forecasting demand responses to changing market conditions and iii) welfare analysis of changing market conditions. The software packages available allow choice modellers to specify very flexible ‘utility’ functions. Most choice modellers are, however, unaware of the intricate relationship between the modelled ‘utility’ functions and the ability to conduct meaningful welfare analysis, especially when it comes to the inclusion of prices and income in the ‘utility’ functions.

The literature on welfare economics and discrete choice modellers is, however, not overly accessible to the choice modelling community. This is a direct consequence of the multi-disciplinary nature of the research conducted, but also due to the assumed knowledge and technical approach of the referred economic papers. The first aim of this paper is to provide an accessible and non-technical review of the connection between welfare economics and the discrete choice modelling literature. We start of by introducing the concept of duality, which allows for a better understanding of the concept of the indirect utility function as modelled in most Random Utility Maximisation (RUM) based discrete choice models.

Our choice models are based on the concept of revealed preferences. That is, we believe that consumers reveal their preferences by selecting the alternatives which maximise their preferences. This assumes a degree of consistency with such utility maximising choice rules at the end of the consumer. When observing choices, we assume that this aspect is satisfied. However, economic theory imposes additional restrictions on the modelled indirect utility functions to solve the integrability problem. In other words, if we wish to trace back from observed choices to preferences our modelled utility functions and corresponding demand functions need to satisfy certain properties to be consistent with economic theory.

Ibanez and Batley (2013) review these properties in the context of utility functions with additive error terms. Based on the work of Ibanez and Batley (2013), Batley and Dekker (2019) conclude that due to the discrete nature of demand costs can only be included in a linear fashion in the indirect utility function. A direct implication is that the marginal disutility of cost needs to be constant across prices and across goods. Batley and Dekker (2019) furthermore highlight the importance of the outside good on welfare effects of price changes. That is, since consumption is restricted to unity any price reductions can only result in an increase in utility when the money saved is spend on other goods (i.e. the numeraire), or when the consumer switches to a different good. The role of the numeraire is often ignored, but since it covers residual budget it implicitly connects the cost and income variable. Batley and Dekker (2019) therefore conclude that income and prices can only be included in the utility function in a linear fashion using λ(M-p), where M-p represent residual budget and λ the marginal utility of income. Notably, this is not consistent with many empirically estimated choice models.

A first contribution of this paper is that we show that the constant marginal utility on income assumption advocated by Ibanez and Batley (2013) and Batley and Dekker (2019) can be relaxed. The reason is that the referred papers have unintendedly assumed that the utility function captures the preferences of the representative consumer, which eliminates all income effects by default. The logic of the numeraire to accommodate welfare effects still prevails and accordingly a functional form based on residual budget h(M-p)needs to be maintained. However, the functional form can be non-linear and in many cases would be assumed to be twice differentiable with ∂h(M-p)/∂M=-∂h(M-p)/∂p>0 and (∂^2 h(M-p))/(∂M^2 )<0, such that additional income increases utility, but at a decreasing rate. Moreover, the functional form of h(M-p) needs to be constant across different alternatives since it captures the utility of the same numeraire good but consumed at different levels due to cost differences across goods. Separate interactions of income or costs with other variables are therefore still not allowed inside the indirect utility function. This theoretical insight allows to connect papers in the literature advocating constant marginal utility with papers studying non-linear income effects (e.g. Herriges and Kling, 1999; Dagsvik and Karlstrom 2005). Once a non-constant marginal utility of income is identified, welfare analysis less straightforward. This paper will set out strategies as to how conduct welfare analysis in these circumstances.

The above analysis applies to choice models with additive error terms. We extend our analysis to the context of multiplicative error terms (e.g Fosgerau and Bierlaire, 2008). Following the same logic of applying the conditions associated with the integrability problem we show that multiplicative nature of the model puts additional constraints on the shape of the utility function for this to be consistent with economic theory. Particularly, it assumes very strong interaction effects between the chosen good and the numeraire.

In the empirical part of this paper we will examine the impact of such restrictions on the emerging welfare measures when alternative functional forms are adopted (not) satisfying the conditions set out by economic theory. A likely candidate database is the 2014/15 UK Value of Time survey, where in addition to multiplicative errors cost and income elasticities were included.

ABSTRACT. Stated preference (SP) survey can be useful as a policy tool as it helps in capturing preferences towards new services that are yet to be introduced. However, it often requires the development of a carefully designed questionnaire and sampling procedure, affecting the reliability and validity of the responses. In order to improve the reliability and validity of SP data, several methods have been proposed to utilize revealed preference (RP) data. In particular, with the development of smart phones and the Internet, it is becoming easier to use RP information in the design of SP surveys. A context-aware SP survey is one of such methods, where users are asked to answer SP questions given a particular RP context they actually faced, such as time constraints. In such a survey, we can ask SP questions without discarding the influence of contextual factors. On the other hand, respondents may find it difficult to precisely recall the RP context with the increase in the timing gap between RP behavior and SP response (response lag) and it may not be able to obtain SP responses that properly take into account the RP context. In order to mitigate these problems, we have developed a survey tool that can reflect the RP context and allow respondents answer SP questions in real time. In the real-time context-aware SP survey, subjects can answer SP questions in real time based on the RP information, which makes it easier for them to recall the RP context and is expected to provide reliable SP answers. However, to the best of our knowledge, there are no existing examples of real-time context-aware SP surveys in the transportation field, except for the one conducted by the authors, and the effectiveness of real-time response have not been confirmed. So, this study formulates the following three hypotheses to confirm the effectiveness of real-time response: 1) the greater the response lag, the greater the systematic bias in the choice result, 2) the greater the response lag, the more difficult it becomes for the user to reconstruct the RP context in answering SP questions, i.e., the response lag acts as a moderator variable that reduces the impacts of the RP attributes on the choice result, and 3) the greater the response lag, the greater the influence of the SP attributes on the choice result, i.e., the response lag acts as a moderator variable that increases the impacts of the SP attributes on the choice result. To confirm these hypotheses, we conduct an empirical analysis using data obtained from a real-time context-aware SP survey conducted in Kumamoto and Hiroshima metropolitan areas of Japan in January and February 2020. The subjects of this survey were 150 drivers selected from each city who regularly pass through or visit the city center by car. In the SP survey, we presented to the users a hypothetical situation in which they would have to pay a certain amount of money (congestion charge) to enter and move around in the city center. First, the trip attributes (purpose, mode of transportation and etc.) were obtained from a smart phone-based RP survey. Next, immediately after the completion of the trip, we calculated the attributes of the SP survey interacted with the RP information and presented the SP alternatives relating to behavioral changes (choice set: 1. pay the charge and continue the present trips under road pricing scheme; 2. cancel the trip; 3. change the time of day; 4. change the destination; 5. change the travel mode; 6. change the route) to the respondents based on the attributes. We analyzed the effects of response lag on the model estimation results by using this survey data. For the model estimation, we use mixed logit model considering unobserved inter-individual heterogeneity. This is because it is possible that multiple responses from the same individual are highly correlated in this survey. For the systematic utility function, we consider two terms: 1) quantify the effect of response delay on choice result by directly adding the response lag as an explanatory variable in order to confirm systematic bias (hypothesis 1), and 2) quantify the effect of the response lag on the effects RP and SP attributes on the choice results by introducing scale parameters for SP and RP attributes, respectively, as functions of the response lag (hypotheses 2 and 3). In addition, in order to visualize the effect of the response lag on the contribution of SP attributes, RP attributes, individual attributes, and the error term towards the total variance of utility differences, we analyzed the variance decomposition. Empirical results show that 1) the greater the response lag, the more likely people are to choose the alternatives that does not change their behavior under road pricing scheme, that is, the greater the systematic bias in the choice result. 2) the greater the response lag, the more it tends to reduce the impacts of RP attributes on the choice result. 3) the greater the response lag, the more it tends to increase the impacts of SP attributes on the choice result. Based on the above trends, it can be argued that the larger the response lag in context-aware SP surveys, larger biases may occur in the results. The results of this analysis will be useful for the construction of a future travel diary survey application. In that survey application, incentives can be provided to reduce response delays when answering SP surveys, which is expected to reduce biases associated with SP surveys.

ABSTRACT. Performing arts, which encompass all creative activity performed in front of an audience, are a source of use as well as passive-use value, including existence and bequest values, among others. To assess the effectiveness of cultural policies aimed at provision of performing arts and to determine their social-welfare maximizing provision, an estimate of the value that performing arts bring to society is necessary. This can be achieved with various economic valuation approaches. Several of them, such as stated and revealed preference methods, rely on conducting primary studies for data collection, which are usually costly and time-consuming. Another approach—benefit transfer—may overcome these challenges and is sometimes the only feasible means to estimate the value when budget or time constraints exist. Its main advantage comes from using existing value estimates from primary studies conducted at one or more (study) sites to calculate values at other (policy) sites. While benefit transfer has been successfully employed in many areas, including health, recreation, and environmental economics, its application in cultural economics has been scarce. Here, we aim to deliver the first benefit transfer for performing arts and the first one, within cultural economics, that is based on estimates from a discrete choice experiment. Moreover, we examine the benefit transfer accuracy when conducted for experienced cultural goods, such as performing art pieces, by investigating the transfer’s validity and reliability.

Our thorough literature review suggests that out of identified five benefit transfer studies for cultural goods, none has considered the value of performing arts (e.g., Tuan et al. 2009). Furthermore, all these studies rely on primary values elicited with the contingent valuation method. Instead, our empirical data comes from a carefully-designed discrete choice experiment survey concerning preferences towards extending the offer of public theatres in Polish provinces within four types of performances: entertainment, drama, children’s and experimental. The data was collected online between May and December in 2018 on a representative sample of 2863 adult residents of Poland.

By examining validity and reliability of benefit transfer for cultural goods, we undertake the unexplored question of whether this valuation technique can be successfully applied to cultural goods, such as performing art pieces. We attempt to identify factors possibly enhancing the transfer’s accuracy for performing arts, by considering several benefit transfer approaches (i.e., function transfer, non-adjusted and income-adjusted unit-value transfer) and various similarity characteristics, among others.

For examining validity and reliability of benefit transfer, we consider six provinces in Poland. We investigate all possible pairs and directions for transfers between these provinces, meaning that each province in our analysis serves as a study site and a policy site. Based on random parameter logit models, we obtain willingness-to-pay values, which are next applied to the inter-provincial benefit transfers. By having the actual value estimate for a given province (based on the primary data) and the transferred value estimate (derived from the value estimate in another province), we are able to assess the transfers’ validity and reliability. We measure validity through a number of statistically indistinguishable willingness-to-pay estimates across sites—this indicates the extent to which the transferred estimates are not biased. Reliability is assessed based on the percentage transfer errors (TEs) and demonstrates the variance.

We find that the number of valid transfers, yielding statistically indistinguishable willingness-to-pay values, appears to be higher for the unit-value transfer (both income-adjusted and non-adjusted) than for the benefit function transfer. For the unit-value transfer, TEs range from low (<10%) to high (>1000%), with a mean of 180%. Instead, the function transfer approach leads to increased TEs and their variability for children’s and experimental performances—for these performances, TEs range from very low (<1%) to very high (>2000%), with a mean of 310%. The analysis of median TEs suggests high transfers’ reliability in four provinces and low in two of them, regardless of whether the province is an assumed policy or study site.

These findings suggest that the choice of the benefit transfer technique may have significant implications for the transfer’s validity and reliability. In our case, socio-economic adjustments do not necessarily improve the accuracy. Although our function transfer controls for factors commonly acknowledged as important predictors of demand for cultural goods (i.e., geographical proximity to cultural goods and higher education), their use reduces the transfer accuracy in our case. Furthermore, our results indicate that the accuracy of benefit transfer in performing arts may significantly hinge on the performance type, with the more known types, involving entertainment and drama performances, being related to higher validity and reliability. Moreover, the results signal that the accuracy may be improved for sites where status-quo conditions (exposure to the theater offer) have been stable over the years.

Overall, we observe that the employed measures of the benefit transfer accuracy, in particular, the TEs, do not diverge substantially from their counterparts reported in other areas, where the benefit transfer method is commonly applied, such as environmental economics (e.g., Rosenberger 2015). Our results suggest that using estimates from discrete choice experiment studies may offer a promising approach for benefit transfer applications in culture. Given the complex nature of cultural goods, extrapolating values of particular attributes rather than a single undecomposed value may reduce a risk of benefit transfer providing irrelevant estimates for the policy site. Altogether, we believe that this analysis can help guide future use of the benefit transfer method when conducted for experienced cultural goods, such as performing art pieces.

Rosenberger, R.S. 2015. “Benefit Transfer Validity and Reliability.” In: R.J. Johnston, J. Rolfe, R.S. Rosenberger, R. Brouwer (eds.), Benefit Transfer of Environmental and Resource Values. Dordrecht: Springer Netherlands. Tuan, T.H., U. Seenprachawong, S. Navrud. 2009. “Comparing Cultural Heritage Values in South East Asia – Possibilities and Difficulties in Cross-Country Transfers of Economic Values.” Journal of Cultural Heritage 10(1):9-21.

ABSTRACT. Hypothetical bias (HB) affects estimated preferences and WTP obtained from discrete choice experiments (DCE). Numerous ex-ante and ex-post mitigation techniques have been proposed to reduce HB (see Penn and Hu 2018 for a review) and the use of the indirect questions (IQ) is an alternative elicitation format to direct questions (DQ). Asking respondents to predict others’ preferences instead of expressing their own should overcome individuals’ incentives to choose alternatives disregarding the cost attribute, thus producing WTP that should be more aligned with the “true” ones.

Implemented in DCE in the late Nineties in the so-called inferred valuation (IV) approach (Lusk and Norwood (2009a, 2009b), IQ have proved to lead to lower WTP estimates than those obtained with DQ for both public and private goods (see e.g., Yadav, van Rensburg and Kelley 2013; Olynk, Tonsor and Wolf 2010; Klaiman, Ortega and Garnache 2016; Menapace and Raffaelli 2020). Menapace and Raffaelli (2020) found that WTP obtained with IV are statistically equivalent to the WTP obtained from revealed preference data. However, WTP distributions are centered to the left for most of the attributes under investigation, suggesting that IV might slightly underestimate WTP obtained from revealed preference data. The possible causes of this underestimation have not been adequately investigated in the DCE literature.

Another aspect that has not been adequately investigated is the potential order effect when respondents are asked to answer both IQ and DQ. The literature on risk preference indicates the presence of an order effect when comparing risk attitudes elicited about self and predicted about others (Hsee and Weber 1997; Faro and Rottenstreich 2006; Li et al. 2017). In the DCE studies employing both DQ and IQ instead, the order of presentation has received scant attention. First evidence of a potential debiasing effect on WTP of IQ asked before DQ is reported in Raffaelli et al. (2021) but samples of respondents were small (92 and 90 tourists, respectively).

The current study aims at contributing to the literature by testing this debasing effect on a bigger sample of 370 respondents interviewed face-to face. The empirical study concerns the management of sheep grazing in a high natural value farming area in Italy (Regional Park of Porto Conte, Sardinia). The attributes are the landscape value of sheep grazing, the effects of sheep management on biodiversity and a cultural-heritage attribute (experiencing life and traditions of shepherds), a private attribute. Choice cards include three unlabeled alternatives where the third alternative depicts the no-action scenario.

All the respondents were presented with six choice cards in which they had to state their own preferences by answering DQ and the same six cards in which they had to predict how other tourists would choose among the three alternatives (IQ). The order of presentation of DQ and IQ was randomized. In order A respondents had to state their personal preference first (DQ) and then predict others’ choices (IQ). In order B respondents had to predict others’ behavior (IQ) before stating their personal preferences by answering DQ.

We estimated four models: i) DQ-order A, ii) DQ-order B, iii) IQ-order A, and iv) IQ-order B by using different modelling approaches (MNL, RPL, preference-space and WTP-space). Mean WTP and the corresponding 95 percent confidence intervals were calculate based on the Krinsky and Robb (1986) parametric bootstrapping and differences in WTP were tested by applying the complete combinatorial approach suggested by Poe et al. (2005).

Consistent with findings from previous studies using IQ, the WTP obtained from IQ-order B (the classical IV) are much lower than those obtained from DQ-order A and the ratio between WTP varies across attributes ranging from 3.03 for the cultural experience (the private attribute) to 5.8 for the landscape (one of the public attributes). This result reveals that respondents predict that other tourists are willing to pay less for all the attributes than themselves, but we cannot ascertain whether this prediction corresponds to the respondents “true” WTP or underestimates it.

More interesting is verifying whether the order of presentation has an effect on WTP estimates. According to the p-values, the order of presentation does not affect answers to IQ and, consequently, WTP estimates obtained from IQ. We can therefore infer respondents asked to predict others’ choices do not change their predictions if they had first to think about their own preferences.

A clear order effect emerges on WTP estimates from DQ. If we compare the WTP obtained from DQ-order A and DQ-order B, p-values suggest the two estimates being statistically different for the public attributes (landscape and biodiversity) and the no-action alternative but not for the private attribute (cultural experience). The ratio between DQ-order A and DQ-order B varies across attribute ranging from 1.45 for the cultural experience to 2.35 for the landscape.

These results suggest that thinking at others’ preference and going through a “prediction exercise” leads respondents to make more “conservative” personal choices that lead to lower WTPs. IQ asked before DQ seem to be able to reduce HB affecting WTP for public attributes, but we cannot ascertain whether they eliminate it. The same debiasing pattern emerges by analyzing choices made by female and men or by graduates or not graduates separately. This seems to reveal the existence of a common cognitive pattern among respondents.

In conclusion, the ordering in terms of magnitude is the following: WTP from IQ-order b (the classical IV) < WTP from DQ-order B < WTP from DQ-order A (the classical DQ). We expect the “true” WTP to fall in the range between WTP from IQ-order B and WTP from DQ-order B.

These findings have practical implications for researchers and practitioners suggesting a wide array of benefits by employing IQ in a within-subjects design. IQ asked first (the classical IV approach) can provide a sort of lower bound WTP estimate, while DQ asked after IQ can provide an attribute-specific calibration ratio alternative to using a mean calibration factor emerging from meta-analysis studies. Further research is needed to unravel the causes of this effect of predictions on own choices.

ABSTRACT. Background

The “Reduction and Innovation Strategy” of the German Federal Ministry for Food and Agriculture (BMEL 2020) aims at reducing the sugar content of food products. In particular, breakfast cereals marketed to children are perceived as having a high content in sugar. As a consequence, the strategy sets the goal of reducing the sugar content in breakfast cereals for children by at least 20% until 2025. Sugar taxes are increasingly regarded as an effective strategy to reduce the intake of sugar and to alleviate associated public health problems such as obesity and diabetes. So far, empirical evidence on the impact of sugar taxes on consumer choices and substitution behavior based on actual market data is still scarce. Studies typically focus on soft drinks and do not account for heterogeneous tastes (see Dubious et al. 2020 for an exception). This study has two major objectives: First, preference towards breakfast cereals marketed to children are studied taking into account unobserved and observed factors. Second, we evaluated heterogeneous responses to different sugar tax regimes and subsequently derived sugar consumption. We thereby also analyze substitution patterns between products when taxes are introduced.

Data

Our empirical analysis is based on household scanner data from the Gesellschaft für Konsumforschung (GfK) in Germany from 2016 and 2017. The data contain quantity and expenditure information of individual purchases at the barcode level, product and store characteristics as well as annually collected socio-demographic characteristics and attitudes of panel households. A common challenge to estimating discrete choice models based on household scanner data is that only information on the product purchased is available (brand, price, other attributes), but not which other products were on shelf and which attributes and prices they had. We regard as the relevant choice set as all children cereal products on shelf in the specific store of an observed purchase (Bonnet and Simioni 2001). As we have a large range of different retailers in the data, we restrict the sample to the two largest supermarket chains and the major hypermarket chain. For each retailer, we assumed those products as part of the choice set that were consistently available with notable sales. Likewise, we also included only products with more than 20 purchase observations over the two-year period. We supplement the data by manually collected information on nutrition values of single products and impute prices based on price strategies regarding discounts, retail chain, geographic area, and week as observed from other households’ purchases. Our final sample includes 6,114 individual choice sets, 13 brands, 60 products and 1,408 households.

Modeling approach

To capture unobserved as well as observed preference heterogeneity, we estimate a series of latent class models. In each class choices are explained by the sugar content, the price and the size of the product. In addition, we include the following characteristics: • 4 private and 9 national brands • 12 variables which capture the style and flavor of the product (e.g. frosted flakes, honey pops, …) • 3 loyalty indicators conveying information on whether the same brand, size, and style / flavor was chosen in the previous choice Class membership is explained by socio-demographic characteristics such as information about children in the household, the households’ income, the place of residence (i.e. the size of the village / city) the household lives in. Model estimates will be used to simulate effects of sugar tax regimes (e.g. excise vs. value-added tax) which will be reflected in price changes of each product.

Results

Preliminary results from a model with three latent classes reveal a large degree of preference heterogeneity. Based on the socio-demographic characteristics implemented in the membership function, the following classes can be distinguished: • Class 1: Households with no children under 15 years are less likely to be found in this class. • Class 2: Households with a high income are less likely to fall in this class. • Class 3: Baseline Mean class allocation probabilities are calculated to be 0.30 (class 1), 0.29 (class 2) and 0.41 (class 3). Taking p =0.05 as a threshold for hypothesis testing we find the following effects for our key variables sugar, price and size: Class 1 is characterized by a clear preference towards cereals with a higher sugar content and lower prices. The package size does not impact utility. Similar to class 1, the probability of choosing a product increases with the sugar content in class 2. Individuals in this class, however, are more price sensitive. The package size has a negative influence on choices. In class 3, the null hypothesis of no influence of the sugar content on product choices cannot be rejected. Price sensitivity is similar to class 2 and smaller packages sizes are clearly preferred. Using these estimates, we well - in the next step – predict class-specific and overall sugar consumption. Then, we will simulate scenarios which will be based on different hypothetical sugar tax regimes. Each time heterogeneous responses to these taxes will be studied, class-specific as well as overall sugar consumption will be predicted, and the effectiveness for these sugar tax regimes will be assessed. Results will be used to provide policy recommendations to effectively meet the reduction goals.

References

Bonnet, C.; Simioni, M. (2001): Assessing consumer response to Protected Designation of Origin labelling: a mixed multinomial logit approach. European Review of Agricultural Economics 28 (4). Bundesministerium für Ernährung und Landwirtschaft (BMEL) (2020): Weniger ist mehr. Zucker, Fette und Salz reduzieren. Hg. v. Bundesministerium für Ernährung und Landwirtschaft (BMEL), Referat 213. BMEL. Bonn. Dubois, Pierre; Griffith, Rachel; O’Connell, Martin (2020): How Well Targeted Are Soda Taxes? In: American Economic Review 110 (11).

ABSTRACT. 1. Introduction

The outbreak of COVID-19 in the world has caused a significant change in people’s mobility patterns as a result of people’s fear of the virus’s consequences, government measures and changes in transport provision (Abdullah et al., 2020; Bin et al., 2021). Although recent literature suggests that the demand for all modes has been altered, the evidence shows that public transport (PT) systems have been one of the most negatively impacted (Vickerman, 2021; Przybylowski et al., 2021; Wielechowski et al., 2020).

Many studies to date have investigated the impact of COVID-19 on travel behaviour. But most of them have been conducted using online surveys, either cross-sectional (Bucsky, 2020) or considering a limited number of waves (Beck and Hensher, 2021). Such online surveys typically have small sample sizes and do not capture the day-to-day variability of people’s mobility over long periods. Passive data (e.g. smart cards, GPS traces, mobile phone records etc.), which have digital mobility footprints of a large number of people over time, have shown a huge potential to evaluate people’s mobility adaptation and analyse the modification in travel habits throughout the outbreak. Specifically, smart card data have been successfully used in many transport planning studies (see Zannat and Choudhury (2019) for a detailed review) and have demonstrated their ability to capture the heterogeneity in public transport passengers behaviour. For instance, He et al. (2018) and El Mahrsi et al. (2017) used anonymous smart card data to classify PT users depending on their mobility profiles. Thus, we postulate that in response to COVID-19 pandemic and associated restrictions, groups of passengers have exhibited homogeneous strategies in terms of making adjustments in their travel behaviour. Modelling the strategies that PT passengers have adopted during the pandemic can help reveal how heterogeneous urban mobility was during this period. In particular, some passengers have consistently travelled less frequently while some may have reverted to pre-pandemic frequencies after the end of the first lockdown; others may have retained the pre-pandemic trip frequencies but adjusted their departure times, etc.

This prompts this research where we aim to identify and characterise discrete mobility profiles of public transport passengers considering the changes in their travel behaviour during key periods of the pandemic and model how they vary depending on the land-use and demographic attributes (e.g. income, employment rate) of their home location. We expand the findings of previous works related to the impact of COVID-19 on travel behaviour by a) using individual-level revealed passive data of smart cards during critical periods of the pandemic, b) revealing hidden mobility profiles of PT passengers during the pandemic, c) considering a broad set of spatial and temporal mobility indicators simultaneously instead of carrying out a one-dimensional analysis, and d) associating explanatory variables to those mobility profiles to infer insights about planning in the post-pandemic era.

2. Data

Smart card data on an individual card level is available for this study, as well as aggregated sociodemographic data from Santiago de Chile. Santiago’s public transport system called Transantiago involves a complex system that integrates urban buses, the underground (the Metro) and an inter-urban rail. The system serves a population of 7 million inhabitants, reaching 3.7 million journeys on average in a workday in 2019.

The data available allows us to track the individual responses of PT passengers during the key periods of the pandemic in Chile: pre-pandemic (first two weeks of March 2020), first measures (last two weeks of March 2020), first lockdown (June and July 2020), after first lockdown (November 2020) and second lockdown (April 2021). In total, more than 145 million trips are available for analysis.

3. Method

A two-step machine learning approach is proposed to achieve the goal of this study. Firstly, a Self-Organized Map (SOM) algorithm is applied to identify discrete mobility profiles. Self-Organizing Maps are an unsupervised data reduction algorithm widely used in cluster analysis. The technique is based on an artificial neural network that allows the reduction of any multi-dimensional data set to a two-dimensional output space, also called a map. Mobility indicators such as number of trips, mode choice, departure time, travel distance, waiting time and number of transfers are used to identify mobility profiles of PT passengers in the key periods of the pandemic in Chile. Secondly, Gradient Boosting Machines (GBM) are applied to model the membership of each passenger to the previously defined clusters which represent a particular mobility profile. GBMs are a supervised machine learning and a tree-based technique that builds an ensemble of successive trees, learning and improving on the previous. A set of explanatory variables such as aggregated socioeconomic characteristics, travel history and type card will be used to explain class membership.

4. Results

It is expected that after applying SOM joint with hierarchical cluster analysis, a proper number of classes is obtained. Each mobility profile will be characterised, emphasising their main differences in terms of their change in travel behaviour during the pandemic. Using GBM will be possible to predict the membership of a passenger to a particular mobility profile, associating explanatory variables. Additionally, GBMs allow the evaluation of the relative importance of each independent variable. Socioeconomic characteristics of the cardholder’s home area such as the share of people above a certain age, gender, density, immigrant population, house type and average income will be tested. Pre-COVID cards activity will also be examined to identify how past travel behaviour is associated with each mobility profile. Variables such as the number of journeys, regularity, frequent departure time, usual travel distance, travel time and modal share will be analysed.

5. Policy implications

The identification and characterisation of mobility profiles of PT passengers will provide better understanding of the changes in travel behaviour during the pandemic. These results will give interesting insights that may aid policy-making processes, which is particularly relevant in countries in the Global South, such as Chile, where more efforts in improving public transport quality and ensuring equity are needed.

ABSTRACT. Motivation and objectives

The COVID-19 pandemic made its way into our lives almost two years ago, and life is yet to return to pre-pandemic conditions. Whether it will at all, and what kind of changes brought on by the pandemic will be permanent is also uncertain. Naturally, how individuals have used their time has continuously shifted, especially during government mandated lockdowns. This is reflected in the daily choices made regarding what kind of activities to engage in (and for how long). An analysis of these changing behaviors is important for policy makers as they make decisions about how best to deal with the spread of the virus while not encroaching too strongly on residents’ autonomy.

Data collected for the MOBIS and later MOBIS:COVID-19 studies (Molloy et al., 2021b) allow for exactly such analyses. Our work until now included a first investigation into time use behaviors from September 2019 to October 2020; that is, from before the pandemic through the first wave and lockdown to post-lockdown times and what ended up being the beginning of the second wave of the pandemic (Winkler et al., 2021; Mesaric et al., 2021). The data are made up of GPS tracks that were passively collected from participants as they went about their everyday routines as part of a mobility pricing study. We aggregated these at individuals’ day-level to understand how everyone spent their 24 hours of a day at different points during the pandemic. Key results from the panel mixed multiple discrete-continuous extreme value (MMDCEV) model applied are discussed below.

The MOBIS:COVID-19 study continues to present day (Molloy et al., 2021a). Since November 2020, a second set of respondents were invited to participate via a market research firm. This dataset differs from the former in that it is more representative of the Swiss population (MOBIS almost exclusively includes car users due to initial study requirements). Effectively, we therefore have over four thousand individuals who have tracked on and off for the past two years whose behaviors can be compared across different periods in the pandemic.

This paper builds upon our first in several important regards. Firstly, we will re-estimate the MMDCEV model applied using the extended dataset that now extends to September 2021 and includes two socioeconomically diverse groups in Switzerland.

Additionally, whether participants worked from home (WFH) and when will be imputed. Time WFH is exactly one of the changes that will likely persist long-term, which affects individuals and their household dynamics, but also how companies will decide whether and where to hold physical office buildings. Survey data and GPS traces will be used in tandem to determine what portions of time spent at home were spent WFH, later an additional alternative to be estimated in the MMDCEV.

We will further extend the models to include unemployed and retired individuals, weekend days, and include a more rigorous measure of risk of infection to take full advantage of the data available and make the results representative for the country.

Data

Our first paper included 23,130 person-days from 867 participants over the course of 13 months. In total, 4,542 MOBIS:COVID-19 participants have generated data on over 600,000 days through September 2021 that will be considered for our next models.

The number of time segments imposed for comparison will also double from those in the first model of pre-pandemic, lockdown, post-lockdown, and new normal, to also include the second wave, the second (partial) lockdown, and relaxed restrictions since April 2021 (similar to the new normal time period).

Model formulation

The main model implemented in the initial paper was a panel MMDCEV model as developed by Bhat (2005, 2008) that included activity purposes home, work, leisure, shopping, travel, and other, with time spent at home as an outside good. The utility function therefore reflects the utility an individual accrues spending time on those activities on a given day.

The model includes translation parameters to allow for corner solutions (i.e., zero consumption) for all activities except for time spent at home that was specified as always occurring. The panel structure of the GPS data is accounted for in the baseline parameters that include a vector that generates correlations for each individual’s choices. Time segments imposed exogenously allowed us to examine the differential impact of the variables in each segment as independent variables in the model.

Preliminary results

Initial model results that focused on the shift in time use during the first wave of the COVID-19 pandemic highlight policy-relevant components that are important to consider when attempting to manage a worldwide pandemic. Switzerland consistently rates highly on metrics like quality of life, and recent ratings say lockdown periods here were among the most bearable (U.S. News & World Report L.P., 2021; Bloomberg L.P., 2021). To those looking at the country from the outside, it seems, as usual, as if everyone were doing just fine.

This was not the case in the first eight months of the pandemic in Switzerland analyzed in our first work. Typical socioeconomic indicators of age, gender, education and income levels all significantly affected propensity to participate in the five out of home activities examined in the first paper. For instance, individuals without a university education were those found to have the highest propensity of engaging in out-of-home work activities during all three pandemic time segments included, as were women in the post-lockdown and new normal periods (all compared to the 2019 baseline).

In other words, men had a lower propensity than women did to travel to a physical workplace in post-lockdown conditions (vs. 2019), which importantly is tied to having the opportunity to do so in the first place. Clearly, it’s crucial that WFH is teased out and analyzed to better understand who (under what conditions) is WFH and in the future ensure that policies are put in place to protect these same sociodemographic groups from exposure to risky situations during global emergencies.

ABSTRACT. Attribute non-attendance has received some attention in the literature investigating the premises of choice experiments. Generally, two approaches are applied: stated non-attendance and inferred non-attendance. The first approach entails direct questioning of respondents on whether they attended an attribute or not in the choice process. The second approach instead uses modelling to reveal whether attributes have been attended or not. Especially the cost attribute is of interest as it plays an essential rule for subsequently calculating marginal and non-marginal welfare measures. Here, we apply a procedure that does not address non-attendance directly. In two surveys, both concerned with different aspects of the Baltic Sea in Germany, we asked respondents for short descriptions of how they had proceeded in the sequence of choice tasks leading to to the choices they made. Answers in the form of keywords only or whole sentences were collected in an open text field. We then went through the answers provided by respondents and analysed the data by coding different types of information processing strategies. Among others, we employed text analysis tools to search for certain keywords in the text provided by respondents. For example, we looked for “cost” or for “gut feeling” (Bauchgefühl) and subsequently created indicators capturing whether respondents mentioned that keyword in their responses or not. Another indicator reflects whether respondents had considered the cost attribute while choosing. The basic assumption behind our approach is that respondents retrospectively would write down what has been activated in their memory while choosing and was driving their decisions. This way, the method is based on the think-aloud approach applied in psychology (e.g. Kuusela & Paul (2000)). Our think-aloud approach was employed to data from two different surveys that took place at different points in time. Both surveys used nationwide samples from online panels. The first survey dealt with the German part of the Baltic Sea and whether respondents (2017, n = 1275) are willing to pay to reach the good ecological status as an environmental quality target. The second survey dealt with coastal adaptation along the German North Sea and the German Baltic Sea coasts. Here, respondents (2021; n = 3400) were asked which type of adaptation given climate change they would prefer for either the North Sea or the Baltic Sea; respondents were randomly assigned to one of the treatments. Overall, we find a divers set of information processing strategies. It ranges from people indicating that they indeed applied a completely compensatory approach to people who simply against protested the whole survey as they think this is beyond their responsibility. Other strategies imply types of semi-compensatory processing by not attending a subset of attributes, by focussing solely on costs or focusing only on the non-monetary attributes. A significant part of respondents even stated that their “gut feeling” guided their choices. In this paper, we focus on cost attendance. Thus, a dummy variable reflecting cost attendance was coded “1” if a respondent had indicated in her or his statement that the cost attribute was accounted for when the choices were stated. We used this information for creating an interaction between the cost attribute and this dummy variable to estimate whether the cost sensitivity of this group of respondents is different from those who did not explicitly mention any accounting of the costs. For both surveys, we find highly significant interaction terms in conditional but also in latent class logit models, suggesting that people who mentioned cost as part of the decision process have a significantly higher cost sensitivity. Calculating subsequently marginal and non-marginal welfare measures shows that these estimates are clearly different from one another. Any policy advice such as those based on cost-benefit analysis would therefore lead to differing recommendations depending on whether all respondents or only those who accounted for cost are considered. This type of think-aloud approach provides rich data that help to understand peoples’ choices and allows to conduct sensitivity analysis. Compared to the so far employed approaches – stated non-attendance or inferred non-attendance - this approach has the advantage that non-attendance as a possible processing strategy is not “suggested” to respondents. We simply told them that we wanted to learn more about how they proceeded when choosing. On the other hand, compared to inferred non-attendance, it gives respondents a larger say and does not only rely on a certain type of model constraint. A drawback, however, is, that analysing and coding the responses requires more time than the so far applied approaches. Also, this approach might be more susceptible to subjective interpretation of responses and subsequent coding. This could be counteracted by using more than one person who is coding the responses. Moreover, strong text mining tools are available today that could be employed here as well. To summarize, we think that this think aloud approach should be applied more broadly to give researchers deeper insides into peoples’ information processing strategies. Our experience has shown that the insights are worth the larger effort.

References:

Kuusela, H., & Paul, P. (2000). A Comparison of Concurrent and Retrospective Verbal Protocol Analysis. The American Journal of Psychology, 113(3), 387–404. https://doi.org/10.2307/1423365

ABSTRACT. Respondents must have sufficient information on the good to be valued for valid stated preference estimates in environmental non-market valuation. Previous studies show that the amount and type of information can affect stated preferences and the validity of value estimates (e.g., Munro and Hanley, 2001). Besides the provision of information, it is also important to ensure that respondents process the information and recall it during preference elicitation.

A small number of studies open the black box of information processing by investigating how respondents’ engagement with the information matters for stated preferences, using visual attendance data from eye-tracking devices (e.g., Balcombe et al., 2017), time spent reading the information (e.g., Vista et al., 2009), or voluntary access to information (e.g., Tienhaara et al., 2021) as indicators for engagement. These studies find correlations between engagement and stated preferences. However, due to potential endogeneity, this correlation does not necessarily imply a causal relationship between engagement and preferences. Hence, we randomly and exogenously manipulate factors of engagement in a stated preference survey.

Drawing on stated preference literature and psychological concepts, we test whether quiz questions and self-reference questions (1) improve engagement with the provided information, (2) increase recall of the information, and (3) affect stated preferences. We use data from a discrete choice experiment survey about the creation of new urban gardens. The study was conducted in September and October 2020 with a sample of 1,686 respondents in the two cities of Berlin and Stuttgart in Germany. Respondents were randomly assigned to one of six versions of a questionnaire in a pre-registered two-by-three between-subject design (link to the pre-registration: https://aspredicted.org/blind.php?x=2aq63r). The questionnaire versions differ in the amount of information provided and whether quiz, self-reference or no questions related to the information were asked.

Our study contributes to the stated preference literature in two ways: First, the recommendation to reinforce respondent engagement in stated preference surveys with quiz questions has, to the best of our knowledge, not been tested empirically. Also, our study is the first to investigate whether self-reference questions affect engagement and value estimates in stated preferences. Second, there has been no studies so far which use a random exogenous manipulation of engagement with the information in stated preferences. Our study design allows for a more rigorous test than previous studies of whether the relation between engagement with information provided and stated preferences is causal.

Using ordinary least squares and count regression models, we investigate the effect of the manipulations on three outcomes: (1) time spent on the information page, (2) recall of the information as measured by correctly evaluated statements about the information, and (3) opt-out choices in the choice experiment. To analyze effects on preferences, we estimate a mixed logit model in willingness-to-pay space with interactions between mean willingness-to-pay and the manipulations. To test whether we can replicate the previously found correlation between engagement and stated preferences beyond our exogenous manipulations, we estimate conditional logit models allowing willingness-to-pay as well as the scale parameter to correlate with the time spent on the information page.

Our results show that respondents spend more time on the information page when confronted with quiz rather than self-reference questions. Although psychological literature found self-reference questions to be more effective for enhancing information processing, our results indicate that this does not transfer to stated preference surveys. For both question types, we do not find effects on recall or stated preferences. This suggests that quiz questions about the provided information may be less helpful for reinforcing the processing of information in stated preference surveys than generally assumed. The findings highlight the importance of testing the understanding and recall of important information in pretests instead.

In contrast to previous studies, our analysis cannot confirm that additional information affects the recall of information or stated preferences. Thus, some caution may be needed when constructing stated preference surveys that rely on information text. Questions or statements about the information similar to those that we employed after the choice task could be used in pretests to assess whether necessary information is understood and recalled. If not, alternative ways of providing important information such as videos, pictures or diagrams could be considered.

While we do not find treatment effects on stated preferences, we find significant correlations of preferences and the time spent on the information page, which is consistent with the findings of previous correlational studies. Our results suggest that the previously found correlations might reflect heterogeneity in respondents characteristics rather than a causal effect of engagement on stated preferences. This may impede enhancing information processing by enforcing engagement with the survey design. Instead, it is important to think carefully about target populations and sampling, as knowledge prior to the survey and interest in the good can be as important as detailed information provided by researchers.

References:

Balcombe, K., Fraser, I., Williams, L. and McSorley, E., 2017. Examining the relationship between visual attention and stated preferences: A discrete choice experiment using eye-tracking. Journal of Economic Behavior & Organization, 144, pp.238-257.

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

Tienhaara, A., Ahtiainen, H., Pouta, E. and Czajkowski, M., 2021. Information use and its effects on the valuation of agricultural genetic resources. Land Economics, pp.090319-0127R1.

Vista, A.B., Rosenberger, R.S. and Collins, A.R., 2009. If you provide it, will they read it? Response time effects in a choice experiment. Canadian Journal of Agricultural Economics/Revue canadienne d'agroeconomie, 57(3), pp.365-377.

ABSTRACT. Stated preference experiments, also called discrete choice experiments (DCEs) or simply choice experiments, are used in many fields, including environmental-, health-, and transport economics, to elicit people’s preferences for everything from environmental protection to preferences for health interventions, and commuting (Bekker‐Grob et al., 2012; Newbold and Johnston, 2020). In DCEs people are asked to make a choice between a finite set of distinct alternatives described by a set of attributes. Often one of these alternatives is a so-called status quo, business as usual or no choice alternative. From here and out, we will use status quo to collectively refer to this type of option. This alternative is “special” in that it conveys what would happen if we do nothing, the situation today or some default existing option (Campbell and Erdem, 2019; Johnston et al., 2017). Welfare changes are then computed as a move away from this alternative. Importantly, we assume that respondents are able and willing to make trade-offs with respect to all alternatives, including the status quo.

In empirical settings, we observe that some respondents always, never, or sometimes choose this option. If such choices are not driven by preferences, but other systematic factors, e.g., design dimensions (Oehlmann et al., 2017), time-of-day (Olsen et al., 2017), type of status quo option, or device used when answering the survey, then our interpretation of choice probabilities, behavior and welfare changes may be incorrect. Within a single study, this type of choice behavior is hard to distinguish from the choice context, e.g., survey fixed effects such as the type of good being valued. However, across multiple surveys, we can control for these context and fixed effects to get a better understanding of the systematic drivers of this type of choice behavior. While this behavior may only be relevant for SP research, given its widespread use in health preference research and environmental economics in particular, improved understanding may have wider consequences for practice in these fields; and in particular how to interpret status quo alternative specific constants and thereby welfare effects associated with moving away from the status quo. In this paper, we have collected existing stated preference studies. The studies span different fields, countries, times, and goods and services under consideration. For each study we have created various indicators aimed at capturing SQ choice behavior such as number of SQ choices, proportion of SQ choices, and whether a respondent always chose the SQ. Each indicator is then regressed on design dimensions, e.g., number of alternatives, attributes, range/dispersion of the price vector including the “jump from zero”; good or service to be valued; type of status quo option, e.g., opt out/no choice and reference alternative; position of the SQ alternative; device type used to answer, e.g., smartphone or desktop computer; time spent on the survey; and survey mode, e.g., was the data collected using mail out surveys, valuation workshops (Sandorf et al., 2016) or internet panels (Lindhjem and Navrud, 2011). In addition, we create indicators to capture the type of good being valued, e.g., consumable, non-consumable, public, private and specific field of application. With this, we aim to (partly) capture some of the motivation underlying the choice of the status quo, which may be very different when asked to choose between a toll or no toll road to work compared to biodiversity conservation efforts. We control for the utility difference of the alternatives using the predicted probabilities from a nested logit model, which recognize that the error variance of the SQ alternative may be different to that of the other alternatives; as well as socio-demographic characteristics of the respondents, and survey-, country- and time fixed effects.

A meta type study such as what is proposed here allow us to better isolate the various effects on choice behavior that goes beyond what is possible in a single study. Ultimately, understanding the extent to which these factors systematically affect SQ choices is important to improve the reliability of the estimates from stated preference surveys.

ABSTRACT. The phenomenon that some respondents to a stated preference survey accept even very high bid levels is known from contingent valuation and called the fat tail problem. A consequence of the fat tail problem is that willingness to pay (WTP) estimates are likely to be inflated. In this paper, we investigate whether the fat tail problem also exists when choice experiments (CE) are employed, and, if so, whether a combination of cheap talk (CT) and an opt-out reminder (OOR) would “nail down” those fat tails, i.e., reduce the potentially distorting influence of these high bids on WTP estimation. Fat tail problems have, to our knowledge, thus far not systematically been investigated in relation to choice experiments. The motivation for this investigation are findings from a recent CE conducted by the authors. In this study, a highest “bid” level (level of the cost attribute) of 800 Euros per year over a ten-year period was chosen to approach the choke price. However, with 19% the share of respondents choosing an alternative where the highest “bid” level was present was much larger than expected. Accordingly, the marginal WTP estimates were inflated . This is what Parson and Myers (2016) call the fat yes-tail problem in stated preference studies. They define a fat tail of a yes-response function in a contingent valuation study as a high and only slowly declining yes-response rate at high bid levels. If fat tail problems are similarly present in CE, the question arises what can be done to mitigate their impact. Fat tail problems can be conceived as a contributor of hypothetical bias. This motivates testing the effectiveness of devices employed to mitigate hypothetical bias to mitigate fat tail problems. Howard et al. (2017) compare the effects of two devices to mitigate hypothetical bias in CE, a CT script shown prior to the choice tasks and honesty priming, with a neutral control group. Howard et al. (2017) found greater sensitivity to cost among choices made immediately after facing the CT script, and that the effectiveness of the CT script decreased for choice made later in the sequence of choice tasks. To counter declining effectiveness of CT scripts, Ladenburg and Olsen (2014) developed a so-called repeated opt-out reminder (OOR) that is available to respondents in each choice set and used in combination with a CT script. The OOR points out that if respondents think that the costs presented in the hypothetical alternatives are higher than the costs they would actually be willing to pay for the respective alternative, they should choose the status quo option, i.e., the no-cost alternative offering no improvements over a defined status quo. In view of the above, we test whether a) WTP estimates are invariant across independent samples with varying ranges of the bid vector levels (i.e., whether a fat tail problem exists); and b) whether responses in treatments employing the same levels of the cost attribute do not differ independently of whether they faced a CT script in combination with an OOR or not (i.e., whether CT and OOR are effective). We use this evidence to appraise the potential of CT and OOR to reduce issues related to fat tails in CE. To test both hypotheses, we use eight independent treatment samples of respondents in a survey conducted in Germany about marine water qualities in the Baltic Sea. Respondents faced the same questionnaire except for two issues: the levels that the cost attribute takes (bid vectors) and whether a device to mitigate the hypothetical bias was shown or not (CT script and OOR). In the first four treatments, the bid vectors, each including eight potential bid levels, vary. While the first four levels are identical across treatments, the last four levels differ and increase to up to 1,800 Euros in the fourth bid vector. The remaining four treatments mirror the bid vectors of the first four treatments but respondents additionally face a combination of CT script and OOR before and while proceeding through the sequence of eight choice tasks. We find that without CT script and OOR the fat yes-tail problem exists, and marginal WTP estimates for some of our attributes are strongly inflated. The marginal WTP estimates also increase as highest cost levels in the bid vectors increase. However, when respondents faced the CT script in combination with the OOR, the WTP estimates are significantly lower and much closer to each other across treatments with varying bid vectors. The bid acceptance for the highest bid is around 20% with the use of the CT script and OOR, and the marginal WTP estimates are close to each other across samples with CT and OOR while they are higher and clearly dependent on the bid vector levels in the samples without CT and OOR. Latent class logit models on a pooled sample indicate that respondents from the treatments without and with CT and OOR have high probabilities to be allocated to different classes, and the class with high probability of no-treatment respondents yields significantly higher marginal WTP estimates. Combining results from debriefing questions and class membership show that respondents who are likely to be in this class on average perceived the cost attribute as less important. Moreover, an analysis of statements about how respondents processed their choices requested after completion of the choice tasks finds that the combined use of CT and OOR increases awareness of the cost attribute and promotes the trading-off of all attributes including cost. We argue that this type of think-aloud protocol reflects what actively guided peoples choices. Overall, the CT and OOR device has been successful in significantly reducing fat tails, even if they were not completely “nailed down”. Moreover, the analysis of respondents’ statements about how they processed their choices provide rich insights into individual decision styles and helps to understand whether such processing is in line with theoretical assumptions (full compensatory) or substantially violates those assumptions. Future work should investigate the potential of automated text analysis to categorise respondents’ choice behaviour.

ABSTRACT. Motivation:

Choice modellers are usually challenged to model more complex and more subtle forms of human decision making. One example of this is Participatory Value Evaluation (PVE). PVE is an elicitation framework based in a choice experiment, in which respondents select combinations of alternatives, subject to resource constraints. In a PVE, respondents face a number of alternatives, the attributes and costs of each alternative, and the available resources [1]. Over the last couple of years, PVE has gained increasing attention from researchers and practitioners alike. Nowadays, PVE is used to elicit citizens’ preferences for COVID-19 measures, healthcare policies, transport infrastructure projects, and climate policy.

Hitherto, data from PVE experiments has been analysed using theory-driven choice models, including MDCEV and utility-based portfolio choice models [2; 3]. These models provide interpretable behavioural information, such as preferences for attributes and marginal rates of substitution, as well as insights for policy appraisal [4]. However, a key limitation of these models is that they rely on rigorous behavioural assumptions, concerning e.g. how utility is derived (e.g. linear-additive) and how decision makers come to their choices (e.g. utility maximisation). Yet, recent exploratory research suggests that respondents participating in a PVE use a broad range of mechanisms to come to their choices [4; 5]. This could be problematic as in case the assumptions underlying current theory-driven choice models poorly describe the actual underlying data generating process, current models yield poor policy advises.

In this work, we explore the use of two well-known data-driven methods to analyse data from PVE experiments. Unlike their theory-driven counterparts, data-driven methods (a.k.a. machine learning) do not rely on rigorous behavioural assumptions. Recent studies have shown data-driven methods can complement to toolbox of choice modellers [6; 7] , and see [8] for a discussion on the topic. The first method we explore in this study is Association Rules (AR) learning; an exploratory approach that aims to identify frequent patterns of interactions between variables of a dataset. AR learning can particularly be used to identify relationships between combinations of alternatives in PVE datasets like positive and negative synergies, or nesting structures that may be overlooked by a choice model. Such insights could then, for instance, support the utility specification of theory-driven choice models. The second method is a Random Forest (RF) model; a predictive machine learning model constructed from an ensemble of decision tree models. RF models are known to be able to capture complex relationships and reach high predictive performance, while also being able to provide the analyst insights on the relative importance of explanatory variables. To demonstrate the use of these data-driven methods for analysing data from complex choice experiments, we use data from a PVE experiment conducted to infer the preferences of Dutch citizens for relaxing COVID-19 restrictions in April-May 2020 [1]. In this PVE respondents were confronted with eight different policy options (alternatives); each policy option generated an increase in the pressure to the healthcare system (resources), and health- and economic-related impacts (attributes). Respondents were asked to select their preferred policy options, without surpassing an increase of 50% of pressure in the healthcare system (constraint).

Preliminary results:

We find that ML methods provide meaningful insights that can complement choice model-based analysis in the context of data from complex choice experiments. Using AR learning, we identified potential positive and negative synergies between chosen policy options without imposing prior behavioural assumptions. For example, policies aiming to re-open businesses and allowing direct contact between relatives are more likely to be chosen together, rather than independently. This insight can be used to guide the specification of theory-driven (portfolio) choice models, quantify the strength of the identified synergies, and derive welfare economic measures. RF models allow to disaggregate the degree of importance of attributes for each individual alternative. This information can guide choice modellers in the variable selection processes and contribute to the development of more parsimonious theory-driven choice models.

More generally speaking, we conclude that machine learning methods provide choice modellers new opportunities for analysis of complex choice behaviour. First, our study support the notion that machine learning models can serve as an informative tool for choice modellers to explore new models for experiments that incorporate complex interactions or alternative decision rules. Second, our findings support the notion that machine learning models and theory-driven choice models are complementary tools, rather than substitutes, to obtain a better understanding of human decision making in complex settings, opening the door for new frameworks that combine the strengths of both approaches.

References

[1] Mouter, N.; Hernandez, J. I. and Itten, A. V. (2021). Public participation in crisis policymaking. How 30,000 Dutch citizens advised their government on relaxing COVID-19 lockdown measures, PloS one 16 : e0250614. [2] Bahamonde-Birke, F. J. and Mouter, N. (2019). About positive and negative synergies of social projects: treating correlation in participatory value evaluation. [3] Dekker, T.; Koster, P. and Mouter, N. (2019). The economics of participatory value evaluation. [4] Mouter, N.; Koster, P. and Dekker, T. (2020). Contrasting the recommendations of participatory value evaluation and cost-benefit analysis in the context of urban mobility investments, Transportation Research Part A: Policy and Practice 144 : 54-73. [5] Volberda, L. (2020). Analyzing citizens’ views on new spatial-infrastructure projects: From the average view towards various clusters within the Participatory Value Evaluation Method. [6] Sifringer, B.; Lurkin, V. and Alahi, A. (2020). Enhancing discrete choice models with representation learning, Transportation Research Part B: Methodological 140 : 236-261. [7] Wang, S.; Wang, Q. and Zhao, J. (2020). Deep neural networks for choice analysis: Extracting complete economic information for interpretation, Transportation Research Part C: Emerging Technologies 118 : 102701. [8] van Cranenburgh, S.; Wang, S.; Vij, A.; Pereira, F. and Walker, J. (2021). Choice modelling in the age of machine learning, arXiv preprint arXiv:2101.11948 .

ABSTRACT. This is a summary of the extended abstract. The full extended abstract is attached as a PDF file as it includes, equations, one table and one Figure.

'This study integrates Gaussian Processes (GPs) into Latent Class Choice Models (LCCMs) to allow for more flexible discrete representation of heterogeneity and improve the overall model fit and prediction accuracy. We propose a GP-LCCM that makes use of GPs to replace the class membership component of the traditional LCCM. The proposed model would rely on GPs as a non-parametric component to probabilistically divide the population into behaviorally homogenous classes while simultaneously relying on random utility models to develop class-specific choice models. We develop an Expectation-Maximization (EM) algorithm for training a GP classification approach as a clustering tool while concurrently learning the parameter estimates of the class-specific choice models. A few studies have addressed this clustering aspect of Machine Learning (ML) within LCCM. Han [1] used Neural Networks (NNs) to specify the classes while Sfeir et al. [2] relied on Gaussian-Bernoulli Mixture Models (GBMs). However, the estimation of NNs is usually complicated since the optimization problem might have several local optima [3]. Moreover, the parametric forms of NNs and GBMs restrict the flexibility of the learning process and might lead to poor prediction accuracies [4]. As for GPs, they are data-driven, do not assume predefined functional forms and consequently are free to learn any functional form. In addition, they are generally easier to handle since their posteriors are convex [3].'

ABSTRACT. Capturing heterogeneity in subjects’ decision making process, as accurate as possible, plays an essential role in choice modeling research. In this paper, we investigate the random taste heterogeneity in travel behavior modeling which is an integral part of decision-making process. In contrast to previous works, we use the Mixture Density Network (MDN) which is built from Neural Network and mixture Gaussian model to identify the latent heterogeneity. We assume that the taste variation of individuals follows a series of distribution with certain mean and standard deviation which are dependent on individual social demographic characteristics. We integrated this machine learning method into the discrete choice model and jointly estimated the parameters. Using the stated preference data of Swiss metro, we applied our proposed model and discovered random taste variations which are highly interpretable. We also compared the model with traditional mixed logit model and found the superiority of the proposed model.

1. Introduction It is an enduring question how to capture the taste heterogeneity of decision makers in choice modeling. In general, heterogeneity is not only associated with some unobservable components in the utility but also interacted with social demographic characteristics of individual people (Greene et al., 2006). Among the existing models to capture heterogeneity, Mixed Logit (MXL) model is the most popular approach. However, in mixed logit models, one has to assume a prior distribution for the random parameter without necessarily a theoretical background. More importantly, the assumption of mixed logit model implies that there exists one single distribution representing the heterogeneity among the population. Similar to the concept of latent class choice models, the heterogeneity between people may vary across the latent groups. Although the latent class mixed logit models somehow capture such heterogeneity, the extent they represent the magnitude of heterogeneity is limited, due to the limited number of classes. In recent years, there has been a growing interest in using machine learning (ML) methods to explore travel behavior. As for modeling taste heterogeneity, neural networks and Gaussian mixture models were used to analyze systematic and unobserved heterogeneity, respectively (Han et al., 2020; Sfeir et al., 2021). While Han estimated an individual’s systematic taste parameter using neural network, Sfeir used a mixture distribution to define classes. However, capturing systematic taste heterogeneity perhaps is insufficient when the heterogeneity is purely randomly varying, and the latent class model is somewhat less flexible than the mixed logit model. In light of the fact that MDN can learn the multiple random distributions from data, we contend that MDN is capable of capturing the random taste heterogeneity from behavioral datasets. Especially when incorporating MDN into the discrete choice models, it would improve the overall prediction accuracy. As a result, it is still similar to the mixed logit model because one network output corresponds to one random mixture distribution. In the current study, we therefore attempt to somehow combine and extend the two previous work by developing a MDN model which identify optimal number of mass points (groups) and estimate a finite mixture distribution. Our proposed model is semiparametric and does not have to assume a prior distribution as that in the mixed logit models. Furthermore, we use reparameterization which allows the joint estimation using machine learning algorithms such as stochastic gradient descent, leading to higher computational efficiency.

2. Methods Our proposed model framework is shown in Figure 1. Assuming N individuals, each facing J alternative for travel modes encompassing M socioeconomic characteristics Z_n^M,n=1,2,…,N. The random taste distribution for each individual can be obtained by a weighted linear mixture of K Gaussian distributions, as shown in Eqn 1. Further, individual’s taste parameters βn are draws from this mixture distribution. Since our total computation is in the neural network architecture, the parameters need to be propagated backward to update the gradient, so we use reparametrization to deal with the sampling. Eqn 1 can also be expressed as:

In this case, we only need to draw from N(0,1) that allows backpropagation to update π_n^k (z), μ_n^k (z), and σ_n^k (z).

As with other general neural networks, the hyper-parameters of this integrated framework include hidden layer size, batch-size, activation function, L1/L2 regularization and training epochs. Additionally, our proposed model also contains particular parameters, such as the number of mixtures and the number of draws.

3. Results Using the open dataset, Swissmetro, we estimated the model and found that the best prediction performance is achieved with 80 hidden units, ReLU activation function, no regularization, 5 mixtures and 500 draws. The prediction accuracy is 73.1%, which represents improvement compared to 71.8% in Han et al. (2020). As shown in Figure 2, we obtained 592 different groups (mass points) within which the taste distribution has the same properties. The top five subgroups contain 450, 252, 180, 180, and 126 individuals, representing 11.11% of the total population. These results suggest that the MDN approach is capable to capture individual random heterogeneity and to reveal homogeneous taste distribution among groups. As expected and shown in Figure 3, the distribution within each group is extremely narrow suggesting semi-homogenous taste within each group.

4. Conclusions This study elaborate the potential of integrating MDN into discrete choice models to capture random taste heterogeneity. We assume that the random taste heterogeneity is correlated with socioeconomic characteristics and that the heterogeneity can be well approximated using finite mixture distribution and numerous mass points. Our model specification uses reparameterization to guarantee the backward gradient update of the parameters. This work provides an operationally feasible case for integrating machine learning methods into the mixed logit model with a nonparametric mixture of distributions. It allows for fast estimates on large data sets and has great model flexibility. Further, it can find groups of rather homogeneous subjects in their tastes, who are not otherwise distinguishable.

ABSTRACT. Real world choice situations are often inherently ordered. For example, a survey respondent could be asked to provide a Likert scale satisfaction rating of a service (e.g., out of very poor, poor, acceptable, good, very good), a customer on an online shopping website could be asked to rate a recent purchase between one and five stars, or an individual might choose how many days per week they work from home. In all of these examples, there are clear correlations between the alternatives: working five days a week from home is more similar to working four days a week from home than not working from home at all, and a service rating of very poor is more similar to poor than it is to good or very good. These correlations between alternatives violate the fundamental i.i.d. error assumption of the logit model, as well as other machine learning models that make use of the logistic (or softmax) function, such as gradient boosting and probabilistic neural networks.

There exist several methods to capture the correlations between ordered alternatives in econometric utility-based choice models, including the use of nested or mixed logit model structures or the (generalised) ordered logit/probit model. However, there are limited investigations which attempt to explicitly capture these correlations in data-driven machine learning models, and in particular with ensemble learning. This is despite the best-in-class predictive performance of ensemble learning models (including gradient boosting) for similar predictive tasks. Whilst there do exist a limited number of studies which investigate ordinal classification using ensemble learning (e.g. [1, 2, 3]), these methods can only be used to perform deterministic (0-1) classification, which is not appropriate for choice modelling. To the best of the author’s knowledge, there exist no studies in the literature which perform probabilistic ordinal classification with ensemble learning.

To address this gap, this paper proposes a new probabilistic ensemble learning model, referred to as the Ordinal Gradient Boosting Machine (OrdinalGBM). The OrdinalGBM model explicitly captures the correlations between ordinal alternatives by generating choice probabilities from a single regression value (similar to the utility-based ordinal logit/probit models). The general approach is to apply a transformation function to the unbound scalar output of an ensemble of additive trees (i.e., a gradient boosting regression model), which maps the ensemble output to a vector of probabilities over the alternatives. This is in contrast with conventional gradient boosting classification, where a separate ensemble is fit to each alternative, before being passed through the logistic (softmax) function. By determining the analytical gradient and Hessian of the cross-entropy loss with respect to the ensemble output, the ensemble can then be fit to training data using numerical optimization in the function space [4].

Two different base transformation functions for OrdinalGBM are implemented, following the approach used in [5]: (i) the binomial distribution, and (ii) the Poisson distribution. Each of these distributions have a fixed unimodal shape, defined only by the continuous ensemble output and the number of alternatives in the choice set. To introduce more flexibility in the fitted distributions, alternative forms of each transformation function are implemented which include a constant additive weight term for each alternative. This allows the models to change the underlying shape of the transformation function to match that of the data. A new training algorithm is implemented that re-estimates the weight terms (using conventional parametric gradient descent) after each boosting iteration.

OrdinalGBM has been implemented as a scikit-learn compliant classification model in Python, making use of the LightGBM library [6]. This allows OrdinalGBM to be systematically evaluated against four groups of models: (i) conventional (nominal) gradient boosting classification, (ii) probabilistic ordinal logit and probit models, (iii) probabilistic ordinal neural networks, and (iv) deterministic ordinal machine learning approaches. A diverse set of performance metrics are used to compare model performance, including the (i) the cross-entropy loss; (ii) the deterministic accuracy, mean absolute error and mean squared error (calculated with maximum probability assignment for comparison with deterministic classifiers); and (iii) two new probabilistic metrics introduced in this paper, the expected mean absolute error and expected mean squared error.

A thorough hyperparameter search is conducted across four benchmark datasets which identifies that, out of the four proposed model specifications, the binomial OrdinalGBM model with additive weights performs best on a wide range of ordinal choice modelling tasks. Furthermore, the systematic evaluation of OrdinalGBM against the benchmark models reveals that the OrdinalGBM model achieves state-of-the-art predictive performance for the four ordinal classification problems. Compared to nominal gradient boosting classification, the OrdinalGBM model achieves similar log-likelihood loss with a far simpler model and achieves better distance-based evaluations (expected and deterministic mean absolute error and mean squared error), representing a better overall fit to the data.

[1] Gonen Singer and Matan Marudi. “Ordinal Decision-Tree-Based Ensemble Approaches: The Case of Controlling the Daily Local Growth Rate of the COVID-19 Epidemic”. In: Entropy 22.8 (2020), p. 871.

[2] Pedro Antonio Gutiérrez et al. “Ordinal Regression Methods: Survey and Experimental Study”. In: IEEE Transactions on Knowledge and Data Engineering 28.1 (2016), pp. 127–146. 6

[3] Joao Costa and Jaime S. Cardoso. “oAdaBoost - An AdaBoost Variant for Ordinal Classification:” in: Proceedings of the International Conference on Pattern Recognition Applications and Methods. International Conference on Pattern Recognition Applications and Methods. Lisbon, Portugal: SCITEPRESS - Science and Technology Publications, 2015, pp. 68–76.

[4] Jerome H. Friedman. “Greedy Function Approximation: A Gradient Boosting Machine”. In: The Annals of Statistics 29.5 (2001), pp. 1189–1232.

[5] Joaquim F. Pinto da Costa, Hugo Alonso, and Jaime S. Cardoso. “The Unimodal Model for the Classification of Ordinal Data”. In: Neural Networks 21.1 (2008), pp. 78–91.

[6] Guolin Ke et al. “LightGBM: A Highly Efficient Gradient Boosting Decision Tree”. In: Advances in Neural Information Processing Systems. Vol. 30. Curran Associates, Inc., 2017.

ABSTRACT. In this study we employ a choice experiment to analyze New York City residents’ preferences for online grocery shopping at the beginning of the COVID-19 pandemic. Specifically, we focus on the role of their attitudes as drivers of behavior.

COVID-19 has affected the lives of people all around the world, forcing them to change their daily habits and behaviors in a rapid fashion. Suddenly, almost every person needed to adjust to a new reality of social distancing, wearing masks, and remote working. Multiple public policies were introduced to limit the spread of the virus, including policies aimed at restricting social life to decrease physical interactions between individuals. For instance, the “New York State on PAUSE” executive order was introduced on March 22 2020, closing all non-essential businesses in the state, including restaurants and retail shops. Our choice experiment survey was conducted at the beginning of May of 2020.

The motivation for this study is twofold. First, the online grocery shopping sector has been steadily growing throughout the last decade, making it an important field of study in its own. This growth has been rapidly hastened by the pandemic. Although, some of the interest in online grocery is likely to fall after the pandemic, some habits developed during this period are likely to continue. Furthermore, multiple large retailers invested in online grocery technology, and continue to do so, even after the number of COVID-19 cases dropped and government restrictions were lifted. We therefore argue that it is of major importance to understand customer preferences for online grocery shopping, to identify attributes that consumers consider important, and to find factors that drive their behavior. Second, because of the pandemic, there are new socio-psychological factors that are likely to affect individuals’ propensity to shop for groceries online, which were not considered before the outbreak. These factors include, among others, perceived vulnerability and risk of infection (Moon et al., 2021), fear of the health-related and economic consequences of the pandemic (Eger et al., 2021), general uncertainty caused by the pandemic, attitude towards social distancing guidelines (Moon et al., 2021), and the perceived risk of formal penalties (e.g., fines) for not complying with the government restrictions. With the pandemic still active, and having in mind possible future pandemics, it is important to understand how these factors may affect individuals’ behavior.

In this study we focus on two attitudes. First is a general concern with COVID-19, which we measure by the self-reported propensity to take measures to mitigate the risk of infection, such as wearing a mask outdoors or using disinfecting wipes. The second is an attitude towards crowding. Social distancing is probably one of the most common mitigation strategies introduced by the governments around the world. At the same time, offline shopping is often connected with crowding, as one of the main advantages of large retail shops is that they can process a large number of customers in a short time. There is a considerable literature on the effect of perceived crowding on retail behavior (Blut and Iyer, 2020). Although the results are mixed, it is likely that the pandemic has strengthened a negative effect, as now crowding can constitute a health hazard. Therefore, we argue that an attitude towards crowding may act as a significant factor pulling individuals from offline shopping to online shopping.

To combine choice experiment data with attitudinal questions from the survey we use a hybrid choice modeling framework. Specifically, we employ a latent class model in which two attitudes enter as latent factors explaining probability of belonging to a given class. The latent class specification is useful in this context, as market segmentation is considered to be essential to better understand shopping behavior (Eger et al., 2021). We identified three distinct latent classes. One of them is an opt-out class, in which individuals are very likely to opt-out from online grocery shopping, irrespectively of the attributes presented in the choice experiment. We find that individuals who belong to this class are mostly older with no previous experience with online grocery shopping. Individuals in the two other classes are likely to buy their groceries online, but differ in terms of their sensitivity to the delivery cost and relative importance of the other attributes. In the class with higher cost sensitivity, no contact delivery occurred to be the most important, whereas for the other class brand variety had the highest willingness-to-pay.

With respect to the two latent variables which we study, we found that both are significant. Individuals who are concerned about COVID-19 are more likely to belong to the class with higher willingness-to-pay, and are less likely to opt-out. Surprisingly, we find that individuals who report a poor health status are less concerned about the pandemic. As for the second latent factor, we find that consumers who have a negative attitude towards crowding are more likely to belong to the class which assigns a high importance to no contact delivery.

Results provided by this study are relevant for the retail industry as well as policymakers who want to limit the spread of the virus. With respect to the latter, we find that individuals who are the most vulnerable (i.e., in poor health and overweight) are either less concerned about COVID-19 or less likely to avoid crowds. This suggests that a strategy of limited restrictions, in which individuals with the highest health risk are assumed to voluntarily take necessary precautions may prove unsuccessful.

References Blut, M., and Iyer, G.R. (2020). Consequences of perceived crowding: A meta-analytical perspective. Journal of Retailing, 96(3), 362-382. Eger, L., Komárková, L., Egerová, D., and Mičík, M. (2021). The effect of COVID-19 on consumer shopping behaviour: Generational cohort perspective. Journal of Retailing and Consumer Services, 61, 102542. Moon, J., Choe, Y., and Song, H. (2021). Determinants of consumers’ online/offline shopping behaviours during the COVID-19 pandemic. International journal of environmental research and public health, 18(4), 1593.

ABSTRACT. Norway and Italy have a quite different car passenger fleet composition and make different car choices. Norway is the world forerunner in PEV (EV+PHEV) adoption, also thanks to the strong incentivizing policies adopted in the last 20 year (Scorrano et al., 2019). Italy, on the contrary, is a European laggard in terms of electrification, with a car fleet still heavily relying on diesel and petrol cars, and more recently on hybrid cars. On the contrary, the PEV share is still low also due to the poor charging infrastructure. In fact, monthly PEV registrations in Norway represent about 90% of the total sales and the PEVs in use as a proportion of all passenger cars on the roads are close to 20%. In Italy, monthly PEV registrations are at 15% and they represent less than 1% of the cars on the road. The aim of the paper is to explore and compare causes and effects of such different innovation adoption paths. Building on previous research (Rose et al., 2009; Hess et al., 2018; Rotaris et al., 2021), we make use of the hybrid mixed choice model to investigate how the revealed and stated choices of car users among alternative propulsion systems are related to vehicle attributes’ preferences, residential characteristics, EV infrastructure and drivers’ attitudes in the two countries. We consider five propulsion systems (petrol, diesel, hybrid, electric and plug-in hybrid cars) and three vehicle attributes (purchase price net of taxes and subsidy, driving range with a full tank/charge and cost of the fuel/electricity per 100 km). We specify the following residential characteristics: neighborhood type, housing tenure, housing type, residential parking, and availability of a photovoltaic and/or a battery system. The EV market is characterized by the following indicators: EV density at city/regional level, High Performance Charging density at city/regional level, the distance (in km) to the nearest EV charging station from residential location, and the availability of a EVSE at work. As for the attitudes and perceptions, we investigate the role of environmental awareness, interest for new technologies, status symbol, social influence, risk aversion, and the perception towards the technical characteristics of EVs. Since the uptake of EVs is so different between the two countries, it is also to be expected that the level of information and knowledge of EVs might play a relevant role, as documented in several papers (Jensen et al., 2013; Danielis et al., 2020; Giansoldati et al., 2020). In order to account for the difference in fleet composition (Italy being characterized by a rather large small market segment due to the density of the urban settlements and the lower income levels), we collect detailed information on the revealed and stated choices distinguishing between two car segments: compact and large cars. This allows us to explore whether EVs are more likely to be used for city travelling or for longer intercity trips. The model is estimated thanks to a recently collected revealed- and stated-preference dataset from a sample of more than 1000 respondents using a specifically designed questionnaire. We find striking differences in future purchase intentions: Norwegians confirm an interest into buying PEVs especially in the small car segment (less so in the large one), while Italians are less prone to give up the conventional propulsion systems with the exception of substituting them with hybrid cars. However, the preference structure and the environmental attitudes appear to be rather similar, apart from the impact of knowledge and peer pressure, indicating that the PEV market uptake in Norway is mainly the result of the strong support resulting from the regulatory and fiscal policies enacted at local and national level (exemptions from VAT, one-off registration tax and annual ownership tax, discounted road toll, public parking fees, ferry fares, and free access to bus lanes).

References Danielis, R., Rotaris, L., Giansoldati, M., & Scorrano, M. (2020). Drivers’ preferences for electric cars in Italy. Evidence from a country with limited but growing electric car uptake. Transportation Research Part A: Policy and Practice, 137, 79–94. https://doi.org/10.1016/j.tra.2020.04.004 Giansoldati, M., Rotaris, L., Scorrano, M., & Danielis, R. (2020). Does electric car knowledge influence car choice? Evidence from a hybrid choice model. Research in Transportation Economics, 80. https://doi.org/10.1016/j.retrec.2020.100826 Hess, S., Spitz, G., Bradley, M., & Coogan, M. (2018). Analysis of mode choice for intercity travel: Application of a hybrid choice model to two distinct US corridors. Transportation Research Part A: Policy and Practice, 116, 547–567. https://doi.org/10.1016/j.tra.2018.05.019 Jensen, A. F., Cherchi, E., & Mabit, S. L. (2013). On the stability of preferences and attitudes before and after experiencing an electric vehicle. Transportation Research Part D: Transport and Environment, 25, 24–32. https://doi.org/10.1016/j.trd.2013.07.006 Rose, J. M., Hensher, D. A., Caussade, S., Ortúzar, J. de D., & Jou, R. C. (2009). Identifying differences in willingness to pay due to dimensionality in stated choice experiments: a cross country analysis. Journal of Transport Geography, 17(1), 21–29. https://doi.org/10.1016/j.jtrangeo.2008.05.001 Rotaris, L., Giansoldati, M., & Scorrano, M. (2021). The slow uptake of electric cars in Italy and Slovenia. Evidence from a stated-preference survey and the role of knowledge and environmental awareness. Transportation Research Part A: Policy and Practice, 144, 1–18. https://doi.org/10.1016/j.tra.2020.11.011 Scorrano, M., Mathisen, T. A., & Giansoldati, M. (2019). Is electric car uptake driven by monetary factors? A total cost of ownership comparison between Norway and Italy. Economics and Policy of Energy and the Environment, 2. https://doi.org/10.3280/EFE2019-002005

ABSTRACT. Food expenditure outlays constitute a third leading category of household budget expenditures in most countries of the world, along with housing and transportation outlays. For example, in the U.S., it has been estimated that about 12.5% of the average U.S. household budget is spent on food. Further, until the COVID pandemic, there was a steady increase in food consumption based on non-domestically cooked meals (NDCM), especially eating meals outside the home at a commercial catering establishment. While the peak COVID period led to a decline in such eat-out activity, the period also led to a surge in eat-in NDCMs, through take-out from an eatery or delivery of fully/partially-prepared meals to the home (for ease in presentation, we will refer to such NDCM food consumption simply as “eat-in”, distinctly different from the traditional domestically-cooked meals or DCM arrangement). In a post-vaccination COVID period, the food industry appears poised to see a surge in both eat-out and eat-in NDCM activity.

The cultural shift away from DCM preparation has been attributed in food sociology studies to multiple factors, including (a) increasing participation of women in the work force and the resulting time rebalance to accommodate this remunerative pursuit, (b) growing acceptance in society that cooking is not primarily the women’s responsibility, (c) rising affordability of NDCMs due to the industrialization of the food industry, (d) delinking of DCMs from its symbolic social, cultural and/or religious significance as a family gathering event, (e) viewing NDCMs as an opportunity to socialize, or an opportunity to experience different cuisines, or an escape from routine, and (f) signaling cultural capital and social distinctiveness.

Many earlier studies of consumers’ NDCM activity have focused on the univariate count of eat-out activity over a given time period, sometimes separated out by time-of-day of the eat-out activity (such as lunch and dinner). The determinant variables used in these count models include socioeconomic factors such as age, gender, employment, income, presence of children in the household, and spatial factors such as the location of work and residence. Other studies examine, at each choice occasion, the eat-out venue (such as fast food outlets, pizza houses, and sit-down restaurants) or the cuisine type (such as by ethnic categorization or other categorizations such as vegetarian versus non-vegetarian) or restaurant choice based on service characteristics (such as decoration and novelty, staff friendliness, reputation, and parking facility). Most studies in this latter category are descriptive or qualitative in nature. In this paper, we contribute to this field of consumer food consumption behavior by examining not only consumer eat-out behavior, but also the relatively unexplored eat-in dimension of NDCM activity behavior. Specifically, we formulate a quantitative model of the NDCM weekday dinner event count in a month by NDCM event state (that is, whether an NDCM episode is eat-in or eat-out, and the venue if eat-out; we confine our attention to dinner meals because lunch meals such as those at work may be less of a choice and more of a biological need). Methodologically, unlike statistically-stitched multivariate count models, we explicitly model (a) the total NDCM count and (b) the event state given an NDCM instance, while also inter-linking these two components of our joint model. For example, a change from working at home to working from the office may reduce the probability of eat-out activity (because the individual may be tired and may want to stay in and relax at dinner time), but also increase the overall frequency of NDCM instances (because of the additional time-crunch created by travel to the work place). In such a situation, while the count of eat-in episodes will increase because of a change from working at home to working from the office, the count of eat-out activity may increase or decrease, depending on whether or not the overall rise in the count of NDCM instances is higher relative to the decrease in eat-out activity given an NDCM episode. Typical statistically-stitched multivariate count models are unable to capture such nuanced and individual effects of a covariate (working from home or not in the example above). Further, an individual’s work location may vary from one weekday to the next (this is particularly the case in the current COVID pandemic existence and beyond), and therefore it becomes necessary that the inter-linkage between the event state and the total NDCM count model recognize time-varying covariates (other such time-varying covariates may include day of the week, working hours during the day, special occasions, weather, and eating companions). To our knowledge, our proposed utility-theoretic linkage structure, which considers time-varying covariates and also accepts the full distribution of the maximum utility from a multinomial probit (MNP)-based discrete event instance model for introduction into the total NDCM count model, is a first in the econometric and consumer choice modeling field.

The proposed approach is applied to examine the number of monthly NDCM dining instances of adults in Central Texas by eat-in or eat-out, and the venue if eat-out. The primary data for the current study will be obtained through a survey, scheduled to be deployed across Texas, US in January, 2022, that will collect information on individuals’ NDCM choices and the time varying attributes corresponding to each NDCM occasion. The results from the study can provide insights into eating-in and eating-out behavior, which can be helpful for the restaurant and catering industry, including in locating food-related businesses and services. Also, food prepared outside the home typically has a higher energy content, and lower nutrient value, potentially having adverse health effects. Thus, the results can provide health and government authorities with useful information for nutritional regulation purposes. Finally, a good understanding of the factors impacting the number of NDCM instances by eat-out versus eat-in can also provide valuable information to travel demand analysts in an environment where NDCM instances are on the rise, and affect travel associated with virtually-driven home meal delivery instances and in-person out-of-home meal instances.

ABSTRACT. Gender inequality in labour market participation is a widespread phenomenon in many countries. The European Commission reports that 67% of women in the EU are currently in employment, while the figure for men is 79%, implying a gender employment gap of 12% (European Commission, 2021). The latest data for the UK place this figure around 10%. While in recent years more women than men have graduated from European Universities, their choice process to participate in the labour market is known to differ from the one that applies to men, as they tend to carry the largest burden in terms of domestic unpaid work as well as child and/or elderly care. The gender employment gap is believed to result in substantial economic loss, estimated at €370 billion per year for the EU.

Studies in labour economics have aimed at understanding the determinants of women’s labour market participation. Standard models include variables such as education, age, marital status as well as household characteristics (e.g. Kanellopoulos & Mavromaras, 2002). While these factors have been shown to significantly affect the choice of employment status, other potential determinants are not considered in traditional model structure.

One such example is attitudes. A wealth of literature across transport and health has shown a link between attitudes and behaviour (e.g. Daly et al. 2012). While many papers have suggested a causal link between the two, Chorus and Kroesen (2014) have warned against the use of cross-sectional data in this context, in favour of panel data.

This study aims to study the link between the change in attitudes towards gender roles in the households and in employment status. This not only contributes to the labour economic literature by considering “soft” factors in addition to the traditionally employed “hard” factors, but also does so with panel survey data, allowing us to draw inferences about the direction of causality between attitudes and behaviour.

The data employed in this paper is Understanding Society – the UK Household Longitudinal Study (UKHLS), a representative sample of over 40,000 households across the UK (University of Essex, 2018). The data include information on social and economic circumstances, attitudes, behaviours and health. Eleven waves are currently available, starting with wave 1 in 2009, which provided data on over 50,000 individuals. The respondents used in our analysis are women who are part of a couple. This is relevant as information about the spouse is also collected, providing a more complete image of the family context.

The specific attitudinal statements of interest are the following: • The husband should earn and the wife stay at home; • Preschool children suffer if the mother works; • The family suffers if the mother works full time; • Husband and wife should both contribute • Employers should help mothers combine employment and childcare.

As the level of agreement with these statements was only collected in three waves (2, 4 and 10), we limit our analysis to these three points in time, focusing on the change in the responses to the statements as well as the change in employment status. At each point in time, women declare if they are not employed (NE) working full time (FT) or working part time (PT), implying 7 options for our model (no change, FT to PT, FT to NE, PT to FT, PT to NE, NE to FT and NE to PT). We also analyse the impact of traditional explanatory variables such as number of children, education, ethnicity, and household income (excluding the participant). Preliminary analyses using waves 2 and 4 were conducted with al sample which includes full answers in two survey waves by 1,819 women. Once the model is developed, adding the third wave of data will be straightforward.

The chosen model structure is a hybrid choice model where the utility of the labour market status transition (i.e. of each of the 7 options listed above) is affected by the so-called “hard” factors, i.e. related to socio-demographics and household characteristics, as well as by the change in attitudes, which in turn explain the responses to the attitudinal statements above. In particular, factor analysis has allowed us to determine a single factor that we have interpreted as egalitarian attitude. We also explore the possibility of the change in this attitude to be explained by socio-demographic characteristics. This allows us to separate the impact of specific variables (e.g. education or age range) on the employment choice and on the attitude.

While the modelling structure is relatively straightforward, this study presents a number of challenges worthy of investigation. The data has revealed that a large proportion of the sample (81%) does not change their employment over the two waves. This is to be expected, together with the fact that most of the changes will be between FT and PT. While the patterns in the data pose some limitations on the analysis, exploring different approaches to face such issues will allow us to potentially advance the use of choice models with panel data of “rare” outcomes. We also hope to demonstrate the applicability of discrete choice models for problems typically linked to strict model structures in the field of economics and demonstrate their flexibility and ability to provide additional insights, possibly encouraging their application to address key societal and economic problems.

References

Chorus, C. G., & Kroesen, M. (2014). On the (im-) possibility of deriving transport policy implications from hybrid choice models. Transport Policy, 36, 217-222.

Daly, A., Hess, S., Patruni, B., Potoglou, D., & Rohr, C. (2012). Using ordered attitudinal indicators in a latent variable choice model: A study of the impact of security on rail travel behaviour. Transportation, 39(2), 267-297.

European Commission (accessed 13/12/2021). Women's situation in the labour market. Url: https://ec.europa.eu/info/policies/justice-and-fundamental-rights/gender-equality/women-labour-market-work-life-balance/womens-situation-labour-market_en

Kanellopoulos, C. N., & Mavromaras, K. G. (2002). Male–female labour market participation and wage differentials in Greece. Labour, 16(4), 771-801.

University of Essex, Institute for Social and Economic Research. (2021). Understanding Society: Waves 1-11, 2009-2020 and Harmonised BHPS: Waves 1-18, 1991-2009. [data collection]. 14th Edition. UK Data Service. SN: 6614, http://doi.org/10.5255/UKDA-SN-6614-15.

ABSTRACT. When consumers are faced with a large number of alternative products from which to choose, they may engage in a two-stage choice process. They first select a subset of alternatives for further evaluation, called the consideration set, and then choose the final option from this consideration set. This two-stage choice process is widely studied in the literature, resulting in different models for the consideration phase and the choice phase. Irrespective of which modeling strategy is used, evidence suggests that including a consideration stage in choice modeling provides a more realistic representation of the choice process and leads to improved forecasts and a better explanation of consumer behaviour.

Modeling the two-stage choice framework depends on whether or not the consideration set is observed. When consideration sets are not observed, researchers either ignore the consideration set generating process by assuming that each individual chooses from the complete choice set or they model the consideration stage as a latent process (referred to as latent CSC model) that is estimated indirectly from the choice data (see for instance Van Nierop et al., 2010). On the other hand, when consideration data is available, it is common to estimate choice models conditional on these data (referred to as observed CSC model). However, estimating models using only the set of considered alternatives might involve inconsistency problems if the chosen alternative is not in the consideration set.

We propose and investigate an extension of the latent consideration set choice model (referred to as extended latent CSC model) to incorporate stated consideration set data. The main motivation is to assess the value of consideration data for recovering the population preference parameters. Unlike the latent consideration model which derives the probability of a consideration set indirectly from choice data, we directly model this probability using the stated consideration data. This approach solves the inconsistency problem. For the sake of simplicity, we consider a compensatory model both for the consideration and the choice phases, as it can provide a reasonably accurate approximation for several non-compensatory processes (Roberts & Lattin, 1991).

We compare the extended CSC model with existing approaches based on a large simulation study. We first generate the individual-specific parameters for the consideration model and for the choice model from multivariate normal distributions. We assume that the individual specific consideration and choice parameters are different but highly correlated as they measure the effect of the same attribute (although we discovered that the value of this correlation does not impact the main results).

First, we simulate the consideration data. The probability of accepting each alternative conditional on individual specific consideration parameters is calculated using a logistic model. Then, these probabilities are in turn used to generate stated consideration data from a Bernoulli distribution. Then we simulate a second set of latent consideration data and the choice data is simulated conditional on the newly simulated consideration data and individual-specific preference parameters. The purpose of the second simulation of consideration data is to create inconsistencies between the latent consideration data on which choices are based and the observed consideration data. The mixed logit model is used to calculate the probability of choosing among those alternatives that are considered. If none of the alternatives is considered, then the opt-out option is chosen with probability one. The choice data is then generated from a multinomial distribution using the calculated probabilities. For each generated dataset, the simulated maximum likelihood approach is used to estimate the parameters of the mixed logit model, the latent CSC model, the observed CSC model and the extended latent CSC model, respectively. The population preference parameter estimates obtained from these approaches are compared for accuracy in terms of the root mean square error (RMSE) and the bias of each of the parameters.

It turns out that the mean preference parameter vector is well recovered by the models that take into account the consideration set screening. Our results confirm what has been shown in several other studies: a one stage mixed logit model does not recover the true mean parameters well when there is screening prior to the choice task. Compared to models that use both consideration and choice data, the RMSE and bias of the mean parameter estimates are only slightly larger for the latent CSC model that only uses choice data.

Contrary to expectation, the proposed extended latent CSC model recovers the true parameters well but slightly worse than the observed CSC model which excludes the approximately 15% inconsistent consideration and choice data in our simulation study. The extra noise induced by modeling the consideration process outweighs the loss of a considerable part of the data.

Regarding the recovery of the heterogeneity parameters, all four models had a very similar performance. It is striking that the mixed logit model which ignores the screening, resulting in serious bias in the estimated mean preference parameters, performs as well as the other models in retrieving the heterogeneity parameters.

From these results we can conclude that when there is consideration set screening, one should explicitly account for the screening process to obtain unbiased estimates of the mean population preference parameters. Although the latent CSC model uses only the choice data, the recovery of the population mean preference parameters is almost as good as by using the models that use also stated consideration data.

References Roberts, J. H., & Lattin, J. M. (1991). Development and testing of a model of consideration set composition. Journal of Marketing Research, 28 (4), 429-440. doi: 10.1177/002224379102800405 Van Nierop, E., Bronnenberg, B., Paap, R., Wedel, M., & Franses, P. H. (2010). Retrieving unobserved consideration sets from household panel data. Journal of Marketing Research, 47 (1), 63-74. doi: 10.1509/jmkr.47.1.63

ABSTRACT. 1 Introduction

A lot of behavioural choice models arise from a tension between normative idealizations and descriptive theories of decision behaviour. As advocated by many authors, we should augment the normative framework with additional descriptive theories that have a solid empirical basis. In the present paper we pursue this goal and develop a consider-then-choose model which assumes decision makers apply screening heuristics before choosing.

Well known heuristics are, for example, lexicographic choice behaviour, elimination or selection by aspects, and disjunctive and conjunctive decision rules. In the present paper, we assume that such heuristics are used as screening rules, which serve to reduce the complexity of a choice task. As such, decision makers first form a consideration set conditionally on which they make a final choice. Over the past decades, research interest in the formation of consideration sets has never declined, and both theoretical and empirical evidence for such screening behaviour has accumulated. In our consider-then-choose model, the conditional choice phase is modelled with a logit model, as compensatory strategies are typically used when the amount of information to process is sufficiently small.

When combining screening heuristics with a compensatory choice phase, it appears to be difficult to reliably estimate both consideration and choice parameters from choice data only, while keeping the models parsimonious and identified. To overcome these difficulties, consider-then-choose models often rely on auxiliary data such as scanner data, consideration data or consider-then-rank data, impose carefully selected assumptions that restrict the number of possible consideration sets, or assume a functional relation between the consideration probability and the choice probability of an alternative. In this paper, we present a two-stage model that solely depends on discrete choice data, in which participants only state their preferred alternative or choose the no-choice option.

2 Relation to existing models

To our knowledge, there exist a few consider-then-choose models that solely require choice data. Gilbride and Allenby (2004) proposed a two-stage model that uses data augmentation and Markov Chain Monte Carlo (MCMC) methods to estimate threshold parameters that allow to identify disjunctive, conjunctive or compensatory screening rules. In addition, they estimate preference parameters to model the compensatory choice stage. Kohli and Jedidi (2005) presented a two-stage model that can be estimated with standard maximum likelihood methods and uses no-choice responses to identify subset-conjunctive screening rules (of which disjunctive and conjunctive rules are special cases). Their model does not take any form of heterogeneity into account and they do not estimate separate preference parameters in addition to the consideration parameters. Van Nierop et al. (2010) developed a two-stage model in which consideration set membership of alternatives is modelled directly with binary probits. Their model is developed for revealed choice data, and as such, does not take no-choice responses into account. Lastly, Cantillo and Ort´uzar (2005) proposed a two-stage model which also makes use of thresholds for accepting attribute levels. The consideration phase is hereby modelled by estimating the distribution of those thresholds.

In this paper, we propose a parsimonious consider-then-choose model that includes a non-compensatory screening phase and a compensatory choice phase. Our model accounts for heterogeneity in consideration screening by allowing respondents to have different thresholds for accepting attribute levels, as is the case in Gilbride and Allenby (2004) and Cantillo and Ort´uzar (2005). By combining such thresholds a variety of heuristics can be mimicked, as the resulting consideration sets would be the same. As such, the model can be seen as a latent class model in which each class represents a unique combination of thresholds (one for each attribute). One particular class (in which all attribute levels are acceptable) represents the case in which no screening is applied. This class is equivalent to a single stage logit model. Our approach differs from Cantillo and Ort´uzar (2005) as we do not use random threshold parameters, but a finite set of predefined possible thresholds for each attribute. Contrary to Van Nierop et al.(2010), and similar to Kohli and Jedidi (2005), we assume that no-choice responses denote empty consideration sets. We do not use MCMC methods such as Gilbride and Allenby (2004), but apply standard maximum likelihood techniques to estimate our model which makes it fairly easy to implement.

3 Contributions

Via a simulation study we show that the model is identified and that both consideration and preference parameters can be reliably estimated. Furthermore, we study the consequences of estimating the MIXL model (which does not take consideration screening into account) when the true data generating mechanism does include consideration sets and vice versa. We show in detail how estimates of compensatory models are affected when screening rules are being used. Conversely we show that, if no-choice responses are mistakenly assumed to be the result of empty consideration sets, the estimates of our model are biased. Importantly is that in all simulations the correct type of model was selected for each generated dataset, indicating that screening behaviour can be detected in discrete choice data. We compared the performance of our model on empirical stated preference data concerning cinemas, with benchmark models such as the conditional logit, mixed logit, and a latent class logit model. We concluded that our model outperformed all other models based on AIC and BIC, and allowed to gain insights which none of the other models could provide. We concluded that 73% of the respondents applied a screening heuristic before making a final choice. This is in line with Gilbride and Allenby (2004), who reported that 92% of the participants in their sample used a conjunctive screening rule.

Bibliography

Cantillo, V. and Ort´uzar, J. d. D. (2005). A semi-compensatory discrete choice model with explicit attribute thresholds of perception. Transportation Research. Part B: Methodological, 39(7):641–657.

Gilbride, T. J. and Allenby, G. M. (2004). A choice model with conjunctive, disjunctive, and compensatory screening rules. Marketing Science, 23(3):391–406. Kohli, R. and Jedidi, K. (2005). Probabilistic subset conjunction. Psychometrika, 70(4):737–757.

Van Nierop, E., Bronnenberg, B., Paap, R., Wedel, M., and Franses, P. H. (2010). Retrieving unobserved consideration sets from household panel data. Journal of Marketing Research, 47(1):63–74.

ABSTRACT. Please view the attached file.

ABSTRACT. The consideration set corresponds to the set of alternatives considered by the individual when she is making a choice. This set plays a fundamental role in choice modeling, since its composition influences the route choice model estimates and the choice probabilities (Bliemer & Bovy, 2008; Prato & Bekhor, 2007). The difficulty of identifying the consideration set arises because this set is latent, i.e., researcher cannot observe it in a real situation. This problem is exacerbated in route choice modelling due to the large number of alternative routes that usually are feasible, especially in a dense network. Up to this date, to the best of our knowledge, little is known about the impact of the composition of the consideration set in public transport modeling. Our study contributes to shedding some light on this subject by comparing various practical approaches that have been proposed in the literature under a common experimental data set constructed from smart card data on public transport route choices for the city of Santiago de Chile. There is a vast literature that has used the construction of the consideration set prior to the route choice model. The seminal theoretical work in this regard corresponds to Manski (1977), who proposes an approach where the consideration set is treated as a latent variable. The practical implementation of this approach involves huge computing costs for realistic problems because requires enumerating a lot of consideration-sets and also defining a function for the probability of considering each one. To solve the latter practical problem, Swait & Ben-Akiva (1987) and Ben-Akiva & Boccara (1995) propose using individual characteristics or restrictions to develop expressions for the consideration-set probability, which, although this approach is intuitively appealing, is still ad-hoc solutions to the problem. Another approach to the problem is aimed to propose practical methods for generating the consideration set, which is usually an algorithm or heuristic that tries to reproduce the behavior that passengers use to choose a route. In this context, the shortest path heuristics are widely used for route choice modelling. These heuristics generate the consideration set by iteratively searching the least cost route, defining “least” under different criteria. Some shortest path heuristics have been applied in the literature of public transport route choice modeling. For example, Guo (2011) worked with the labeling approach proposed by Ben-Akiva et al. (1984). Anderson et al. (2017) adapted the double stochastic method to estimate the choice set in a large-scale multimodal public transport network. Tan et al. (2015) used a combination of four choice set generation approaches: link elimination, labeling, k-shortest path, and simulation. Another practical approach used in the last decade to tackle the consideration set problem has been to impute it from historical data, i.e., previous choices made in a similar situation (Jánošíkova et al., 2014; Kim et al., 2020; Raveau et al., 2011, 2014; Yap et al., 2020). We denominate this method as the Historical/Cohort approach, which can be formally defined as building a practical consideration set from observed choices that occurred in some past instances of the traveler or observed choices of other users in the same cohort in cross-section data. The cohort in cross-sectional data may correspond to trips that were performed on the same OD pair, period, trip purpose, by individuals of the income group, and household composition. In this line, Villalobos & Guevara (2021) investigated the impact of different heuristics to construct the consideration set in the estimation of route choice models, using Monte Carlo experiments, and they evidenced that the Historical/Cohort approach was the only one of the methods analyzed that was able to recover the population parameters in the experiments. In this paper, we extend the empirical results of Villalobos & Guevara (2021) in two ways. We first revisit a formal theoretical demonstration (Guevara, 2021) that shows that, under appropriate assumptions, the Historical/Cohort approach to build the consideration set is able to provide consistent estimators of the population parameters. This demonstration is based on an adaptation of the theorem of sampling of alternatives (McFadden, 1978), in which each historical/cohort choice is understood as a draw from the true consideration set and the sampling correction is canceled out when there are many observations. In this context, we hypothesize that the route choice models that use the Historical/Cohort approach for identifying the consideration set obtain the same or better results than any other consideration set generation approach. To evaluate the research hypothesis, we use data from the public transport system in Santiago, Chile to estimate route choice models with different consideration set generation approaches: the Historical/Cohort and six approaches based on shortest paths searching: the Labeling, the Link elimination, the Link penalty, the K-shortest paths, the Simulation, and the Combination of all above approaches. We use revealed preference data built with smart card’s transaction data of three weeks representing the actual route choice behavior of passengers that travel in 258 OD pairs. For the analysis, we use two specifications for the choice model: the multinomial logit (MNL) model and the path size logit (PS) model, which accounts for the correlation of routes. With those specifications, we evaluate first the impact of different consideration set generation approaches by assessing the plausibility of the model. Then, we conclude by studying the out-of-sample prediction performance attained by each method on the fourth week of data. The results show that all PS logit models obtained a better fit than the MNL models. Focusing on PS logit models, the estimated parameters suggest that all consideration set generation approaches can represent the perception of passengers well for all attributes, except for the walking time in transfer, which is well represented only by the Historical/Cohort approach. Additional to this, the comparison of prediction accuracy across different consideration set generation approaches suggests that the Historical/Cohort approach allows estimating models with better performance in the prediction sample.

ABSTRACT. Financial advisers can make their clients’ lives much easier. Being considered experts in their domain, they can provide clients economies of scale in information acquisition. This is of particular relevance as more and more products such as credit cards, mortgages and investment products require consumers to take responsibility for their financial decisions. At the same time, studies from around the world overwhelmingly document that many consumers are not financially literate enough to make sensible decisions. Whilst financial advisers can thus provide their clients with guidance, and alleviate the anxiety often associated with financial decisions, evidence from the Australian regulator’s (ASIC) shadow shopping suggests that even though 86% of clients think they receive good financial advice, only 6% of advice is indeed good (ASIC 2012).

One possible explanation for this discrepancy between clients’ perceptions and the reality of poor quality advice may lie in smart catering strategies that advisers use. Given that the importance of making a good first impression is well known to businesses, financial advisers may use a catering strategy that consists of starting the relationship with good advice on an easy topic, thus creating a positive first impression. Once this first impression has been made and clients have formed beliefs about the quality of the adviser, confirmation bias may influence the clients’ subsequent evaluation of the adviser. More specifically, confirmation bias, which is defined as” the seeking or interpreting of evidence in ways that are partial to existing beliefs” (Nickerson 1998), will lead to any subsequent advice on difficult topics to be viewed as positive advice, thus reinforcing the positive image the adviser created during the first impression. In this paper we investigate whether such a strategy can indeed explain why clients’ perceptions of their financial adviser deviate so much from the actual quality of these advisers. More specifically, we are interested in whether consumers’ willingness to pay for additional advice from their financial adviser is a function not only of the first impressions the adviser made, but also of subsequent confirmation bias, and whether some clients are more prone to paying for bad advice than others.

To answer our research question, we use an incentivized online survey where 2,003 respondents viewed video advice from two different advisers related to four financial topics. For each topic, one adviser presented the good advice and the other presented the bad advice. Attributes of the financial advisers varied between respondents (certification/not). After the advice for each topic was given, respondents were asked “Whose advice would you be most likely to follow?” and after all 4 choices were made, respondents were asked whether they would be willing to pay $x for a one hour session with the advisors (with x varying across respondents). We chose topics based on whether they had one clear ‘right’ answer for all people, that were commonly faced around the world, and that had been studied in prior literature. This led us to topics related to management fees in index funds, retirement account consolidation, investment diversification, and debt. In addition to the advice choice and willingness to pay question, we also asked respondents questions that helped us assess their financial competence and numeracy, their knowledge and awareness of financial products, their trust in advisers as well as questions related to their demographics, personality traits and risk attitudes.

To test whether confirmation bias is indeed influencing clients’ willingness to pay for financial advisers, we test two ways that individuals can treat ambiguous information: First, consumers can form opinions “rationally”, ignore ambiguous signals (i.e. advice on hard topics), and not update their evaluation of the adviser. Second, consumers can use an iterative process of Bayesian updating with limited memory where they interpret ambiguous signals in favour of their current evaluation of the adviser (Fryer, Harms, and Jackson 2019), Whilst the first way of treating ambiguous information will lead to an evaluation of the adviser that is independent of the sequence in which the advice was given, the second way of treating ambiguous information leads clients to judge a financial advisor not only based on the objective quality of his or her advice, but also on the interaction of quality and perceived difficulty.

We use a latent mixture choice model to test which of these two ways of treating ambiguous advice consumers use. More specifically, we assume that the utility of an adviser is influenced by the belief that the adviser is good as well as by other variables. This belief depends on prior beliefs regarding the quality of the adviser as well on the (to the researcher latent) way the client treats ambiguous advice, and the sequence of advice quality and signal clarity. Using the information obtained from both the choice of advice data as well as the willingness to pay question, we are able to classify respondents into those likely to use “rational” updating versus confirmation bias updating, and link this classification to observed characteristics of the respondents. Our results show that younger, more trusting, more impulsive, less financially literate and less numerate participants are most vulnerable to paying a poor-quality adviser. We show that if an adviser chose to use observable characteristics to identify such vulnerable clients and strategically offer advice to them, the probability of these clients paying more than resilient clients is further increased. Our results thus provide important insights for policy makers who are interested in identifying and educating vulnerable individuals, and managers who want to measure the monetary impact of making a good first impression and resulting confirmation bias.

Australian Securities and Investment Commission (ASIC), 2012. Shadow Shopping Study of Retirement Advice, Report 279 (ASIC, Sydney).

Fryer, Roland G., Philipp Harms, and Matthew O. Jackson, 2019, Updating beliefs when evidence is open to interpretation: Implications for bias and polarization, Journal of the European Economic Association 17, 1470–1501.

Nickerson, Raymond S., 1998, Confirmation bias: A ubiquitous phenomenon in many guises, Review of General Psychology 2.2, 175-220.

ABSTRACT. Background: Rates of online gambling, particularly on horse racing and sport, have increased significantly over the last 15 years as new technologies have become available. This increase was particularly evident during the COVID pandemic with restrictions on in-venue gambling in many countries. Online gambling increases the risk of problem gambling relative to in-venue gambling. Understanding online gambling behaviour is important for implementing interventions to reduce the prevalence and severity of problem gambling. However, there is currently limited knowledge about what influences peoples’ choices when gambling online. Real-world gambling data are of limited use for this purpose being unlikely to allow for identification of the separate impact of different features of the online gambling market, experimental methods such as random allocation, or observation of the full set of options people are choosing between. Stated choice methods could circumvent these issues, but there are several unique challenges with undertaking a stated choice experiment in gambling. First, real-world gambling is complex and many features of a betting market may not be obviously visible to the gambler. To minimise hypothetical bias the stated choice experiment needs to capture this complexity. Second, there are multiple stages of choice when people decide what bet to place (for example in horse racing: which race, which horse, and how much to bet). These may be made sequentially as separate decisions or simultaneously as one decision. Third, in the real world people can bet on multiple events and each bet may be influenced by the result of previous bets. We attempt to address these issues while conducting a Discrete Choice Experiment in the context of online gambling on horse racing. This is a novel application of a DCE and could be a gateway for future research to further explore online gambling behaviour to inform policy change.

Method: This study asked participants to imagine they have an online gambling account, to which they have just logged in, and use it to bet on horse racing. Each participant was offered a series of six choice sets; each presented two hypothetical horse races, each with 8 horses. Participants were asked to make two choices: (i) on which race they would bet or if they would not bet on either and (ii) if they opt in, on which horse they would bet. The study had 2000 participants, randomised to one of two arms. In Arm 1 they were asked to imagine they were considering betting $10 of their own money, and in Arm 2 that they were considering betting with a $10 Bonus Bet. The amount bet was fixed to reduce cognitive burden and the sequence of choices. The race results were not revealed, and each choice set is framed as independent, to reduce the influence of previous bet results. The attributes and levels for the experimental design were features of each race, not each horse. Each race was described by five attributes: (i) time until the race starts, (ii) set of odds, and whether (iii) information on horse’s recent form, (iv) expert tips, and (v) inducements were provided. To maintain the complexity of real-world gambling, attributes and levels were presented less explicitly to participants than they might be in another DCE. We will initially model the two choices, race and horse, separately using latent class scale heterogeneity models. We will then model them together using a novel hurdle model. We will also conduct a series of Swait & Louviere poolability tests to explore whether mental health, Bonus Bets, and gambling experience affect gambling behaviour.

Anticipated results: We do not have results for this study at the time of submitting this abstract. Results of this study will provide information on how features of online gambling on horse racing influence peoples’ decisions, and how personal characteristics like mental health affect this. From this, we can infer traits about peoples’ gambling behaviour and how this may affect their likelihood of developing problem gambling. For example, a higher preference for the shortest time until the race starts may signal gambling as an escape, a lower preference for races with form and a higher preference for races with a clear favourite or an expert tip may signal that less thought is put into the bet, and a higher preference for races with an inducement may signal they are more influenced by marketing offers. By analysing the horse chosen we will also be able to explore participants’ risk preferences and how inducements affect their perception of risk. These are important traits to explore because existing literature suggests that they can influence the likelihood of developing problem gambling.

Discussion: The results of this study have important policy implications for litigation and law reform in the online gambling space. Specifically, inducements are at the forefront of online gambling regulation in Australia, so analysis of their effect on gambling behaviour, and the populations most influenced by them, will provide important evidence for this policy relevant question. Analysis of the effect of personal characteristics on gambling behaviour is important for informing policy to reduce problem gambling for high-risk groups, such as people with mental health problems, because it can help to identify the mechanisms causing the increased risk and hence interventions can be targeted more effectively. This study also has important implications for future research to explore gambling behaviour. There have been few stated choice experiments in the gambling space, but they have enormous potential to provide information not otherwise obtainable. This study could be a gateway to studies exploring different aspects of online gambling such as the effect of bet size, betting market, the result of previous bets, and the betting environment on choices, and the factors which affect peoples’ choice of betting market and bet type. This presentation will discuss the challenges and potential solutions to the use of stated choice methods in this application.

ABSTRACT. Classical portfolio choice theory is built around the notion of the risk-return tradeoff, where risk refers to the variance or second-order moment of the uncertain asset returns. A growing body of empirical findings, however, suggest that higher-order moments of the asset returns are also priced in financial markets (Chabi-Yo, 2012; Holzmeister et al., 2020). In particular, stocks with “lottery-like” features are known to be overpriced, resulting in their underperformance relative to passive investment portfolios over time (Bali et al., 2017; Eraker and Ready, 2015). Whilst many alternative measures of lottery likeness are available such as jackpot probabilities, bankruptcy probabilities, and maximum daily returns, a common thread across those measures is the recognition that stocks with more right-skewed return distributions (i.e., greater third-order moments) display more lottery-like features (An et al., 2020). The empirical asset pricing literature has attributed the overpricing of lottery-like stocks to non-standard investor preferences. For example, given an inverse-S shaped probability weighting function, as commonly assumed in applications of Prospect Theory (e.g., Barberis et al. (2016)), the attractiveness of such stocks is enhanced as investors tend to overweight small probabilities of earning extremely good returns. Plausible as such explanations may be, there is no empirical evidence as yet that directly links individual investors’ preferences to the lottery-like features of their stock holdings.

Our objective is to estimate individual investors’ risk preference structurally, and evaluate whether investors with certain preference profiles are more likely hold financial portfolios that display lottery-like features. For example, by estimating the Rank-Dependent Utility or Prospect Theory model that accounts for the effects of probability weighting and comparing the estimated probability weighting function for an individual investor to their stock holdings, we will be able to test the hypothesis that investors who overweight small probabilities of good outcomes tend to chase lottery-like stocks. To identify the relevant preference parameters, we plan to utilise data from discrete choice experiments conducted as part of the American Life Panel (ALP). The experiments prompted each individual to make choices between risky prospects with alternative payoff distributions, and enable us to estimate the random coefficients models of non-standard risk preferences along similar lines as Conte et al. (2011) and Harrison et al. (2020). We will use the results to infer the individual-level preference parameters (i.e., posterior mean coefficients for each individual) in a similar manner to Fiebig et al. (2010), and explore how the inferred aspects of investor preferences relate to the same individuals’ stock holdings, which are reported as part of the survey component of the ALP. Our findings will be among the first to shed light on the extent to which conventional wisdom about lottery-related anomalies in asset pricing is supported by individual-level choice behaviour.

References

An, Li; Wang, Huijun; Wang, Jian, and Jianfeng, Yu, “Lottery-Related Anomalies: The Role of Reference-Dependent Preferences, Management Science, 66(1), 2020, 473-501.

Bali, Turan G.; Brown, Stephen J.; Murray, Scott, and Tang, Yi, “A Lottery-Demand-Based Explanation of the Beta Anomaly,” Journal of Financial and Quantitative Analysis, 52, 2017, 2369-2397.

Barberis, Nicholas; Mukherjee, Abhiroop, and Wang, Baolian, “Prospect Theory and Stock Returns: An Empirical Test,” Review of Financial Studies, 29(11), 2016, 3068-3107.

Chabi-Yo, Fousseni, “Pricing Kernels with Stochastic Skewness and Volatility Risk,” Management Science, 58(3), 2012, 624-640.

Conte, Anna; Hey, John D., and Moffatt, Peter, “Mixture Models of Choice under Risk,” Journal of Econometrics, 162(1), 2011, 79-88.

Eraker, Bjørn, and Ready, Mark, “Do investors overpay for stocks with lottery-like payoffs? An examination of the returns of OTC stocks,” Journal of Financial Economics, 115, 2015, 486-504.

Fiebig, Denzil G.; Keane, Michael P.; Louviere, Jordan, and Wasi, Nada, “The Generalized Multinomial Logit Model: Accounting for Scale and Coefficient Heterogeneity,” Marketing Science, 29, 2010, 393-584.

Harrison, Glenn W.; Lau, Morten I., and Yoo, Hong Il, “Risk Attitudes, Sample Selection and Attrition in a Longitudinal Field Experiment,” Review of Economics and Statistics, 102(3), 2020, 552-568.

Holzmeister, Felix; Huber, Jürgen; Kirchler, Michael; Lindner, Florian; Weitzel, Utz, and Zeisberger, Stefan, “What Drives Risk Perception? A Global Survey With Financial Professionals and Lay People,” Management Science, 66(9), 2020, 3977-4002.

ABSTRACT. Please find the attached PDF file.

ABSTRACT. There is substantial interest in the land use-transportation literature on disentangling associative effects versus causal effects in the impacts of residential built environment on traveler behavior. In this direction of research, motivated by the potential of influencing individual’s travel-related choices through built environment (BE) configurations and policies, it is critical to understand whether the co-movement of BE characteristics and the travel-related variables reflects “true” causality, or is simply a spurious correlation stemming from intrinsic attitudinal factors that lead individuals to live in specific built environments and also partake in a distinctive set of activity-travel patterns. For example, in the case of residential choice, auto-ownership and travel mode preferences, individuals with a high green lifestyle propensity (GLP) may choose to locate in relatively dense regions (to reduce motorized travel), own fewer cars, and have a high inclination toward the use of non-motorized and transit modes. If this effect of green lifestyle propensity (say an unobserved factor) on dense residential living and low car ownership/non-drive alone mode use is ignored, it may incorrectly manifest itself in an exaggerated residential BE impact on traveler behavior choices. Thus, it is important to explicitly model the jointness in (that is, correlation in unobserved factors impacting) residential living choice and traveler behavior choices.

In our current effort, we contribute to this thread of land use-transportation relationship by investigating residential location effects on auto-ownership levels and rank-based travel mode preferences of individuals, within a hypothetical futuristic autonomous vehicle (AV) landscape. Rank-based preference surveys can be exploited to achieve a certain desired precision in choice model estimation with a much smaller sample size, making ranked data surveys more cost-effective than traditional first-choice surveys. Moreover, recent studies show that the rank-ordered probit (ROP) model (even if using IID kernel error terms) is far superior than the commonly used rank-ordered logit (ROL) framework for the analysis of ranked data. Conceptually speaking, the ROL model is an “impossible” structure for ranking data analysis and must be avoided. Indeed, the difference between the IID ROP and ROL models for ranking data is not the same as the difference between an IID multinomial probit model and a multinomial logit model in the context of first-choice data analysis, but substantially more dramatic. Accordingly, and, to our knowledge, for the first time in the literature, we adopt an ROP structure for travel mode choice preferences, even as we jointly model ranked modal preferences with residential choice and auto ownership choices to understand BE effects in a future AV environment. To do so in a parsimonious manner, we consider a set of stochastic latent constructs as common determinants of residential choice, auto-ownership, and modal preferences within a Generalized Heterogeneous Data Model (GHDM) framework. The latent constructs (or psycho-social factors) used in the GHDM include Green Lifestyle Propensity (GLP), Luxury Lifestyle Propensity (LLP), and Mobility Control (MC). The first latent construct relates to a general consciousness and concern about the environment; the second LLP construct is characterized by a penchant for consuming more, marked by a desire for privacy, spaciousness, and signaling exclusivity; and the final MC construct is associated with a need for control over spatial-temporal movement. The resulting GHDM is estimated using a maximum approximate composite marginal likelihood (MACML) approach, which is particularly suited to rank-ordered dependent outcomes within a high-dimensional multivariate mixed model system.

The data for this study is drawn from the 2019 multi-city Transformative Technologies in Transportation (T4) Survey for the city of Austin. The survey elicited information on place of residence, auto ownership levels, and several user characteristics/choices associated with traveler behavior, including attitudes and preferences, current travel patterns, and perceptions toward mobility-on-demand services. Additionally, through the use of a stated preference (SP) question, respondents were asked to rank, in the context of a future autonomous world, their mode choice preferences (from most preferred to least preferred) for non-work/non-mandatory trips for seven mode choice alternatives: private vehicle (human driven or autonomous), bicycle, public transport (bus/rail), human-driven private ride-hailing, human-driven pooled ride-hailing, autonomous vehicle (AV) private ride-hailing, and AV pooled ride-hailing. The SP experimental design was characterized by three trip attributes - wait time, in-vehicle travel time, and total trip cost.

The results from our study should provide important insights regarding possible neo-urbanist neighborhood design policies (for example, investments in bicycling and walking infrastructure or improvements in transit station design) aimed at encouraging a “greener” way of traveling and reduced dependence on personal automobiles. Furthermore, by evaluating the average treatment effect of residential density, we are able to quantify the contribution of “true” BE effects and “spurious” BE effects on auto-ownership and mode choice behavior. Finally, sociodemographic impacts on the main outcomes can be partitioned based on their direct effect as well as effects mediated through each of the three latent constructs, allowing policy-makers to effectively strategize campaigns aimed at specific sociodemographic groups to promote environmentally-friendly living and travel behavior.

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

Background

COVID-19 vaccines are viewed as the most effective way to end the COVID-19 pandemic and the associated public health crisis. The success of mass vaccination campaigns depends critically on the decisions of individuals to get vaccinated. Preferences for COVID-19 vaccines are likely influenced by vaccine attributes and person-specific characteristics.

Methods and data

In this study, we analyse individual preferences for COVID-19 vaccines using data from a nationwide stated choice survey (N=1421), which we conducted in the US in March 2021. The survey featured a discrete choice experiment consisting of a choice between two hypothetical COVID-19 vaccines and an opt-out alternative. The vaccines in the discrete choice experiment were described by nine attributes, namely the out-of-pocket costs, the effectiveness, the protection period, the incubation period, the risk of severe side effects, the risk of mild side effects, and the number of required doses, whether the vaccine has a booster against variants and the origin of the vaccine.

For the analysis of the stated choice data, we formulate a new normal error components mixed logit model (Walker et al., 2007) in which parts of the systematic utility are represented using the Choquet integral (Grabisch et al., 2008; Grabisch and Labreuche, 2010), a flexible aggregation operator which captures attribute interactions while ensuring interpretability and monotonicity of preferences. The Choquet integral also quantifies the relative importance of individual attributes and the degree of interaction of attributes (i.e. the Choquet integral identifies to what extent two or more attributes are independent, synergistic or redundant).

Results

Effectiveness is the most important attribute, followed by severe side effects and protection period. Mild side effects is the least important attribute, followed by out-of-pocket costs and incubation period. Effectiveness and severe side effects are significantly more important than the average, whereas mild side effects and incubation period are significantly less important than the average. These findings suggest that improving the availability of highly effective vaccine with minimal severe side effects is the comparatively most effective way to improve vaccine uptake.

The interaction analysis reveals that the non-pecuniary vaccine attributes are synergistic and should thus be well satisfied together in order to maximise vaccine attractiveness. Out-of-pocket costs are independent of effectiveness, incubation period and mild side effects but exhibit moderate synergies with the remaining attributes.

Our analysis of preferences for the opt-out alternative in the discrete choice experiment offers insights into the factors that are likely associated with vaccine (non-)adoption. We estimate that vaccine adoption is significantly more likely among individuals who identify as male, have obtained a bachelor’s degree or a higher level of education, have a high household income, support the democratic party, had COVID-19, got vaccinated against the flu in winter 2020/21 and have an underlying health condition. By contrast, individuals who belong to the Baby Boomer generation or an older generation and are black or African-American are significantly more likely to select the opt-out alternative.

Conclusion

Our analysis suggests that people’s preferences should be considered in the design of information campaigns, vaccine procurement and the development of new vaccines. For example, information campaigns aimed at improving vaccine acceptance should emphasise vaccine attributes which are perceived as most important by respondents (i.e. effectiveness, risk of severe side effects and protection period). Information campaigns should also explicitly target socio-demographic groups with a lower likelihood of vaccine adoption. In addition, our findings suggest that the likelihood of widespread vaccine adoption can be increased by improving the availability of vaccines that satisfy important attributes. Due to the synergistic interactions between vaccine attributes, the most effective way to maximise vaccine adoption is to improve the availability of vaccines that perform well across all non-pecuniary vaccine attributes. These insights should be exploited in the procurement of vaccines and the development of new vaccines.

ABSTRACT. Discrete choice experiments (DCE) assume that people have stable preferences that can be elicited by asking the right questions (Hanley & Barbier, 2009). Psychological research has long demonstrated that preferences are not stable, but rather constructed, and that they can be influenced by aspects of the environment in which a response is elicited (Slovic, 1995). Here we focus on priming, which concerns how the situational context - in a passive and unintended way - influences the accessibility of information that comes to mind, and how this, in turn, impacts how individuals’ think, feel, and behave (Bargh & Chartrand, 2000). The origins of priming research date back to the late sixties and were inspired by the spreading activation model of semantic memory according to which when a concept is activated the activation spreads to other semantically related concepts (Collins & Loftus, 1975). For example, activating the concept of nature could also activate associated concepts such as environmental protection. There are many different ways in which memory constructs can be primed, such as by having participants unscramble sentences connected to a particular theme like the environment (Verplanken & Holland, 2002), or by infusing the location where judgment takes place with a particular fragrance congruent with the target issue (e.g., Bonini et al., 2015). Here, we are interested in a much less artificial and inescapable priming manipulation - the location in which people are surveyed. A previous study examined the effect of polling location on people’s voting behavior (Berger et al., 2008). That study found that people voting in schools are more likely to fund a school initiative than people voting in churches. More pertinent to the present purposes, another study found that the building in which a behavior is observed can have an impact on it (Wu et al., 2016). For example, participants tested in a “green” building were more likely to recycle waste than participants tested in a less environmentally friendly building (Wu et al., 2013.) Motivated by this work, we examine whether location can also impact how people value environmental interventions. While priming has been investigated in many psychological areas, to the best of our knowledge our study is the first to explore the influence of contextual priming on the formation of preferences measured by DCE. Moreover, many priming manipulations used in psychological research are artificial, such as exposing interviewees to words, pictures, etc., which could make them aware of the manipulation. The priming manipulation used in this study is natural, it concerned the location where the interviewees were approached, and in a sense inescapable. Our case study concerns conservation measures of a Nature Park in the eastern Italian Prealps. Specifically, we examine whether the mere context in which a survey takes place, whether interviewees are approached in areas related or unrelated to a target issue being evaluated, may affect their choices and, consequently, their assumed preferences. The questionnaire was administrated on site face-to-face to a random sample of 819 visitors. Combinations of attribute levels were arranged into choice cards by means of a sequential D-efficient design (Ferrini & Scarpa, 2007; Bliemer et al., 2008), prepared in NGene software (ChoiceMetrics, 2012). Selected attributes were flora biodiversity, conservation of the yellow-bellied toad, development of trials for visitors and development of local organic products. To test the priming effect we selected specific points inside the nature park to approach interviewees: we interviewed people close to meadows to consider their priming effect on flora biodiversity, close to mountain puddles where the toad lives to test the contextual priming for the yellow-bellied toad, close to mountain trails to test their priming effect, and close or in huts and shelters where food is available to test priming for local organic products. In order to be able to account for both preference and scale heterogeneity, analyses were conducted by means of a Generalized Mixed Logit (GMX) model (Fiebig et al., 2010; Hensher et al., 2015). To test whether the priming effect leads to differences in observed choices, each parameter was interacted with a dummy representing being interviewed with the corresponding priming stimuli (e.g. flora biodiversity was interacted with being interviewed close to meadows, etc.). In general, respondents showed positive preferences and WTP for all attributes The GMX model produced larger WTP for trails, which was anticipated because of the tourism-oriented characteristics of the study area. Moving to the interactions for the priming treatments, the coefficients for local organic products and flora biodiversity are positive and statistically significant, indicating a priming effect for people interviewed in huts and shelters where food is available and close to meadows, while the non-significant coefficients for the other interaction terms indicate there is not a direct priming effect for protection of the yellow-bellied toad or trails. Apparently, the stimuli related to the availability of food and flowery meadows have proved to be the strongest stimulus for our respondents. The only statistically significant standard deviations for interaction terms is that of the coefficient associated with the trails treatment, suggesting limited preference heterogeneity across the sample when counting for the priming effect. Finally, our results show the presence of scale heterogeneity across individuals. WTPs that account for contextual priming turn out to be higher for local organic product and biodiversity. Governmental agencies and international institutions spend millions to survey people’s preferences. The results of such surveys are supposed to reveal people’s real preferences about the target issues, and thus provide input to litigations, political decisions and ultimately guide policy. We provide preliminary evidence that people’s responses in such surveys may be systematically swayed by the subtlest of cues: the location in which the interviewees happened to be surveyed. Nevertheless, the results of this study are far from being conclusive and we therefore encourage further studies to bring additional knowledge to this important topic. If replicated and extended, one manner in which the present results could be utilized is by keeping track of the survey location and statistically accounting for potential location effects when aggregating data across different survey locations.

ABSTRACT. Global warming is a reality of global concern that represents the greatest threat to the sustainability of life on the planet. This is exacerbated by human activity, with the carbon footprint produced by humans being 50% higher in 2021 than the corresponding level in 1990 (UNDP, 2021). Changes in human behaviour and consumption habits can contribute to reducing greenhouse gases and therefore human impact on climate change. As a result, planners must meet the growing consumption needs of the population by defining sustainability goals that require strategies to achieve the widespread adoption of a sustainable consumption pattern by the population. The individual plays a decisive role, since their actions, habits and consumption can contribute to the reduction or increase of the carbon footprint which is responsible for much of the abrupt changes in climate (Ivanova et al., 2020). Understanding human behaviour and the psychology of the individual is difficult, since it can be influenced by many factors that can be difficult to quantify, in the present context most notably the level of commitment to environmental beliefs. Previous studies have sought to understand pro-environmental preferences (and their role in individual decision making) on the basis of observable characteristics of the individual (e.g., gender, age), less easily quantifiable components such as beliefs and attitudes (Klöckner, 2013) and finally characteristics of goods and services, including the effect of incentives and advertising suggestions (Donmez-Turan & Kiliclar, 2021). However, a factor that has not been widely studied in the choice modelling literature is that preferences, as well as changes in consumption habits, may be motivated or influenced by past behaviour of the individual. In psychology, changes in environmental behaviour motivated by other behaviour by the same actor are known as pro-environmental behaviour spillover – PBS (Truelove et al., 2014). There is not yet a consensus about the factors that drive PBS, but it has been associated amongst others with individuals’ values (Thøgersen & Ölander, 2003), moral norms (Thøgersen, 2004), the difficulty or inconvenience of behaviour (Juhl et al., 2017), perceived social norms (Sintov et al., 2019), and cognitive dissonance (Thøgersen, 2004), among others (Truelove et al., 2014). In the present paper, we argue that a failure to consider the role of PBS may give a limited picture of human behaviour in the environmental domain and could therefore cause bias due to omitted variables (Antonakis et al., 2010). At the same time, we identify a major limitation in that most previous studies of PBS have not sufficiently addressed the role of individual heterogeneity in driving spillover effects. In other words, we draw on the spillover literature to inform existing choice modelling approaches, and at the same time we draw on the choice modelling literature to improve our understanding of behavioural spillover effects and PBS in particular. We develop an econometric approach to address both the lack of attention to PBS in past choice modelling studies, and the absence of treatment of heterogeneity in PBS work in general. We design a survey instrument and apply it to an online sample (n = 300). We use two related experiments. The first experiment (A) presents individuals with hypothetical choice tasks looking at different green actions that require varying levels of commitment/effort (e.g., reduced use of motor vehicles, changes in diet) and lead to different levels of reduction in the carbon footprint. Two versions of experiment A are used, with A1 showing individuals the impact of these actions only in terms of CO2 reductions, while A2 explains the reductions also in terms of the equivalent number of trees that would need to be planted to achieve the same CO2 reductions. The second experiment (B) focuses on choices between different donation amounts for the planting of trees to offset the person’s carbon footprint by different amounts. The order of the three experiments (A1, A2, B) is varied across individuals. Econometric methods belonging to the family of discrete choice models, and specifically those based on the theory of random utility, will help us disentangle the different influences on choice. However, in most common settings, the basis for these models is mainly that the probability of choosing a specific option among mutually exclusive alternatives increases in the presence of desirable characteristics and decreases in the presence of undesirable characteristics. In our work, we seek to go further by testing how changes in behaviour may be due to the additional presence of heterogeneity of individuals driven by their level of environmental commitment. An important point to note is that if an individual chooses to make a large donation after previously choosing to significantly reduce CO2 emissions by adopting a vegan diet, then this by itself does not imply a causal link but could simply be a reflection of strong pro-environmental preferences. A key feature of our work lies in disentangling spillover effects from heterogeneity in preferences, which is made possible by the experimental setup with variations in the order of experiments across individuals. The econometric starting point for our analysis is a hybrid modelling framework that combines random utility theory with structural equation models and incorporates heterogeneity in attitudes and sensitivities. The model is used to determine the strength of the underlying attitudes of individuals and understand their role in the choices the individuals make. In addition, the model incorporates lagged effects to study the impact on choices both of previous decisions and the exposure to earlier scenarios, thus measuring any spillover effects, where special care will be taken to disentangle causality from heterogeneity in preferences, allowing us to test the influence of the characteristics and the order of the situations participants face on the response to quantitative environmental attributes in a related action, i.e., donations to plant trees and reduce the carbon footprint. The model makes use of additional indicators in the form of questionnaire items relating to psychological attitudes to calibrate the role of underlying attitudes, which are then also used to explain the heterogeneity in preferences, the degree of influence of information, and the susceptibility to spillover effects.

ABSTRACT. In face of the environmental crisis, investments taking environmental, social, and/or governance (ESG) criteria into account, so-called sustainable investments, are necessary to reach the goals of the Paris Agreement. Investments by institutional investors and governments alone are, however, not sufficient to reach these goals. The aim is thus to mobilize individual investors to invest sustainably, as high information costs and little knowledge about sustainable investments still represent barriers inhibiting investments (e.g. Gutsche and Zwergel, 2020). Yet, it is not clear whether pecuniary or non-pecuniary information are more suitable to decrease information costs and lead to a behavior adaption of individual investors (e.g. Riedl and Smeets, 2017). We therefore use a novel choice modelling application combined with the approach as in Glac (2009) and Døskeland and Pedersen (2016) to examine the causal effect of different (pecuniary vs. non-pecuniary) information frames on individual investment decisions with real world investment products. Specifically, we look at a) financial information, b) information about investor impact, and c) information about the investment behavior of other individual investors. Using these frames, this study aims to empirically analyze I.) which different information (frames) affect sustainable investment at the individual level and II.) which investor types (in terms of financial literacy, environmental awareness, and trust) react to which type of information.

This study is based on a representative online survey including an incentivized investment experiment among about 1600 households’ financial decision makers in Germany conducted between May and August 2021. After a short description of the experiment, respondents were randomly assigned to one of four different groups, namely the a) control group, b) financial information group, c) impact information group, and d) social norm group. The control group obtained no information prior to the investment decision. The financial information group received information that sustainable investments can perform superior compared to conventional investments (Friede et al., 2015). Respondents in the impact information group were given the information that by considering ESG criteria, investors can encourage companies to act more sustainably (Kölbel et al., 2020). Finally, respondents in the social norm information group received the information that investors often consider sustainability criteria when investing (Gutsche, 2019). Afterwards, all respondents are endowed with 500 Euros and choose six times among four real funds that are traded on the market. In an unlabeled design, they see information on the fees, the strength of sustainability, returns in the past two years, and the share of issuers of bonds from the European Union which correspond to the actual values for the 16 funds. They also learn that, for 10 randomly selected respondents one chosen fund is randomly selected and bought, and that the payoff is the value of the fund minus fees after a one year holding period.

This setting is analyzed using mixed logit models. We find a positive willingness to pay (WTP) for sustainability over all four groups. Individuals in the financial information and in the impact information group have a significantly higher WTP for sustainability compared to the control group, but not the social norm information group. We also analyze heterogeneous treatment effects using split samples, and find that financial literacy, as well as trust in scientists impact the responsiveness to certain frames. Respondents with moderate to low financial literacy have a higher WTP for impact and social norm information, whereas respondents that have low trust in scientists have a higher WTP for all three information. Environmental awareness does not impact responsiveness.

This study makes three key contributions. First, we disentangle the relevance of pecuniary and non-pecuniary motives for (sustainable) investment decisions of individuals in a framed field experiment. We extend previous studies that were either based on student samples (Glac, 2009) or did not include a control group (Døskeland and Pedersen, 2016). Second, similar to, for example, Allcott (2011) in the energy consumption context, we examine the causal effect of social norm information on (sustainable) investment decisions. Third, our study is the first to give a broader overview over which type of investor is most responsive to certain information. Døskeland and Pedersen (2021) show that wealth affects responsiveness towards different information. We are adding on to their work by further including financial literacy, environmental awareness, and trust in our analysis. The results of this study can be used to advise policy makers and practitioners on the kind of information they should use to inform individual investors to increase sustainable investments. Furthermore, our results can be used to increase reactions towards sustainable investments by targeting specific groups of investors with certain information.

References Allcott, H., 2011. Social norms and energy conservation. Journal of Public Economics, 95 (9-10), pp. 1082-1095. Døskeland, T. and Pedersen, L. J. T., 2016. Investing with Brain or Heart? A Field Experiment on Responsible Investment. Management Science, 62(6), pp. 1632-1644. Døskeland, T. and Pedersen, L. J. T., 2021. Does Wealth Matter for Responsible Investment? Experimental Evidence on the Weighing of Financial and Moral Arguments. Business & Society, 60(3), pp. 1-34. Friede, G., Busch, T., and Bassen, A., 2015. ESG and financial performance: aggregated evidence from more than 2000 empirical studies. Journal of Sustainable Finance & Investment, 5(4), pp. 210-233. Glac, K., 2009. Understanding Socially Responsible Investing: The Effect of Decision Frames and Trade-off Options. Journal of Business Ethics, 87, pp. 41-55. Gutsche, G., 2019. Individual and Regional Christian Religion and the Consideration of Sustainable Criteria in Consumption and Investment Decisions: An Exploratory Econometric Analysis. Journal of Business Ethics, 157, pp. 1155-1182. Gutsche, G. and Zwergel, B., 2020. Investment Barriers and Labeling Schemes for Socially Responsible Investments. Schmalenbach Business Review, 72, pp. 111-157. Kölbel, J. F., Heeb, F., Paetzold, F., and Busch, T., 2020. Can Sustainable Investing Save the World? Reviewing the Mechanisms of Investor Impact. Organization & Environment, 33(4), pp. 554-574. Riedl, A. and Smeets, P., 2017. Why Do Investors Hold Socially Responsible Mutual Funds?. Journal of Finance, 72(6), pp. 2505-2550.

ABSTRACT. Design and implementation of effective health policies require a proper understanding of public preferences involving individual and collective interests as well as relative concerns. Knowledge about such often competing and overlapping interests is imperative to leverage healthcare resources which are usually limited in supply relative to respective demand. Consider, for example, the case of treatment waiting times allocation decision for a particular medical condition subject to healthcare capacity constraints. Such a decision is mostly informed by analysis that assumes that the decision of every individual (patient) is driven solely by rational self-interest, i.e., the individual’s own waiting times. Despite its popular use to understand patients’ preferences also in several other cases, two strands of the literature on individual decision-making behavior challenge this conventional assumption of homo economicus.

One body of this literature emphasizes people’s relative (or positional) concerns. The central idea in this branch of the literature is that, people derive satisfaction (experience disutility) merely from having more (less) access to resources than relevant others. With reference to the example of waiting times, this implies that just having a shorter (longer) waiting time than others results in a utility (disutility) for some individuals. Another strand of the literature relates to the idea that some decision makers have concerns over the distribution of pay-offs across (relevant section of) the society. The latter variant of the literature explains the relative importance of fairness considerations as well as inequality aversion and maximin preferences. For example, in relation to waiting times, some individuals’ choice of waiting time allocation can be guided by the shape of the distribution of waiting times across people. The difference between positional concerns and distributional considerations is that the latter is independent of the decision maker’s own position in the distribution (i.e., the decision maker’s choice is made from behind a metaphoric ‘veil of ignorance’).

This study, to our knowledge, is the first to jointly recognize and empirically explore the relative importance of self-interest (SI), positional concerns (PC) and distributional considerations (DC), in individual decision-making related to access to healthcare services. We contribute to the literature on individual decision making in general and on patient preference in particular in three ways.

First, using a variety of advanced choice models, we provide a rigorous empirical investigation into the relative importance of SI, PC and DC in decision making related to medical treatment waiting times. In connection with PCs, we examine potential asymmetry between a PC due to disutility from disadvantageous inequality (i.e., having longer waiting time relative to others) and utility from advantageous inequality (i.e., having shorter waiting time of equal magnitude relative to others). In addition, we identify several distributional principles namely expected value, minimax principle, Shannon's entropy and generalized entropy (GE) inequality index to capture various types of DC with appropriate metrics. Furthermore, we distinguish between decision makers’ ideal distributional preferences based on choices made behind a “veil of ignorance” and genuine distributional preferences that result from choices made when decision makers know their position in the distribution of waiting times.

Our second contribution relates to the comparison we make concerning the relative importance of the three respective motivations, between decision makers choosing for themselves and decision makers choosing for their most loved ones. By doing so, we provide valuable insights for framing of interventions by comparing the weights decision makers attach to the different motivations when they choose for themselves and when they choose for their relative or friend.

The third contribution lies in carrying out our empirical investigations in two countries with very different healthcare systems, namely the US and the UK. We compare the strengths of SI, PC and DC in the context of preferences for waiting time allocations, between citizens of the US and the UK. By doing so, we establish the robustness of our findings not only to differences in healthcare systems but also to potential cultural discrepancies and political orientations.

The empirical investigation in this study is based on an extensive and unique tailor-made stated choice experiment conducted in the US and the UK. Our experiment was designed in such a way that the three types of motivations could be disentangled and identified in an econometrically efficient manner, from choices made between healthcare policies resulting in different distributions of waiting times across society, and different positions of a decision-maker on those distributions. Across the two countries, the sample was split equally, and roughly half of the respondents in each country made choices for themselves while the other half indicated their choice for their close relative or friend. Our analysis involves estimation of a series of discrete choice models on data from each sub-sample. We find that SI, PC and DC are all significant in explaining respondents’ choices in both countries, the weight for DC being larger than the weights for the former two. Our results indicate no evidence of asymmetry between advantageous and disadvantageous inequality. Of all the considered distributional principles to represent DC, Expected value appears to be the most suitable metric. We did not find significant differences in relative importance of the different motivations between choices made for loved ones and choices for oneself in the UK. However, US respondents who chose for their relative or friend attach significantly larger weight to the (absolute) levels as well as to distributions of waiting times (i.e., DC) than respondents who chose for themselves. Moreover, among those who made choices for themselves, we find that respondents in the UK attach a relatively larger weight to SI and DC than those from the US while respondents from the latter country display stronger positional concerns than those from the former. Also, UK respondents who chose for their relative or friend give a larger weight for DC than their US counterparts. We discuss the relevance of these findings for health policy.

ABSTRACT. Data is one of the key resources of the modern economy. In exchange for access to their services, digital service providers collect personal data about the users and their online behaviour, which can help tailor the services to the users and increase the revenue from advertising, among others. The user receives personalised service and pays for it, often unawarely, with personal data. A different approach may be considered to increase online privacy protection where users could have more control over their data. The question is if it pays off for the users—is the value of enhanced online privacy worth bearing its cost? A possible means to measure value of digital goods and services that are not subject to monetary market transactions, which could approximate their values, is to apply non-market valuation techniques, such as stated preference methods (e.g., Brynjolfsson et al., 2019; Corrigan et al., 2018). Some inquiries particularly attempt to value the privacy and personal data shared on the Internet (e.g., Winegar and Sustein, 2019). Here, we present results from a discrete choice experiment study conducted to estimate the value of increased privacy protection in online platforms. The valuation scenario considers reductions in personal data and information about users’ behaviour on the Internet collected by digital service providers, such as Facebook and Google. This is a novel approach to apply a discrete choice experiment to assess the value of protecting the access of online platforms to personal data. Our findings deliver potential guidance to policymakers and platform operators on people’s preferences toward different levels of privacy protection on the Internet.

In the discrete choice experiment study, we focus on two large platform providers: Facebook and Google. The valuation scenario proposes a new tool to allow users to have more control over the service provider’s access to their data in return for a monthly fee. We treat various attributes of privacy separately to measure preferences toward different levels of personal data protection. For each of the two platforms, we consider three attributes: data access range, profiling, and advertising. The first one distinguishes between several scopes of data on users’ online behaviour available to the service provider—precisely, the platform can track only the data actively given by its user (e.g., uploaded photos) versus any data, including that passively shared (such as a contact list or localisation). The attribute on profiling captures preferences toward being profiled—whether the platform provider can create a user’s profile (i.e., classify users based on political or other views, sexual orientation, health, etc.). The third attribute concerns application of the collected data for advertising. It helps assess respondents’ preferences toward personalised and non-personalised advertisement versus no advertisement at all. The final attribute is a monthly cost to the platform user.

The survey was administered as computer-assisted web interviews in September 2021 to a representative sample of 944 residents of Poland. Every respondent was presented with 6 choice tasks for each of the two platforms (Facebook and Google). Each task involved two alternatives tied to increased privacy protection and a status quo, representing a default privacy protection level at a zero cost (this option assumed creating a profile, personalised advertising, and a maximum scope of shared data, which reflects common practices online). Hypothesising strong preference heterogeneity, we employ mixed (random-parameter) logit (MXL) and latent class (LC) models to understand respondents’ preferences toward online privacy.

The MXL results indicate that respondents experience, on average, negative utility from: the current status quo, a large scope of tracked behavioural data, profiling, and advertising. Mean willingness to pay (WTP) for limiting the tracking of users’ behaviour is about 3 EUR monthly for strongly improved privacy (no access to users’ behavioural data) for Google and 3.7 EUR for Facebook. The WTP for partially improved privacy (the platform can access only the data actively provided by a user on the platform) is 1.5 EUR monthly for Google and 2.7 EUR for Facebook. The marginal WTP for not being profiled is lower but still statistically significant (0.4 EUR for Google, 0.8 EUR for Facebook). Finally, we observe substantial heterogeneity in preferences toward advertising. In general, respondents’ utility is negatively affected by advertisement, and they are willing to pay for removing advertisement entirely from the platforms. The results suggest when advertising is in place, the preference is slightly toward personalised compared to non-personalised advertising. Summing up, public benefits from the largest improvement in online privacy protection in Poland, as described in the evaluated scenario, amount to approximately 80 M EUR monthly for each platform, assuming that almost all of the 27.7 M Internet users in Poland use Google browser and that Facebook has 21.5 M members in Poland (data taken from statistics for Poland).

Results of the LC model indicate that two groups can be distinguished with substantially different preferences. One class of the size of 57% has ambivalent preferences toward privacy, with most WTP values being insignificantly different from zero or at a very low level, and the other class (constituting 43% of the population) with strongly negative preferences toward little online privacy protection.

Our research demonstrates that Internet users value their online privacy and are willing to pay non-negligible amounts to reduce the use of their personal data by online service providers. Moreover, we show how stated preferences methods can be applied to evaluate public preferences toward privacy on the Internet. The results may provide important input to a public debate on restricting personal data usage or help revise business models of Internet platforms operators.

Brynjolfsson, E., Collis, A., Eggers, F. (2019). Using massive online choice experiments to measure changes in well-being. Proceedings of the National Academy of Sciences of the United States of America, 116(15):7250-7255. Corrigan, J.R., Alhabash, S., Rousu, M., Cash, S.B. (2018). How much is social media worth? Estimating the value of Facebook by paying users to stop using it. PLoS ONE, 13(12):e0207101. Winegar, A.G., Sunstein, C.R. (2019). How Much Is Data Privacy Worth? A Preliminary Investigation. Journal of Consumer Policy, 42(3):425-440.

ABSTRACT. Short abstract Healthcare is under immense pressure due to rising expenditures, an ageing population, and high prices of new medical treatment options. As health resources are scarce, policymakers are often confronted with moral dilemmas in their (public) health decision-making. Understanding moral preferences and choice behaviours of citizens, patients, and medical professionals enables better decisions in healthcare by avoiding sub-optimal policy decisions and trial-and-error implementations. However, current choice models in healthcare (and beyond) fall short to unravel and understand moral preferences and choice behaviours. This paper – containing a detailed comprehensive literature review – proposes a research agenda towards integrating machine learning (data-driven) into choice modelling (theory-driven) methods in the context of moral decision-making in healthcare. We argue that once this research agenda has been addressed, this approach leads to more accurate choice predictions and behaviourally realistic models than extant models permit, and so makes an important step for using data-driven choice models in health policymaking and beyond.

Extended abstract Moral choices are part of everyday decision-making in public health and healthcare. When considering scarce health resources, policy decisions can have as potential consequences that one patient gains more life-years while another patient does not. Think of situations where one chooses to treat a critically ill patient over a severely ill one, to prioritize younger patients over older patients, or to withhold or withdraw treatment from a cost-effectiveness perspective. Such diabolical dilemmas have especially come to the forefront in the current Covid-19-pandemic, which has urged morally salient (public) health decisions to be made. To build effective and morally acceptable policies supported by different stakeholders, it is key for health policymakers to understand the moral preferences and choice behaviours of citizens, patients, and medical professionals.

Discrete choice modelling (DCM), rooted in micro-econometrics and the behavioural sciences, is an increasingly and widely used method for understanding choice behaviour and developing behaviour-informed policies in healthcare and beyond. However, standard DCMs (e.g., linear additive logit) are inadequate to apply in morally sensitive decision contexts. These conventional models, for instance, deem agents (e.g., citizens, patients, medical professionals) to be selfish and rational in their decision-making, by assuming compensatory choice behaviour. It is also assumed that agents know their own preferences. While this may be true for some agents, others use intuitive reasoning where they attempt to rationalize their decision-making but might fail due to strong moral reactions. Specifically, moral decisions are often based on social heuristics and emotions rather than the type of trade-offs witnessed in everyday (consumer) decision-making. Moreover, agents may be highly uncertain about their moral preferences (“what should I do?”) in morally sensitive situations they have not encountered before. This raises the question of whether integrating data-driven techniques (i.e., machine learning methods) into choice modelling (theory-driven) methods might be a solution for understanding moral preferences and choice behaviours, as data-driven techniques can capture the data-generating process more accurately.

The aim of this paper is twofold. Firstly, to obtain an overview of current machine learning and choice modelling practices in the context of moral decision-making. Secondly, to generate a research agenda that reflects the research directions for an integrated paradigm of choice modelling and machine learning methods in the context of moral decision-making in healthcare.

To achieve these objectives, a comprehensive detailed review was conducted of articles published between 1950 and 2021 concerning machine learning practice and choice modelling in the context of moral decision-making. Two reference datasets were generated: (i) core dataset, which was gathered by following a systematic search strategy based on PRISMA guidelines, and (ii) supplementary dataset, which follows a scoping search strategy. While the supplementary dataset contains articles that were come across while screening reference lists and considered influential, the core dataset used four databases: PubMed, Scoping, Web of Science, and ArXiv. After removing all duplicates and articles without abstracts and/or identifiers, 4,539 unique articles were collected based on keyword searches. After screening titles and abstracts, 732 unique papers were read full-text. Two-hundred ninety-five articles – in both the core and supplementary datasets combined – met the inclusion criteria and were subject to data extraction and analysis. The outcomes of the comprehensive detailed review were translated into a research agenda that highlights how choice modelling and machine learning methods can be integrated to optimally inform health policy decisions and beyond.

The preliminary results showed that there is a growing interest in studying how choice modelling (theory-driven) and machine learning (data-driven) methods relate to each other. It was found that integrating choice modelling and machine learning (ML) methods has indeed the potential to be helpful to optimally inform (health) policymaking when moral dilemmas play a role. However, there are some theoretical and methodological challenges that must be overcome in order to achieve data-driven choice models for moral choice analysis. There exist persistent misconceptions that ML methods can only be used for prediction, a correlational concept, rather than behavioural inference. This view can occur given the fact that policy analysis (i.e., by using conventional choice models) demands to answer questions related to causation, which requires data to be combined with theory-driven hypotheses. However, choice modellers may not know a priori the underlying decision-making processes. Using the wrong model specification potentially biases the estimated preference weights by not accounting for bounded rational processes; i.e., adverse effects caused by model misspecifications. Data-driven methods – containing algorithmic techniques from subfields like deep learning, computer vision, and natural language processing - can overcome challenges of theory-driven models in the search for the optimal model specification. Therefore, it is argued that integrating choice modelling and ML methods into a combined approach – while considering the moral aspects of choice behaviour – has the potential to enhance the predictive power and behavioural realism of the models. Even though the comprehensive detailed review examines the potential of data-driven choice models for moral choice analysis, it should be noted that the results can be used outside the scope of moral decision-making as well. Thus, this paper makes an important step for using data-driven choice models in health policymaking and beyond.

ABSTRACT. 1.Introduction

The multinomial logit(MNL) model is widely used in discrete choice modelling, despite the rigid independently and identically distributed (IID) assumption of the error terms, that leads to the restrictive independence of irrelevant alternatives(IIA) property. The IIA property creates limitations in modelling when the error terms of different alternatives are correlated. McFadden showed that the family of nested logit models which relies on multivariate extreme value(MEV) distributions, relax the IIA assumption partially. In 2001, Wen and Koppelman introduced Generalized Nested Logit(GNL) models with higher degree of flexibility in substitution or cross-elasticity between pairs of alternatives. This paper studies the specification of the error portion of utility function in generalized nested logit models. We show that all MEV models can be re-written as GNL models for more flexibility while maintaining or improving model performance. Examples are multinomial logit(MNL), nested logit(NL), paired combinatorial logit(PCL) and then to extend the PCL model, we introduce a paired GNL(PGNL) model, where all alternatives in the GNL generating function are paired in nests. Later, we explain that to maintain flexibility and introduce exchangeability, one identical nest in PGNL is added to the generating function.

2.Methodology

The methodology comprises two parts. The first part explains how to write MEV models as GNL models. In this part we also introduce the PGNL model that is more flexible than PCL model. In the second part we explain how to introduce exchangeability to GNL models.

2.1.MEV models as GNL models

McFadden’s family of nested logit models allows for more flexibility in substitution pattern of alternatives in a logit model by placing alternatives that are correlated in error term in the same nest. These models are based on the multivariate extreme value (MEV) distributions and defined by a generating function that is a non-negative and homogeneous of-degree-one function of where. The probability function for choosing alternativei is then given by. introduced the GNL model that is a MEV model derived from the function

where Nm is the set of all alternatives in nest m, αn0m the allocation parameter that measures the portion of alternative n0 assigned to the nest m, µm is the scale or dissimilarity parameter for nest m and Yn0 characterizes the value for each alternative n0.

We show that any MEV model can be written as a GNL model as some examples are provided in. We introduce PGNL model that includes properties of PCL and GNL models combined. In PGNL all alternatives are paired in nests while including the allocation parameters. The alternatives with IID error terms in a nest will have.

2.2.Exchangeability

According to exchangeability theorem, exchangeability exists in a sequence of IID variables, if the joint probability distribution of the sequence is invariant under re-ordering of all permutations of each variable. In a MEV generating function with Jalternatives, iandj are exchangeable if.

To add more flexibility to the GNL model, it is possible to add an identical nest of exchangeable alternatives in the generating function. The scale parameter and the allocation parameter are the determinants of exchangeability in GNL models. If alternatives iandj are paired with similar differing alternativek in nests and, the scale and allocation parameters for iandj must be the same for these alternatives to be exchangeable.

3.Application and Results

The application is focused on introducing exchangeability to GNL models with a mode choice case. We use the data from Swedish national survey 1995-1998 with 5761 observations to implement the models in. Six modes are included in the modeling: car, car passenger, bus, train, walk and bike. The first model is a simple MNL to make grounds for comparison. The variables are travel time and cost for all modes. The second model is a NL model with two nests of car and car passenger, and bus and train. The nests are formed based on an intuitive expectation of the data structure. As shown in, in MEV models the alternatives with IID error terms are exchangeable. This translates to same values of the scale parametersλ and allocation parametersα for the exchangeable alternatives in nests. In an attempt to investigate exchangeability in the PCL model, fifteen nests for all combinations of pairs of alternatives in total are considered with a block diagonal weight matrix for exchangeable modes and another block for all other combinations with the weight value, 0.15 which is an arbitrary small number for computational reasons. The block diagonal structure of the nests suggests that car and bike have IID error terms, in addition to the slight improvement of the log-likelihood which suggests that car and bike are exchangeable in this case. In the PGNL model, which is formed by introducing the allocation parameters to PCL model, the allocation parameters for the nest with alternatives that are exchangeable in the PCL model are estimated and the allocation parameters for alternatives in all other nests are set to 1 to maintain symmetry in the nesting structure. The result of PGNL model shows that log-likelihood of PCL is further improved by introducing the allocation parameters, while it resulted in different allocation parameters(α) suggesting that car and bike are not exchangeable if PCL is written as PGNL. In cases where it is preferred to include exchangeability, it is possible to add an identical nest to the generating function of the PGNL where the exchangeable alternatives have fixed α values (here).

4.Concluding Comments

In this paper we show that any MEV model can be written as GNL. We introduce PGNL model that includes the properties of PCL model while is more flexible as it includes the allocation parameter from the GNL model. We explain how to introduce exchangeability in the GNL models to increase flexibility while maintaining model performance. This is empirically similar to a conjoint analysis problem when several alternatives are identical except for one unlabeled or exchangeable attribute. One example of a conjoint analysis problem with exchangeable attributes is selection of higher education courses where reputation for academic quality of the course is exchangeable and course content and course location are non-exchangeable.

ABSTRACT. Motivation:

Before and during the Covid-19 pandemic, public transportation agencies have begun implementing on-demand, van-based microtransit services to reduce operating costs in areas with low levels of fixed-route bus ridership. To identify service design improvements for public microtransit, agencies need to understand the factors underlying riders’ choice to use microtransit relative to other modes, as well as how these factors have shifted in the pandemic. These agencies also need to evaluate the societal impacts of these services, including changes in vehicle-miles-traveled (VMT), travel/wait times for riders, and the equity implications of how travel/wait time impacts vary by demographic group. Our approach to evaluating these impacts is unique in the breadth of its dataset, with stated-preference survey data from microtransit riders at 14 public transportation agencies across the United States, as well as the methodology, including the first mode choice model to incorporate public microtransit and the application of a new approach from Aboutaleb, Danaf, Xie, and Ben-Akiva (2021) to identify the block diagonal covariance matrix best supported by the data—a critical step to inform substitution to/away from microtransit when estimating impacts of the services.

Data and analysis methods:

From April to July 2021, we surveyed 700 frequent microtransit riders and 386 infrequent or non-riders across 14 public transportation agencies to capture their stated preferences in a total of 3042 pre-pandemic trip scenarios and 2803 pandemic trip scenarios. The microtransit services encompass variation in service area size, population/job density, income and racial demographics, service design (cost, hours, number of vehicles, and on-demand algorithms provided by three different private companies), Covid-19 conditions (cases and vaccination rate) and Covid-19 response (increasing, maintaining, and decreasing service). From the survey data, we estimated a latent-class logit mixture model; it is the first mode choice model to include public microtransit. The model also includes ride-alone Uber/Lyft, pooled Uber/Lyft, personal vehicle, fixed-route bus, and first-mile/last-mile public microtransit and Uber/Lyft as fixed-route transit connections.

Because the covariance matrix determines substitution between microtransit and the other modes when applying the model to forecast impacts or estimate replaced modes, we needed to identify the covariance matrix structure best supported by the survey data. A restricted diagonal covariance matrix could ignore statistically significant correlations between the model’s coefficients, while a full covariance matrix may not yield the best fit when adjusted for the number of parameters. A block diagonal matrix used in nested or cross-nested models also imposes restrictions that may miss statistically significant correlations or may not yield the best adjusted fit. Rather, we use a new method from Aboutaleb et al. (2021) to search across potential combinations of correlated coefficients to determine the covariance matrix that is best supported by the survey data based on log-likelihood and adjusted for the number of parameters. Aboutaleb et al. (2021) first apply a mixed-integer optimization program to Markov Chain Monte Carlo posterior draws from the full covariance matrix to identify an optimal block diagonal structure at each sparsity level and then determine the optimal sparsity level through out-of-sample validation.

In the first application of this novel method to a complex choice model with numerous modes and explanatory variables, we are currently using their approach on our latent-class logit mixture model to inform the removal of the microtransit choice from the model. We will then: 1) apply the mode choice model with and without the microtransit option to estimate replaced modes for all microtransit trips taken at the participating agencies from January 2020 to July 2021, and 2) use the adjusted mode choice model to forecast mode shares with and without the microtransit option under a range of Covid-19 scenarios and microtransit service designs. We will then use origin/destination, time of day, and transportation network information to calculate VMT and travel/wait times for the projected mode splits with and without the microtransit service. Equity impacts stem from analyzing how travel/wait times and costs shift by demographic group.

Obtained or anticipated results:

Findings from the base mode choice model (before estimating the sparse covariance matrix) include: 1) wait times for public microtransit and the Uber/Lyft modes are perceived less negatively than wait times for the modes involving fixed-route bus service; 2) reductions in 30-minute delays of public microtransit would benefit frequent microtransit users, who have less access to personal vehicles, whereas reductions in 15-minute delays may attract new, higher-income riders; 3) aversion to fixed-route bus service emerges at lower levels of Covid-19 spread than aversion to public microtransit; and 4) trip cost and walk times have less of an impact on modal choice during the Covid-19 pandemic.

Our anticipated results include lessons from the sparse covariance matrix about correlation between preferences for the given modes and attributes. We will also estimate VMT, travel/wait time, and equity impacts from 14 microtransit services in the U.S. before and during the pandemic. Finally, we will use microtransit trip data and mode share data to calibrate the mode choice models to real-world contexts and then apply the models to forecast mode share shifts from changes in service or in Covid-19 spread and vaccination conditions.

Application potential and policy relevance:

The results provide guidance to public transportation agencies on how to tailor microtransit services to increase ridership while reducing single-occupancy vehicle use and supporting fixed-route transit services in a socially equitable way. This includes identifying potential ridership gains from changes in service area size, hours, travel/wait times, and cost, as well as how to distribute resources between fixed-route bus and on-demand microtransit service based on VMT impacts, population/job density, and riders’ preferences under a range of Covid-19 conditions. The results also suggest which service improvements to prioritize to benefit specific demographic groups, such as low-income riders. Public transportation agencies can incorporate our mode choice model into their full activity-based demand model framework for a more complete understanding of impacts including induced demand and other behavioral shifts. Finally, our methodology can also inform application of complex covariance matrices in non-transportation settings.

References:

Aboutaleb, Y.M., et al. "Sparse covariance estimation in logit mixture models." The Econometrics Journal 24.3 (2021): 377-398.

ABSTRACT. The decoy effect is a discrete choice behavioral phenomenon that consists in the addition of a third alternative to favor one of the preexistent options over the other (Huber et al., 1982). The alternative added is known as the decoy, the one favored by the decoy is known as the target, and the third alternative is known as the competitor. In the case of the asymmetrically dominated decoys, studied in this research, the decoy is worse in all features compared to the target, but not by the competitor, what somehow facilities the choice by making the target relatively more appealing. This “irrational” behavior implies a violation of the regularity assumption, and thus cannot be modeled with traditional Random Utility Models (RUM) like the logit but has been found to be reproducible by prospect theory models like the Random Regret Minimization Models (RRM) (Guevara and Fukushi, 2016). This phenomenon has been found in different contexts, from presidential elections to animal behavior, but, to the best of our knowledge, it has never been studied in public transportation. Due to the importance of public transportation in the sustainable developments of cities, and the difference between users and systems equilibrium on transportation systems, it is important to study the decoy effect in this context as a potential way to influence users’ behavior. This article has three main objectives. The first is to study if the decoy effect can influence the choices of public transportation options, specifically for bus and train. The second corresponds to study, for the first time, to the best of our knowledge, if the distance between the competitor and the target has an impact on the size of the decoy effect. The third is to contribute to the literature on the susceptibility and valuation of the decoy effect using discrete choice models. To pursue these goals, we developed two stated preferences surveys. Both surveys inquired about a hypothetical trip from the Chilean cities of Santiago (the capital) and Chillán (located some 400 Km south). In the first survey, answered by 224 adults, the profiles considered trips by bus or train and the latter alternative was defined as the target. In the second survey, answered by 212 adults, all the profiles considered only trips by bus. Reference travel time for this trip is around 5 hours, and the cost of the ticket by bus or train is about US$14 (11,000 CLP). To isolate the impact of the decoy effect, the interviewees were randomly distributed into two groups. The competitor and target´s attributes stay fixed in both groups, but the decoy alternative was only added to one of the groups. Under this experimental setting, given the random assignment, the decoy acted as a treatment which effect can then be measured as any significant change between both groups. Besides, to study the impact that the distance between the target and the competitor may have in the decoy, half of the profiles of the second survey (the one with bus only) considered a systematic variation of that feature. The distance variation was implemented as a combination of differences in travel time and travel cost, ranging from 20 to 60 min for the former, and the equivalent cost for a representative given value of time. Three main results are derived from this study. The first is that the decoy effect can influence public transportation's demand, at least under the hypothetical scenarios presented in these experiments. For all the various cases analyzed the share of the target increased when the decoy was present and in 43% of the cases that difference surpassed a 5% critical value for a Chi-squared test. The second result achieved was the confirmation of the hypothesis that the distance between the competitor and the target has an impact on the decoy effect. When the distance increased, the impact of the decoy effect also increased. The intuition for this result is that, as the target gets further away from the competitor, the relative impact of a decoy (which is closer to the target) becomes greater, that is not described in previous literature to the best of our knowledge. For further validation of the empirical result about the target-competitor distance, we explored it under a data generation process (DGP) based on the RRM, which has been shown to produce outcomes that are compatible with the decoy effect and is thus a possible underlying choice behavior that can be behind this phenomenon. The simulation analysis confirmed that, under an RRM DGP, the impact of the decoy grows with the distance between the target and the competitor. The third result consists in the study of susceptibility to and the valuation of the decoy effect using a latent class emergent value model. The model has two classes, one that is affected by the decoy and the other that does not. This is implemented by incorporating a dummy variable to capture the impact of the decoy effect in the correspondent class. With this model it was found that the decoy effect was equivalent to US$ 6.2 (4900 CLP) of fare or 88 minutes of travel time reduction, which compare to the US$14 and the 5 hours reference attributes for the profiles. Besides, in line with Fukushi et al. (2021), belonging to the class that is affected by the decoy was found to be more likely for interviewees that took less time to answer the survey. On the contrary, the decoy effect was found to work better with the older interviewees, the opposite what of what was found by Fukushi et al. (2021). This latter result suggests that the nature of the choice context may have an impact on the susceptibility to the decoy. The finding of the decoy effect on public transportation choices and the result about the impact of the distance target-competitor, response time, and individual's age on the decoy, may be used in the design of decoy settings that can influence public transportation choices to favor sustainable development.

ABSTRACT. Use of the multinomial logit (MNL) functional form is widespread in location choice models (LCMs) of travel and land use demand, both in the research literature and in applied settings. As the state-of-the-art in land use and travel demand modeling continues to evolve awar from aggregate forecasting towards activity-based models and microsimulation, MNL models are being asked to accommodate increasingly large choice sets. This trend imposes significant challenges on the different phases of modelling: specification, estimation, and prediction. Depending on the particular phase with which one is concerned, different strategies exist to address these challenges, none perhaps as commonly employed as sampling of alternatives. McFadden’s (1978) demonstration of the uniform conditioning property of MNL models showed that modelers could obtain consistent LCM parameter estimates while using only a random subset of non-chosen alternatives. Nerella and Bhat (2004) examined the influence of sample size on model efficiency and accuracy in model estimation. Further research by Guevara and Ben Akiva (2013), among others, developed approaches to improve model efficiency while retaining consistency, also in the context of estimation. To the best of our knowledge, there is no published research that systematically assesses the impact of sampling of alternatives on MNL models in a predictive context. That is the gap this paper addresses.

We conduct a numerical analysis of the effect of sampling of alternatives in discrete choice models within the context of disaggregate location choice forecasting. The goal of this paper is not only to define and describe this problem, but also to provide a procedure to aid practitioners who, faced with finite computational resources, must evaluate the tradeoff between smaller sample rates of more disaggregate alternatives and larger sample rates of more aggregate alternatives. Of particular interest is quantifying the extent of this problem as it presents itself in activity-based models (e.g. ActivitySim, CT-RAMP) or microsimulation platforms (e.g. UrbanSim (Waddell, 2002)), in which choice sets can number into the millions of alternatives.

In contrast to the process of estimation, in which choice sets are constructed by appending observed choices to randomly sampled alternatives, prediction of future choice scenarios means that there are no “observed” choices to speak of. This fact has two related consequences: 1) model error cannot be measured by comparing simulated choices to observed choices (or aggregate shares of observed choices); 2) sampling of alternatives may result in choice sets comprised entirely of unattractive alternatives, the likelihood of which is inversely proportional to the sample size. In practice, we find that this latter consequence results in an overdispersion of the aggregate predicted probability of alternatives across choosers, with choosers more likely to select sub-optimal (lower utility) alternatives than if they were choosing from the true universe of alternatives. In a model of residential location choice, for example, this phenomenon, which we term “dispersion error”, might negatively bias predicted population densities. Meanwhile, the absence of observed choices makes the severity of this overdispersion somewhat difficult to measure.

In this paper we provide a mathematical definition of dispersion error in multinomial logit predictions with sampling of alternatives, and attempt to characterize its empirical relationship to sample rate. To achieve this, we conduct a series of experiments using synthetic data generated from MNL models of our own design. These same models are used to simulate choice probabilities using sample sizes between 2 and N-1, where N corresponds to the total number of alternatives (i.e. 100% sample rate). Dispersion error is then measured by comparing the aggregate distribution of probability across choosers for each alternative to the distribution obtained with a sample rate of 100%. We define the bounds of this relationship by performing these experiments while varying both the size of N and the size (variance) of the model error. We subsequently explore how alternative techniques to random sampling of alternatives can be used to reduce dispersion error.

In general we observe a non-linear relationship between sample rate and dispersion error, whereby increasing the sample rate from 10 to 20% yields a greater reduction in error than we see when increasing the rate from 20 to 30%. The degree of non-linearity appears to be greatest for models with fewer total alternatives, with more linear trends observed for models that approach the scale of microsimulation. We also find that the total amount of dispersion error depends on both the size of the universe of alternatives as well as the size of the model error, with the greatest levels of overdispersion observed in models with fewer total alternatives and greater statistical precision.

Lastly, we contextualize these findings by estimating a location choice model of discretionary activities using real data taken from the California Household Travel Survey (CHTS). We estimate a model of the form commonly found in some of the most popular microsimulation platforms in use today in order to show how our experimental results can be used to estimate both the severity of overdispersion in a given model and the benefit one can expect to yield through the use of larger sample sizes.

ABSTRACT. This paper aims to understand accessibility to quality education in Santiago, Chile. Usually, accessibility in education is studied through 'school choice,' i.e., the access to quality schools is understood as the opportunities that different families face when they must choose a school for their children (see Elacqua et al., 2006, McEwan et al., 2008, Elacqua et al., 2012, and Gallego and Hernando, 2010 for examples in Chile). However, little has been said about how accessible each school is. Using a spatial econometric model, we try to determine what factors (school's performance measures, school's characteristics, and school's built environment) dictate the spatial coverage of each school. In other words, we postulate here that a different and complementary lens can be used to understand education accessibility and that "offer" is as important as "demand."

Chile's educational system is a singular case since its nationwide voucher system implementation in the 1980s (Chumacero et al., 2011). With such a system, parents are free to decide among several educational alternatives. The system has been recurrently criticized for allowing discrimination and social segregation (Elacqua, 2012; Carrasco et al., 2021; Gayo et al., 2019; Valenzuela et al., 2014) through school's student selection system and co-payment instruments. The Chilean system is similar to Friedman's original work and imposed little to no restrictions to the theoretical model in comparison to other countries' implementations (Hofflinger et al., 2020). The current study provides insight into the relation between school and residential location choice in a context where these two choices are not perfectly intertwined, as might be the case in other global south countries. This contrasts with the school system in countries like the USA, where about 75% of children attend their "assigned" local neighborhood public school (National Center for Education Statistics, 2016). Although many have recognized that school selection is an important factor in residential location (Lareau and Goyette, 2014; Owens, 2020), even affecting housing prices (Nguyen-Hoang and Yinger, 2011; Ely and Teske, 2015) and forcing informal practices from parents (Roberts, 2012), such as using a relative's address for registration, the dynamics of these two decisions are seemingly present in the US households, but they can only be explicitly captured in a case like the Chilean system.

The analysis is made using enrollment data (provided by the Ministry of Education of Chile) from all the public schools and private schools subsidized by the government in Santiago, Chile. The proposed framework is based on the spatial multiple discrete-continuous model (Bhat et al., 2015). The coverage area of each school is described using "distance"-based rings which the school itself as the center. We describe each school enrollment as the percentage of students that live in each of the different rings. For example, 16.19% of students of Instituto Nacional (one of the most emblematic public schools in Chile) live in the surrounding neighborhood (distance less than 0.5 km in public transportation), but only a few live in intermediary rings: 0.38% live between 0.5 km and 1 km, 1.51% live between 1 km and 2 km, 5.99% live between 2 km and 4 km, 8.88% live between 4 km and 7.5 km. Finally, it is not a surprise that 67.04% of enrolled students live farther than 7.5 km (in public transportation) from school, highlighting both Instituto Nacional's attractiveness and the spatial segregation of the city. Our model can characterize the coverage of each school and determine precisely the impact of each school and built environment characteristic, providing valuable results for the evaluation of mobility and education policies. Additionally, spatial correlation is included in the model, capturing what can be understood as "competition" between neighboring schools.

Preliminary results show that primary schools tend to capture students living in nearer areas (in comparison to secondary schools), private schools have fewer (in comparison to public schools) students living closer, single-gender schools attract more students from distant places, schools with better results in standard tests capture students from more distant areas, and public transport access plays a prominent role on school accessibility.

ABSTRACT. Demand side response involves dynamically changing electricity demand depending on the state of the electric grid to respond to volatile renewable generation and new electric loads in smart grids. Residential demand side response (RDSR) has been considered challenging due to heterogeneity in costumer engagement. Moreover, RDSR typically relies on pricing mechanisms to nudge consumers to shift or reduce demand, which raise equity concerns. This present paper analyses choice behaviour of participants to the residential DSR trial Greenwich Energy Hero (GEH) that uses rewards that are converted into vouchers for participants themselves to spend in the local community or donate to local charities. The GEH trial for residential energy demand side response that took place in 2020 in Greenwich, London. This was implemented through a smartphone app and residential energy meters connected to a backend. In this paper we explore the factors underlying participants choice to take part to demand side response events during the GEH trial.

*Experimental design* Amongst the potential factors affecting users participation in DSR events, 4 variables were considered as design variables (treatment) to deliberately vary during the initiative to assess their effect on events participation: -The type of event: “Constraint” event (C-event)in which participants would need to reduce their energy consumption over a period of 1h; or “Time of Use” event (ToU-event), essentially 3 CE events over 3 consecutive hours, during which the number of hero points to be gain might change -The maximum number of "Hero points" (rewards) to be gained in each 1-h event (70, 150 and 300) -The day of the week of the event (weekday or weekend) -The time of the day of the event (morning vs afternoon) The events were organised over 3 waves or tranches over the initiative period. While the above design variables varied across events and across waves, only in the third tranche of C-events and ToU-events characteristics were set using a D-efficient choice experiment design approach. In tranches 1 and 2 the design variables were varied across events, but not following a pre-determined experimental design approach, but simply ensuring some variability across events.

*Data* The full set of planned events across the three tranches consist of 78 1-hour event, including each of the three hours of ToU- event. 73 events were successful. In this paper we analyse the effect of event characteristics and participants’ characteristics on the intent to participate in a DSR event, as recorded from the GEH App, but do not consider if participants actually reduce demand at the event time. In particular, we consider that a GEH participant intends to participate in an event if they declare that they will participate in the event when prompted by an alert from the App. A GEH user declaration that they will not participate in the event or no record of response to the event alert are treated as a lack of intention to participate.

*Modelling DSR event participation intention* The binary choice of whether intending to participate or not in an event is modelled as a random effect binary logit model, in which an error component varying across individuals but not across observations is added to the specification of the liner predictor. This captures the correlation between repeated observation from the same GEH participants with a “random intercept” is assumed to vary across individuals, but constant across observations from the same individual. Response records are available for 52 GEH participants. Participants' characteristics included as explanatory variables are: dwelling characteristics, socio-demographics, most important attribute considered when evaluating home appliances, and in-home activities profile. The last two items are intended to characterise the in-home energy consumption profile of a relevant decision-maker in the household. The in-home activity profile is obtained through an exploratory factor analysis of weekly frequencies of in-home, energy consuming activities.

*Results* The Random Effect Logit model detailed above was estimated using event participation response data for each single hour event (i.e. also each hour from ToU-events is considered as a separate observation). Based on the significance estimated effects we can make the following observations: -We cannot reject the null hypothesis that “Hero points” are NOT associated with participation -Time of day seems to be associated with participation (early evening is associated with a higher likelihood to participate) -Socio-demographics are associated with participation: Higher education, part-time employment, Number of children -Appliance characteristics preferences are associated with participation -In-home activities profile is associated with participation The observation that the number of “Hero points” do not seem to affect the intention to participate in DSR events is unsurprising. Indeed, individuals who decided to engage with the GEH initiatives are likely to be motivated to participate independently from the points to be gained, as they are self-selected participants. Random selection of participants in future similar trials may enable identify groups that are sensitive to the non-monetary incentive. On the contrary, the characteristic of the event that appears to which participants seem sensitive is the time of day of the event. This suggests that behavioural approaches such as GEH to harness residential demand response for grid services might be only effective at specific times of the day. The fact that the intention to participate increases between 4pm and 7pm suggest that such an approach might contribute to alleviating electricity demand in the proximity of the evening peak. However, tests on much larger and representative samples should be carried out to verify this hypothesis stemming from the small exploration in this GEH trial. Finally, individual characteristics and circumstances are strongly associated with participation intent. Therefore, a successful scale-up of initiatives such as GEH is likely to be highly dependent on the circumstances and habitual in-home activity behavioural patterns and preferences in a household. This means that either the profiles that are likely to be more responsive need to be selected to achieve a high level of responsiveness, or systematic exploration of effective incentivisation mechanisms for the less responsive profiles need to be undertaken for effective deployments of this type of residential demand response approach.