Showing papers in "Journal of choice modelling in 2019"

Apollo: A flexible, powerful and customisable freeware package for choice model estimation and application

Abstract: The community of choice modellers has expanded substantially over recent years, covering many disciplines and encompassing users with very different levels of econometric and computational skills. This paper presents an introduction to Apollo, a powerful new freeware package for R that aims to provide a comprehensive set of modelling tools for both new and experienced users. Apollo also incorporates numerous post-estimation tools, allows for both classical and Bayesian estimation, and permits advanced users to develop their own routines for new model structures.

Airline itinerary choice modeling using machine learning

Abstract: Understanding how customers choose between different itineraries when searching for flights is very important for the travel industry This knowledge can help travel providers, either airlines or travel agents, to better adapt their offer to market conditions and customer needs This has a particular importance for pricing and ranking suggestions to travelers when searching for flights This problem has been historically handled using Multinomial Logit (MNL) models While MNL models offer the dual advantage of simplicity and readability, they lack flexibility to handle collinear attributes and correlations between alternatives Additionally, they require expert knowledge to introduce non-linearity in the effect of alternatives’ attributes and to model individual heterogeneity In this work, we present an alternative modeling approach based on non-parametric machine learning (ML) that is able to automatically segment the travelers and to take into account non-linear relationships within attributes of alternatives and characteristics of the decision maker We test the models on a dataset consisting of flight searches and bookings on European markets The experiments show our approach outperforming the standard and the latent class Multinomial Logit model in terms of accuracy and computation time, with less modeling effort

In-store or online shopping of search and experience goods: A hybrid choice approach

Abstract: This paper aims at explaining the choice between online and in-store shopping for typical search and experience goods (standard electronic appliances and groceries) within an artificial experimental setting assuming no privately owned cars. We present the first alternative-specific integrated choice and latent variable (ICLV) model using stated preference data in the field of shopping behavior research, explicitly asking respondents to trade-off attributes specific to each shopping channel. Respondents with pro-online shopping attitudes have a higher shopping cost sensitivity, which can be explained by the expanded choice set when effectively considering both purchasing channels. They also exhibit a higher choice probability of online shopping for groceries compared to electronic appliances, given the nature of experience goods being preferably purchased in-store, while the pleasure of shopping shows no substantial effect on choice behavior. Results reveal a user profile of pro-online shoppers that is mainly characterized by a technology-oriented generation of younger and well-educated men. Also, given the relatively high value of travel time compared to the value of delivery time, we show that especially for electronic appliances, avoiding a shopping trip produces more benefits than waiting for the delivery of ordered products.

Simulation error in maximum likelihood estimation of discrete choice models

Abstract: Maximum simulated likelihood is the preferred estimator of most researchers who deal with discrete choice. It allows estimation of models such as mixed multinomial logit (MXL), generalized multinomial logit, or hybrid choice models, which have now become the state-of-practice in the microeconometric analysis of discrete choice data. All these models require simulation-based solving of multidimensional integrals, which can lead to several numerical problems. In this study, we focus on one of these problems – utilizing from 100 to 1,000,000 draws, we investigate the extent of the simulation bias resulting from using several different types of draws: (1) pseudo random numbers, (2) modified Latin hypercube sampling, (3) randomized scrambled Halton sequence, and (4) randomized scrambled Sobol sequence. Each estimation is repeated up to 1 000 times. The simulations use several artificial datasets based on an MXL data generating process with different numbers of individuals (400, 800, 1 200), different numbers of choice tasks per respondent (4, 8, 12), different number of attributes (5, 10), and different experimental designs (D-optimal, D-efficient for the MNL and D-efficient for the MXL model). Our large-scale simulation study allows for comparisons and drawing conclusions with respect to (1) how efficient different types of quasi Monte Carlo simulation methods are and (2) how many draws one should use to make sure the results are of “satisfying” quality – under different experimental conditions. Our study is the first to date to offer such a comprehensive comparison. Overall, we find that the number of the best-performing Sobol draws required for the desired precision exceeds 2 000 in some of the 5-attribute settings, and 20,000 in the case of some 10-attribute settings considered.

Specification of mixed logit models assisted by an optimization framework

Abstract: Mixed logit is a widely used discrete outcome model that requires for the analyst to make three important decisions that affect the quality of the model specification. These decisions are: 1) what variables are considered in the analysis, 2) which variables are to be modeled with random parameters, and; 3) what density function do these parameters follow. The literature provides guidance; however, a strong statistical background and an ad hoc search process are required to obtain an adequate model specification. Knowledge and data about the problem context are required; also, the process is time consuming, and there is no certainty that the specified model is the best available. This paper proposes an algorithm to assist analysts in the search of an appropriate specification in terms of explanatory power and goodness of fit for mixed logit models. The specification includes the variables that should be considered as well as the random and deterministic parameters and their corresponding distributions. Three experiments were performed to test the effectiveness of the proposed algorithm. Comparison with existing model specifications for the same datasets were performed. The results suggest that the proposed algorithm can find adequate model specifications, thereby supporting the analyst in the modeling process.

Context-aware stated preferences with smartphone-based travel surveys

Abstract: Stated preferences surveys are most commonly used to provide behavioral insights on hypothetical travel scenarios such as new transportation services or attribute ranges beyond those observed in existing conditions. When designing SP surveys, considerable care is needed to balance the statistical objectives with the realism of the experiment. This paper presents an innovative method for smartphone-based stated preferences (SP) surveys leveraging state-of-the-art smartphone-based survey platforms and their revealed preferences sensing capabilities. A random experimental design generates context-aware SP profiles using user specific socioeconomic characteristics and past travel data along with relevant web data for scenario generation. The generated choice tasks are automatically validated to reduce the number of dominant or inferior alternatives in real-time, then validated using Monte-Carlo simulations offline. In this paper we focus our attention on mode choice and design an experiment that considers a wide range of possible existing mode alternatives along with a new alternative on-demand mobility service that does not exist in real life. This experiment is then used to collect SP data or a sample of 224 respondents in the Greater Boston Area. A discrete mode choice model is estimated to illustrate the benefit of the proposed method in capturing current context-specific preferences in response to the new scenario.

Flexible estimates of heterogeneity in crowding valuation in the New York City subway

Abstract: This paper aims at better understanding passenger valuation of subway crowding in New York City. To this end, we conducted a stated preference survey with a discrete choice experiment where New Yorkers chose an alternative from a set of two hypothetical unlabeled subway routes based on occupancy levels and other attributes. We used the collected data to estimate crowding multipliers that quantify the trade-off between travel time and standee density. The previous studies have resorted to parametric heterogeneity distributions in analyzing preference variations in crowding multipliers, which can lead to misspecification issues. The contribution of this study is thus to estimate crowding multipliers using state-of-the-art semi-nonparametric models – logit-mixed logit (LML) and mixture of normals multinomial logit (MON-MNL), and compare them across different parameter spaces. The estimated distribution of crowding multiplier of LML and MON-MNL coincide below median, but the former underestimates and the latter overestimates above median. Even though these flexible logit models can be useful for a comprehensive economic analysis of transit service improvements, these differences in estimates make model selection an important avenue for future research.

An R package and tutorial for case 2 best–worst scaling

Abstract: Case 2 (profile case) best–worst scaling (BWS) is a question-based survey method for measuring preferences for attribute levels. Several existing R packages help to implement the construction of Case 2 BWS questions (profiles) and the discrete choice analysis of the responses to the questions. Structuring the dataset for Case 2 BWS analysis is, however, complicated: there are several model variants for the analysis, and independent variables are set according to the variants. This complexity makes it difficult for non-expert users to prepare datasets for Case 2 BWS analysis. To improve the capability of R with respect to Case 2 BWS and facilitate easier data analysis, the package support.BWS2 has been developed. The package provides a function to map raw survey data into a format suitable for analysis, and also includes other useful functions, such as a function to calculate count-based BWS scores. A free online tutorial for Case 2 BWS in R has also been made available. These works make it easier for those new to Case 2 BWS to complete research using R, and facilitate the use of Case 2 BWS in various research fields.

A web survey application of real choice experiments

Abstract: This research note presents the first study to implement a real choice experiment in a web survey. In a case study on ethical food consumption, we find statistically significant lower willingness-to-pay values for the attributes “organic production” and “fair trade” in a choice experiment involving real payments compared to a choice experiment without real payments. This holds only true for respondents who are prepared to provide their personal details in order to deliver the product (83% of the sample), providing further evidence that lack of consequentiality can be an important source of validity problems. The implementation of a real choice experiment online proves useful and can form the baseline for future tests of the effectiveness of ex ante approaches such as cheap talk or honesty priming as well as consequentiality scripts in web-based choice experiments.

New software tools for creating stated choice experimental designs efficient for regret minimisation and utility maximisation decision rules

Abstract: At the time of creating an experimental design for a stated choice experiment, the analyst often does not precisely know which model, or decision rule, he or she will estimate once the data are collected. This paper presents two new software tools for creating stated choice experimental designs that are simultaneously efficient for regret minimisation and utility maximisation decision rules. The first software tool is a lean, easy-to-use and free-of-charge experimental design tool, which is dedicated to creating designs that incorporate regret minimisation and utility maximisation decision rules. The second tool constitutes a newly developed extension of Ngene – a widely used and richly featured software tool for the generation of experimental designs. To facilitate the use of the new software tools, this paper presents clear worked examples. It focusses on practical issues encountered when generating such decision rule robust designs, such as how to obtain priors and how to deal with alternative specific parameters. Furthermore, we analyse the robustness of the designs that we created using the new software tools. Our results provide evidence that designs optimised for one decision rule can be inefficient for another – highlighting the added value of decision rule robust designs.

Investigating users' preferences for Low Emission Buses : Experiences from Europe's largest hydrogen bus fleet

Abstract: The Health Economics Research Unit is funded by the Chief Scientists Office (CSO) of the Scottish Government Health and Social Care Directorate. The views expressed paper in this paper are those of the authors and not necessarily those of the CSO. Data collection for this study was funded by the Henderson Economics Research Fund. The funders had no role in study design, data collection and analysis, or preparation of the manuscript.

Modelling taste heterogeneity regarding offence location choices

Abstract: One of the central topics in crime research, and one in which discrete choice modelling has been relatively recently introduced, is the study of where offenders choose to commit crime. Since the introduction of this approach in 2003, it has become relatively popular and used in over 25 published studies covering a range of crime types and study areas. However, in most of these analyses the conditional logit has been used which assumes offenders are homogenous in their offence location preferences. This is despite various research finding offenders vary in their decision-making criteria. As such, while three recent studies (Townsley et al., 2016; Frith et al., 2017; Long et al., 2018) used the mixed logit and found some evidence of preference heterogeneity between offenders, there are still open questions regarding its nature. To this end, this study uses the latent class (and mixed and conditional logit) to examine the offence location choices of serious acquisitive crime offenders in York (UK). In particular, to understand how the spatial preferences differ between offenders and if there are any observable sources. Like the previous studies, this analysis identifies the presence of preference heterogeneity. This study also finds that the latent class and mixed logit equally fit the data though there are some differences in the results. These findings and other factors therefore raise questions for future crime location choice research regarding the appropriate model for these types of analyses and the true underlying nature of offender preferences.

Destination choice modeling using location-based social media data

Abstract: Destination choice models play a critical role in transportation demand analysis. However, collecting individual destination choices at a large scale is costly and time consuming. In this context, the availability of location based social media (LBSM) data gives us the opportunity to gather destination choice behavior of a large number of people in a continuous basis. In this paper, we present methods to extract and analyze large-scale data collected from Twitter for modeling travelers' destination choice behavior. We have adopted filtering steps to remove social bots from the dataset and prepare a reliable sample for analysis. We have created a joint database combining social media data with traditional census tract based socio-economic, land-use and infrastructure data. Using this dataset, we have estimated a Panel Latent Segmentation Multinomial Logit (PLSMNL) model which offers better insights on individual destination choices compared to machine learning/data mining methods. Estimated parameters indicate that the proposed PLSMNL intuitively assign destinations by trip purpose (shopping, recreational and other), gender, weekday (or weekend) and home zone land use measures. The results offer intuitive insights and highlight the applicability of social media data for destination choice analysis. Thus, this study demonstrates how we can potentially complement traditional travel survey-based data collection efforts with emerging social media data. [Pre-print of the article published in the Journal of Choice Modeling. The published article is available at https://doi.org/10.1016/j.jocm.2019.03.002]

Using panel data for modelling duration dynamics of outdoor leisure activities

Abstract: This paper examines the effects of socioeconomic characteristics, trip characteristics and life events on outdoor leisure activities and leisure duration in the Netherlands, based on 14 554 observations from three waves of data from The Netherlands Mobility Panel (in Dutch: MobiliteitsPanel Nederland). A standard mixed logit as well as a ‘zero-leisure’ scaled model was estimated to cover interpersonal and intrapersonal heterogeneity, The model was estimated for weekends, weekdays, transport mode choice of the activity, and specific leisure activity. The results show that travel time and transport mode choice for leisure trips have significant links with activity duration. Walking and cycling are dominant modes for short-duration activities and public transport for long-duration activities, and activity duration increases with travel time. The probability of short-duration leisure activities is higher on workdays. Certain life events positively affect the duration of leisure activities, whereas accessibility and bicycle ownership have no effect on leisure activity duration. The scaled model shows that the utility of any duration is about 10% larger for respondents who reported at least one day without leisure activities (‘zero leisure’). Leisure activities undertaken during the same week are significantly correlated, representing significant intrapersonal variation. The paper highlights the importance of analysing duration of activities for different activity types and days of the week and underlines the strong link of temporal (week, year) and spatial (activity type location) dimensions with transport mode choice.

‘Computer says no’ is not enough : Using prototypical examples to diagnose artificial neural networks for discrete choice analysis

Abstract: Artificial Neural Networks (ANNs) are increasingly used for discrete choice analysis, being appreciated in particular for their strong predictive power. However, many choice modellers are critical – and rightfully so – about using ANNs, for the reason that they are hard to diagnose. That is, for analysts it is hard to see whether a trained (estimated) ANN has learned intuitively reasonable relationships, as opposed to spurious, inexplicable or otherwise undesirable ones. As a result, choice modellers often find it difficult to trust an ANN, even if its predictive performance is strong. Inspired by research from the field of computer vision, this paper pioneers a low-cost and easy-to-implement methodology to diagnose ANNs in the context of choice behaviour analysis. The method involves synthesising prototypical examples after having trained the ANN. These prototypical examples expose the fundamental relationships that the ANN has learned. These, in turn, can be evaluated by the analyst to see whether they make sense and are desirable, or not. In this paper we show how to use such prototypical examples in the context of choice data and we discuss practical considerations for successfully diagnosing ANNs. Furthermore, we cross-validate our findings using techniques from traditional discrete choice analysis. Our results suggest that the proposed method helps build trust in well-functioning ANNs, and is able to flag poorly trained ANNs. As such, it helps choice modellers use ANNs for choice behaviour analysis in a more reliable and effective way.

Explore preference heterogeneity for treatment among people with Type 2 diabetes: A comparison of random-parameters and latent-class estimation techniques

Abstract: There has been an increasing interest in studying patient preference heterogeneity to support regulatory decision-making. While the traditional mixed logit (MXL) and the latent class logit (LCL) models have been commonly used to analyze preference heterogeneity in discrete choice data, they have limitations. This study empirically compares a random effects latent class logit (RELCL) model to the traditional approaches using preference data from a discrete-choice experiment among patients with Type 2 diabetes. Each survey contained 18 pairs of hypothetical diabetes medications that differed in six attributes. Sensitivity analysis is also performed to explore under what circumstances RELCL outperforms LCL. Significant preference heterogeneity was found in all models. The 2-class RELCL has the lowest BIC (8350.64) and prediction error (11.6%) compared to MXL (BIC = 9345.40; pred. err. = 13.0%) and the 5-class LCL (BIC = 8440.30; pred. err. = 16.4%), indicating improved model fit. Allowing random effects also reduces the number of classes from five in LCL to two, both having significant policy and clinical implications. RELCL provides the flexibility of LCL and the parsimony of MXL. Both our empirical results and sensitivity analysis shows that when there is significant preference heterogeneity among patients that cannot be captured by a small number of clusters, RELCL may be used to generate more accurate predictions and more parsimonious results that are policy-relevant.

Using Firth's method for model estimation and market segmentation based on choice data

Abstract: Using maximum likelihood (ML) estimation for discrete choice modeling of small datasets causes two problems. The first problem is that the data may exhibit separation, in which case the ML estimates do not exist. Also, provided they exist, the ML estimates are biased. In this paper, we show how to adapt Firth's penalized likelihood estimation for use in discrete choice modeling. A powerful advantage of Firth's estimation is that, unlike ML estimation, it provides useful estimates in the case of data separation. For aggregates of six or more respondents, Firth estimates have negligible bias. For preference estimates on an individual level, Firth estimates show little bias as long as each person evaluates a sufficient number of choice sets. Additionally, Firth's individual-level estimation makes it possible to construct an empirical distribution of the respondents' preferences without imposing any a priori population distribution and to effectively predict people's choices and detect market segments. Segment recovery may even be better when individual-level estimates are obtained using Firth's method instead of hierarchical Bayes estimation under a normal prior. We base all findings on data from a stated choice study on various forms of employee compensation.

Cost attribute in health care DCEs: Just adding another attribute or a trigger of change in the stated preferences?

Abstract: This paper investigated the effects of including an opt-out option and the cost attribute on the elicited preference structure and response error variance in a discrete choice experiment (DCE) valuing the preferences for the delivery of dental care. The mixed logit framework was used for testing the effects of survey design features on respondents' preferences and scale. The standard practice of testing these effects was further expanded by using the structural choice modelling (SCM) framework. Recent studies have suggested that not offering respondents an opt-out option may distort the utility estimates. However, the influence of the opt-out option, framed as a ‘substitute care provider’ alternative, on the preferences for dental care and response error variance was not significant. When the cost attribute was added to the choice sets, the rank order of the attributes remained the same, and overall preferences did not differ significantly. This indicates that respondents did not change their decision rule. However, they were not very consistent in their preferences for all attributes. Including an extra cost attribute significantly increased the response error variance. The findings indicate that the cost attribute could be safely used, at least in similar contexts, without concerns for disturbing the preferences for other attributes. We did not find strong evidence that the effect of including an extra cost attribute is any different from the expected effect of including any other choice attribute; therefore, its influence may not be as relevant as some of the previous studies may have suggested.

Willingness to pay functions for emergency ambulance services

Abstract: In the case of a medical emergency, a failure to provide timely care could have serious health consequences and even cause death. When the waiting time increases, the patient perceives the loss of well-being associated with the lack of primary care. Considering the willingness to pay (WTP) for avoiding human suffering in the context of emergency medical services (EMS) within the impact assessment of prehospital care is a relatively new research topic. This paper evaluates, through discrete choice models, the WTP for avoiding the externality associated with the human suffering experienced by the patient who does not receive immediate attention in a medical emergency. The cost of the service, the waiting time of the service, the time elapsed at the time of making the decision, the patient's condition and the time to aggravate are relevant to define the WTP in the context of EMS. Heterogeneity in the WTP was captured by the inclusion of some socioeconomic attributes and proxy variables to evaluate several levels of patient severity. The results are intended to demonstrate the relevance that this externality could have in the planning, operation, management of the emergency system and in the coordination of public policies aimed at providing efficient services with a better distribution of resources.

Best-Worst Scaling with many items

Abstract: Best-worst scaling (BWS) has become so useful that practitioners feel pressure to include ever more items in their experiments. Researchers wanting more items and enough observations of each item by each respondent to support individual respondent-level utility models may greatly increase the burden on respondents, resulting in respondent fatigue and potentially in lower quality responses. Wirth and Wolfrath (2012) proposed two methods for creating BWS designs that allow for large numbers of items and respondent-level utility estimation, Sparse and Express BWS. This study aims to uncover the recommended approach when the goal is recovering individual respondent-level utilities and intends to do so by comparing the relative ability of Sparse and Express BWS to capture the utilities that would have resulted from a full BWS experiment, one with at least three observations of each item by each respondent. The current study repeats previous comparisons of Sparse and Express BWS using a new empirical data set. It also extends previous findings by collecting enough observations from each respondent for both a full experiment and one of the proposed methods, Express BWS and Sparse BWS. The results replicate and extend previous findings regarding the superior ability of the Sparse BWS methodology, relative to Express, to reproduce “known” utilities or utilities that result from a full BWS design.

Analysis of the economic impact of water management policy on residential prices: Modifying choice set formation in a discrete house choice analysis

Abstract: This paper extends discrete residential choice models by incorporating choice set formation. Most discrete residential choice models make relatively arbitrary assumptions about the choice set – the set of houses to be considered by the purchaser. In this paper we explore several formulations of endogenous choice sets in which the decision maker's selection of a choice set is based on certain attributes and the final selection is made from this reduced choice set. The proposed approach is empirically applied to a housing transaction dataset and welfare measures are generated for non-marginal changes associated with a water management policy. A comparison of models across different temporal windows to define an individual's choice set shows that model parameters are sensitive to the assumptions used to define the choice sets. The models that approximate choice set formation improve the efficiency of estimation and influence estimated welfare measures suggesting the importance of choice set formation in the context of discrete housing choice models.

Information theoretic-based sampling of observations

Abstract: Due to the surge in the amount of data that are being collected, analysts are increasingly faced with very large data sets. Estimation of sophisticated discrete choice models (such as Mixed Logit models) based on these typically large data sets can be computationally burdensome, or even infeasible. Hitherto, analysts tried to overcome these computational burdens by reverting to less computationally demanding choice models or by taking advantage of the increase in computational resources. In this paper we take a different approach: we develop a new method called Sampling of Observations (SoO) which scales down the size of the choice data set, prior to the estimation. More specifically, based on information-theoretic principles this method extracts a subset of observations from the data which is much smaller in volume than the original data set, yet produces statistically nearly identical results. We show that this method can be used to estimate sophisticated discrete choice models based on data sets that were originally too large to conduct sophisticated choice analysis.

Aggregation biases in discrete choice models

Abstract: This paper examines the common practice of aggregating choice alternatives within discrete choice models. We carry out a Monte Carlo study based on realistic vehicle choice data for sample sizes ranging from 500–10,000 individuals. We consider methods for aggregation proposed by McFadden (1978) and Brownstone and Li (2017) as well as the more commonly used methods of choosing a representative disaggregate alternative or averaging the attributes across disaggregate alternatives. The results show that only the “broad choice” aggregation method proposed by Brownstone and Li provides unbiased parameter estimates and confidence bands. Finally, we apply these aggregation methods to study households’ choices of new 2008 model vehicles from the National Household Travel Survey (NHTS) where 1120 unique vehicles are aggregated into 235 make/model classes. Consistent with our Monte Carlo results we find large differences between the resulting estimates across different aggregation methods.

On the effect of HB covariance matrix prior settings: A simulation study

Abstract: The authors conduct an extensive simulation study to substantially contribute to the question how HB prior parameter settings (i.e. the prior variance and the prior degrees of freedom) affect the performance of HB-CBC models. The statistical performance of HB is evaluated under experimentally varying conditions based on six experimental factors using criteria for goodness-of-fit, parameter recovery and predictive accuracy. The results indicate that the prior degrees of freedom play a negligible role as there is not any noticeable impact on the performance of HB when varying that factor. For increasing prior variance levels overfitting problems occur that markedly affect parameter recovery and model fit, and a number of related interaction effects with regard to the settings for the prior variance can be observed both at the upper and lower level of the HB model. Perhaps the most striking finding however is that the predictive performance of HB-CBC is hardly affected by an increase of the prior variance. Many of our findings regarding the parameter settings of the inverse Wishart prior contrast those reported in previously proposed empirical studies.

A practical approach to designing partial-profile choice experiments with two alternatives for estimating main effects and interactions of many two-level attributes

Abstract: A method is presented which facilitates the practical construction of designs for stated choice experiments in which the choice sets are pairs of partial profiles and where, for a potentially large number of two-level attributes, the main effects and two-factor interactions are to be estimated. Although partly heuristic, the approach has a sound theoretical basis and can be used to generate utility-neutral designs for the multinomial logit model which possess a high statistical efficiency. Applying the method neither requires expert knowledge of design theory nor specialized software and is illustrated with several examples.

Selection bias in estimation of peer effects in product adoption

Abstract: It is well recognized that consistent estimation of peer effects faces formidable identification challenges. The confounding factors the literature usually focuses on are endogenous group formation, correlated unobservables and simultaneity. In this paper, I show that another significant source of bias arises when a researcher examines peer effects in product adoption. When people differ in their valuations of a product, people with higher valuations tend to adopt earlier so that only those with increasingly lower valuations comprise the set of potential adopters as time progresses. Such an endogenous attrition over time, if not correctly accounted for, will lead to inconsistent estimates of peer effects. I present simulations to numerically demonstrate the presence and extent of such a selection bias. I also propose a simple solution to remove the bias and examine its performance.

Arriving at a decision: A semi-parametric approach to institutional birth choice in India

Abstract: The Multinomial Logit (MNL) model is popular, but a semi-parametric specification of its link/utility function has seldom been used in empirical applications. This is primarily because of the resource intensive nature of semi-parametric estimation. In this paper we propose and implement a parallel computation algorithm to estimate the semi-parametric kernel MNL model. This algorithm reduces model estimation time by a factor of 2–10, depending on the size of the dataset and the available resources for computation. These computational gains make the estimation of this model feasible for large datasets. Additionally, using a Monte Carlo study we show that the kernel MNL outperforms the traditional linear MNL model in terms of fit and predicted choice probabilities. We demonstrate how kernel-based specification can unearth important heterogeneities in the effect of covariates through an empirical exercise. We use data from a nationally representative household survey (N = 157,804) to analyze the factors associated with institutional births (as opposed to home births) in India. Our revealed-preference results indicate that maternal education, household assets, distance to formal health facility, and birth order play an essential role in determining birth location choice. Although the directions of impact are similar across both the linear and the kernel MNL specifications, there are significant differences in the marginal effects of different factors across the two models. These differences, which arise due to the flexibility afforded by the semi-parametric specification, potentially bring additional nuance to policy discussions.

Valuation of labour market entrance positions among (future) apprentices - Results from two discrete choice experiments

Abstract: In this paper, we estimate the relative value of different employment characteristics when choosing between apprenticeship and job offers. Further, we test assumptions derived from sociological rational choice theory on preference heterogeneity by individual and context characteristics. For this purpose, we analyse data from two discrete choice experiments, one focusing on the choice of an apprenticeship position and the other on the choice of first employment position after vocational training. The experiments were conducted as part of the DAB panel study on educational and occupational trajectories of adolescents in German-speaking Switzerland on students respectively one year prior to leaving compulsory school, and during vocational training. Our findings show that the most relevant aspect when choosing a labour market entrance position is, that the job should match the desired occupational specialisation. Furthermore, considerable preference heterogeneity is found, which can partly be accounted for by individual- and labour market-specific subjective utility.

On consistency of the MACML approach to discrete choice modelling

Abstract: In this paper the properties of the maximum approximate composite marginal likelihood (MACML) approach to the estimation of multinomial probit models (MNP) proposed by Chandra Bhat and coworkers is investigated with respect to asymptotic properties. It is shown that, if the choice proportions are normalized to sum to one, a variant of the method provides consistent estimates of the choice proportions for a number of approximation methods. Furthermore it is argued that each approximation method leads to a particular mapping of regressors to choice proportions which is close - but not identical - to the map induced by the probit model. If the data are in fact generated according to this mapping then standard asymptotics, that is consistency and asymptotic normality of the estimators, hold. If the data are, however, generated by a probit model and the approximations are used for estimation, then the corresponding estimators are not guaranteed to be consistent. Different approximation methods are subsequently analyzed with respect to their asymptotic biases and additionally with respect to finite sample properties. It is shown that normalization of the choice proportions is essential for obtaining consistent estimates of the choice proportions. Normalization also decreases biases in parameter estimates.

Heterogeneity and college choice: Latent class modelling for improved policy making

Abstract: The huge increase in higher education coverage in many developing countries has gone hand-in-hand with an additional supply of private colleges and with the enrolment of low to middle-class students, previously excluded from a historically elitist education segment. The larger diversity of both “suppliers and consumers”, unseen a few years ago, calls for methodological approaches that recognize heterogenous tastes and eventually, to classify individuals into mutually exclusive groups, something that can improve the design of public policy. The importance of college choice in educational systems using voucher schemes, makes it relevant to know what are the main variables determining such choice and whether they differ among different groups. Chile, one of the countries with the most extensive voucher system in education, experienced a significant increase in higher education enrolment (over 250% over the last 15 years), and has faced fierce political controversy due to the high heterogeneity in college quality. We use a latent class choice model to capture differences between groups, an approach that performs significantly better than simpler models previously used in this area. We found that groups of individuals with different selection patterns exist, which could not be defined ex ante, and that those groups are not only differentiated by their income level but mainly, by how they performed at high school. From the different sensitivities to college characteristics such as cost, quality, and location, identifying these groups allows us to derive different policy prescriptions.