Semiparametric multinomial logit models for analysing consumer choice behaviour

The multinomial logit model (MNL) is one of the most frequently used statistical models in marketing applications. It allows one to relate an unordered categorical response variable, for example representing the choice of a brand, to a vector of covariates such as the price of the brand or variables characterising the consumer. In its classical form, all covariates enter in strictly parametric, linear form into the utility function of the MNL model. In this paper, we introduce semiparametric extensions, where smooth effects of continuous covariates are modelled by penalised splines. A mixed model representation of these penalised splines is employed to obtain estimates of the corresponding smoothing parameters, leading to a fully automated estimation procedure. To validate semiparametric models against parametric models, we utilise different scoring rules as well as predicted market share and compare parametric and semiparametric approaches for a number of brand choice data sets.     

Flexible Covariance Structures for Categorical Dependent Variables Through Finite Mixtures of Generalized Extreme Value Models

A new class of finite mixture discrete choice models, denoted FinMix (fīn m?ks), is introduced. These arise from the combination of a finite number of core Generalized Extreme Value (GEV) models to achieve more flexible functional forms, particularly in terms of error covariance structures. Example members of the class include combinations of (1) Multinomial Logit (MNL) models with differing scales, (2) multinomial logit with nested MNL models, (3) tree extreme value models with differing preference trees, and so on. Compatibility of FinMix models with utility maximization is easily determined, which permits empirical investigation of the suitability of specific model forms for economic evaluation exercises.     

THE ANALYSIS OF DISCRETE CHOICE EXPERIMENTS WITH CORRELATED ERROR STRUCTURE

In a stated preference discrete choice experiment each subject is typically presented with several choice sets, and each choice set contains a number of alternatives. The alternatives are defined in terms of their name (brand) and their attributes at specified levels. The task for the subject is to choose from each choice set the alternative with highest utility for them. The multinomial is an appropriate distribution for the responses to each choice set since each subject chooses one alternative, and the multinomial logit is a common model. If the responses to the several choice sets are independent, the likelihood function is simply the product of multinomials. The most common and generally preferred method of estimating the parameters of the model is maximum likelihood (that is, selecting as estimates those values that maximize the likelihood function). If the assumption of within-subject independence to successive choice tasks is violated (it is almost surely violated), the likelihood function is incorrect and maximum likelihood estimation is inappropriate. The most serious errors involve the estimation of the variance-covariance matrix of the model parameter estimates, and the corresponding variances of market shares and changes in market shares.

In this paper we present an alternative method of estimation of the model parameter coefficients that incorporates a first-order within-subject covariance structure. The method involves the familiar log-odds transformation and application of the multivariate delta method. Estimation of the model coefficients after the transformation is a straightforward generalized least squares regression, and the corresponding improved estimate of the variance-covariance matrix is in closed form. Estimates of market share (and change in market share) follow from a second application of the multivariate delta method. The method and comparison with maximum likelihood estimation are illustrated with several simulated and actual data examples.

Advantages of the proposed method are: 1) it incorporates the within-subject covariance structure; 2) it is completely data driven; 3) it requires no additional model assumptions; 4) assuming asymptotic normality, it provides a simple procedure for computing confidence regions on market shares and changes in market shares; and 5) it produces results that are asymptotically equivalent to those produced by maximum likelihood when the data are independent.     

A Diagnostic Test for the Multinomial Logit Model

I propose a Lagrange multiplier test for the multinomial logit model against the dogit model (Gaudry and Dagenais 1979) as the alternative hypothesis. In view of the well-known drawback of the restrictive property of independence from irrelevant alternatives implied by the multinomial logit model, a specification test has much to recommend it. Finite sample properties of the test are studied using a Monte Carlo experiment, and the test's power against the nested multinomial logit model and the multinomial probit model is investigated. The test is found to be sensitive to the values of the regression parameters of the linear random utility function.     

Small area estimates of labour force participation under a multinomial logit mixed model

Summary.  A new methodology is developed for estimating unemployment or employment characteristics in small areas, based on the assumption that the sample totals of unemployed and employed individuals follow a multinomial logit model with random area effects. The method is illustrated with UK labour force data aggregated by sex–age groups. For these data, the accuracy of direct estimates is poor in comparison with estimates that are derived from the multinomial logit model. Furthermore, two different estimators of the mean-squared errors are given: an analytical approximation obtained by Taylor linearization and an estimator based on bootstrapping. A simulation study for comparison of the two estimators shows the good performance of the bootstrap estimator.     

Data augmentation,frequentist estimation,and the Bayesian analysis of multinomial logit models

This article describes a convenient method of selecting Metropolis– Hastings proposal distributions for multinomial logit models. There are two key ideas involved. The first is that multinomial logit models have a latent variable representation similar to that exploited by Albert and Chib (J Am Stat Assoc 88:669–679, 1993) for probit regression. Augmenting the latent variables replaces the multinomial logit likelihood function with the complete data likelihood for a linear model with extreme value errors. While no conjugate prior is available for this model, a least squares estimate of the parameters is easily obtained. The asymptotic sampling distribution of the least squares estimate is Gaussian with known variance. The second key idea in this paper is to generate a Metropolis–Hastings proposal distribution by conditioning on the estimator instead of the full data set. The resulting sampler has many of the benefits of so-called tailored or approximation Metropolis–Hastings samplers. However, because the proposal distributions are available in closed form they can be implemented without numerical methods for exploring the posterior distribution. The algorithm converges geometrically ergodically, its computational burden is minor, and it requires minimal user input. Improvements to the sampler’s mixing rate are investigated. The algorithm is also applied to partial credit models describing ordinal item response data from the 1998 National Assessment of Educational Progress. Its application to hierarchical models and Poisson regression are briefly discussed.     

Kernel-based semiparametric multinomial logit modelling of political party preferences

Conventional, parametric multinomial logit models are in general not sufficient for capturing the complex structures of electorates. In this paper, we use a semiparametric multinomial logit model to give an analysis of party preferences along individuals’ characteristics using a sample of the German electorate in 2006. Germany is a particularly strong case for more flexible nonparametric approaches in this context, since due to the reunification and the preceding different political histories the composition of the electorate is very complex and nuanced. Our analysis reveals strong interactions of the covariates age and income, and highly nonlinear shapes of the factor impacts for each party’s likelihood to be supported. Notably, we develop and provide a smoothed likelihood estimator for semiparametric multinomial logit models, which can be applied also in other application fields, such as, e.g., marketing.     

Bayesian inference for Poisson and multinomial log-linear models

Categorical data frequently arise in applications in the Social Sciences. In such applications, the class of log-linear models, based on either a Poisson or (product) multinomial response distribution, is a flexible model class for inference and prediction. In this paper we consider the Bayesian analysis of both Poisson and multinomial log-linear models. It is often convenient to model multinomial or product multinomial data as observations of independent Poisson variables. For multinomial data, Lindley (1964) [20] showed that this approach leads to valid Bayesian posterior inferences when the prior density for the Poisson cell means factorises in a particular way. We develop this result to provide a general framework for the analysis of multinomial or product multinomial data using a Poisson log-linear model. Valid finite population inferences are also available, which can be particularly important in modelling social data. We then focus particular attention on multivariate normal prior distributions for the log-linear model parameters. Here, an improper prior distribution for certain Poisson model parameters is required for valid multinomial analysis, and we derive conditions under which the resulting posterior distribution is proper. We also consider the construction of prior distributions across models, and for model parameters, when uncertainty exists about the appropriate form of the model. We present classes of Poisson and multinomial models, invariant under certain natural groups of permutations of the cells. We demonstrate that, if prior belief concerning the model parameters is also invariant, as is the case in a ‘reference’ analysis, then the choice of prior distribution is considerably restricted. The analysis of multivariate categorical data in the form of a contingency table is considered in detail. We illustrate the methods with two examples.     

Pooling choices or categories in multinomial logit models

The estimation of multinomial logit models today is routine. With this increased use has also come a need for testing. A test to determine whether choices can be combined is important. This paper presents a likelihood ratio test for combining choices in multinomial logit models. The use of the test is demonstrated with a simple example.     

Regression Analysis of Multivariate Fractional Data

The present article discusses alternative regression models and estimation methods for dealing with multivariate fractional response variables. Both conditional mean models, estimable by quasi-maximum likelihood, and fully parametric models (Dirichlet and Dirichlet-multinomial), estimable by maximum likelihood, are considered. A new parameterization is proposed for the parametric models, which accommodates the most common specifications for the conditional mean (e.g., multinomial logit, nested logit, random parameters logit, dogit). The text also discusses at some length the specification analysis of fractional regression models, proposing several tests that can be performed through artificial regressions. Finally, an extensive Monte Carlo study evaluates the finite sample properties of most of the estimators and tests considered.     

A comparison of artificial neural network and multinomial logit models in predicting mergers

A merger proposal discloses a bidder firm's desire to purchase the control rights in a target firm. Predicting who will propose (bidder candidacy) and who will receive (target candidacy) merger bids is important to investigate why firms merge and to measure the price impact of mergers. This study investigates the performance of artificial neural networks and multinomial logit models in predicting bidder and target candidacy. We use a comprehensive data set that covers the years 1979–2004 and includes all deals with publicly listed bidders and targets. We find that both models perform similarly while predicting target and non-merger firms. The multinomial logit model performs slightly better in predicting bidder firms.     

The negative multinomial logit model

A polychotomous logit model is defined for negative multinomial frequency counts within independent populations. An efficient estimator of the model parameters and estimator covariance matrix is given in closed form. Minimum chi-square and Wald tests are presented.     

Model-Robust Design of Conjoint Choice Experiments

Within the context of choice experimental designs, most authors have proposed designs for the multinomial logit model under the assumption that only the main effects matter. Very little attention has been paid to designs for attribute interaction models. In this article, three types of Bayesian D-optimal designs for the multinomial logit model are studied: main-effects designs, interaction-effects designs, and composite designs. Simulation studies are used to show that in situations where a researcher is not sure whether or not attribute interaction effects are present, it is best to take into account interactions in the design stage. In particular, it is shown that a composite design constructed by including an interaction-effects model and a main-effects model in the design criterion is most robust against misspecification of the underlying model when it comes to making precise predictions.     

Bayesian inference and model selection in latent class logit models with parameter constraints: an application to market segmentation

Latent class models have recently drawn considerable attention among many researchers and practitioners as a class of useful tools for capturing heterogeneity across different segments in a target market or population. In this paper, we consider a latent class logit model with parameter constraints and deal with two important issues in the latent class models--parameter estimation and selection of an appropriate number of classes--within a Bayesian framework. A simple Gibbs sampling algorithm is proposed for sample generation from the posterior distribution of unknown parameters. Using the Gibbs output, we propose a method for determining an appropriate number of the latent classes. A real-world marketing example as an application for market segmentation is provided to illustrate the proposed method.     

Analysing Financial Returns by Using Regression Models Based on Non-Symmetric Stable Distributions

The daily evolution of the price of Abbey National shares over a 10-week period is analysed by using regression models based on possibly non-symmetric stable distributions. These distributions, which are only known through their characteristic function, can be used in practice for interactive modelling of heavy-tailed processes. A regression model for the location parameter is proposed and shown to induce a similar model for the mode. Finally, regression models for the other three parameters of the stable distribution are introduced. The model found to fit best allows the skewness of the distribution, rather than the location or scale parameters, to vary over time. The most likely share return is thus changing over time although the region where most returns are observed is stationary.     

Stochastic variational inference for large-scale discrete choice models using adaptive batch sizes

Discrete choice models describe the choices made by decision makers among alternatives and play an important role in transportation planning, marketing research and other applications. The mixed multinomial logit (MMNL) model is a popular discrete choice model that captures heterogeneity in the preferences of decision makers through random coefficients. While Markov chain Monte Carlo methods provide the Bayesian analogue to classical procedures for estimating MMNL models, computations can be prohibitively expensive for large datasets. Approximate inference can be obtained using variational methods at a lower computational cost with competitive accuracy. In this paper, we develop variational methods for estimating MMNL models that allow random coefficients to be correlated in the posterior and can be extended easily to large-scale datasets. We explore three alternatives: (1) Laplace variational inference, (2) nonconjugate variational message passing and (3) stochastic linear regression. Their performances are compared using real and simulated data. To accelerate convergence for large datasets, we develop stochastic variational inference for MMNL models using each of the above alternatives. Stochastic variational inference allows data to be processed in minibatches by optimizing global variational parameters using stochastic gradient approximation. A novel strategy for increasing minibatch sizes adaptively within stochastic variational inference is proposed.     

A self-consistency approach to multinomial logit model with random effects

The computation in the multinomial logit mixed effects model is costly especially when the response variable has a large number of categories, since it involves high-dimensional integration and maximization. Tsodikov and Chefo (2008) developed a stable MLE approach to problems with independent observations, based on generalized self-consistency and quasi-EM algorithm developed in Tsodikov (2003). In this paper, we apply the idea to clustered multinomial response to simplify the maximization step. The method transforms the complex multinomial likelihood to Poisson-type likelihood and hence allows for the estimates to be obtained iteratively solving a set of independent low-dimensional problems. The methodology is applied to real data and studied by simulations. While maximization is simplified, numerical integration remains the dominant challenge to computational efficiency.     

Modeling the Distribution of Price Sensitivity and Implications for Optimal Retail Pricing

This article focuses on the distribution of price sensitivity across consumers. We employ a random-coefficient logit model in which brand-specific intercepts and price-slope coefficients are allowed to vary across households. The model is estimated with panel data for two product categories. The implications of the estimated model are deduced through an optimal retail pricing analysis that combines the panel data with chain-level cost figures. We test parametric distributional assumptions using semiparametric density estimates based on series expansions.     

Simple inference in multinomial and ordered logit

This paper provides some simple methods of interpreting the coefficients in multinomial logit and ordered logit models. These methods are summarized in Propositions concerning the magnitudes, signs, and patterns of partial derivatives of the outcome probabilities with respect to the exogenousvariables. The paper also provides an empirical example illustrating the use of these Propositions.     

Small sample properties of the logit maximum likelihood estimator:a simulation study

This paper explores the effect of sample size, scale of parameters and size of the choice set on the maximum likelihood estimator of the parameters of the multinomial logit model. Data were generated by simulations under a three-way factorial experimental design for logit models containing three, four and five explanatory variables. Simulation data were analyzed by analysis of covariance and a regression model of the performance measure, the log root mean-squared error (LRMSE), fitted against the three factors and their interactions. Several important conclusions emerged. First, the LRMSE improves, but at a decreasing rate, with increases in the model's degrees of freedom. Second, the number of choice alternatives in the decision makers' choice sets has a significant impact on the LRMSE; however, heterogeneity in the choice sets across the sample has little or no impact. Finally, the scale of parameters and all of its two-way interactions with the other two factors significantly affect the LRMSE. Using the regression results, a family of iso-LRMSE curves are derived in the space of model degrees of freedom and scale of parameters. Their implications for researchers in choosing sample size and scale of parameters is discussed.