@article{2017:elshiewy:multinomia,
title = {Multinomial Logit Models in Marketing - From Fundamentals to State-of-the-Art},
year = {2017},
note = {One of the fundamental elements of marketing research is the analysis
of consumer choice behaviour. Hereby, researchers aim to identify
important determinants that affect the choice among different
alternatives. These valuable insights can show how to stimulate brand
and product choice in order to enhance brand performance. In addition,
they can also strengthen the understanding of psychological and
behavioural phenomena, such as reference price formation and brand
loyalty. From this background, our paper gives an overview of the most
important aspects and models when it comes to analysing brand choice.
We hereby focus on the most prominent discrete choice model, namely
the multinomial logit (MNL) model. Starting with the theoretical
foundation of choice behaviour, we provide a discussion of the basic
models and present the state-of-the-art extensions that account for
more realistic brand choice behaviour. We supplement each model
description with an empirical example to emphasise the advantage of
each approach. After providing the fundamentals, we demonstrate how to
estimate, test, and interpret the β coefficients of the explanatory
variables that are assumed to influence the brand choice behaviour in
a basic MNL model. We further show how to calculate elasticities to
account for the dimension of the explanatory variables and compare the
effects in terms of percentage changes. After introducing the basic
MNL model, we discuss its three limitations: (i) proportional
substitution patterns, (ii) no random taste variation due to common
β coefficients across consumers, and (iii) no consideration of
unobserved factors over time in repeated choice situations. To provide
guidance how to address the three limitations of the basic MNL model,
our paper summarises the state-of-the-art MNL models that relax all
three limitations. The common approach of these extensions is to allow
the β coefficients to vary across consumers. This variation is
interpreted as consumer heterogeneity in sensitivity to changes in the
marketing-mix instruments (or sensitivity in behavioural drivers, such
as loyalty) and introduces the more realistic random taste variation
in brand choice behaviour. In addition, the individual-level β
coefficients allow estimating disproportional substitution patterns
and capturing unobserved factors of consumers over the time. To
illustrate the advantage of allowing consumer heterogeneity, we
estimate the Latent-class/Finite-mixture MNL, the Mixed MNL and
Hierarchical Bayesian MNL with normally distributed β coefficients as
well as a mixture of Normals heterogeneity distribution for β. We
show how accounting for heterogeneous β coefficients affects the
parameter estimates compared to the basic MNL, and how the knowledge
of heterogeneous marketing-mix sensitivities across consumers provides
valuable insights for understanding, predicting, and influencing brand
choice behaviour. From this, we emphasise the importance of using the
state-of-the-art approaches, even if only interested in aggregated
parameter estimates. Researchers and practitioners who want to
replicate our results or analyse their own research questions using
the models in this paper will find the estimation code in the web
appendix. Building on the current body of knowledge in discrete choice
modelling, our paper concludes with avenues for future research in
marketing.},
journal = {Marketing ZFP},
pages = {32--49},
author = {Elshiewy, Ossama and Guhl, Daniel and Boztug, Yasemin},
volume = {39},
number = {3}
}