ArticlePDF Available

Abstract and Figures

Discrete choice experiments have emerged as the state-of-the-art method for measuring preferences, but they are mostly used in cross-sectional studies. In seeking to make them applicable for longitudinal studies, our study addresses two common challenges: working with different respondents and handling altering attributes. We propose a sample-based longitudinal discrete choice experiment in combination with a covariate-extended hierarchical Bayes logit estimator that allows one to test the statistical significance of changes. We showcase their use in studies about preferences for electric vehicles over six years and empirically observe that preferences develop in an unpredictable, non-monotonous way. We also find that inspecting only the absolute differences in preferences between samples may result in misleading inferences. Moreover, surveying a new sample produced similar results as asking the same sample of respondents over time. Finally, we experimentally test how adding or removing an attribute affects preferences for the other attributes.
Content may be subject to copyright.
Longitudinal Study in Preferences for Electric Vehicles
1
Sample-based longitudinal discrete choice experiments:
preferences for electric vehicles over time
Katharina Keller
1
Christian Schlereth
2
Oliver Hinz1
Forthcoming: Journal of the Academy of Marketing Science
Keywords: Adoption; Electric vehicles; Complementary mobility services; Discrete Choice
Experiment; Dual Response; Sample-based longitudinal study
Acknowledgements: The authors gratefully thank Luigi Bianco for his support in data collection
for the 2013 study and Prof. Dr. Bernd Skiera for his help and feedback during the colloquium.
This work has been funded by the German Research Foundation (DFG) within the Collaborative
Research Center (CRC) 1053 - MAKI.
1
Katharina Keller, Oliver Hinz, Goethe University Frankfurt, Chair of Information Systems and
Information Management, Theodor-W.-Adorno Platz 4, 60323 Frankfurt am Main, Germany,
Phone: +49-69-798-34656, Fax: +49-69-798-33910, kakeller@wiwi.uni-frankfurt.de,
ohinz@wiwi.uni-frankfurt.de.
2
Christian Schlereth, WHU Otto Beisheim School of Management, Chair of Digital Marketing,
Burgplatz 2, 56179 Vallendar, Germany, Phone: +49-261-6509-455, Fax: +49-261-6509-509,
christian.schlereth@whu.edu.
Longitudinal Study in Preferences for Electric Vehicles
2
Sample-based longitudinal discrete choice experiments: preferences for
electric vehicles over time
Abstract
Discrete choice experiments have emerged as the state-of-the-art method for measuring
preferences, but they are mostly used in cross-sectional studies. In seeking to make them applicable
for longitudinal studies, our study addresses two common challenges: working with different
respondents and handling altering attributes. We propose a sample-based longitudinal discrete
choice experiment in combination with a covariate-extended hierarchical Bayes logit estimator
that allows one to test the statistical significance of changes. We showcase this method’s use in
studies about preferences for electric vehicles over six years and empirically observe that
preferences develop in an unpredictable, non-monotonous way. We also find that inspecting only
the absolute differences in preferences between samples may result in misleading inferences.
Moreover, surveying a new sample produced similar results as asking the same sample of
respondents over time. Finally, we experimentally test how adding or removing an attribute affects
preferences for the other attributes.
Longitudinal Study in Preferences for Electric Vehicles
3
Introduction
At present in the year 2020, the fictive manufacturer ACME is working hard to ensure that one of
its recent innovations will become a mainstream technology in the foreseeable future. As the
management at ACME know, innovations do not break through overnight. Instead, they pass
through a long process until they become widely adopted (Rogers 1962). For example, it took 40
years for the TV to reach a market penetration of 50 million owners. The e-bike, which was
developed before 1900, entered the mass market with almost 1 million units sold in 2018 (Statista),
enduring peaks and troughs in between. Even the iPhone needed more than three years after the
market launch to sell more than 50 million units.
One decisive factor for an innovation’s success is whether manufacturers have carefully aligned
these innovations with consumer preferences (Reinders et al. 2010). To measure preferences,
practitioners have come to rely on discrete choice experiments (e.g., Gensler et al. 2012; Louviere
et al. 2000; Papies et al. 2011; Schlereth and Skiera 2017). However, such tests have mostly been
used cross-sectionally (i.e., at just one point in time), which makes it impossible to derive
conclusions about changes in preferences over time. Only a few discrete choice experiments have
tackled preference measurement in a longitudinal setting (Ambos et al. 2019; Jensen et al. 2014;
Meeran et al. 2017). However, all of them retained the same set of attributes across the duration of
their studies and, with the exception of Ambos et al. (2019), they also surveyed the same samples
of respondents, which might explain the low number of longitudinal studies.
Discrete choice experiments face two challenges in longitudinal studies. First, it is often impossible
to survey the same sample respondents over time, particularly over several years: some
respondents may have changed their (email) address, lost their willingness to participate, classified
Longitudinal Study in Preferences for Electric Vehicles
4
the emails as spam, or just entered a different life situation. Meanwhile, a representative sample at
one point in time might not be representative after a few years: all subjects in the sample become
older and some may no longer be available. Second, innovative technologies evolve, with some
attributes (e.g., price or certain capabilities) changing between two or more consecutive studies.
For example, the iPhone originally had a screen size of 3.5” in its first four years. It has since then
nearly doubled in sizea development that Apple considered impracticable at the beginning.
Firms might also introduce new features in later stages of the innovation process, such as in the
case of NFC technology, which has become a prerequisite for mobile payment. Naturally,
researchers may want to adapt later studies to evolving market needs without losing the ability to
compare respective preferences.
To overcome these challenges, this paper proposes a sample-based longitudinal discrete choice
experiment. To this end, we developed a covariate-extended hierarchical Bayes logit estimator that
combines the estimation of multiple discrete choice experiments within a single estimation.
Thereby, we make the following main contributions to the literature: First, despite surveying
different respondents, our approach enables researchers to estimate not only the parameter values
that represent individual respondents’ preferences, but also which preferences have changed over
time and to what extent. Second, we formally outline how to interpret the results when continuous
attribute levels, such as price, have changed over time, new attributes have been added, or
attributes from earlier studies have been removed. These two main contributions provide a new
level of flexibility for researchers and practitioners to employ discrete choice experiments. Many
companies might have ready access to this kind of data from their past applications of discrete
choice experiments. For them, our approach reduces the requirements of linking together already
collected data to identify trends.
Longitudinal Study in Preferences for Electric Vehicles
5
We showcase the utility of sample-based longitudinal discrete choice experiments by eliciting
preferences for electric vehicles over more than six years. Electric vehicles are an interesting focal
point because they have yet to achieve mainstream adoption. Besides highlighting the use of our
proposed method, we examine a range of associated questions: First, we experimentally test how
adding an attribute affects preferences for all other attributes. We find that only a subset of
attributes proportionally lose importance in the face of a newly added attribute, and some attributes
are statistically unaffected. Second, we experimentally compare the results of surveying a new
sample versus asking the same respondents over time. The general purchase intention remained
mostly the same. Lastly, we demonstrate how to handle changes in continuous attribute levels. We
also compare our model’s internal and predictive validity against a model that separately estimates
preferences for each year and a model that assumes a single normal distribution for the population
of all respondents.
From a theoretical perspective, we demonstrate that an innovation process does not necessarily
follow a monotonously increasing pattern; it might fluctuate through peaks and troughs, thereby
underscoring the benefits of using a sample-based longitudinal study. From a managerial
perspective, we conclude by performing a suite of counterfactual exercises that provide managerial
guidance on an essential pricing decision problem.
In the next section, we provide a literature overview of the related streams of research in discrete
choice experiments and models describing the evolution of innovation processes. Then, we
formally describe how to reflect changes to the list of attributes and attribute levels over time.
Finally, we outline the study context on electric vehicles, the setup, and the results. We conclude
with managerial implications, conclusions, and limitations.
Longitudinal Study in Preferences for Electric Vehicles
6
Related research
In the following, we establish a theoretical framework for embedding our new modeling approach.
We start by summarizing two streams of research related to longitudinal studies in the context of
innovations: One stream on the behavioral and analytical aspects of discrete choice experiments,
and the other on product innovation growth.
Related research on discrete choice experiments
Discrete choice experiments ask respondents to make hypothetical choices between multiple sets
of alternatives. They are backed up by a long-standing theory (e.g., Louviere et al. 2000) and
parallel real-world purchase decisions. These experiments can satisfactorily explain actual
purchasing behavior, even when the product or service of interest does not exist or was recently
launched. Table 1 summarizes a non-exhaustive list of related research, mostly on discrete choice
experiments that inspired our study; we will cover these in more detail below. For a broader
overview of recent accomplishments, we refer to, e.g., Agarwal et al. (2015).
Longitudinal use of discrete choice experiments We are aware of few longitudinal studies that
use discrete choice experiments. One is Jensen et al.’s (2014) study, which asked respondents to
decide between electric vehicles and conventional cars. In the first round, the researchers surveyed
respondents without any experience; then, the respondents had the opportunity to use an electric
vehicle for three months and were surveyed again afterward with the same questionnaire. The
researchers compared the choice proportions of choosing an electric vehicle before and after
respondents’ experience in real life and found that after the trial phase, respondents chose the
electric vehicle only half as often as before. Yet, their perceptions about the electric vehicles’
driving performance had improved.
Longitudinal Study in Preferences for Electric Vehicles
7
Study
Study Size
Longitudinal study
Key findings
Longitudinal Use of Discrete Choice Experiments
Jensen et al. (2014)
196 respondents in Denmark
Yes,
Time frame: 3 months
same attributes & levels
Respondents chose electric vehicles after trial phase only half as
often as before
Meeran et al.
(2017)
161 respondents
Yes
Time frame: 6 months
same attributes & levels
For products with rapid technological change and short life
cycles, consumer preferences are more likely to shift
Ambos et al.
(2019)
238 managers from Germany,
Austria, and Switzerland
Yes
Time frame: 4 years
same attributes & levels
Managerial preferences vary over time, but heuristics for
decision-making remain stable
Theory of Unseen Attributes
Bradlow et al.
(2004)
130 undergraduate students
No
Consumers deduce missing attribute levels from missing and non-
missing attribute information
Yang et al. (2015)
70 respondents
No
Improves predictions outside the sample and differentiation
between characteristics, allows shorter questionnaires
Gilbride et al.
(2016)
895 shoppers
No
Inferring ex-ante value of attribute level dominates
Covariate-extended Hierarchical Bayes Modelling and Combining Multiple Data Sources
McCullough and
Best (1979)
100 undergraduate students
Yes
Time frame: 3 days
One attribute exchanged
High temporal stability and structural reliability, despite changes
in the list of attributes.
Teas (1985)
148 respondents
Yes
Time frame: 1 week
One attribute deleted
Repetition of earlier findings: high temporal stability and
structural reliability, negligible interaction effects
Lenk et al. (1996)
Study 1: 100 synthetic
respondents, study 2: 179
students
No
Response quality is unchanged by questionnaire length
Swait and
Andrews (2003)
Scanner Panel Data: 526
panelists, choice experiment:
384 Chicago residents
No
Joint model shows better prediction performance on holdout data
set compared to pure scanner panel model.
Orme and Howell
(2009)
326 respondents
No
Usage of too many covariates in the model should be avoided
Kurz and Binner
(2010)
10 studies, 30,000+
interviews
No
Mostly, covariates did not improve results
Sentis and Geller
(2010)
5 studies, 8,445 interviews
No
Fit and predictive accuracy unchanged
Liakhovitski and
Shmulyian (2011)
160 synthetic respondents
No
No significant improvement in the hit rate by using covariates, but
partly better distributed utilities + aggregated metrics
Ellickson et al.
(2019)
510 survey respondents,
actual purchase data from
4,288 customers
No
The approach increases the accuracy of forecasts for subsequent
actual purchases and overcome selection and contextual bias
Table 1. Non-exhaustive overview of related research
Longitudinal Study in Preferences for Electric Vehicles
8
Meeran et al. (2017) measured consumers’ changing preferences across four different product
categories over six months. They conducted the same survey three times, questioning the same
respondents and estimating the separate and joint aggregate parameters using Maximum
Likelihood Techniques. Like in Jensen et al. (2014), the authors used the same attributes and their
levels between studies. Meeran et al. (2017) found that consumer preferences are more likely to
shift for products that experience rapid technological change and short lifecycles.
In another longitudinal discrete choice experiment, Ambos et al. (2019) studied management
decisions regarding international locations at two points in time. The authors used hierarchical
Bayes to separately estimate the preferences and found that they varied over time, although the
decision-making heuristics remained stable. Like in Jensen et al. (2014) and Meeran et al. (2017),
the authors used the same attributes and levels across studies.
In sum, all these studies were conducted within a relatively short time (three months, six months,
four years), so preferences may not have changed that much. Observing greater changes may
require a longer study period. This short time-span may be due to their inherent challenges:
ensuring answers from the same respondents and keeping the focal product the same. Concerning
the former, it is unclear whether a study benefits from surveying the same sample of respondents.
Certain biases might arise, such as the mere exposure effect (Zajonc 1968), whereby respondents
assess an originally neutral product or service more positively when repeatedly asked about it.
Concerning the second challenge, the shifting nature of innovation processes makes it difficult to
measure preferences for the same set of attributes over a long period. With our proposed sample-
based longitudinal discrete choice experiment, we aim to overcome these challenges.
Longitudinal Study in Preferences for Electric Vehicles
9
Capturing the longitudinal nature of the data Comparing the preferences of different samples
of respondents over time requires the joint estimation of multiple discrete choice experiments. We
propose a covariate-extension of the generic hierarchical Bayes logit estimator, which explicitly
allows the sample preferences to differ. We base our implementation on Lenk et al.’s (1996) work
concerning a more flexible definition of the upper-layer model in hierarchical Bayes. Whereas the
generic hierarchical Bayes model assumes a single multivariate normal distribution for all
respondents, we explicitly allow for multiple distributions that depend on when the data collection
occurred.
Often, we think of covariates such as common demographics like gender or age; however,
according to Orme and Howell (2009), these variables have a low correlation with preferences
within choice contexts. The use of covariates is more meaningful when researchers are concerned
that the traditional shrinkage of individual parameters toward the population mean goes against
their expectation of different, distinguishable segments. Another reason for their use is when a
segment of respondents is oversampled: Without including covariates, the shrinkage toward the
population means can bias the estimates of the undersampled group.
Researchers who apply covariate modeling emphasize that improving predictions is not the main
motivation. Several studies have independently found that improvements in predictive validity are
rather modest (if any) in comparison to the generic hierarchical Bayes modeling with the single
distribution assumption (Kurz and Binner 2010; Liakhovitski and Shmulyian 2011; Orme and
Howell 2009; Sentis and Geller 2010). Nevertheless, covariates are meaningful because they
“allow us to test more formally the differences between segments and the part-worth” (Orme and
Howell 2009). Thus, understanding and explicitly capturing developments in preferences in
longitudinal settings is the primary benefit of covariates for our study.
Longitudinal Study in Preferences for Electric Vehicles
10
Besides using a model that assumes homogeneity on the population level, previous studies applied
two other techniques in the case of multiple data sources. The first is to estimate each study
separately (like in, e.g., Ambos et al. (2019)); however, changes in parameter values might not
always be apparent, especially for altered lists of attributes over time.
The stream on data enrichment, which jointly estimates scanner data with discrete choice
experiments, provides a second technique (c.f., Louviere et al. 2000, chapter 8). This stream
combines two data sources with only a partial overlap in the attributes (e.g., Swait and Andrews
2003; Ellickson et al. 2019) by distinguishing two sets of attributes: a common set across both data
sources and an uncommon one. Ellickson et al. (2019) incorporated a mean additive preference
shifter and a multiplicative variance shifter for common attributes, assuming that both data sources
share the same covariance matrix. For the uncommon attributes, they separately captured
preferences with separate covariance matrices. Nevertheless, their study context differs from ours
because they always fused exactly two data sources; generalizing this technique to more than two
can quickly become challenging, as the number of distinguishable sets of attributes rises
exponentially. Some elements, such as the mean additive shifter, served as inspiration for our
sample-based longitudinal discrete choice experiment.
In the empirical study, we will compare the performance of the covariate-extended hierarchical
Bayes logit estimator against the performance of the generic and the separate estimation. More
importantly, we will outline how the covariate-extended model’s estimation output (i.e., the
posterior) directly provides a simple, but effective way of testing whether changes in parameters
across longitudinal samples are significant or just noise.
Longitudinal Study in Preferences for Electric Vehicles
11
Theory of unseen attributes Besides asking different samples of respondents, longitudinal studies
for innovations require solutions to address changes in market-relevant attributes. In this respect,
we rely on the theory of unseen attributes to account for situations where attributes are added or
removed, such that they are unseen in some of the studies. A prerequisite is that preferences for
attributes are structurally reliable, even though the set of attributes does not overlap completely
across studies (see also McCullough and Best 1979; Teas 1985). This assumption complies with
common theory in design generation techniques (Street and Burgess 2007), i.e., that a rigorously
constructed design improves the quality of the parameter estimates (i.e., the standard deviations of
the beta parameters), but it does not affect the location of the beta parameters (i.e., their means).
Based on the behavioral literature on unseen attributes, we identified three lines of arguments on
how respondents process them (c.f., Bradlow et al. 2004; Gilbride et al. 2016): The first argument
assumes a null effect for missing attributes, i.e., that respondents exclusively focus on the
information that is directly available in the choice tasks. This line of argument builds on bounded
rationality theory and assumes that respondents tend to ignore information that is not directly
available (Bettman et al. 1998).
The second argument is that respondents substitute “market means” for missing attributes in the
choice sets. Used in cases like Yang et al. (2015), it builds on experimental findings wherein
respondents treated unseen attributes as if they had the average perceived utility in the market.
Branco et al. (2012) broadened the assumption of one commonly known market average and
proposed that each respondent inferred an “expected individual value” that can differ across
respondents, but they did not specify the process of setting the value.
Longitudinal Study in Preferences for Electric Vehicles
12
The third argument fills the lack of an explicitly specified process by assuming that respondents
impute expected individual values through a “pattern-matching learning model” (Bradlow et al.
2004). This means that respondents infer values based on seen information; for example, they
conclude that a certain, yet unseen, attribute is available in case of a higher price. Using partial
profile rating-based conjoint, Bradlow et al. (2004) showed that their model featured better internal
and predictive validity than the “null effect” model. However, unlike in their study, where
respondents were aware that some of the information was missing, ours does not employ partial
profiles. For this reason, we will concentrate our modeling on the first two assumptions and show
that they eventually result in the same model.
Related research on product innovation growth
Growth models Another stream of research that complements the analysis of preference evolution
over time deals with product innovation growth. We distinguish between two growth processes:
The first one is the Gartner hype cycle, an annually updated model by the American research and
advisory firm Gartner, Inc (Linden and Fenn 2003), which assesses the commercial viability of an
emerging technology. It distinguishes five phases, starting with the innovation trigger. The
subsequent phase, the peak of inflated expectations, is followed by a negative hype caused by
decreasing interest (the trough of disillusionment), as consumers come to understand that the
innovation will not fulfill all expectations. Companies can overcome this low-point phase by
developing second-generation products that make the innovation’s utility more visible and
understandable. These efforts lead to the final phases: slope of enlightenment and plateau of
productivity. The innovation begins to achieve wide acceptance from the masses, and mainstream
adoption begins. The theoretical basis of this non-monotonous concept is an over-expectation of
the technology’s potential, coupled with excessive confidence in its functionality on the part of
Longitudinal Study in Preferences for Electric Vehicles
13
developers, stakeholders, and early adopters. Several factors cause these initial perceptions, such
as the sheer optimism of stakeholders about promising, but not yet proven, abilities, as well as the
intended (exaggerated) publicity as a means of facilitating rapid adoption (Lucker et al. 2018). The
Gartner hype cycle’s third phase, trough of disillusionment, aligns with the expectation
disconfirmation theory, according to which individual expectations interact with objective criteria
to determine people’s product satisfaction (Bhattacherjee 2001; Oliver 1980).
Market diffusion models provide the second representation of growth processes. Fourt and
Woodlock (1960), for example, assumed that the cumulative sales curve of an innovation follows
an exponential, i.e., monotonously increasing shape. Rogers (1962) and other researchers predicted
a bell-shaped frequency curve, with a steep rise in the early adopter segment and a decline after
reaching a peak due to market saturation. The corresponding cumulative curve is S-shaped, again,
monotonously increasing, but with a low gradient at the beginning as well as the end and a higher
gradient in the middle.
Market diffusion models and models like the Gartner hype cycle can complement each other and
are not necessarily in conflict. In contrast to the market diffusion models, the Gartner hype cycle
captures the evolution of expectations rather than sales. Thus, its conceptual foundation aligns
more with preferences: Preferences translate into sales when the innovation reaches the slope of
enlightenment or the plateau of productivity. Hence, the Gartner hype cycle’s two final stages
reflect the beginning of the market diffusion models.
Longitudinal Study in Preferences for Electric Vehicles
14
Reasons for changes in preferences over time Product assessmentsand the accompanying
consumer acceptance of an innovationare largely influenced by consumer preferences. Thus, it
is important to understand why these preferences may change over time. One reason is that
consumers face a trade-off between the monetary costs of a new technology and its perceived
benefits (Dodds et al. 1991). If the ratio between value and price is high, customers’ usage intention
and purchase intention will increase (Venkatesh et al. 2012). This ratio improves with decreasing
prices, as is common for technological innovations. The technology also becomes more attractive
if an innovation accumulates more benefits over time. If both occur, the cost-benefit ratio greatly
improves.
Another reason is that unfamiliarity with new products, lacking experience, and limited
information may influence consumer preferences over time (Coupey et al. 1998). According to the
bounded rationality theory, which acknowledges that people have limited capacity for processing
information when making decisions (Simon 1955), unfamiliarity with certain attributes may result
in situations where consumers only consider a subset of attributes. The more that customers learn
about those products, the more they enhance their experience; over time, they may change the
subset of attributes used for decision-making, and assign different levels of importance to the
attributes (Meeran et al. 2017). Also, the certainty of these decisions may change when more
information becomes available. The bounded rationality theory aligns with decision researchers
growing opinion that consumer preferences are often not well-defined ex-ante; rather, preferences
are constructed during decision-making (Bettman et al. 1998). Taking these arguments together,
we will focus on capturing the shift in preferences. Later, in the Web Appendix, we will test
whether the model benefits from explicitly accounting for differences in choice consistency.
Longitudinal Study in Preferences for Electric Vehicles
15
Summary of potential changes
In sum, innovation processes can evolve in manifold ways, and it is of great importance to analyze
this evolution, for which sample-based longitudinal studies are particularly well suited. Taking
into account the argumentation in the aforementioned literature streams, Table 2 presents the
potential changes that need to be considered in sample-based longitudinal studies. Some of these
changes are induced by the respondents (cases #1 and #2) and some of these changes are induced
by the researchers (cases #3 - #8). Table 2 also summarizes the related assumptions in handling
these changes and reveals our core findings, which we explain in more detail in the following
sections.
Model and estimation
Covariate-extended hierarchical Bayes logit estimator
We assume a utility-maximizing respondent and base our estimation on the random utility theory
(Thurstone 1927). We decompose the utility uh,i of respondent h and alternative i into a
deterministic part vh,i that contains observable, experimentally manipulated elements and a
stochastic, i.e., unobservable part εh,i : uh,i = vh,i + εh,i. The stochastic part accounts for the Thurstone
(1927) realization that respondents make errors in their choices that cannot be explained by the
attributes and levels in the deterministic part alone. Let vh,0 = 0 be the deterministic utility of the
reference alternative, i.e., an unobserved outside option. We assume an extreme value distribution
for εh,i and obtain the choice probability Prh,i of respondent h for alternative i in choice set a, with
Ia being the set of alternatives in choice set a as follows (c.f., Train 2009, p. 34f):
Longitudinal Study in Preferences for Electric Vehicles
16
Case
Type of Change
Induced by
Main Assumptions
What we Show
Addressed in
Empirical Study
#1
Changes in respondents’ preferences for attributes
Respondent
- Changes in preferences can be
captured through additive shifter on
population level
- Samples are comparable
Section Model and Estimation:
How to capture and test for the significance of such shifts
through a covariate-extended model
Yes
#2
Changes in respondents’ consistency in decision
making through better understanding of attributes
Respondent
- Changes in consistency can be
captured through a variance shifter on
population level
- Samples are comparable
Robustness section:
Adding a variance shifter for each year or for each
sample did improve neither internal nor predictive
validity. We conclude that changes in consistency played
a minor role when combining multiple discrete choice
experiments.
Yes
#3
Changes in experimental design
Researcher
Using the same design generation
process ensures structural reliability of
measured preferences
Not in the study focus, but c.f., Street and Burgess (2007)
Yes
#4
Adding attributes, which existed at times of earlier
studies, but were not included
Researcher
“Null effect” or expected individual
value” assumption for missing attributes
Section Model and Estimation:
For discrete choice experiments: “Null effect” or
expected individual value assumption for missing
attributes eventually leads to the same model for the
estimation
Not in the focus
#5
Adding attributes, which did not exist at times of
earlier studies, but are now included
Researcher
“Null effect” assumption for missing
attributes
Yes (experimentally
tested in 2019)
#6
Removing attributes, which remain as part of the
product but are no longer in the focus of the study
Researcher
“Null effect” or expected individual
value” assumption for missing attributes
Not in the focus
#7
Removing attributes, which are no longer part of
the product
Researcher
“Null effect” assumption for missing
attributes
Yes (experimentally
tested in 2019)
#8
Changing values of continuous attributes levels
(e.g. price)
Researcher
Linearity in preferences for a continuous
attribute
Changes by an additive constant shift of
level values to avoid biases referred to as
attribute range effects
Section Model and Estimation:
Shift does not affect interpretation of the beta parameter
that is associated with the continuous attribute, but rather
it impacts constant. Alternative solution on handling this
change in robustness section: by dummy coding
continuous values
Yes (demonstrated
between 2013 and
2017)
Table 2. Summary of potential changes in longitudinal studies
Longitudinal Study in Preferences for Electric Vehicles
17
 

 (hH; iIa). (1)
We estimate the parameters for the preferences at two layers. The upper layer captures
respondents behavior aggregated over the population, i.e., it assumes that a multivariate normal
distribution describes the behavior of all respondents. The lower layer captures respondents
individual behavior. We assume that, given a respondent’s parameter values, his/her probability
of choosing an alternative i is governed by the multinomial logit model.
Without sacrificing generalizability, we consider two discrete choice experiments at time t1 and
t2 (in the empirical study, we expand the model to more points in time). On the upper layer, we
account for the change in preferences across multiple studies through a covariate matrix Zh, which
is of size |H|x2 and consists of |H| (the number of respondents) rows zh, each with two columns.
The first column is 1 (i.e., it captures preferences at time t1) for all respondents, while the second
column is 1 for responses in t2 and 0 for the reference study in t1.
On the lower layer, we use an additive model for the individual deterministic utility vh,i consisting
of a |P|-dimensional vector of parameter values, βh, times the design vector, xiI of product i, where
|P| is for the number of parameters to estimate. Following Lenk et al. (1996), we link the lower
and upper layers through the following regression model:
   (hH), (2)
with ςh following a normal distribution   and the matrix θ including the upper layer
regression parameters to express the sensitivity of the parameter values βh to the study-related
covariates zh. Equation (2) does not include a time index because the individual index h already
Longitudinal Study in Preferences for Electric Vehicles
18
contains the information that a respondent participated at time t. θ is of size |P|x2, i.e., one row and
two columns for each estimated parameter. The vector βh follows a multivariate normal
distribution     with means    and covariance matrix Σ. The Technical Appendix
(see Web Appendix) provides an outline of the estimation procedure.
To sum it up, βh captures the individual preferences of respondent h on the lower layer. On the
upper layer, the first column of θ captures the preferences in t1 as a p-dimensional vector  and
the second column captures the change in preferences  in t2 in comparison to the reference
study (case #1 in Table 2), such that    .
For more than two studies, let b quantify the number of parameters in the vector βh and n be the
number of longitudinal studies. Then, we estimate n·b+b·(b+1)/2 parameters on the upper layer,
i.e., one set of parameters for each longitudinal study and one covariance matrix. In contrast, the
generic estimation with the single, normal population assumption estimates b+b·(b+1)/2
parameters (one set of parameters and one covariance matrix), and the separate estimator uses
n·b+n·b·(b+1)/2 parameters (n sets of parameters and n covariance matrices). Because the
covariate-extended model explicitly accounts for heterogeneity across studies but assumes one
common covariance matrix, it makes the estimation more parsimonious than separating the
samples (Orme and Howell 2009, p. 3). Next, we explain how changes in the attributes and levels
affect the interpretability of the results.
Adding and removing attributes over time
We first consider the cases #4, #5, #6, and #7 from Table 2, i.e., that researchers added or removed
some attributes in later studies. For illustration purposes, we subsequently consider a product class
with two attributes in time t1 and t2 as well as a third attribute, present only in t2. Let all attributes
Longitudinal Study in Preferences for Electric Vehicles
19
have two levels, such that only one parameter per attribute is estimated. Respondent h’s
deterministic utility for that product i is then:
, ,0 ,1 ,1 ,2 ,2 ,3 3h i h h i h i h
v x x x
 
= +  + +
(hH; iIa). (3)
Assuming an expected individual value for the missing attribute (cases #4 and #6) implies that
3 ,3h
xx=
and that
,3 ,3hh
x
is constant across alternatives at time t1. Hence, the unseen attribute
does not affect the choice decision when using the logit model. Yet, it affects the decisions to buy
or not buy the product, which is captured in the new parametrization of the constant at time t1, i.e.,

   
,3h
x
, resulting into the deterministic utility function for:  
 
    . Thus, we do not need to make explicit assumptions about the level of
3
x
because
the new constant implicitly captures preferences for the unseen information.
Assuming a null effect for missing attributes (cases #4, #5, #6, and #7) implies that    at
time t1. The main difference to the previous assumption is the interpretation of , i.e., whether
it contains unobserved information or not. The estimation is the same for both lines of arguments.
To illustrate the estimation, let Table 3 contain the exemplary design matrices Xt for the studies in
t1 and t2 for the example in the previous section. Each row represents an alternative shown to
respondents, and the columns contain the effects-coded attribute levels. Since only the study at
time t2 showed the third attribute, the column representing the unseen attribute is zero at time t1.
To account for unseen attributes in sample-based longitudinal studies, we modified the standard
covariate-extended hierarchical Bayes logit estimator that turned out to be effective. Assuming a
normal distribution for parameters that are related to zero-columns in the design matrix (i.e., in our
example for βh,3 in t1) would add noise to the estimation. Instead, we set this parameter in the
Longitudinal Study in Preferences for Electric Vehicles
20
Metropolis Hastings-step to zero. We also adjusted the draw of the -matrix (see the Technical
Appendix in the Web Appendix for details). We applied the two modifications solely to parameters
that relate to a zero column in the design matrix. In simulations, we found that the ability to recover
parameter values substantially increased with this change.
Respondents in t1
Respondents in t2
Constant
Attribute 1
Attribute 2
Attribute 3
Constant
Attribute 1
Attribute 2
Attribute 3
1
1
1
0
1
-1
1
-1
1
1
-1
0
1
1
1
1
1
1
1
-1
1
0
1
-1
-1
1
Table 3. X-matrices in case of an unseen attribute in t1
Adjusting continuous levels of an attribute over time
In some circumstances, researchers wish to change continuous levels of attributes over time (case
#8 in Table 2). Take the attribute price, which rarely stays constant over the years. Adjusting it in
later studies based on market prices makes the experiment more realistic. One solution for handling
changes in the continuous levels is to treat the attribute as two separate attributes: as one newly
added attribute (the new prices) and as one removed attribute (the previous prices). A more elegant
solution may be possible if the researcher can reasonably assume a linear relationship of the
preferences for the continuous levels. Scholars have often made this linearity assumption when
dealing with the price attribute (e.g., Papies et al. 2011; Meyer et al. 2018; Völckner 2008).
Subsequently, for illustration purposes, we consider a product class with two attributes in t1 and
t2. The first attribute is the same in both studies and contains two (effects-coded) levels xi,1. The
second attribute contains (e.g., four) continuous levels, for which we assume linearity in the
corresponding preferences. Furthermore, we assume that all level values in t2 have been shifted
Longitudinal Study in Preferences for Electric Vehicles
21
by , i.e.,     . This shift by a constant factor ensures that we do not modify
the attribute range, such that we can confidently rule out biases, which are commonly associated
with the attribute range effect (Liu et al. 2009; Ohler et al. 2000; Verlegh et al. 2002).
On the individual layer, the deterministic utility function  for alternative i can be written as:
          i,2,t1     (hH; iIa), (4)
where  refers to the individual preferences for the continuous attribute, and w is a binary
indicator that is 1 for observations at time t2 and 0 otherwise. On the upper layer, we write the
deterministic utility vi,t at time t as:
                          
(iIa; t{t1, t2}), (5)
For t1, (i.e., w = 0), equation (5) shrinks to
, 1 0, 1 1, 1 ,1 2, 1 ,2, 1i t t t i t i t
v x x
 
= +  +
. For t2, we substitute
     for j{1;2} and obtain
                (iIa). (6)
Equation (6) outlines two implications on how shifts by the factor of
j
x
affect the results: First,
despite the shift in the level values, the parameter  still reflects the change in preferences for
the updated continuous attribute over time. Consequently, its interpretation of the parameter value
is not affected by the shift. Second, the changes in the level values affect the constant of the
deterministic utility function, such that it must be readjusted. For the study at t1, the constant is
 on the upper layer. For the study at t2, we must account for the shift by readjusting the
Longitudinal Study in Preferences for Electric Vehicles
22
constant according
0, 2 0, 1 0 2, 2 2t t t x
 
= +  +  
, but we can interpret  as if the original values
have been used. Hence,  on the lower layer is also comparable over time.
Empirical study
Background
We chose electric vehicles for the study context because they represent an innovative technology
that has yet to penetrate the market. Electric vehicles are an opportunity to reduce CO2 emissions
and offer a promising alternative to vehicles running on gasoline or diesel in the face of limited oil
resources and climate change. The push for electric vehicle diffusion has been one of the leading
environmental topics for several decades (Graham-Rowe et al. 2012).
In Germany, which constitutes the focal region of our studies, the electric vehicle market was only
1.8% of the total market share in 2019, according to the German Federal Automotive Office. Like
in many other countries, Germany’s federal government has set an official target of at least one
million registered electric vehicles by 2020. In 2016, the government started a financial incentive
program that offered 4,000€ for each bought electric vehicle with a net-list price below 40,000€.
In February 2020, the government increased the financial incentive to 6,000€ and, in June 2020,
to 9,000€. Of the 9,000€, the government contributes 6,000€, and the manufacturer pays the
remaining 3,000€. Nevertheless, given that there were only 179,473 newly registered electric
vehicles as of February 2020 (according to the German Federal Office of Economics and Export
Control), the official target will be missed by the end of 2020. Hence, it is important to observe
and analyze preference development over time to understand consumers’ perceptions of electric
Longitudinal Study in Preferences for Electric Vehicles
23
vehicles and generate reliable estimates of the market potential. After we discuss the market study
results, we will revisit this case and clarify how the findings support managerial decision-making.
Attribute selection
We started with a literature search (see Table A1 in the Web Appendix), focusing on consumer
reports on electric vehicles as external sources, which resulted in a list of the ten most frequently
used attributes. To narrow down the list, we employed the dual questioning approach proposed by
Myers and Alpert (1968). Thereby, we surveyed 251 respondents in advance of the empirical study
and asked for the perceived importance of and difference in the proposed attribute levels (c.f. Hinz
et al. 2015). Based on this, we included the attributes purchase price, electricity cost per 100km,
range per charge, charging time, and motor power. For further information, we refer to the paper
of Hinz et al. (2015), which was published in a prestigious journal.
When discussing electric vehicles with industry experts, we found that IT-enabled complementary
mobility services can substantially foster user acceptance by offering a unique driving experience.
Few studies have included the availability of electric charging stations or permissions to use bus
lanes with an electric vehicle (Hackbarth and Madlener 2016). Still, these attributes are exogenous
to manufacturers. Together with a consultancy agency, we created a list of nine exemplary
complementary services that could provide a unique driving experience. Using the method of best-
worst scaling, case 1 (Louviere et al. 2013), which we included after the dual questioning task, we
ranked these services according to their importance in stimulating purchases.
The top four services, in order from most to least important, are: IT-based parking space and
payment”, “intelligent charging station”, “augmented reality services via head-up displays”, and
“remote diagnostics and update supply”. IT-based parking systems guide drivers to parking spaces
Longitudinal Study in Preferences for Electric Vehicles
24
and enable automatic payment. Intelligent charging stations simplify charging the battery by
automatically identifying drivers and billing the consumed energy. Augmented reality services
project relevant information on the windshield via head-up display, such as navigation, or nearby
charging stations prices and locations. The remote diagnostics and update supply enable the
remote detection of defects during operation to initiate the right measures immediately. Moreover,
a decentralized update supply for the software provides a fundamentally different user experience
compared to conventional cars, for which updates are often carried out locally in workshops. Even
though some of these services might also be available with conventional cars, they embody the
exclusive experience of using electric vehicles and thus have the potential to raise their
attractiveness.
Study setup
We conducted the study in April 2013, June 2017, and November 2019.
3
The questionnaire was
the same with the following exceptions: In 2017, we adjusted the level values of two continuous
attributes to reflect changes in the market offering. In 2019, we randomly assigned respondents to
one of three experimental conditions: (1) either they saw the same survey from 2017, (2) the study
contained an additional attribute, or (3) the study featured both an additional and removed attribute.
In a fourth experimental condition, we separated all respondents who participated in 2017 and re-
invited them to take the same survey. Re-inviting respondents from 2013 was not possible because
changes in the database system prevented the market research firm from matching respondent IDs.
3
For the year 2013, we use the data from Hinz et al. (2015). The studies in 2017 and 2019 are replications of the one
in 2013 to demonstrate our approach and to study the aspects of sample-based longitudinal discrete choice
experiments.
Longitudinal Study in Preferences for Electric Vehicles
25
According to their expert opinion, the overlap was at most 10%. We implemented and executed
the questionnaire with the dual response task (for illustration, see Figure A1 in the Web Appendix)
using the online survey platform DISE (Schlereth and Skiera 2012). Table 4 summarizes the setup
of the studies.
In 2013 and 2017, we used only the top three of the four IT-enabled complementary mobility
services. In 2017, we accounted for the fact that, due to technological improvements, range per
charge improved substantially compared to 2013, while electric vehicles became slightly more
expensive (case #8 in Table 2). Accordingly, to create realistic products in our discrete choice
experiment, we increased all levels of range per charge by 150km and the purchase prices by
5,000€, thereby keeping the attribute ranges constant. In 2019, we kept the levels the same, but
added the attribute “remote diagnostics and update supply and removed “augmented reality
services via head-up displays” in some versions of the study (case #5 and #7 in Table 2).
We constructed 14 choice sets using a D-optimal (4∙2∙2∙4∙4∙2∙2∙2) fractional factorial design that
was specialized for measuring main effects (Street and Burgess 2007). We used this design in all
studies, except for the one version with the additional attribute, which required a different D-
optimal (4∙2∙2∙4∙4∙2∙2∙2∙2) fractional factorial design (case #3 in Table 2). Each choice set
(illustrated in Figure A1 in the Web Appendix) presented three electric vehicles. After choosing
the preferred electric vehicle, respondents had to answer whether they would buy the chosen
electric vehicle or not.
The two types of questions are known as dual response (Brazell et al. 2006; Schlereth et al. 2018).
This variant overcomes a common disadvantage of including the no-purchase option in every
choice set because we now also observe respondents’ trade-offs between the attributes even in
cases where respondents would rather refrain from buying one of the presented products.
Longitudinal Study in Preferences for Electric Vehicles
26
2013
2017
2019_1_same
2019_2_add
2019_3_add_remo
ve
2019_retakers
Description
Original study
Changes in the
continuous levels
range per charge
and purchase price
Same study as the
one in 2017; new
respondents
One attribute has
been added in
comparison to study
2017; new
respondents
One attribute has
been added and one
removed in
comparison to study
2017; new
respondents
Same study as the
one in 2017; sent
out to the same
respondents
Number of Respondents
327
553
764
680
740
305
No intention in
2017: 172
Intention in 2017:
133
Number of Respondents
with General Purchase
Intention (%)
168 (51.38%)
326 (58.95%)
301 (39.40%)
281 (41.32%)
312 (42.16%)
123 (40.33%)
No intention in
2017:
14 (10.53%)
Intention in 2017:
109 (63.37%)
Attributes
Unit
Levels
Levels
Levels
Levels
Levels
Levels
Range per charge
km
100; 175; 250; 325
250; 325; 400; 475
250; 325; 400; 475
250; 325; 400; 475
250; 325; 400; 475
250; 325; 400; 475
Charging time
h
1; 4
1; 4
1; 4
1; 4
1; 4
1; 4
Motor power
kW
40; 80
40; 80
40; 80
40; 80
40; 80
40; 80
Purchase price
15,000; 20,000;
25,000; 30,000
20,000; 25,000;
30,000; 35,000
20,000; 25,000;
30,000; 35,000
20,000; 25,000;
30,000; 35,000
20,000; 25,000;
30,000; 35,000
20,000; 25,000;
30,000; 35,000
Electricity cost per 100km
1; 3; 5; 7
1; 3; 5; 7
1; 3; 5; 7
1; 3; 5; 7
1; 3; 5; 7
1; 3; 5; 7
IT-based parking space and
payment
supported;
not supported
supported;
not supported
supported;
not supported
supported;
not supported
supported;
not supported
supported;
not supported
Intelligent charging station
supported;
not supported
supported;
not supported
supported; not
supported
supported;
not supported
supported;
not supported
supported;
not supported
Augmented reality services
via head-up displays
supported;
not supported
supported;
not supported
supported;
not supported
supported;
not supported
-
supported;
not supported
Remote diagnostics and
update supply
-
-
-
supported;
not supported
supported;
not supported
-
Table 4. Description of our studies, number of respondents, attributes, and attribute level
Longitudinal Study in Preferences for Electric Vehicles
27
The questionnaire showed some general information about electric vehicles and asked respondents
about their gender, age, and income. Afterward, they were asked: “Can you imagine purchasing
an electric vehicle?” Only respondents with a general purchase intention for electric vehicles
entered the discrete choice experiment. If they claimed that they could not imagine purchasing an
electric vehicle or chose instead to use one as part of a car-sharing service, their survey directly
ended. We assumed that their willingness-to-pay would be lower than the lowest purchase price in
the experiment. We hired the same market research firm for all studies to select respondents with
the same demographic criteria to create comparable samples.
Results
General purchase intention
Table 4 summarizes respondents general purchase intention in each study. In 2013, we collected
327 completed questionnaires, and 51.38% of the respondents claimed to have a general purchase
intention. In 2017, of the 553 respondents, 58.95% entered the discrete choice experiment. In 2019,
we randomly assigned 2,184 respondents to one of the three study versions, whereby only 40.96%
indicated their general purchase intention. This strong decline is consistent across all the 2019
questionnaires (39.40%, 41.32%, and 42.16%).
We can only speculate about the reasons, but the observed decrease is backed up by the Gartner
hype cycle for connected vehicles from July 2019, in which experts placed electric vehicles at the
trough of disillusionment (Ramsey 2019). In 2019, electric vehicles still had a low market share of
1.8% in the focal country. As more information about the experience of using electric vehicles
became available, their practical problems also became more apparent. As noted by Jensen et al.
(2014), the purchase intention for electric vehicles decreased after the individual trial periods.
Longitudinal Study in Preferences for Electric Vehicles
28
We also note the extent to which we succeeded in re-inviting the 2017 respondents. In 2019, 387
of the 553 respondents from 2017 were still active in the panel, and 305 respondents accepted the
invitation (55.15%). Their general purchase intention substantially declined to 40.33%a share
that aligns with the percentage of the new respondents in 2019. The percentage is 63.37% among
those respondents who also indicated a general purchase intention in 2017, and 10.53% among the
others. We conclude that, in a time when electric vehicles became more and more available,
respondents’ tendency to purchase them decreased more than increased.
We next tested for differences in gender, age, and income. Using multiple t-tests with a Bonferroni
correction, we found no significant differences for gender in any comparisons (all p>.1).
Concerning age, the 2019 group of retakers was significantly older than they were in 2017 (p<.01);
while we did expect this result, we found no significant differences in age when compared to the
three other groups of new respondents in 2019. Further, for income, we did not observe any
significant difference between the 2017 and 2019 retakers (p>.1), as well as between the three
groups of new respondents in 2019 (all p>.1). Yet the 2019 retakers indicated that they earned
significantly more (p<.01) than the other respondents in 2019, such that among all the tests, the
only one that could indicate a problem of the sample’s representativeness is the significant income
difference between 2017 and 2019. It could also indicate some changes over time in the general
population. In the Web Appendix, we provide robustness tests, which rule out that the differences
in income systematically biased our results.
Model comparison
We assumed a linear relationship for the purchase price and range per charge (c.f., Table A1 in the
Web Appendix). We estimated purchase price per 10,000€ and range per charge per 100km to
Longitudinal Study in Preferences for Electric Vehicles
29
ensure that all X-matrix values vary with a similar magnitude and that prior distributions are
consistent across parameters. In total, we ran 7,000 iterations, of which we used the last 2,000
iterations to estimate the posterior. We visually assessed the convergence of the posterior in the
burn-in phase using trace plots of the likelihood function.
We compared the log-marginal density (LMD) of the covariate-extended hierarchical Bayes model
against the separate estimation and the generic model. The estimation that used most of the
parameters on the upper layer, i.e., the separate estimation with 475 parameters (= 5·11 + 1·12
parameters plus 5·(11*(11+1)/2) + 12*(12+1)/2 for the six covariance matrices), performed best
concerning internal validity (LMD of -16297, see Table A2 in the Web Appendix). It is followed
by the model with the second most parameters, i.e., the covariate-extended model (145 parameters;
LMD of -16410) and the model with the least number of parameters, i.e., the generic one (90
parameters; LMD of -16676) performs worst.
Yet an improvement in fitting the choices that were used for the estimation does not necessarily
equal a better model. Overfitting can occur if these parameters do not improve the fit to the choices
held out for validation (Ellickson et al. 2019; Orme and Howell 2009). For the two holdouts, the
covariate-extended model (-4765) edges out the separate (-5953) and generic (-4940) models. The
separate estimation performs worst, such that the improved internal model fit does not translate
into better predictions. To assess the size of the differences, we adapted a test from Ellickson et al.
(2019), which entails leaving out 100 respondents and predicting the mean absolute differences of
their observed choice shares. Even though the covariate-extended model is best here, the
differences are rather small (between 6.73% and 6.87%).
Orme and Howell (2009, p. 19) wrote that predictions are not the main reason for using covariate
models, but they “allow us to test more formally […] the differences between segments and the
Longitudinal Study in Preferences for Electric Vehicles
30
part-worth.” And indeed, we can see the main advantage of using the covariate-extended model,
as this modeling approach best aligns with the goals of a sample-based longitudinal study. These
goals are to link multiple studies within one estimation (which the separate estimation cannot do)
and to explicitly capture differences in preferences across samples (which the generic model
cannot do). Besides, it enables researchers to formally test whether changes in preferences across
samples are substantial or merely noise: We simply inspect the signs of the posterior draws for
each parameter in the θ-matrix (Orme and Howell, 2009). The matrix contains the upper layer’s
changes in preferences relative to a reference study (Equation 2); hence, when >95% or <5% of
draws are positive, these covariate weights are significantly different from zero, at or better than
the 90% confidence level (two-sided test). We will subsequently demonstrate the value of such a
test.
Changes in preferences
In Table 5, we report the results of the covariate-extended model. All signs and magnitudes are
consistent and reasonable, indicating high face validity. Table 5 also reports the percentage of
positive and negative draws with the 2017 study as a reference, when the general purchase interest
in electric vehicles peaked. For example, when comparing the 2013 study results to the ones of
2017, the preferences that largely changed related to the electric vehicle-specific attributes, but not
to the complementary mobility services. Respondents in 2013 were more price-sensitive, with a
greater emphasis on charging time (the 1h attribute level had a positive change and 4h a
negative one) and less emphasis on motor power (the less-preferred 40KW had a positive change
and the more-preferred 80KW had a negative one).
Longitudinal Study in Preferences for Electric Vehicles
31
Table 5 demonstrates the benefit of inspecting the signs of the posterior draws of the θ-matrix: A
pragmatic approach would have been to compare the absolute or relative differences in parameter
values between two samples. However, in Table 5, we observe that large differences do not
automatically imply that the two samples differ in preferences. For instance, a rather large
difference has the constant between 2017 and the 2019 retakers (i.e., 0.65 = -1.34 (-1.99)), and
a rather small difference has the attribute intelligent charging station (i.e., 0.10 = 0.63 0.53). Yet,
the θ draws only statistically support that the preferences for the complementary mobility service
intelligent charging stations decreased; there was no statistical support for a change of the
constant. Hence, looking only at the value of parameters (as with the separate estimation) might
provide misleading insights.
Figure 1 compares and graphically displays the average attribute importance weights (Louviere
and Islam 2008) over time. Economic considerations (i.e., electricity costs and purchase price)
dominated choice decisions. This is particularly interesting because electricity costs, which were
the most important attribute throughout the three studies, are rather neglected in communication
about electric vehicles. Range per charge came in third: It gained importance in 2017, but lost
some in 2019. Obviously, it was initially important to be able to overcome a certain range, which
was still too small in 2013. Considering that we have adjusted the range per charge between 2013
and 2017, the benefit of the longer range decreases after 2017 as the market-observed ranges
increase further. Overall, the importance weights for the attributes remained mostly stable for over
six years. There were some position changes, but rarely by more than one rank.
Longitudinal Study in Preferences for Electric Vehicles
32
2013
2017
(reference)
2019_1_same
2019_2_add
2019_3_add_remove
2019_4_retakers
Attributes
Attribute
level
Parameter
Values
Change
Parameter
Values
Parameter
Values
Change
Parameter
Values
Change
Parameter
Values
Change
Parameter
Values
Change
Constant
-1.35
-1.99
-1.54
-1.20
+
-1.01
++
-1.34
Range per charge
Per 100 km
0.71
0.70
0.60
-
0.63
0.62
0.69
Purchase price
per 10,000€
-1.18
-
-0.94
-1.10
-
-1.18
--
-1.15
--
-1.15
Charging time
1h
0.29
++
0.14
0.24
+
0.26
++
0.28
+++
0.28
++
4h
-0.29
--
-0.14
-0.24
-
-0.26
--
-0.28
---
-0.28
--
Electricity cost
per 100 km
1 €
1.07
0.99
0.96
0.93
1.04
0.90
3 €
0.48
++
0.34
0.53
+++
0.44
+
0.42
+
0.45
5 €
-0.30
-0.32
-0.43
-
-0.27
-0.38
-0.33
7 €
-1.25
--
-1.02
-1.06
-1.10
-1.08
-1.02
Motor power
40 kW
-0.41
+
-0.52
-0.43
+
-0.25
+++
-0.38
+++
-0.42
80 kW
0.41
-
0.52
0.43
-
0.25
---
0.38
---
0.42
IT-based parking
space and
payment
Supported
0.42
0.50
0.40
--
0.29
---
0.36
---
0.44
Not
supported
-0.42
-0.50
-0.40
++
-0.29
+++
-0.36
+++
-0.44
Intelligent
charging station
Supported
0.58
0.53
0.53
0.32
---
0.47
0.63
+
Not
supported
-0.58
-0.53
-0.53
-0.32
+++
-0.47
-0.63
-
Augmented
reality services
via head-up
displays
Supported
0.20
0.22
0.22
0.24
0.15
-
Not
supported
-0.20
-0.22
-0.22
-0.24
-0.15
+
Remote
diagnostics and
update supply
Supported
0.21
+++
0.26
+++
Not
supported
-0.21
---
-0.26
---
Note: posterior assessment of change in parameters in comparison to 2017 study: +: >90%; ++: >95%; +++:>99%; -:<10%; --: <5%; ---: <1% positive Δ draws
Table 5. Average parameter values based on covariate-extended model
Longitudinal Study in Preferences for Electric Vehicles
33
Note: Lower and upper bound of 95% confidence intervals in [brackets]
Figure 1. Average importance weights and key findings
Summary of the Key Findings:
Importance weights remain mostly stable over six years
Economic considerations (electricity cost, purchase price) dominate choice decisions; Range per charge ranked third
The most important attribute is electricity costs, which is rarely mentioned in the communication of an electric vehicle
Additional attribute does not draw importance from the higher-ranked attributes, but mainly from the lower-ranked attributes
Retakers: higher purchase price sensitivity but lower for electricity costs in 2019 compared to 2017
Longitudinal Study in Preferences for Electric Vehicles
34
Adding and removing an attribute
Next, we examine how the inclusion of the ninth attribute affected preferences for the other
attributes. Thereby, we used only the experimental conditions of the 2019 study and reran the
estimation, using 2019_1_same as a reference. Table A3 in the Web Appendix contains the
posterior mean for each study, together with a test on the percentage of positive and negative θ
draws indicating whether differences are substantial or mere noise.
We observe that the fourth complementary mobility service (2019_2_add vs. 2019_1_same)
proportionally drew its importance weight mostly from the four lower-ranked attributes, i.e., the
other three complementary mobility services and motor power. All other attributes are unaffected.
This may be because the new variable itself is one of the lower-ranked attributes. When the fourth
complementary mobility service replaced the head-up display attribute (2019_3_add_remove vs.
2019_1_same), the other importance weights did not change on a significant level.
Preferences of retakers vs. new sample of respondents
Using Table 5 and Table A3 (in the Web Appendix), we compare the preferences of the retakers
(2019_4_retakers) against their preferences in 2017 and the new sample of respondents
(2019_1_same). In comparison to 2017, the attribute charging time is the one that changed on a
substantial level by gaining importance. There were also some changes for the complementary
mobility services intelligent charging station” and “augmented reality services via head-up
displays; however, we consider these changes as minor because the former one slightly gained in
importance and the latter lost some. When comparing the sample of retakers against the same study
with completely new respondents (2019_1_same), we detected nearly no differences in the
attribute parameters.
Longitudinal Study in Preferences for Electric Vehicles
35
Overall, we conclude that it is valuableif possibleto invite the same respondents again. There
is great value in the potential to modify certain aspects in the selected list of attributes and levels,
as proposed by the sample-based longitudinal discrete choice experiment. However, if it is not
possible, then entirely new respondents provided similar insights in our study.
Changes in purchase probabilities
In the last column of Table 6, we report the share of purchase decisions. Whereas the general
purchase intention decreased from 2017 to 2019, the share of purchase decisions increased
monotonously among those respondents with a general purchase intention.
Status Quo
40,000€
→ 36,000€
(government’s
financial incentive
2016)
40,000€
→ 31,000€
(government’s
financial
incentive 2020)
5€ per 100
km
3€ per
100 km
250 km range
325 km
range
IT-based
parking space
and payment
2013
4.60%
5.04%
5.89%
5.76%
5.29%
5.96%
2017
7.21%
7.51%
8.12%
8.13%
8.30%
9.25%
2019
(average across
the 3 versions)
4.38%
4.68%
5.23%
5.39%
5.02%
5.28%
2019_retakers
3.87%
4.35%
5.09%
4.57%
4.61%
5.27%
Intelligent
charging
station
Augmented
reality services
via head-up
displays
Remote
diagnostics and
update supply
With top
two
mobility
services
Share of Purchase Decisions
2013
6.37%
5.25%
4.60%
8.07%
42.64%
2017
9.29%
8.30%
7.21%
11.47%
45.33%
2019
(average across
the 3 versions)
5.57%
4.80%
4.88%
6.67%
46.03%
2019_retakers
5.77%
4.67%
3.87%
7.51%
45.88%
Status Quo: range per charge: 250km; charging time: four hours; motor power: 40kW; purchase price: 40,000€;
electricity cost per 100km: 5€, no complementary mobility services
Table 6. Counterfactual simulations on purchase probabilities
The subsequent counterfactual simulation combines the joint effect of the two trends. We created
a stylized scenario, in which we report the purchase probabilities for a “status quo” scenario. We
derived this scenario by analyzing a basic electric vehicle offer (here: range per charge of 250km;
Longitudinal Study in Preferences for Electric Vehicles
36
charging time of four hours; motor power of 40kW; purchase price of 40,000€; electricity cost per
100km of 5€, and no complementary mobility services) and then alternately varied several attribute
levels one by one. We summarize the results in Table 6.
4
For each scenario, we averaged the
individual purchase probability by employing the second term in Equation (A1) (c.f., Technical
Appendix in the Web Appendix) and then multiplied this probability by the average purchase
intention of the respective year. For better readability, we combined the three versions in 2019.
In all scenarios, purchase probabilities increased from 2013 to 2017, but decreased in 2019. In
2013, only 4.60% of respondents would buy the status quo product, compared to 7.21% in 2017
and 4.38% in 2019. Considering the 4,000€ governmental incentive in 2016, we observed a
purchase probability of 4.68% in 2019, but 5.23% when the subsidy increased further to 9,000€ in
June 2020. The additional 5,000€ subsidy increases the probability of buying an electric vehicle
by +.55% or, in absolute figures, by 19,855 electric vehicles (i.e., .55% * 3.61 million newly
registered vehicles in Germany in 2019, according to Statista).
Comparing the predictions of the counterfactual between the new 2019 samples and the retakers,
we also observed only minor differences. These results further suggest that surveying a new sample
vs. asking respondents from previous studies provides comparable insights.
4
We replicate the counterfactual simulation for the separate and generic estimation. The market shares of the generic
estimation are close to the ones of Table 6 with a mean absolute difference in the percentages of .39% (std.dev. .18%).
For the separate estimation, the percentage differences between each simulated condition are similar in size compared
to the ones of Table 6. However, the overall market shares discriminate more between samples: the separate 2017
shares are on average .78% (std.dev. .79%) higher and the other market shares are 1.57% (.47%) lower.
Longitudinal Study in Preferences for Electric Vehicles
37
Model-based managerial decision
Finally, we used the 2019 estimates to present an illustrative optimization problem for automobile
manufacturers. Since June 2020, the government supports the purchase of an electric vehicle with
a discount of 9,000€. This discount comprises a 6,000€ federal subsidy and an obliged
manufacturer’s share of 3,000€. The market simulation in Table 6 assumed that the car
manufacturer passes on the whole discount directly to the consumer. A more realistic view is that
the manufacturer would have offered a voluntary discount, even without the governmental
program, and thus must decide on how much of the 9,000€ it should pass on to its customers to
maximize profits. This maximization problem is also relevant beyond the context of electric
vehicles. Essentially, the 6,000€ federal subsidy represents a reduction of the manufacturer’s
variable costs. The underlying general decision problem is: What percentage should a
manufacturer pass on, in case the variable costs decrease?
We assumed a list price l and an average discount d, which would have been offered without an
incentive program, such that the actual purchase price p is p = l - d. Since manufacturers vary in
their discounts, we examined the optimal decision-making of a car manufacturer under different
pricing strategies. While Tesla, for example, communicates a zero-discount strategy, other
manufacturers usually grant a certain discount, although the exact amount is not known and varies
between manufacturers. Thus, we distinguished three pricing strategies in order to determine the
effects of a subsidy. We considered a manufacturer who, with a previous discount of 11,000€, has
already granted a discount of more than 9,000€ before the government subsidy. As a second pricing
strategy, we assumed a previous discount of less than 9,000€ (i.e., 7,000€). Finally, we also
considered a manufacturer that does not offer a discount at all.
Longitudinal Study in Preferences for Electric Vehicles
38
Assuming a manufacturer’s margin m, the variable costs cvar without governmental support are cvar
= p x (1 m). The new regulations of the federal government reduce the variable costs by 6,000€,
such that cvar* = p* x (1 m) 6000€, where p* is the new purchase price corresponding to the new
discount d*. Let customer h’s purchase probability equals (c.f., Equation (A.1)):
( )
exp( ( ))
Pr ( ) 1 exp( ( ))
h
hh
vp
pvp
=+
. For better readability, we omitted the index i for an electric vehicle.
The maximization problem is:
         
      
 (7)
s.t.  (8)
  (9)
The decision variable in the model (7) - (9) is the new discount d*. The objective function (7)
maximizes the profit π of an electric vehicle manufacturer. The constraint in (8) ensures that at
least 9,000€ will be passed on to customers, and (9) ensures that the manufacturer offers at least
the discount it would have offered, even without governmental support.
Besides the pricing strategy (i.e., the discount), we also vary the margins in two levels (5% and
20%). Thus, we examine the maximization problem for a total of six settings. For this analysis, we
map the characteristics of one of the best-selling electric vehicles in Germany in 2019 (according
to Statista) to the corresponding attribute levels of our study design (List price: 35,900€; range per
charge: 300km, charging time: 1h; motor power: 80 kW; electricity cost per 100km: 7€, none of
the complementary mobility services). In Table 7, we report the results for the (3×2=) six settings.
Longitudinal Study in Preferences for Electric Vehicles
39
Purchase price p
without
governmental
support
Discount d
without
governmental
support
Margin
per
electric
vehicle
Original average
purchase
probability
Pr(p)
Optimal new
purchase
price p*
Optimal
discount d*
Change in
average purchase
probability
New margin per
electric vehicle
Required change in
purchase probability to
compensate for passing
on an additional 1,000€
24,900€
11,000€
5%
18.05%
24,900€
11,000€
+0%
29.10%
+2.89%
24,900€
11,000€
20%
18.05%
24,900€
11,000€
+0%
44.10%
+1.81%
28,900€
7,000€
5%
16.10%
26,900€
9,000€
+0.93%
20.24%
+3.83%
28,900€
7,000€
20%
16.10%
26,900€
9,000€
+0.93%
36.36%
+1.94%
35,900€
0€
5%
13.62%
26,900€
9,000€
+3.41%
-4.48%
not applicable
35,900€
0€
20%
13.62%
26,900€
9,000€
+3.41%
15.54%
+5.35%
Table 7. Optimal discount strategy for manufacturers with new governmental program
Longitudinal Study in Preferences for Electric Vehicles
40
The model recommends the minimum discount in all settings, determined by the constraints (2) and
(3). Thus, manufacturers with previous discounts lower than 9,000€ (e.g., 7,000€ or no discount at
all) should now offer a discount of 9,000€, and manufacturers with a previous discount greater than
9,000€ should stick to their previous discount (i.e., 11,000€). The pricing policy of the manufacturer
determines how much it benefits from the governmental incentive program. Manufacturers, who
charged list price without a discount before the program, face a decrease in their margin or even losses.
Their rationale for taking part in such a program is probably rather due to reasons of competitiveness.
Those with originally large discounts benefit from the federal subsidy. The larger the original margin,
the more they benefit.
The results indicate that the purchase probability does not increase sufficiently if manufacturers partly
pass on the federal subsidy. This observation implies that the price sensitivity is too low (respectively,
the parameter for the purchase price is too close to zero). The electric vehicle’s features and the
consumers’ general attitudes toward buying one mainly drive the decision to buy an electric vehicle
the price is less of an issue, thereby indicating a price-inelastic demand. To compensate for a passed-
on discount of 1,000€, the average purchase probabilities would have to change by at least 1.81% to
increase the manufacturer’s profits (see the last column in Table 7).
Although the results of our longitudinal study indicate an increase in the price sensitivity between
2017 and 2019 (see Table 5), a sufficient increase in purchase probabilities is currently out of reach.
From today's perspective, the governmental incentive program encourages consumers to buy the
electric vehicle before the program runs out, but it is barely able to stimulate additional purchases.
Nevertheless, it is important to observe whether changes will occur to the price sensitivity in the near
future.
Longitudinal Study in Preferences for Electric Vehicles
41
Conclusion
Methodological contribution and contribution to theory
Discrete choice experiments are well known, but they are rarely used to study preferences over time.
In this research, we proposed a sample-based longitudinal discrete choice experiment, together with
the covariate-extended hierarchical Bayes logit estimator, to track changes in preference over time
using different samples of respondents. We also structured all elements that can change from a
respondent’s and the researchers perspective in order to guide researchers in their modeling. Hence,
the sample-based longitudinal discrete choice experimental approach can be easily transferred to other
research topics.
When examining the performance of the covariate-extended model, we found that its utility is
motivated more by theoretical arguments and its statistical testing ability than by its potential gains in
internal and predictive validity (in line with Orme and Howell 2009; Sentis and Geller 2010). We
consider its testing ability to be the core benefit because it answers a range of questions that
researchers typically have when conducting longitudinal studies. One of these questions is whether it
is appropriate to survey different samples of respondents. Here, we statistically observed that upper
layer preferences do not differ for the new sample relative to the respondents from previous studies.
While we cannot generalize from this single study, this finding suggests that interviewing new
participants is viable. At the same time, we demonstrated the need to ask different respondents in later
studies: After 2.5 years, only 55.15% of the respondents from 2017 responded to our survey invitation
in 2019. Without our approach, it would have been necessary to gather large, costly sample sizes at
the beginning of the studies to guarantee an appropriate sample size for later studies. Our approach
relaxes this requirement. Thus, it reduces the expenses for conducting longitudinal studies and even
Longitudinal Study in Preferences for Electric Vehicles
42
allows researchers to expand sample sizes in later studiesfor example, to experimentally test for
different attribute specifications, as we did in the 2019 study.
Concerning the testing ability of the covariate-extended model, we showed that statistically testing
for differences is necessary; looking only at absolute or relative changes in parameter values can be
misleading in detecting significant changes. Thereby, our approach can be more general than
comparing groups of respondents over time: It enables a new level of flexibility in analyzing multiple
studies that only partially overlap. For example, a globally operating company has data sets available
at not only several points in time, but also for different groups (e.g., different market regions like U.S.
vs. European car buyers) that the management would like to combine and evaluate. For these groups,
different attributes or attribute characteristics are conceivable depending on the country, and
researchers can now account for them separately in our proposed model.
Moreover, adding and changing attributes enables researchers to react to future changes in the market.
There are manifold reasons why the ability to change attributes may be beneficial to a longitudinal
study (c.f., Table 2). For example, the importance of some attributes may not be recognized until later,
or new technology features arise after the longitudinal study begins. We explicitly considered this
prospect in our modeling and provided a theoretical base, with explicit assumptions, about how
respondents of earlier studies take unseen information into account. We also help explain changes in
the continuous levels of attributes by outlining that they mostly affect the constant, but not the
parameter related to the changed attribute. Concerning adding attributes, we experimentally
demonstrated that their percentage in importance weights were not drawn proportionally from all other
attributes; rather, some of the attributes were unaffected. Having knowledge about which attributes
are affected can support product managers in their decision-making and communication.
Longitudinal Study in Preferences for Electric Vehicles
43
A final finding relates to the evolution of preferences for innovations. As one of the very few who
have conducted a longitudinal study over a longer period of time, we show that preferences do not
necessarily evolve monotonously. Accordingly, forecasting models that extrapolate the development
of preferences into the future are rather difficult. Longitudinal studies are all the more suitable for this
purpose.
Managerial contribution
Let us return to the previously mentioned manufacturer ACME, which might use these analyses and
the corresponding information for future management decisions. Based on the first two studies in 2013
and 2017, ACME might have made larger investments since 2017, as the curve showed a strong
growth trend promising a constant and rapid breakthrough of the technology. However, since we
observed an inverted U-shaped curve with a peak in 2017 as the overall effect over the entire period,
these investments would not have paid off to the extent hoped for so far. The sales market for electric
cars has continued to grow, but not to the extent that firms expected. With the third major study in
2019, the manufacturer would gain more up-to-date information about the market and adjust its
marketing measures accordingly. According to the Gartner hype cycle, innovation diffusion can
fluctuate in the pre-mass market phase, which is why not all investments should be discontinued. It
may be that the breakthrough to the plateau of productivity is imminent.
Notably, the preferences for the different attributes remained largely constant. Price sensitivity was
relatively low, although electricity costs and purchase price were the most important attributes. From
this, we can deduce that manufacturers should more directly communicate information about
electricity costs, as these are consistently among the most important attributes. The low price
sensitivity suggests that the purchase price is not that important: People either have strong preferences
Longitudinal Study in Preferences for Electric Vehicles
44
toward owning an electric vehicle or continuing to drive a conventional fuel-based vehicle. A potential
reason for this is that the media excessively covers the problems of electric vehicles. For example,
people might hear about the low number of charging stations in Germany, or that small accidents can
damage the battery—the car’s main value—and lead to a total loss relatively quickly. Moreover,
prospective buyers were concerned about batteries setting fire to the entire electric vehicle or the
garage in the house and that the monetary advantage of an electric vehicle may be lost when charging
at commercially operated charging stations. Finally, critical reports question the environmental
friendliness of electric vehicles, e.g., if the electricity is generated in coal-fired power plants. In sum,
researchers can use sample-based longitudinal studies to analyze the overall perception of an
innovation within the population and, above all, test possible solutions that will achieve acceptance
among end customers.
Limitations
Like any research, our study features some limitations. First, we only studied one product in our
longitudinal study. We encourage future studies to examine other products over a longer time to
observe possible similarities or differences in terms of product development or market penetration.
By including different product categories or product lifecycles, future studies could address cases #4
and #6 (Table 2), which we left out empirically.
Second, the available options were characterized solely by their attribute levels. The preference and
purchase decisions were, therefore, hypothetical. This is exacerbated by the fact that the market
penetration of electric vehicles in the focal country is still very low, so our respondents may have felt
a lack of experience with the product under consideration.
Longitudinal Study in Preferences for Electric Vehicles
45
Third, in using a D-efficient design, we only measured the main effects and could not observe
interaction effects. This limitation probably applies to almost all discrete choice experiments. Some
authors have proposed designs that can also handle a selected set of attribute-interaction effects (e.g.,
Yu et al. 2006). However, we are not aware of any research that applies them in an empirical setting,
as they require a substantially higher number of choice setsa situation we wanted to avoid for our
study. Follow-up studies could conceivably address this limitation.
Finally, our study results suggest that surveying the same respondents over time provides similar
insights compared to questioning a new sample of respondents. Future research could challenge the
boundary conditions and the generalizability of this finding. Research on biases, such as the mere
exposure effect (e.g., Zajonc 1968), could serve as a suitable starting point together with dynamic
models that allow for individual-level heterogeneity around an aggregated trend (e.g., Liechty et al.
2005). Another boundary condition is the number of changed attributes and levels, i.e., how much
change is possible to guarantee a certain level of temporal stability and structural reliability of the
results.
Acknowledgements The authors gratefully thank Luigi Bianco for his support in data collection for
the 2013 study and Prof. Dr. Bernd Skiera for his help and feedback during the colloquium. This work
has been funded by the German Research Foundation (DFG) within the Collaborative Research Center
(CRC) 1053 - MAKI.
References
Agarwal, J., DeSarbo, W. S., Malhotra, N. K., & Rao, V. R. (2015). An Interdisciplinary Review of
Research in Conjoint Analysis: Recent Developments and Directions for Future Research.
Customer Needs and Solutions, 2, 1940.
Longitudinal Study in Preferences for Electric Vehicles
46
Ambos, T. C., Cesinger, B., Eggers, F., & Kraus, S. (2019). How does de-globalization affect
location decisions? A study of managerial perceptions of risk and return. Global Strategy
Journal, 16, 210236.
Bettman, J. R., Luce, M. F., & Payne, J. W. (1998). Constructive Consumer Choice Processes.
Journal of Consumer Research, 25, 187217.
Bhattacherjee, A. (2001). Understanding Information Systems Continuance: An Expectation-
Confirmation Model. MIS Quarterly, 25, 351370.
Bradlow, E. T., Hu, Y., & Ho, T.-H. (2004). A Learning-Based Model for Imputing Missing Levels
in Partial Conjoint Profiles. Journal of Marketing Research, 41, 369381.
Branco, F., Sun, M., & Villas-Boas, J. M. (2012). Optimal Search for Product Information.
Management Science, 58, 20372056.
Brazell, J. D., Diener, C. G., Karniouchina, E., Moore, W. L., Séverin, V., & Uldry, P.-F. (2006).
The No-Choice Option and Dual Response Choice Designs. Marketing Letters, 17, 255268.
Coupey, E., Irwin, J. R., & Payne, J. W. (1998). Product Category Familiarity and Preference
Construction. Journal of Consumer Research, 24, 459468.
Dodds, W. B., Monroe, K. B., & Grewal, D. (1991). Effects of Price, Brand, and Store Information
on Buyers' Product Evaluations. Journal of Marketing Research, 28, 307319.
Ellickson, P. B., Lovett, M. J., & Ranjan, B. (2019). Product Launches with New Attributes: A
Hybrid ConjointConsumer Panel Technique for Estimating Demand. Journal of Marketing
Research, 56, 709731.
Fourt, L. A., & Woodlock, J. W. (1960). Early Prediction of Market Success for New Grocery
Products. Journal of Marketing, 25, 3138.
Gensler, S., Hinz, O., Skiera, B., & Theysohn, S. (2012). Willingness-to-pay estimation with choice-
based conjoint analysis: Addressing extreme response behavior with individually adapted
designs. European Journal of Operational Research, 219, 368378.
Gilbride, T. J., Currim, I. S., Mintz, O., & Siddarth, S. (2016). A Model for Inferring Market
Preferences from Online Retail Product Information Matrices. Journal of Retailing, 92, 470485.
Graham-Rowe, E., Gardner, B., Abraham, C., Skippon, S., Dittmar, H., Hutchins, R., et al. (2012).
Mainstream consumers driving plug-in battery-electric and plug-in hybrid electric cars: A
qualitative analysis of responses and evaluations. Transportation Research Part A: Policy and
Practice, 46, 140153.
Hackbarth, A., & Madlener, R. (2016). Willingness-to-Pay for Alternative Fuel Vehicle
Characteristics: A Stated Choice Study for Germany. Transportation Research Part A: Policy
and Practice, 85, 89111.
Hinz, O., Schlereth, C., & Zhou, W. (2015). Fostering the adoption of electric vehicles by providing
complementary mobility services: a two-step approach using BestWorst Scaling and Dual
Response. Journal of Business Economics, 85, 921951.
Jensen, A. F., Cherchi, E., & Dios Ortúzar, J. de. (2014). A long panel survey to elicit variation in
preferences and attitudes in the choice of electric vehicles. Transportation, 41, 973993.
Kurz, P., & Binner, S. (2010). Added Value through Covariates in HB Modeling? Proceedings of
the Sawtooth Software Conference.
Lenk, P. J., DeSarbo, W. S., Green, P. E., & Young, M. R. (1996). Hierarchical Bayes Conjoint
Analysis: Recovery of Partworth Heterogeneity from Reduced Experimental Designs. Marketing
Science, 15, 173191.
Liakhovitski, D., & Shmulyian, F. (2011). Covariates in Discrete Choice Models: Are They Worth
the Trouble? ART Forum.
Longitudinal Study in Preferences for Electric Vehicles
47
Liechty, J. C., Fong, D. K. H., & DeSarbo, W. S. (2005). Dynamic Models Incorporating Individual
Heterogeneity: Utility Evolution in Conjoint Analysis. Marketing Science, 24, 285293.
Linden, A., & Fenn, J. (2003). Understanding Gartner's Hype Cycles. Retrieved January 28, 2020
from
https://www.bus.umich.edu/KresgePublic/Journals/Gartner/research/115200/115274/115274.pdf.
Liu, Q., Dean, A., Bakken, D., & Allenby, G. M. (2009). Studying the level-effect in conjoint
analysis: An application of efficient experimental designs for hyper-parameter estimation.
Quantitative Marketing and Economics, 7, 6993.
Louviere, J., Lings, I., Islam, T., Gudergan, S., & Flynn, T. (2013). An introduction to the
application of (case 1) bestworst scaling in marketing research. International Journal of
Research in Marketing, 30, 292303.
Louviere, J. J., Hensher, D. A., & Swait, J. D. (2000). Stated choice methods: Analysis and
applications. Cambridge: Cambridge University Press.
Louviere, J. J., & Islam, T. (2008). A comparison of importance weights and willingness-to-pay
measures derived from choice-based conjoint, constant sum scales and bestworst scaling.
Journal of Business Research, 61, 903911.
Lucker, J., Hogan, S. K., & Sniderman, B. (2018). Fooled by the hype: Is it the next big thing or
merely a shiny new object? Deloitte Review, 23, 8495.
McCullough, J., & Best, R. (1979). Conjoint Measurement: Temporal Stability and Structural
Reliability. Journal of Marketing Research, 16, 2631.
Meeran, S., Jahanbin, S., Goodwin, P., & Quariguasi Frota Neto, J. (2017). When do changes in
consumer preferences make forecasts from choice-based conjoint models unreliable? European
Journal of Operational Research, 258, 512524.
Meyer, J., Shankar, V., & Berry, L. L. (2018). Pricing hybrid bundles by understanding the drivers
of willingness to pay. Journal of the Academy of Marketing Science, 46, 497515.
Myers, J. H., & Alpert, M. I. (1968). Determinant Buying Attitudes: Meaning and Measurement.
Journal of Marketing, 32, 1320.
Ohler, T., Le, A., Louviere, J., & Swait, J. (2000). Attribute Range Effects in Binary Response
Tasks. Marketing Letters, 11, 249260.
Oliver, R. L. (1980). A Cognitive Model of the Antecedents and Consequences of Satisfaction
Decisions. Journal of Marketing Research, 17, 460469.
Orme, B., & Howell, J. (2009). Application of Covariates Within Sawtooth Software’s CBC/HB
Program: Theory and Practical Example. Proceedings of the Sawtooth Software Conference.
Papies, D., Eggers, F., & Wlömert, N. (2011). Music for free? How free ad-funded downloads affect
consumer choice. Journal of the Academy of Marketing Science, 39, 777794.
Reinders, M. J., Frambach, R. T., & Schoormans, J. P. L. (2010). Using Product Bundling to
Facilitate the Adoption Process of Radical Innovations. Journal of Product Innovation
Management, 27, 11271140.
Rogers, E. M. (1962). Diffusion of innovations (Social science). New York: Free Press.
Schlereth, C., & Skiera, B. (2012). DISE: Dynamic Intelligent Survey Engine. In A.
Diamantopoulos, W. Fritz, & L. Hildebrandt (Eds.), Quantitative marketing and marketing
management (pp. 225243). Wiesbaden: Gabler Verlag.
Schlereth, C., & Skiera, B. (2017). Two New Features in Discrete Choice Experiments to Improve
Willingness-to-Pay Estimation That Result in SDR and SADR: Separated (Adaptive) Dual
Response. Management Science, 63, 829842.
Longitudinal Study in Preferences for Electric Vehicles
48
Schlereth, C., Skiera, B., & Schulz, F. (2018). Why do consumers prefer static instead of dynamic
pricing plans? An empirical study for a better understanding of the low preferences for time-
variant pricing plans. European Journal of Operational Research, 269, 11651179.
Sentis, K., & Geller, V. (2010). The Impact of Covariates on HB Estimates. Proceedings of the
Sawtooth Software Conference.
Simon, H. A. (1955). A Behavioral Model of Rational Choice. The Quarterly Journal of Economics,
69, 99118.
Street, D. J., & Burgess, L. (2007). The construction of optimal stated choice experiments: Theory
and methods. Hoboken, NJ: Wiley-Interscience.
Swait, J., & Andrews, R. L. (2003). Enriching Scanner Panel Models with Choice Experiments.
Marketing Science, 22, 442460.
Teas, R. K. (1985). An Analysis of the Temporal Stability and Structural Reliability of Metric
Conjoint Analysis Procedures. Journal of the Academy of Marketing Science, 13, 122142.
Thurstone, L. L. (1927). A Law of Comparative Judgment. Psychological Review, 34, 273286.
Train, K. (2009). Discrete choice methods with simulation. Cambridge: Cambridge University Press.
Venkatesh, Thong, & Xu. (2012). Consumer Acceptance and Use of Information Technology:
Extending the Unified Theory of Acceptance and Use of Technology. MIS Quarterly, 36, 157
178.
Verlegh, P. W. J., Schifferstein, H. N. J., & Wittink, D. R. (2002). Range and Number-of-Levels
Effects in Derived and Stated Measures of Attribute Importance. Marketing Letters, 13, 4152.
Völckner, F. (2008). The dual role of price: decomposing consumers’ reactions to price. Journal of
the Academy of Marketing Science, 36, 359377.
Yang, L., Toubia, O., & DeJong, M. G. (2015). A Bounded Rationality Model of Information
Search and Choice in Preference Measurement. Journal of Marketing Research, 52, 166183.
Yu, J., Goos, P., & Vandebroek, M. L. (2006). The Importance of Attribute Interactions in Conjoint
Choice Design and Modeling. Working paper 0601, Department of Decision Sciences and
Information Management, Katholieke Universiteit Louven.
Zajonc, R. B. (1968). Attitudinal effects of mere exposure. Journal of Personality and Social
Psychology, 9, 127.
... Conjoint analysis is one of the most widely applied preference measurement techniques (Keller et al., 2021;Orme & Chrzan, 2017;Pachali et al., 2023). To shape product innovation, pricing, and market penetration decisions, the industry relies on this technique to understand consumers' product and service requirements (e.g., Papies et al., 2011;Voleti et al., 2017). ...
... Likewise, researchers are left alone to decide which principle to apply. This issue is significant as academia views incentive-aligned CBC as a promising conjoint methodology (e.g., Keller et al., 2021;Wlömert & Eggers, 2016;Yang et al., 2018), while management practice increasingly relies on adaptive designs, particularly ACBC (Sawtooth Software Inc. 2022b). ...
... CBC supports managerial decision-making in many fields (e.g., Keller et al., 2021;Schmidt & Bijmolt, 2020). Its broad and ongoing popularity leads to steady improvements to this research toolbox. ...
Article
Full-text available
Choice-based conjoint (CBC) analysis features prominently in market research to predict consumer purchases. This study focuses on two principles that seek to enhance CBC: incentive alignment and adaptive choice-based conjoint (ACBC) analysis. While these principles have individually demonstrated their ability to improve the forecasting accuracy of CBC, no research has yet evaluated both simultaneously. The present study fills this gap by drawing on two lab and two online experiments. On the one hand, results reveal that incentive-aligned CBC and hypothetical ACBC predict comparatively well. On the other hand, ACBC offers a more efficient cost-per-information ratio in studies with a high sample size. Moreover, the newly introduced incentive-aligned ACBC achieves the best predictions but has the longest interview time. Based on our studies, we help market researchers decide whether to apply incentive alignment, ACBC, or both. Finally, we provide a tutorial to analyze ACBC datasets using open-source software (R/Stan).
... Few longitudinal studies using discrete choice experiments have been undertaken [90], therefore analyses will be largely exploratory. Because time between assessments (12 months) is short, the same DCE instrument will be used at each timepoint [90,91]. ...
... Few longitudinal studies using discrete choice experiments have been undertaken [90], therefore analyses will be largely exploratory. Because time between assessments (12 months) is short, the same DCE instrument will be used at each timepoint [90,91]. Following best practices, we will analyze models at each timepoint using multinomial logit modeling [91]. ...
... By including demographic, socioeconomic, and care-related questionnaire data into our models, we will be able to explore how external factors, and especially changes in external factors such as income status, care use/disuse, and diabetes management and distress, are associated with changes in preferences over time. In the unlikely case that we experience greater loss to followup and attrition than expected, we will use repeated cross-sectional DCE approaches (such as a covariate extended model) to measure if differences in preferences across samples are significant [90]. Analyses will be completed in Stata v. 17 (College Station, TX) using the Choice Models (CM) suite of commands [92]. ...
Article
Full-text available
In Samoa, adult Type 2 diabetes prevalence has increased within the past 30 years. Patient preferences for care are factors known to influence treatment adherence and are associated with reduced disease progression and severity. However, patient preferences for diabetes care, generally, are understudied, and other patient-centered factors such as willingness-to-pay (WTP) for diabetes treatment have never been explored in this setting. Discrete Choice Experiments (DCE) are useful tools to elicit preferences and WTP for healthcare. DCEs present patients with hypothetical scenarios composed of a series of multi-alternative choice profiles made up of attributes and levels. Patients choose a profile based on which attributes and levels may be preferable for them, thereby quantifying and identifying locally relevant patient-centered preferences. This paper presents the protocol for the design, piloting, and implementation of a DCE identifying patient preferences for diabetes care, in Samoa. Using an exploratory sequential mixed methods design, formative data from a literature review and semi-structured interviews with n = 20 Samoan adults living with Type 2 diabetes was used to design a Best-Best DCE instrument. Experimental design procedures were used to reduce the number of choice-sets and balance the instrument. Following pilot testing, the DCE is being administered to n = 450 Samoan adults living with diabetes, along with associated questionnaires, and anthropometrics. Subsequently, we will also be assessing longitudinally how preferences for care change over time. Data will be analyzed using progressive mixed Rank Order Logit models. The results will identify which diabetes care attributes are important to patients (p < 0.05), examine associations between participant characteristics and preference, illuminate the trade-offs participants are willing to make, and the probability of uptake, and WTP for specific attributes and levels. The results from this study will provide integral data useful for designing and adapting efficacious diabetes intervention and treatment approaches in this setting.
... The popularity of electric vehicles brings a large number of charging demand, promoting the upgrading and transformation of the power distribution network (Yin et al., 2023;Buzna et al., 2021). EVs have the dual attributes of load and power source at the same time, and access to charging piles will change the operation mode of the power grid, and the distribution network will become a complex multi-power network interconnected with users (Dong et al., 2022;Solanke et al., 2020;Keller et al., 2021). ...
Article
Full-text available
The development of electric vehicles (EVs) reduces dependence on fossil fuels, promotes energy conservation and emissions reduction, and facilitates the transition to clean energy sources in the power grid. However, subjective charging behavior among EV owners can lead to blind charging practices, compromising the reliability of the distribution network by widening the peak-to-valley difference. To address safety concerns during the charging process, this paper proposes hardware and software systems for an experimental verification system. The network architecture, focused on charging safety, is examined. Analysis of the system’s operation data reveals that it enables bidirectional interaction between electric vehicles and the power grid. This solution proves ef-fective for integrating a large number of EVs in peak-shaving and valley-filling efforts, laying a technical foundation for their inclusion in the power grid for peak shaving, valley filling, as well as providing standby and frequency regulation services.
... In a separate analysis, we examined what differentiates the subgroup who are not currently providing intrapartum care from those who offer such care (either in hospitals or out-of-hospital). For that, we applied the covariate-extended hierarchical Bayes multinomial logit estimator to the larger data set that includes both subgroups (Keller et al., 2021;Lenk et al., 1996). Results are Working Paper Jessica Rheindorf, Christian Hagist, Christian Schlereth, Hannah Petry (2024) presented in Table A2 in the supplementary material. ...
Article
Full-text available
Background: There is a severe global shortage of midwives, and the situation worsens when qualified professionals leave their jobs because of inadequate working conditions. Hospitals have increasing difficulties in filling vacancies for midwives. In the case of Germany, midwives tend to give up birth assistance after an average of seven years working in delivery rooms, which are usually led by physicians. Objective: We aim to provide concrete recommendations on encouraging qualified professionals to work in maternity wards by examining the job preferences of midwives who currently do not provide such services. These insights shall help policy makers and hospital managers to fill vacancies more quickly and provide adequate care to more women. Design: Discrete choice experiment Setting(s): Online survey promoted through email and social media to midwives in Germany Participants: 415 midwives participated; we examine the subgroup of 241 midwives who do not offer birth assistance Methods: We obtain individual parameter estimates through a multinominal logit analysis with hierarchical Bayes estimation techniques, calculate importance weights, and simulate uptake probabilities of different hypothetical job offers that include birth assistance. Results: Participants want to provide birth assistance but fiercely reject doing so under physicians’ supervision. With a 15% increase in income, however, 16% would accept this least preferred setting. Forty-four percent, however, would choose to offer birth assistance if they could work in a midwife-led unit. An additional increase in income of 5% (15%) could even lead to uptake probabilities of 67% (77%). Conclusions: There is a common understanding that midwife-led care is a safe and effective option for healthy women. Policy makers are advised to further extend their initial support for such units to fill vacancies quicker and enable comprehensive healthcare for more childbearing women.
... The average utility scores must be interpreted relatively within the context of the study and the tested attributes (Orme 2019). A direct comparison of average importance scores with prior work is not possible unless an identical experiment has been already conducted pre-COVID-19, for instance, as part of a longitudinal study (Keller et al. 2021). To discuss possible preference changes due to the COVID-19 pandemic, we compare presented results with similar pre-COVID-19 choice studies from Chiambaretto (2021) and Hinnen et al. (2015), who surveyed the French and Swiss market respectively for their preferences on a long-and medium-haul flight. ...
Thesis
Door-to-door (D2D) air travel is gaining momentum for airlines, airports, and feeder traffic providers. The mobility industry and researchers are broadening their scope to include the entire travel chain, from origin to final destination. Intermodal mobility products are already on the market. At the same time, widespread trends affect transport service providers (as the suppliers) and passengers (regarding demand). Acquiring a better understanding of future D2D air travel trends is crucial for the mobility sector for long-term planning, product adaptation, the services provided and the pricing of these, and improvements in the passenger experience. Focusing on the European market, the overall objective of this doctoral thesis is to identify and understand the future trends of D2D air travel. It is divided into three parts; these provide different perspectives on trends and employ a range of methods that lead to results that develop from each other. In Part One, the Delphi technique is utilized to identify future travel trends. The study considers projections of European air passengers and their requirements for their entire air travel chain, including airport access, a long-haul flight, and airport egress. The research focuses on 2035 and is based on a two-round Delphi survey involving 38 experts from the transport industry, academia, and consultants. The Delphi survey is supplemented with findings from a preliminary study, consisting of a literature review, interviews with 18 experts in the field of air travel, and a workshop attended by experts. Results reveal that digitalization and personalization will be the main drivers in 2035 and that passengers might demand value-added use of their travel time. In addition, environmentally friendly travel products are considered desirable but only somewhat probable by 2035. Passenger type, age, origins, and travel budget will still be influential factors in 2035. Based on the results from a hierarchical cluster analysis, Part One presents three possible future scenarios: (1) personalized D2D travel, (2) integrated D2D travel, and (3) the game-changer. A technical chapter elaborates on the Delphi technique and individual research steps. Part Two explores the supply aspect and to what extent transport service providers consider strategically relevant trends. The scope of D2D air travel is adapted by applying multi-labeled text classification models to 52 corporate reports from a sample of transport service providers that operate in the European market. Trends identified in the first Delphi study and from an additional literature review are used to develop seven classes. Two prototype models are developed: a dictionary-based classifier and a supervised learning model using the multinomial naive Bayes and linear support vector machine classifiers. The latter yields the best model output, revealing which trends have a higher, medium, or lower relevance on the supply side. The results show that providers consider environmentally friendly air transport and related products to be highly relevant while disruption management, leveraging passengers' data, and improving airport feeder traffic through innovative mobility initiatives are considered to be of medium relevance. Part Three explores air passengers' preferences and willingness to pay for ancillary services in the current transition into the new normal, brought about by the ongoing COVID-19 pandemic, high uncertainty, and changing market dynamics. A choice-based conjoint analysis is used to test six attributes within a hypothetical travel scenario for a long-haul one-way air trip. Choice data from 269 German business and leisure passengers are analyzed using a hierarchical Bayes estimator. Results reveal that the total ancillary service upgrade price influences passengers' choices the most, followed by a seat upgrade for greater comfort and the CO 2-compensation of a flight. Hygiene-related ancillar-ies bring low utilities. Female and senior passengers care more for environmentally friendly ancillaries. Confirming previous research, business passengers and frequent flyers care more for onboard comfort. Download: https://opus4.kobv.de/opus4-whu/frontdoor/index/index/docId/934
... Familiarity with some performance attributes of EVs such as driving comfort, acceleration, usage cost, charging infrastructure, etc. enhances its perceived benefit/usefulness, value orientation toward EVs (Gnann et al., 2018;Liu et al., 2018), which in turn fosters a positive attitude toward EVs and consequently leads to adoption intention. Moreover, a strong association is observed between knowledge (Keller et al., 2021) about the benefits, advantages of performance attributes, and perceived usefulness (Jaiswal et al., 2021a;Su et al., 2020). Attitude towards EVs adoption and adoption intentions are also influenced by subjective norms and these norms refer to one's perception of the social pressure from the reference group to which he/she belongs (Ajzen, 1991). ...
Article
Full-text available
Market segmentation becomes a crucial tool for evolving transportation technology such as electric vehicles (EVs) in emerging markets to explore and implement for extensive adoption. EVs adoption is expected to grow phenomenally in near future as low emission and low operating cost vehicle, and thus, it drives a considerable amount of forthcoming academic research curiosity. The main aim of this study is to explore and identify distinct sets of potential buyer segments for EVs based on psychographic, behavioral, and socioeconomic characterization by employing an integrated research framework of 'perceived benefits-attitude-intention', The study applied robust analytical procedures including cluster analysis, multiple discriminant analysis and Chi-square test to operationalize and validate segments from the data collected of 563 respondents using a cross-sectional online survey. The findings posit that the three distinct sets of young consumer groups have been identified and labelled as 'Conservatives', 'Indifferents', and 'Enthusiasts' which are deemed to be buddying EV buyers The implications are recommended, which may offer some pertinent guidance for scholars and policymakers to encourage EVs adoption in the backdrop of emerging sustainable transport market.
... Eq. (4) assumes that preferences for the price are linear. This linearity assumption has often been made when dealing with price ( Keller et al., 2021 ;Papies et al., 2011 ;Meyer et al., 2018 ;Völckner, 2008 ). This assumption also makes our model more parsimonious: An alternative would have been to use a parthworth model that requires two additional parameters for the estimation. ...
Article
Background Mothers in Germany are entitled to midwifery care; however, they face a lack of skilled professionals. While the reliability of the access to midwifery is of great public interest, we know little about clients’ preferences. Objectives We conduct a discrete choice experiment to study preferences and willingness to accept copayment for the entire scope of midwifery care (pregnancy, delivery, and postnatal). Thereby, we aim to provide policy recommendations for priority settings in times of scarcity. Furthermore, we evaluate to what extent midwives’ education matters to parents and assess the degree of support for the latest Midwifery Reform Act that transfers education from vocational schools to universities. Design Discrete choice experiment with separated adaptive dual response. Settings Online Survey promoted through Facebook to parents in Germany. Respondents 2,080 respondents completed the experiment. They all have or are expecting at least one natural child, mainly born between 2018 and 2020 (87%). The average respondent is female (99%), 33 years old, with a university degree (50%). Methods We use a D-optimal fractional factorial design and obtain individual parameter estimates through a Multinomial Logit analysis with Hierarchical Bayes estimation techniques. We calculate willingness to pay and importance weights and simulate uptake probabilities for different packages of care. To avoid extreme choice behavior, we apply separated adaptive dual response. Results Home visits during the postnatal phase are most important (importance weight 50%); online support is demanded when no personal support is available. We find that 1:1 care during delivery is highly preferred, but one midwife supporting two women intrapartum is still acceptable. The midwife´s education plays a minor role with an importance weight of 3%; however, we find a preference for midwives trained at vocational schools rather than at universities. Conclusions In times of scarcity, postnatal care in the form of home visits should be prioritized over pregnancy counseling, and online services should be promoted as an add-on but not as a substitute for personal support. There is a high level of willingness to accept co-financing to ensure the availability of services usually covered by health insurance. Tweetable abstract How much are #mothers willing to pay to ensure the availability of a #midwife? Doi:
Article
This literature survey explores the potential avenues for the design of a green auto asset‐backed security (Green Auto ABS) by focusing on the European auto securitization market. In this context, we examine the entire value chain of the securitization process to understand the incentives and interests involved at various stages of the transaction. We review recent regulatory developments, feasibility concerns, and potential designs of a sustainable securitization framework. Our study suggests that a Green Auto ABS could be based on both a green use of proceeds and a green collateral‐based methodology.
Article
By customizing smart assistance systems to customer needs, car manufacturers can improve their systems and create additional benefits for users. However, it is still unclear which characteristics car drivers perceive as favorable and useful. To examine this, we employ a mixed-method approach. In our first study, we conduct a survey (N $=$ 301) to investigate general user perception antecedents of smart assistant systems in cars. We analyze the indirect effects of different system quality characteristics mediated by user perception on their usage intention. In our second study (N $=$ 270), we use a discrete choice experiment to measure the effects of concrete system attributes on user acceptance of IT-based parking systems representing a concrete instantiation of a smart assistance system. Consistent with our first study, we observe that the system quality factors user interface intuitiveness, full language flexibility, and system error occurrence significantly influence the consumers’ intention to use the technology. Accordingly, car manufacturers should put a particular focus on these factors when developing and implementing their smart assistance systems in cars. For IT-based parking systems, in particular, consumers are very price-sensitive. However, by implementing additional technical features, manufacturers can significantly increase the systems’ price value and thus the purchase probability.
Article
Full-text available
The authors propose and empirically evaluate a new hybrid estimation approach that integrates choice-based conjoint with repeated purchase data for a dense consumer panel, and they show that it increases the accuracy of conjoint predictions for actual purchases observed months later. The key innovation lies in combining conjoint data with a long and detailed panel of actual choices for a random sample of the target population. By linking the actual purchase and conjoint data, researchers can estimate preferences for attributes not yet present in the marketplace, while also addressing many of the key limitations of conjoint analysis, including sample selection and contextual differences. Counterfactual product and pricing exercises illustrate the managerial relevance of the approach.
Article
Full-text available
Time-variant pricing plans in electricity markets aim to mitigate mismatches between demand and supply by incentivizing consumers to shift their demand from costly peak to cheaper off-peak times. Their implementation can be manifold; they could depend statically on the time of the day (i.e., time-of-use pricing) or adjust prices dynamically in nearly real time (real-time pricing). If consumers reduced demand in peak times, then they would realize lower prices and providers would operate at lower costs. Still, consumers frequently refuse time-variant pricing plans. The authors develop a new conceptual framework to study and explain this behavior. It supports the optimal choice of time-variant pricing plans by jointly considering price fairness and economic antecedents. In a discrete choice experiment, the authors use a hierarchical Bayes covariate extended logit estimation to measure respondents’ probability of switching from a time-invariant pricing plan to a time-variant pricing plan. The results show that economic antecedents, such as price consciousness and flexibility, have a stronger effect on the choice of a time-variant pricing plan than price fairness considerations; cost insurance is a promising instrument for increasing acceptance of dynamic pricing plans. The results also suggest new ways to target prospective customers.
Article
Research Summary Our understanding of how managers take international location decisions is still scarce. Building on the microfoundations view, we explore managers’ perceptions of risk and return in a discrete choice experiment with 2,618 decisions in 2013 (a globalizing world) and 2017 (a de‐globalizing world). While managerial perceptions vary over time due to economic and political changes, such as the current de‐globalization trend, decision heuristics remain remarkably stable: locations perceived as least risky offer the highest expected returns. We also find that distance is a good proxy for managerial perceptions. Investigating the microfoundations of decision‐making we show that international experience, risk‐taking propensity and shareholder status affect heuristics. In sum, our study provides novel insights into the microfoundations of location decisions and extends the behavioral perspective on internationalization. Managerial Summary While many researchers in the field of international business and global strategy have studied internationalization, surprisingly few of them put the actual decision‐maker in the spotlight. This study investigates the decision heuristics of managers who are confronted with international location decisions. We show that their perceptions of risk and return in the context of a globalizing world (in 2013) and in the context of a de‐globalizing world (in 2017) vary. However, the underlying heuristic – locations that are perceived as least risky offer the highest expected returns – remains stable. Our study also indicates that distance is a good proxy for managerial decisions and that managers’ international experience, risk‐taking propensity and shareholder status affects their decision‐making patterns.
Article
The authors report a study of the effects of price, brand, and store information on buyers’ perceptions of product quality and value, as well as their willingness to buy. Hypotheses are derived from a conceptual model positing the effects of extrinsic cues (price, brand name, and store name) on buyers’ perceptions and purchase intentions. Moreover, the design of the experiment allows additional analyses on the relative differential effects of price, brand name, and store name on the three dependent variables. Results indicate that price had a positive effect on perceived quality, but a negative effect on perceived value and willingness to buy. Favorable brand and store information positively influenced perceptions of quality and value, and subjects’ willingness to buy. The major findings are discussed and directions for future research are suggested.
Article
Reliability and stability of preference patterns obtained through conjoint measurement are examined by perturbing the stimulus structure and by repeating measurements after the passage of time. Conjoint measurement appears to be very robust to perturbation and reasonably stable over time.
Article
A model is proposed which expresses consumer satisfaction as a function of expectation and expectancy disconfirmation. Satisfaction, in turn, is believed to influence attitude change and purchase intention. Results from a two-stage field study support the scheme for consumers and nonconsumers of a flu inoculation.
Article
This article describes novel methods of using consumer panel statistics to predict the success of new grocery products. Since about four out of five such products fail, even modest improvements in early identification of successes is important. The authors show how greater attention to successive waves of repeat buying and to intervals between purchases make significant improvement practical. They link these somewhat neglected statistics to a mathematical model of penetration.
Article
Which attitudes are related to purchasing decisions? In this article the authors argue that out of many possible attitudes only a few really relate to or “determine” buying behavior. These attitudes are defined in this article and methods of measuring them are discussed.
Article
Many companies are increasingly selling hybrid bundles, which comprise one or more goods and one or more services. Hybrid bundle pricing depends on understanding consumer willingness to pay (WTP) for the bundle, which rests on trade-offs among the benefits from four key drivers: service autonomy, complementarity, service quality variability, and overall bundle quality (basic vs. premium). The effects of these drivers and their interactions on the WTP of hybrid bundles are unknown. The authors develop hypotheses and test them rigorously using incentive-aligned choice-based conjoint and hierarchical Bayesian analysis. The results offer important guidelines for developing appropriate hybrid bundles. If a typical firm under budget constraint has to offer either of two hybrid bundles, one with high complementarity or one with service autonomy, the results suggest that it should offer the bundle with high complementarity. Furthermore, contrary to the conventional wisdom of minimizing service quality variability for premium quality bundles relative to basic quality bundles, the results recommend lowering service quality variability for basic quality bundles but maintaining it for premium bundles.
Article
Forecasting the sales or market share of new products is a major challenge as there is little or no sales history with which to estimate levels and trends. Choice-based conjoint (CBC) is one of the most common approaches used to forecast new products’ sales. However, the accuracy of forecasts based on CBC models may be reduced when consumers’ preferences for the attributes of products are labile. Despite this, there is a lack of research on the extent to which lability can impair accuracy when the coefficients estimated in CBC models are assumed to be constant over time. This paper aims to address this research gap by investigating the prevalence of lability for consumer durable products and its potential impact on the accuracy of forecasts. There are reasons to expect that lability may be particularly evident where a product is subject to rapid technological change and has a short product life-cycle. We carried out a longitudinal survey of the preferences of 161 potential consumers relating to four different types of products. We established that for both functional and innovative products: (i) the CBC models vary significantly over time, indicating changes in consumer preferences and (ii) such changes may cause large differences in forecasts of the probabilities that consumers will purchase particular brands of products. Hence employing models where coefficients do not change over time can potentially lead to inaccurate market share forecasts for high-tech, short life-cycle products that are launched even a short time after the choice-based modelling has been conducted.