Content uploaded by Daniel Guhl
Author content
All content in this area was uploaded by Daniel Guhl on Feb 26, 2018
Content may be subject to copyright.
Please note: The authors would like to thank two anonymous referees
for their helpful and constructive comments as well as the co-editor of
the special section, Udo Wagner, for his detailed feedback and sugges-
tions. Furthermore, the authors benefitted from critical discussions with
participants at the 5th French-Austrian-German Workshop on Consumer
Behavior. Daniel Guhl gratefully acknowledges support by the Deutsche
Forschungsgemeinschaft (DFG) through CRC TRR 190.
Research Articles
Friederike Paetz, Clausthal Uni-
versity of Technology, Institute
of Management and Econom-
ics, Department of Marketing,
Julius-Albert-Str. 6, 38678
Clausthal-Zellerfeld, Germany,
Phone: +49 5323 727682,
Fax: +49 5323 727659,
E-Mail: friederike.paetz@
tu-clausthal.de.
*Corresponding Author
Daniel Guhl, Humboldt Univer-
sity Berlin, School of Business
and Economics, Institute of
Marketing, Spandauer Str. 1,
10178 Berlin, Germany,
Phone: +49 30 2093 5884,
E-mail: daniel.guhl@hu-berlin.de.
Understanding Differences in Segment-specific
Willingness-to-pay for the Fair Trade Label
By Friederike Paetz and Daniel Guhl
We conduct a discrete choice analysis in Ger-
many to study consumer behaviour for or-
ange juice in a Fair Trade context. In particu-
lar, we estimate Finite Mixture-Multinomial
Logit models for analysing the segment-spe-
cific willingness-to-pay for the Fair Trade la-
bel. The average willingness-to-pay is 48 Eu-
rocent (relative price premium of 24 %), and
we find substantial heterogeneity in values
between segments (10 to 95 Eurocent). A
profiling of the segments reveals these differ-
ences can be attributed to differences in seg-
ment members’ gender, consumption fre-
quency, and consciousness-for-fair-con-
sumption. Segments with a large proportion
of women, a high level of consciousness-for-
fair-consumption and frequent consumption
of orange juice have a higher willingness-to-
pay for the Fair Trade label.
1. Motivation and literature review
Currently, the increasing competition in saturated mar-
kets of fast-moving consumer goods drives companies to
highly sophisticated marketing strategies. Increasing
brand equity and reducing prices are two tools that are
frequently used to increase consumer preferences and
purchase intentions and, therefore, to increase compa-
nies’ sales and ultimately profits (cp. Cobb-Walgren
et al. 1995). A product enhancement with a social attri-
bute such as the Fair Trade (FT) label could be interpret-
ed as an additional tool. If respondents gain a utility sur-
plus from the social product attribute and honour it mon-
etarily, the company’s profits may also increase. Hence,
focusing on consumers’ social preferences is an interest-
ing research field for marketing managers and academ-
ics. In particular, assessing consumers’ willingness-to-
pay (WTP) for the social attribute is highly interesting: a
positive WTP serves as a necessary condition for the
price premium and, therefore, for increasing profits be-
cause adhering to FT standards may result in additional
costs for commodities (Fair Trade Deutschland 2014, p. 4).
In marketing practice, profit-oriented companies use the
prevailing trend of socially responsible consumption and
add social product features to differentiate their products
from their competitors’. For example, Starbucks prom-
ises their customers that 100 percent of their coffee is
ethically sourced (certified through external audit sys-
tems such as FT), which they achieved in 2015 (Star-
bucks 2015). This research has proven interesting not on-
ly from a practical point of view but also from an aca-
demic research perspective. Understanding consumer be-
haviour in the context of fair consumption has emerged
as a vast research field in recent years, as supported by
the literature reviews of Andorfer and Liebe (2012) and
Tully and Winer (2014). In particular, the consumption
of FT products as a component of socially responsible
consumption has drawn high interest in practice and mar-
keting theory. Within the context of FT in marketing the-
ory, many studies have focused on consumers’ WTP for
the FT label. They found that an aggregated assessment
of consumers’ WTP for the FT label attribute might be
overly restrictive: some consumers may have a highly
positive WTP while others may yield a small or even
negative WTP for the FT label attribute. Hence, an ag-
gregated view could result in biased WTP estimates and,
38 MARKETING · ZFP · Volume 39 · 4/2017 · P. 38–47
finally, a loss in profits for companies that derive price
premia from these biased WTPs.
The problems caused by neglecting WTP heterogeneity
(as is done within aggregated approaches) have been
identified in the relevant literature, and several studies
have highlighted the importance of considering WTP
heterogeneity in the FT context. The seminal study of
Basu and Hicks (2008), for example, compared differ-
ences in the WTP for FT coffee between German and US
consumers. They found that German consumers are more
sensitive than US consumers, leading to a stronger de-
cline in German consumers’ WTP for increasing relative
inequality between participants and non-participants in
the FT program (cf. Basu and Hicks 2008, p. 13). Rotaris
and Danielis (2011) identified socio-demographic differ-
ences in price premia for the FT label in an Italian sam-
ple in the product category of coffee. Consumers who are
female or young or who regularly buy FT products are
more sensitive to FT issues. Yang et al. (2012) deter-
mined individual WTP within the coffee category and
observed gender and consumption-related differences in
a Chinese sample. Female and long-term consumers (or
those who intend to increase their coffee consumption)
were willing to pay high price premia for the FT label at-
tribute. De Pelsmacker et al. (2005a) asked Belgian re-
spondents about their preferences for FT coffee. They
found that the respondents to whom the FT label attribute
had the highest importance when making a purchase de-
cision were willing to pay the highest relative price pre-
mium of 35 %. In contrast, those respondents who most-
ly cared exclusively about brand or flavour were only
willing to pay a relative price premium of 5 %. On aver-
age, De Pelsmacker et al. (2005a) calculated a relative
price premium of 10 %. This discrepancy shows, once
again, that if WTP heterogeneity is neglected, marketing
managers are in danger of losing profits when deciding
price premia based on aggregate WTP calculations.
So far, the relevant literature has used varying ap-
proaches to evaluate the drivers of differences in consu-
mers’ WTP for the FT label attribute. The recent litera-
ture apparently uses non-aggregated approaches, such as
individual or segment-level analyses, to incorporate con-
sumers’ WTP heterogeneity. In particular, segment-level
WTP estimates are commonly used in marketing practice
because they provide a sound basis for companies’ (seg-
ment-specific) pricing differentiation strategies and,
hence, contribute to companies’ revenue. Furthermore,
managers prefer segment-level results because they are
easy to communicate and comprehend. However, in the
context of preference-based segmentation approaches for
determining WTP for the FT label, the so-called one-
stage segmentation approaches (simultaneous market
segmentation and estimation of segment-specific prefer-
ences) are less popular in the marketing literature. These
approaches, however, allow a deeper understanding and
a more precise prediction of consumer behaviour com-
pared to two-stage (i. e., estimation of individual prefer-
ences and subsequent clustering) segment approaches
(Ramaswamy and Cohen 2007, p. 297) and may thus al-
so contribute to strategic marketing decisions in practice.
In the context of ethical consumption, the study of Auger
et al. (2008) is an example that actually uses a one-stage
segmentation approach. The authors found three con-
sumer segments that differed in the importance that they
attached to several social product features (child labour,
minimum wage, dangerous working conditions and liv-
ing standards). However, they neither classified their re-
sults to the FT context nor derived segment-specific
WTP for social product attributes. To contribute to the
scarce academic research within this field, we conduct a
discrete choice analysis for orange juices and estimate a
preference-based segmentation model, i. e., Finite Mix-
ture-Multinomial Logit (FM-MNL) model, calculating
segment-specific WTP for the FT label attribute. We ad-
dress the following research questions: Do differences
exist in the WTP for the FT label between different pref-
erence-based consumer segments? How can these differ-
ences be explained in terms of the socio-demographic or
psychographic characteristics of segment members?
Usually, the influence of demographic variables such as
gender or age is examined, but the current results are
conflicting: Rotaris and Danielis (2011) identified an in-
fluence of demographic variables on consumers’ WTP
for an Italian sample, and Yang et al. (2012) did so for a
Chinese sample, but De Pelsmacker et al. (2005b) did
not observe demographic differences in consumer prefer-
ences for the FT label attribute in a Belgian sample. Be-
cause these conflicting results were found for samples in
different cultures, cultural differences in the demograph-
ic drivers of consumer WTP are likely. We focus on a
German sample.
In addition to these demographic variables, we use the
psychographic variable “Consciousness-for-fair-con-
sumption” (cfc) (cp. Balderjahn et al. 2013) to test its in-
fluence on consumers’ WTP. Considering psychographic
variables to be potential drivers of WTP differences
seems reasonable because De Pelsmacker et al. (2005a,
p. 366) already stated that demographic variables might
not be sufficient to fully describe the socially responsible
consumer. Psychographic variables may also affect con-
sumers’ purchase behaviours. The cfc construct serves as
a psychographic variable and is tailored to the FT con-
text. Such variables are highly underrepresented in rele-
vant studies so far. The work of Arnot et al. (2006) is one
of the few studies to consider a related variable: “famil-
iarity with the concept of FT coffee”. The authors found
that respondents who are familiar with the concept of FT
coffee are more likely to choose a FT coffee. Balderjahn
and Peyer (2012) examined the cfc variable in the FT
context and discovered a relationship between consu-
mers’ cfc levels and the FT label’s influence on consu-
mers’ purchase decisions for a German sample. Hence,
this further investigation of the cfc variable’s potential as
a driver for WTP differences contributes to the recent lit-
erature.
Paetz/Guhl, Understanding Differences in Segment-specific Willingness-to-pay for the Fair Trade Label
MARKETING · ZFP · Issue 4 · 4. Quarter 2017 39
Finally, we address consumers’ consumption frequency
(e. g., weekly, monthly, or less often consumption) as a
behavioural construct. Consumption-related variables,
such as “generally buying FT products: yes/no” (cp. Ro-
taris and Danielis 2011) or “long-term consumption”
(i. e., more than five years) (cp. Yang et al. 2012), have
been considered in the recent literature. However, these
variables relate primarily to consumers’ familiarity with
the (FT) product category than to the actual consumption
frequency of the focal product. However, the frequency
of purchase and, therefore, the amount of money spent
for the focal product may influence consumers’ WTP in
general.
The remainder of this paper is structured as follows: We
explain data collection and analysis methods in Sec-
tion 2. We summarize the empirical results of our dis-
crete choice experiment for the estimated FM-MNL
model in Section 3, including the profiling of consumer
segments. We close in Section 4 by drawing conclusions
and discussing limitations as well as future research is-
sues.
2. Methods
2.1. Data collection
To obtain the data, we conducted a discrete choice exper-
iment via an online survey. For the following analysis,
we can distinguish two different types of data: choice da-
ta and respondent data. Specific details regarding the da-
ta of our empirical study are presented in the next sec-
tion; here, we focus on how the data were collected.
Choice data
The survey contained a discrete choice experiment that
consisted of 16 choice sets. We used the first 14 choice
sets of each respondent for model estimation and the last
two choice sets for model validation. Note that this num-
ber is well below the number of 20 choice sets, which
Johnson and Orme (1996) suggested as an upper bound
before the data quality begins to degrade. Each choice set
included three (hypothetical) orange juice alternatives as
well as a ‘no purchase’ option. The latter is essential to
precisely measure WTP, because a forced choice situa-
tion (i. e., a choice set without a ‘no purchase’ option)
may cause an underestimation of price sensitivity (Allen-
by et al. 2014). To describe the orange juice alternatives,
we used the following four attributes (with the corre-
sponding levels in parentheses): brand (Albi, Granini,
Hohes C, Valensina), type of packaging (Tetra Pak (car-
ton), PET bottle (plastic)), display of an FT label (yes,
no), and price (per litre) (1.09 , 1.39 , 1.69 , 1.99 ).
We followed the guidelines of Orme (2002). For the
price levels and different types of packaging, we used
those levels that are prevalent in German grocery stores.
For the brand levels, we chose leading national orange
juice brands from the German market (Statista 2016).
Note that we use these attributes and levels to create or-
ange juice alternatives that are hypothetical but which re-
spondents nevertheless perceive as realistic choice op-
tions. The specific variation of attributes in our experi-
mental design enables us to measure their effects on utili-
ty, including the impact of a currently not available FT
label for the top four national orange juice brands. The
online survey was created and administered using Saw-
tooth Discover. The individual-specific fractional facto-
rial design, without any attribute prohibitions, is near or-
thogonal and has high D-efficiency. In particular, our
4x2x2x4 design has a D-efficiency of over 99 % com-
pared to a design created using the complete enumeration
approach of Sawtooth’s SSI Web module, which has op-
timal efficiency for main effect designs. However, our
design also contains a certain amount of overlap within
and between attributes and avoids dominant alternatives
(see Sawtooth Software 2014). The use of a discrete
choice experiment tackles the problem of the often-dis-
cussed gap between consumer attitudes to buy FT prod-
ucts due to social desirability and their actual behaviour
in real purchase situations. It is therefore advisable to use
a realistic choice situation, in which respondents must
trade-off among several product attributes (cf. DePels-
macker et al. 2005, p. 513). Discrete choice experiments
closely mimic real choice situations and are therefore
known to relax the overestimation of the influence of so-
cial product attributes, such as the FT label attribute (cf.
Auger and Devinney 2007).
Respondent data (background variables)
As mentioned before, we aim to analyse (potential) WTP
differences for the FT label across consumer segments.
To this end, we employ several promising respondent-
level background variables, which we derived from prior
literature. In particular, we use demographics (age and
gender), orange juice consumption frequency, and cfc.
While questions eliciting the former variables are
straightforward (e. g., “how frequently do you drink or-
ange juice”), some clarification of the cfc scale is in or-
der. Balderjahn et al. (2013, p. 546) presented a scale
measuring the “latent disposition of consumers to prefer
products that are produced and traded in compliance
with fair labour and business practices”. This psycho-
graphic construct is directly related to preferences for FT
products: Consumers with higher cfc should also have
higher WTP for a FT label. Balderjahn et al. (2013) test-
ed the cfc scale in three independent studies and found
high validity and reliability across all samples. Hence,
we are confident in the quality of this scale. The respon-
dents’ cfc levels are measured on a 7-point rating scale,
with six items each corresponding to a specific labour
standard (i. e., compliance with workers’ rights; freedom
from forced labour; abolition of illegal child labour; non-
discrimination in the workplace; compliance with inter-
national statutory labour standards; fair wages for work-
ers). For each item, a belief component Bjl (“I only buy a
product if I believe that in its production ...”) and an im-
Paetz/Guhl, Understanding Differences in Segment-specific Willingness-to-pay for the Fair Trade Label
40 MARKETING · ZFP · Issue 4 · 4. Quarter 2017
Fig. 1: Main steps of the analysis
portance component Ijl (“How important is it for you per-
sonally that in companies ...”) are measured, and the
model for the cfc is cfcj=l=1
6
ΣBjl ·Ijl,wherejindicates
respondents and lindicates items (see Balderjahn et al.
2013, p. 548 for more details). Finally, we standardise
the cfc values to ease interpretation.
2.2. Analysis
Our analysis consists of several parts. The main steps are
depicted in Figure 1 and briefly discussed. The appendix
contains technical details regarding model, estimation,
and segmentation. The data described before can be sep-
arated into 3 data sets that serve different purposes in our
analysis: (1) choice data (in-sample), which we use for
model estimation; (2) choice data (out-of-sample), which
we use for model validation and selection; and (3) re-
spondent data (background variables), which we use for
profiling segments.
We employ the FM-MNL model to determine segment-
specific preferences. This model is rooted in random util-
ity theory and allows us to account for preference hetero-
geneity (see, e. g., Elshiewy et al. 2017, for an over-
view). The FM-MNL model, as a one-stage segmenta-
tion approach, permits a simultaneous market segmenta-
tion and estimation of segment-specific preferences (see,
e. g., Vriens et al. 1996 for a comparison of one-stage
and two-stage segmentation approaches). The determin-
istic part of the utility function is linear-additive in the at-
tribute levels of an alternative, and parameters can vary
across segments. All non-price attributes are dummy-
coded; hence, the corresponding parameters measure the
utility of an attribute with respect to the reference level
(i. e., brand “Albi”, packaging “plastic”, FT label “no”).
The (negative) effect of price on utility is assumed to be
linear, i. e., a vector utility function (Orme 2007, p. 2).
Although the use of partworth utility functions may ac-
count for nonlinearities in price, we decided to use a line-
ar price function for several reasons. First, nonlinearities
in the price function are negligible in the case of our da-
ta, so a more complicated model with additional parame-
ters is unnecessary. Second, the use of a linear price pa-
rameter simplifies our subsequent calculation of respon-
dents’ WTP because WTP is then price-independent
(Louviere et al. 2000, p. 280).
Because the number of segments is unknown before the
analysis, we estimate FM-MNL models with 2 to 5 seg-
ments to cover the typical number of segments (3–4 seg-
ments) in marketing literature and practical applications
(Tuma and Decker 2013, p. 11). We also add a version
with one segment only (i. e., the regular MNL model) as
a benchmark. We use several starting values in the maxi-
mum likelihood estimation to avoid convergence to local
optima (cf. Grün 2008).
After model estimation, we must select the final model
for analysis (cf. Melnikov and Maitra 2010, p. 88). We
employ multiple measures to this end (see, e. g., Wedel
and Kamakura 2000, for an overview) and also consider
the interpretability of resulting segments. We evaluate
model fit using log-likelihood (LL) values and first
choice hit rates. For the latter, the two extra choices (out-
of-sample) are used. In addition, we use the Bayesian in-
formation criterion (BIC) and Entropy. Whereas the BIC
helps in model selection by penalizing model fit with
model complexity (number of parameters), Entropy in-
forms a model that contains a particular number of seg-
ments of how distinct the assignment of respondents to
the segments is.
Once we have decided which model is best according to
the different measures described above, we can obtain
segment-specific estimation results. For each segment,
we can obtain parameter estimates and the segment size.
Paetz/Guhl, Understanding Differences in Segment-specific Willingness-to-pay for the Fair Trade Label
MARKETING · ZFP · Issue 4 · 4. Quarter 2017 41
# Segments 1 2 3 4 5
# Parameters 7 15 23 31 39
LL -5323.63 -4936.28 -4656.41 -4472.77 -4332.60
BIC 10706.94 10000.44 9508.91 9209.82 8997.67
Entropy .85 .88 .89 .90
Hit rate .54 .58 .63 .64 .65
Note: The proportional chance criterion (Morrison 1969) for the out-of-sample
data is .26.
Tab. 1: Number of parameters and performance measures
Furthermore, we can derive segment-specific attribute
importance (i. e., the utility of an attribute relative to the
range of utility values) and WTP values for the FT label
(i. e., the utility of the FT label rescaled in monetary
units). These measures are common in discrete choice
analysis and help us to understand and compare raw pa-
rameter estimates. The latter is based on the marginal
rate of substitution between the FT label attribute and
price. For our models with a dummy-coded FT label at-
tribute and a linear price effect, we have the following
for each segment (Tully and Winer 2014):
WTPFT label =–
U(xFT label =1)– U(xFT label =0)
⭸U⭸p
=–
β
FT label
α
,
(1)
where
β
FT label measures the utility difference of an alter-
native with and without a FT label and
α
is the marginal
effect of price pon utility U(see Small and Rosen 1981
for a general discussion of welfare measures in a discrete
choice analysis). Note that we assume an economically
meaningful negative price-effect (i. e., all else being
equal, consumers prefer to pay less). Additionally, this
measure is not an elasticity measure because we are not
interested in the relative change in utility caused by a rel-
ative change in an independent variable, such as price.
We are interested in the absolute change in utility due to
the presence of the FT label, and we express this change
in monetary units (using
α
) for ease of interpretation. Fi-
nally, in contrast to direct approaches where respondents
state their WTP directly, indirect methods suffer less
from an overrating of respondents’ WTP (see, e. g., Brei-
dert et al. 2006 for further details).
After selecting the final model, we calculate respon-
dent’s probability of belonging to a particular segment
based on the parameter estimates and individual choices.
Then, we assign each respondent to the segment where
he/she has the maximum (so-called) posterior member-
ship probability (“discrete segmentation”) (see DeSarbo
et al. 1995). Intuitively, the potential segmentation error
is smaller with non-overlapping segments if the solution
has high Entropy (i. e., well-separated segments).
Because our research aims not only to identify differ-
ences in segment-specific preferences but also to under-
stand whether respondents’ characteristics can explain
these differences, we “profile” these segments after as-
signing each respondent to a single segment. This means
we use respondent-level background variables and ana-
lyse their distributions post hoc across segments (cp. We-
del and Kamakura 2000, p. 145). Such an approach is
simple to perform by using standard methods of multi-
variate analysis, and the communication of the results is
straight forward. In particular, for categorical variables,
such as gender or age-class,
χ
2-tests of the correspond-
ing contingency tables are appropriate (e. g., H0: the fre-
quency of women are the same across segments). For
continuous variables, such as cfc-values, F-test based on
a one-way ANOVA are adequate (e. g., H0: mean cfc-val-
ues are the same across segments).
3. Empirical study
The questionnaire was distributed online at two German
universities in spring 2015. Overall, 360 respondents
completed the survey: 58 % were female, and most re-
spondents were 25 years old or younger (64 %). Regard-
ing consumption frequency, many respondents con-
sumed orange juice ‘at least once a week’ (41 %), fol-
lowed by ‘one to three times per month’ (36 %). Only
22 % of the respondents stated they consume orange
juice ‘less often than once per month’.
The 360 respondents made 5040 and 720 discrete choice
decisions in- and out-of-sample, respectively. The in-
sample choice shares are very similar for the three or-
ange juice alternatives (each about 29 %), which was an
expected result because the alternatives were “unla-
beled” in the experiment (i. e., all attributes varied freely
over all alternatives in each choice set). The choice share
for the “no purchase” option is 13 %. The choice shares
for the out-of-sample data are 29 %, 28 %, and 26 % for
the orange juice alternatives and 17 % for the ‘no pur-
chase’ option. Although these values are quite compara-
ble to the corresponding in-sample values, the choice
share for the ‘no purchase’ option is slightly higher.
The cfc-scale has high reliability (Cronbach’s
α
= .96),
and the results of an exploratory factor analysis indicate
the scale’s unidimensionality (e. g., all factor loadings
.85; proportion of variance explained = .80). The results
are very similar to those of Balderjahn et al. (2013), and
we conclude that the scale measures the cfc construct
well.
3.1. Model selection
Using in-sample choice data, we estimated several FM-
MNL models with a varying number of segments. As ex-
plained earlier, we also estimate as a benchmark a simple
MNL model with one segment. Tab. 1 displays the num-
ber of parameters to be estimated and the statistical mod-
el fit (i. e., LL value, BIC statistic, Entropy, and first
choice hit rates).
Apparently, the LL value decreases with an increasing
number of segments. This was expected because consid-
ering more segments improves model fit and therefore
causes higher LL values. To select a model that both fits
the data well and is parsimonious, it is advisable to in-
spect the BIC statistic, which accounts for the model’s
LL value and the model-specific number of parameters.
Paetz/Guhl, Understanding Differences in Segment-specific Willingness-to-pay for the Fair Trade Label
42 MARKETING · ZFP · Issue 4 · 4. Quarter 2017
Segment 1 Segment 2 Segment 3 Segment 4
No purchase -3.21* (.33) -11.26* (.73) -3.13* (.32) -2.06* (.53)
Granini -.40* (.17) .09 (.20) .06 (.11) 1.28* (.29)
Hohes C .27 (.15) .65* (.20) .40* (.10) 2.13* (.31)
Valensina -.37* (.17) .35* (.19) .25* (.12) .68* (.28)
FT label (yes) 1.62* (.14) .74* (.23) 1.73* (.12) .35* (.13)
packaging (carton) 1.59* (.20) .02 (.15) 1.02* (.09) -.53* (.12)
price -4.07* (.24) -7.29* (.60) -1.84* (.14) -1.26* (.18)
seg. size (%) .21* (.01) .27* (.01) .34* (.01) .19* (.01)
WTP .40* (.03) .10* (.03) .95* (.08) .27* (.12)
Note: * indicates significance at p< .05. Standard errors in parenthesis.
Attribute
Segment 1 Segment 2 Segment 3 Segment 4
Brand .09 .08 .08 .51
Price .49 .82 .34 .27
FT label .22 .09 .36 .08
Packaging .21 <.01 .21 .13
Tab. 2: Segment-specific
estimates of parameters and
FT label WTP
Tab. 3: Segment-specific attribute importances
Here, the BIC statistic also decreases monotonically with
an increasing number of segments, but the decline slows
between the four- and five-segment solutions (BIC-dif-
ference of 212.15) compared to the three- to four-seg-
ment solutions (BIC-difference of 299.09). The entropy-
based measure saturates at quite high levels (88 % and
90 % for solutions with more than two segments), which
indicates well-separated segments. The first choice hit
rates indicate an excellent out-of-sample fit compared to
the values of a trivial benchmark (proportional chance
criterion, see Morrison 1969). All models increase the fit
by more than 100 %, but again, we see that the model
with five segments is only marginally better than the
four-segment solution. Based on the measures reported
in Tab. 1 one may favour the five-segment-solution.
However, a closer inspection of this solution reveals that
it contains several non-significant parameter estimates
indicating overfitting. Furthermore, two (of the five) seg-
ments are quite similar in their estimates. This prevents a
unique interpretation of segments, which is crucial from
a managerial point of view. Because it is a very good
trade-off between model fit, predictive validity, and in-
terpretability of segments, we select the four-segment so-
lution. This solution also provides a sound basis for our
subsequent profiling task.
3.2. Estimation results
Tab. 2 displays the associated segment-specific parame-
ter estimates of the four-segment solution. The segments
2, 3, and 4 most prefer the brand Hohes C, but the prefer-
ence order of the other brands differs across segments
(e. g., segment 1: Hohes C/Albi, Valensina, and Granini;
segment 4: Hohes C, Granini, Valensina, and Albi). Ad-
ditionally, all segments prefer orange juice with the FT
label. Furthermore, the linear price parameter has a rea-
sonable negative sign in all segments, implying that re-
spondents’ preferences decrease as prices increase. Seg-
ment 4 favours the PET bottle, in contrast to segments 1
and 3 that prefer a carton (Tetra Pak). Furthermore, all
segments have a meaningful size (19 % to 34 %), which
indicates that the four-segment solution does not suffer
from overfit.
The WTP for the FT label attribute is positive and signif-
icant in all segments and ranges between 10 Eurocent
(segment 2) and 95 Eurocent (segment 3). Standard er-
rors of the WTP estimates are computed using a paramet-
ric bootstrap with 10,000 draws (Krinsky and Robb
1986). The aggregated (i. e., segment-size weighted)
WTP for the FT label within our orange juice example is
approximately 48 Eurocents (95 % CI [42.00, 54.93]),
which translates into a relative price premium of approx-
imately 24 %, based on an average estimated WTP of
1.96 for orange juice alternatives in the experiment.
The magnitudes of the WTP estimates are reasonable
given recent meta-analysis results from Tully and Winer
(2014), who evaluated 80 studies dealing with socially
responsible consumption. If we subset their results (see
Appendix A of Tully and Winer 2014, pp. 267–271) to
food-related studies where people are the beneficiary of
the FT label (as in our study), the relative price premium
is approximately 29 %. Thus, our estimate is slightly
lower but still comparable with the results from litera-
ture, which strengthens our confidence in the validity of
our results.
An inspection of segment-specific attribute importance
in Tab. 3 reveals further insights that aid in explaining
these segment-specific differences in consumers’ WTP
for the FT label attribute.
While segment 2 almost exclusively cares about the
price of orange juice (82 %), the brand attribute is highly
important (51 %) for segment 4. Segment 1 attaches the
greatest importance to the price attribute (49 %), but it
also (nearly equally) cares about packaging (21 %) and
the FT label (22 %). Members of segment 3 care primari-
ly about the FT label (36 %) but also consider the price
attribute (34 %) in their purchase decision process. This
behaviour coincides with the results of Devinney et al.
(2012), who concluded that companies should (also)
heavily focus on appropriate pricing for their (socially
enhanced) products because even consumers who care
Paetz/Guhl, Understanding Differences in Segment-specific Willingness-to-pay for the Fair Trade Label
MARKETING · ZFP · Issue 4 · 4. Quarter 2017 43
Fig. 2: Proportions of individual
background variables in seg-
ments
about social product attributes (e. g., the FT label) also
care about core product attributes (cf. Auger et al. 2008,
p. 190). This calls for a holistic and integrated mixed
marketing strategy including investments in socially re-
sponsible production practices.
Segment-specific preference characteristics are also pre-
sent in the corresponding WTP values. While the price-
sensitive segment 2 yields the smallest WTP for the FT
label of 10 Eurocents, social segment 3 exhibits the high-
est FT label WTP of 94 Eurocents. In sum, this answers
our first research question on WTP differences between
different preference-based segments.
3.3. Profiling segments
Based on the four-segment solution for the FM-MNL
model from Section 3.2, we now assign respondents to
particular segments and profile the segments using indi-
vidual background variables. This reveals further valu-
able insights concerning our second research question of
how these differences can be explained in terms of the
socio-demographic or psychographic characteristics of
segment members.
First, we analyse the cfc variable. The group means for
segments 1 to 4 are .22, -.23, .19, and -.27, respectively.
A one-way ANOVA clearly shows that the cfc-means are
not the same across groups (F= 6.17, df =3,p< .01).
Segments with higher average cfc-values also show
higher FT label WTP (e. g., segment 1 and 3). This repli-
cates the results of Balderjahn and Peyer (2012), who re-
port a positive influence of consumers’ cfc level on con-
sumers’ WTP for the FT label.
Figure 2, furthermore, illustrates gender-specific differ-
ences between segments. The proportion of women in
segment 1 (61 %) and segment 3 (67 %) exceeds those in
segment 2 (51 %) and segment 4 (51 %). Hence, seg-
ments with a higher WTP for the FT label include higher
proportions of women. These gender-specific differences
between segments is (weakly) significant based on a con-
tingency analysis (
χ
2= 7.39, df =3,p= .06). This mir-
rors the findings of gender’s influence on consumers’
WTP found in Italian (Rotaris and Danielis 2011) and
Chinese (Yang et al. 2012) samples.
The contingency analysis did not find significant differ-
ences in age (classes) between segments (
χ
2
= 1.21, df =3,
p= .75), which is in line with the results of Yang et al.
(2012). However, a contingency analysis discovered seg-
ment-specific differences in the frequency of orange juice
consumption (
χ
2
= 13.62, df =6,p= .03), which mirrors
the results of Rotaris and Danielis (2011) in the Italian
sample. Segments 1 and 3 contain higher proportions of
members who consume orange juice at least once a week
than do segments 2 and 4, which exhibit a lower WTP for
the FT label. Hence, those segments consuming orange
juice more often are willing to pay higher price premia for
the FT label than those segments consuming less often.
4. Conclusion
In this contribution, we focused on segment-specific dif-
ferences in consumers’ WTP for the FT label as well as
on determinants for these WTP differences. We conduct-
ed a discrete choice experiment for orange juice at two
German universities and estimated FM-MNL models.
We selected a four-segment solution and calculated seg-
ment-specific WTPs for the FT label by using the mar-
ginal rate of substitution between the FT label and price.
We found substantial differences in segment-specific
WTP for the FT label: between 10 and 95 Eurocents. The
average value in the sample is 48 Eurocents, which cor-
responds to a relative price premium of 24 %. The profil-
ing of segments with individual background variables re-
vealed that these differences could be related to differ-
ences in segment members’ gender, consumption fre-
quency, and cfc level. Segments with a larger proportion
of women, a higher cfc level or a more frequent con-
sumption of orange juice have, all else being equal, a
higher WTP for the FT label. However, consumer’s age
did not affect their WTP in our sample.
Paetz/Guhl, Understanding Differences in Segment-specific Willingness-to-pay for the Fair Trade Label
44 MARKETING · ZFP · Issue 4 · 4. Quarter 2017
Our results also have relevant managerial implications. In
contrast to a priori segmentation, which is commonly
based on one individual background variable, preference-
based one-stage segmentation approaches with the subse-
quent segment profiling with several individual back-
ground variables contribute to a deeper understanding of
social consumer segments and are easy to conduct. Mar-
keting managers can now identify that the target segment
for an FT orange juice consists of women with a high cfc
level who consume orange juice frequently. Mixed mar-
keting strategies, e. g., managerial decisions on product
line extensions, price differentiation or tailored promo-
tional campaigns, could be built on this information and
potentially lead to increasing company profits because
consumers are addressed more appropriately.
Our study is based on a convenience (student) sample,
where young and well-educated respondents are over-
sampled. Because students are known to be more recep-
tive to the FT context in general (cp. Yang et al. 2012, p.
24), our WTP estimates may be upwardly biased. Addi-
tionally, we used stated choices from a discrete choice
experiment, rather than revealed market data, which may
also contribute to inflated WTP estimates (cp. Völkner
2006). However, no top German orange juice brand has a
FT label so far, which renders impossible the use of re-
vealed choice data for WTP calculations in our specific
application. Furthermore, our results in terms of WTP
value seem to be conservative compared with the results
of the recent meta-analysis of Tully & Winer (2014). Ad-
ditionally, the corresponding general price elasticities of
the FM-MNL model are elastic, reasonable in magnitude
(e. g., between -6 and -1.5) and comparable to other stud-
ies using market data (cf. Weber and Steiner 2012).
Hence, we are confident that our WTP estimates for the
FT label are not (heavily) overestimated. Furthermore,
our main findings primarily concern the segment-specif-
ic WTP differences in consumers’ gender, cfc level and
consumption frequency, and we are confident that our
findings are replicable as a representative German sam-
ple.
We identified several drivers for WTP differences, which
were either already verified or discarded for being sam-
ples of varying cultures. Since the recent (sparse) litera-
ture already observed differences in WTP between dif-
ferent cultures, future research should focus more heavi-
ly on cultural differences in WTP determinants. In this
context, the investigation of emerging versus developed
countries may constitute an especially interesting re-
search field and furthermore relates to prospective inves-
tigations of consumer income as a WTP driver. Some re-
searchers (e. g., Sonnier et al. 2007) also advocate speci-
fying and estimating discrete choice models directly in
WTP spaces instead of preference spaces and then, as we
did, computing WTP values using equation (1). Differ-
ences between both approaches might be relevant if a
continuous specification of consumer heterogeneity is
used (i. e., mixed logit models, see Elshiewy et al. 2017).
In our application, the differences for the FM-MNL
should be negligible. However, we still think that em-
ploying a discrete choice model in WTP space is a fruit-
ful avenue for future research in the context of analysing
WTP for FT labels.
Appendix
In this appendix, we provide technical details regarding the meth-
ods (model, estimation, and segmentation) used in the analysis.
We follow closely DeSarbo et al. (1995) and Wedel and Kamakura
(2000). See also Elshiewy et al. (2017) for an overview of multi-
nomial logit models in marketing, incl. the Finite Mixture-Multi-
nomial Logit (FM-MNL) model used here.
We assume that respondents come from a population composed of
several unobserved segments and that they have the following
general (indirect) utility function:
Ujit
⏐
s=k=1
K
Σ
β
ksxjitk +
α
spjit +
ε
jits,(A1)
with:
j=1,...,Jrespondents,
i=1,...,Ichoice alternatives,
t=1,...,Tchoice sets,
s=1,...,Slatent segments,
k=1,...,Kattributes and dummy variables excluding price,
xjitk: for respondent j,thekth dummy variable of alternative iin
choice set t,
pjit: for respondent j, the price variable of alternative iin
choice set t,
β
ks:effectofthekth attribute for segment s,
α
s: effect of the price for segment s,
yjit: choice variable, equals 1 if respondent jchooses alterna-
tive iin choice set tand 0 otherwise,
ε
jits: random components, assumed to be i.i.d. extreme value
type I distributed.
Hence, utility Ujit
⏐
sis linear additive in the effects of the covariates
xjitk and pjit as well as the error term
ε
jits. Furthermore, respondents
are utility maximisers, i. e., respondent jpicks an alternative iif
no other alternative q≠iin choice set thas a higher utility. To-
gether with the particular choice for the distribution of the random
components of the model (
ε
jits) this leads to the well known multi-
nomial logit (MNL) model for the choice probabilities Pr:
Prjit
⏐
s=exp(Ujit
⏐
s)
i⬘=1
I
Σexp(Uji⬘t
⏐
s).(A2)
These choice probabilities are conditional on respondent jbelong-
ing to segment s. The unconditional choice probability for alterna-
tive iis:
Prjit =s=1
S
Σ
π
sPrjit
⏐
s.(A3)
Here the conditional choice probabilities (A2) are multiplied by
the unknown size of a segment
π
s, which can be interpreted as the
apriorprobability of finding a respondent in segment s:
π
s=exp(
γ
s)
s⬘=1
S
Σexp(
γ
s⬘).(A4)
Hence, for the segment sizes, we also assume a MNL (sub)model,
where
γ
sare parameters to be estimated, with
γ
1= 0 as additional
restriction for identification. Note that s=1
S
Σ
π
s=1and0<
π
s<1.
Model parameters are estimated using maximum likelihood, and
we employ the following log-likelihood function of the FM-MNL
model:
Paetz/Guhl, Understanding Differences in Segment-specific Willingness-to-pay for the Fair Trade Label
MARKETING · ZFP · Issue 4 · 4. Quarter 2017 45
LL(
θ
)=j=1
J
Σln⎛
⎝s=1
S
Σ
π
st=1
T
Πi=1
I
ΠPjit
⏐
s
yjit
⎞
⎠,(A5)
where the vector
θ
=(
α
1,...,
α
S,
β
11,...,
β
KS,
γ
2,...,
γ
S)con-
tains all parameters to be estimated. Note that we exploit the struc-
ture of the data in (A5), where due to the data collection design all
respondents have Tobservations, because the whole product of
conditional choice probabilities for the observed sequence of
choices for each respondent is weighted by the segment sizes. We
estimate all (K+2)·S– 1 parameters simultaneously using gradi-
ent-based methods. Specifically, we use the Broyden-Fletcher-
Goldfarb-Shanno (BFGS) algorithm (see Greene 2008, p. 1071)
implemented in the gmnl package in R (Sarrias and Daziano
2017). Because mixture models may have multiple local optima
(see Wedel and Kamakura 2000), multiple starting values should
be tested to find the global optimum (cf. Grün 2008, p. 235).
Once the parameters have been estimated, Bayes’ theorem can be
used to compute the posterior probability for respondent jbelong-
ing to segment s:
τ
ˆ
js =
π
ˆ
st=1
T
Πi=1
I
ΠP
ˆjit
⏐
s
yjit
s⬘=1
S
Σ
π
ˆ
s⬘t=1
T
Πi=1
I
ΠP
ˆjit
⏐
s⬘
yjit .(A6)
In (A6), the estimated prior probabilities are re-weighted using the
estimated likelihood of each respondent jconditional on segment s.
These probabilities can be directly used for a “fuzzy” (i. e., probabi-
listic) segmentation, where respondents can be fractional members
of multiple segments (cf. Wedel and DeSarbo 1994). However, we
follow DeSarbo et al. (1995) and apply a non-overlapping segmen-
tation where we form discrete segments by assigning each respon-
dent jto the segment swhere the value for
τ
ˆ
js
is highest.
Because the number of segments Sis unknown before the analy-
sis, we repeat the model estimation for multiple models with vary-
ing numbers of segments and use multiple measures to choose a
specific value for S. In particular, we evaluate model fit using the
LL-value (in-sample) and hit rates (out-of-sample). The latter is
defined as the proportion of correctly predicted choices not used
for estimation. Apparently, this measure ranges between zero and
one, where a higher percentage conveys a better predictive validi-
ty. However, for interpretation, this measure should be compared
to a meaningful benchmark, such as the proportional chance crite-
rion or the max chance criterion (Morrison 1969). Further, we use
the Bayesian information criterion (BIC, cf. Jain et al. 1994, p.
320):
BIC =–2·LL +ln(J·T)·((K+2)·S–1) (A7)
and the Entropy (Wedel and Kamakura 2000, p. 92; Ramaswamy
et al. 1993):
ES=1+ j=1
J
Σs=1
S
Σ
τ
ˆ
jsln(
τ
ˆ
js)
Jln(S).(A8)
BIC supports model selection by penalizing model fit with model
complexity (i. e., number of parameters), so good fitting but parsi-
monious models should be preferred. On the other hand, ESin-
forms about how distinct the assignment of respondents to the seg-
ments for a particular choice of Sis and can have a values between
0 (i. e., all posterior probabilities are the same) and 1 (i. e., perfect
distinction). A better fitting FM-MNL model is not necessarily
preferable if the segments overlap significantly. Our non-overlap-
ping segmentation works particular well if ESis high.
References
Allenby,G.M.,Brazell,J.D.,Howell,J.R.,&Rossi,P.E.(2014).
Economic Valuation of Product Features. Quantitative Market-
ing and Economics, 12(4), 421–456.
Andorfer, V. A., & Liebe, U. (2012). Research on Fair Trade Con-
sumption – A Review. Journal of Business Ethics, 106(4), 415–
435.
Arnot, C., Boxall, P. C., & Cash, S. B. (2006). Do Ethical Consu-
mers Care about Price? A Revealed Preference Analysis of Fair
Trade Coffee Purchases. Canadian Journal of Agricultural Eco-
nomics, 54(4), 555–565.
Auger, P., & Devinney, T. M. (2007). Do What Consumers Say
Matter? The Misalignment of Preferences with Unconstrained
Ethical Intentions. Journal of Business Ethics, 76(4), 361–383.
Auger, P., Devinney, T. M., Louviere, J. J., & Burke, P. F. (2008).
Do Social Product Features have Value to Consumers? Interna-
tional Journal of Research in Marketing, 25(3), 183–191.
Balderjahn, I., & Peyer, M. (2012). Das Bewusstsein für fairen
Konsum: Konzeptualisierung, Messung und Wirkung. Die Be-
triebswirtschaft, 72(4), 343–364.
Balderjahn, I., Peyer, M., & Paulssen, M. (2013). Consciousness
for Fair Consumption: Conceptualization, Scale Development
and Empirical Validation. International Journal of Consumer
Studies, 37(5), 546–555.
Basu, A., & Hicks, R. (2008). Label Performance and the Willing-
ness to Pay for Fair Trade Coffee: A Cross National Perspec-
tive. International Journal of Consumer Studies, 32(5), 470–
478.
Breidert, C., Hahsler, M., & Reutterer, T. (2006). A Review of
Methods for Measuring Willingness-to-Pay. Innovative Market-
ing, 2(4), 8–32.
Cobb-Walgren, C. J., Ruble, C. A., & Donthu, N. (1995). Brand
Equity, Brand Preference, and Purchase Intent. Journal of Ad-
vertising, 24(3), 25–40.
De Pelsmacker, P., Driessen, L., & Rayp G. (2005a). Do Consu-
mers Care about Ethics? Willingness to Pay for Fair-Trade Cof-
fee. Journal of Consumer Affairs, 39(2), 363–385.
De Pelsmacker, P., Janssens, W., Sterckx, E., & Mielants, C.
(2005b). Consumer Preference for the Marketing of Ethically
Labelled Coffee. International Marketing Review, 22(5), 512–
530.
DeSarbo, Wayne S., Ramaswamy, V., & Cohen, S. H. (1995).
Market Segmentation with Choice-Based Conjoint Analysis.
Marketing Letters, 6(2), 137–147.
Devinney, T. M., Auger, P., & Eckhardt, G. (2012). Can the So-
cially Responsible Consumer be Mainstream?. Journal for
Business, Economics & Ethics, 13(3), 227–235.
Elshiewy, O., Guhl, D., & Boztug, Y. (2017). Multinomial Logit
Models in Marketing – From Fundamentals to State-of-the-
Art. Marketing ZFP – Journal of Research and Management,
39(3), 32–55.
Fair Trade Deutschland (2014). Statement Fairer Handel am Bei-
spiel Orangensaft, http://alt.fairtrade-deutschland.de/fileadmin/
user_upload/ueber_fairtrade/fairtrade-themen/fairtrade_state-
ments_saft.pdf (accessed August 18, 2017).
Greene, W. H. (2008). Econometric Analysis.6
th Ed, New Jersey:
Pearson Prentice Hall.
Grün, B. (2008). Identifiability of Finite Mixtures of Multinomial
Logit Models with Varying and Fixed Effects. Journal of Clas-
sification, 25(2), 225–247.
Jain, D. C., Vilcassim, N. J., & Chintagunta, P. K. (1994). A Ran-
dom-Coefficients Logit Brand-Choice Model Applied to Panel
Data. Journal of Business & Economic Statistics,12(3),317–
328.
Johnson, R. M., & Orme, B. (1996) How Many Questions Should
You Ask in Choice-Based Conjoint Studies? https://www.saw
toothsoftware.com/download/techpap/howmanyq.pdf (accessed
August 26, 2017).
Krinsky,I.,&Robb,L.(1986),OnApproximating the Statistical
Properties of Elasticities. Review of Economics and Statistics,
68 (4), 715–719.
Louviere, J. J., Hensher, D. A., & Swait, J. D. (2000). Stated
choice methods, Cambridge: Cambridge University Press.
Melnikov, V., & Maitra, R. (2010). Finite Mixture Models and
Model-Based Clustering. Statistical Surveys, 4, 80–116.
Paetz/Guhl, Understanding Differences in Segment-specific Willingness-to-pay for the Fair Trade Label
46 MARKETING · ZFP · Issue 4 · 4. Quarter 2017
Morrison,D.G.(1969).OntheInterpretation of Discriminant
Analysis. Journal of Marketing Research, 6(2), 156–163.
Orme, B. (2002). Formulating Attributes and Levels in Conjoint
Analysis. https://www.sawtoothsoftware.com/download/tech-
pap/formatt.pdf (accessed August 25, 2017).
Orme, B. (2007). Three Ways to Treat Overall Price in Conjoint
Analysis. https://www.sawtoothsoftware.com/download/tech-
pap/price3ways.pdf (accessed August 18, 2017).
Ramaswamy, V., & Cohen, S. H. (2007). Latent Class Models for
Conjoint Analysis. In A. Gustafsson, A. Hermann, & F. Huber
(Eds.), Conjoint Measurement,4
th Ed., Berlin: Springer-Verlag,
295–319.
Ramaswamy, V., DeSarbo, W., Reibstein, D. J., & Robinson, W. T.
(1993). An Empirical Pooling Approach for Estimating Market-
ing Mix Elasticities with PIMS data. Marketing Science, 12(1),
103–124.
Rotaris, L., & Danielis, R. (2011). Willingness to Pay for Fair
Trade Coffee: A Conjoint Analysis Experiment with Italian
Consumers. Journal of Agricultural & Food Industrial Organi-
zation, 9(1), 1–20.
Sarrias, M., & Daziano, R. A. (2017). Multinomial Logit Models
with Continuous and Discrete Individual Heterogeneity in R:
The gmnl Package. Journal of Statistical Software, 79(2), 1–46.
Sawtooth Software (2014). Discover-CBC: How and Why It Dif-
fers from SSI Web’s CBC Software. https://www.sawtoothsoft
ware.com/support/technical-papers/169-support/technical-pape
rs/cbc-related-papers/1268-saas-delivered-cbc-for-the-classroom
-how-and-why-it-differs-from-ssi-web-s-cbc-software (access-
ed August 25, 2017).
Small, K. A., & Rosen, H. S. (1981). Applied Welfare Economics
with Discrete Choice Models. Econometrica, 49(1), 105–130.
Sonnier, G., Ainslie, A., & Otter, T. (2007). Heterogeneity Distri-
butions of Willingness-to-Pay in Choice Models. Quantitative
Marketing and Economics, 5(3), 313–331.
Starbucks (2015). Global Responsibility Report 2015, https://glo
balassets.starbucks.com/assets/ee8121c1a6554399b554d12622
8d52ed.pdf (accessed May 18, 2017).
Statista (2016). Inlandsabsatz der führenden Hersteller von
Fruchtsäften und Nektaren in Deutschland im Jahr 2016 (in
Millionen Liter), https://de.statista.com/statistik/daten/studie/
76900/umfrage/hersteller-fruchtsaft-fruchtnektar-getraenke-in-
landsabsatz/(accessed May 18, 2017).
Tully, S. M., & Winer, R. S. (2014). The Role of the Beneficiary in
Willingness to Pay for Socially Responsible, Products: A Meta-
Analysis. Journal of Retailing, 90(2), 255–274.
Tuma, M. N., & Decker, R. (2013). Finite Mixture Models in Mar-
ket Segmentation: A Review and Suggestions for Best Prac-
tices. Electronic Journal of Business Research Methods, 11(1),
2–15.
Völkner, F. (2006). An Empirical Comparison of Methods for
Measuring Consumers’ Willingness to Pay. Marketing Letters,
17(2), 137–149.
Vriens, M., Wedel, M., & Wilms, T. (1996). Metric Conjoint Seg-
mentation Methods: A Monte Carlo Comparison. Journal of
Marketing Research,33(1), 73–85.
Weber, A., & Steiner, W. J. (2012). Zur Berücksichtigung von He-
terogenität versus funktionaler Flexibilität in Absatzreaktions-
modellen: Eine empirische Studie auf Basis von Handelsdaten.
Zeitschrift für Betriebswirtschaft, 82(12), 1337–1365.
Wedel, M., & DeSarbo, W. S. (1994). A Review of Recent Devel-
opments in Latent Class Regression Models. In R. Bagozzi
(Ed.), Advanced Methods of Marketing Research, Cambridge:
Blackwell Pub., 352–388.
Wedel, M., & Kamakura, W. (2000). Market Segmentation: Con-
ceptual and Methodological Foundations,2
nd Ed., Boston et al.:
Kluwer.
Yang, S.-H., Hu, W., Mupandawana, M., & Liu, Y. (2012). Con-
sumer Willingness to Pay for Fair Trade Coffee: A Chinese
Case Study. Journal of Agricultural and Applied Economics,
44(1), 21–34.
Keywords
Fair Trade Label, Discrete Choice Experi-
ment, Finite Mixture-Multinomial Logit Model.
Paetz/Guhl, Understanding Differences in Segment-specific Willingness-to-pay for the Fair Trade Label
MARKETING · ZFP · Issue 4 · 4. Quarter 2017 47
