Food Quality and Preference 39 (2015) 28–39

Contents lists available at ScienceDirect

Food Quality and Preference

journal homepage: www.elsevier.com/locate/foodqual

Investigating individual preferences in rating and ranking conjoint

experiments. A case study on semi-hard cheese
Valérie Lengard Almli a,b,⇑, Arnstein Øvrum c, Margrethe Hersleth a,b, Trygve Almøy b, Tormod Næs a
a
Nofima AS, PO Box 210, 1431 Ås, Norway
b
The Norwegian University of Life Sciences, Department of Chemistry, Biotechnology and Food Science, PO Box 5003, N-1432 Ås, Norway
c
Norwegian Agricultural Economics Research Institute, PO Box 8024 Dep, 0030 Oslo, Norway

a r t i c l e i n f o a b s t r a c t

Article history: Stated preference conjoint experiments and self-explicated measures based on rating and ranking
Received 16 December 2011 approaches were conducted to investigate Norwegian consumers’ choices among healthier and organi-
Received in revised form 15 June 2014 cally produced semi-hard cheeses. In the conjoint experiments, one group of participants (n = 114) per-
Accepted 17 June 2014
formed a rating task of eight cheeses whereas the other group (n = 105) performed a ranking task of the
Available online 25 June 2014
same cheeses, all based on pictorial stimuli only. Then, all participants performed self-explicated rating
and ranking evaluations of the cheese attributes. Conjoint rating data were analysed by mixed model
Keywords:
ANOVA, while conjoint ranking data were analysed by mixed logit. The different approaches are com-
Method comparison
Conjoint
pared in terms of data analysis methodologies, outcomes and practicalities for the experimenter as well
Self-explicated measures as for the respondents. Rather than average population effects, focus is brought on individual preferences
Mixed logit and consumer segmentation. Findings reveal that the two conjoint experiments lead to similar popula-
Consumer segmentation tion effects and consumer segments. Consumers on average prefer cheeses of new (healthier) fat compo-
Cheese sition, organic production and lower price to cheeses of regular fat composition, conventional production
and higher price. Two consumer segments are investigated. Consumers in the New fat segment are
health-conscious, whereas consumers in the Regular fat segment are attracted by conventional cheese
and lower prices. Self-explicated ratings of the cheese attributes corroborate these findings.
Ó 2014 Elsevier Ltd. All rights reserved.

Introduction variations and consumer segmentation. Further, conjoint experi-

ments may often be complex to design, time-consuming to perform
Experimental approaches are widely used to study consumer and costly to carry-out (Sattler & Hensel-Börner, 2003). A second
responses to food products. A first level of research on consumer aspect of this paper is thus to compare conjoint approaches with
experimental methods concerns the selection of a methodology, self-explicated approaches, where the consumer is plainly asked
comparing for example experimental auctions to conjoint studies about preference levels for a product’s attributes (Sattler &
(Grunert et al., 2009; Sichtmann & Stingel, 2007), or combining such Hensel-Börner, 2003).
methods (Combris, Bazoche, Giraud-Heraud, & Issanchou, 2009). A
second level of research concerns possible options within one meth-
Rating and ranking scales
odology. This paper addresses the latter by comparing an accep-
tance rating test to a preference ranking test in a conjoint study
Several rating and ranking scales have been developed and are
on generic unbranded semi-hard cheese. More specifically, focus
commonly used in consumer testing (Hein, Jaeger, Carr, &
is brought on modelling strategies with regard to the different nat-
Delahunty, 2008). We will here focus on the types utilised in the
ure of rating and ranking data. As preference heterogeneity is a very
present conjoint study: acceptance rating with a 9-point category
relevant and natural element of food choice research, described as
scale ranging from 1 to 9, and preference ranking with no ties
‘‘a key and permanent feature of food choices’’ (Combris et al.,
allowed (forced choice). In acceptance rating, consumers evaluate
2009), emphasis is made on studying inter-individual preference
each product separately and rate these according to their degree
of appreciation. Rating generates an indirect measure of product
⇑ Corresponding author at: Nofima AS, PO Box 210, 1431 Ås, Norway. Tel.: +47 64 distances. In preference ranking, consumers order products accord-
97 03 05; fax: +47 64 97 03 33. ing to their preferences from best to worst. Ranking involves
E-mail address: valerie.almli@nofima.no (V.L. Almli). performing a succession of product choices where the consumer

http://dx.doi.org/10.1016/j.foodqual.2014.06.011
0950-3293/Ó 2014 Elsevier Ltd. All rights reserved.
 V.L. Almli et al. / Food Quality and Preference 39 (2015) 28–39 29

is forced to discriminate between products, but no information (Ortúzar, 2010). Mixed logit is an advanced discrete choice model
regarding the degree of appreciation is obtained (Hein et al., where one may freely include random parameters of any distribu-
2008). Rating and ranking methods have previously been com- tions and correlations between random factors. This flexibility
pared in a number of studies (see e.g. Hein et al., 2008; Kozak & allows writing models that better match real-world situations. By
Cliff, 2013; Lagerkvist, 2013; Villanueva, Petenate & Da Silva, including random parameters, mixed logit intrinsically models
2000, 2005), often with a general focus on mean population results preference heterogeneity, i.e., inter-individual preference varia-
comparisons. In a comprehensive method comparison study, Hein tions. Further, mixed logit acknowledges the fact that any food
et al. (2008) tested five common acceptance and preference choice decision in the experiment, in this case any product ranking,
methods based on rating and ranking approaches: 9-point hedonic may be dependent on the consumer’s previous decisions. Even
scale, labelled affective magnitude scale, unstructured line scale, though discrete data is common in sensory and consumer science,
best–worst scaling and preference ranking. Their main finding is there is no tradition in sensometrics for mixed logit, which was
that all five methods lead to the same conclusions regarding the recently introduced to the field by Barreiro-Hurlé, Colombo, and
products, with slight performance differences observed in product Cantos-Villar (2008), Jaeger and Rose (2008) and Ortúzar (2010).
discrimination power, ease of use and perceived accuracy in favour We refer to the latter for a sound introduction to the mixed logit
of the best–worst scaling method. However these authors worked model and to Train (2009) for a comprehensive description.
with hedonic tests involving real food stimuli and the results may Following the study of mean population effects, a study of
not necessarily generalise to other contexts, such as pictorial stim- preference heterogeneity is often required to identify trends within
uli in a web-based survey. Further, their study neither investigated subgroups of the consumer sample. Various methods of consumer
conjoint factors, nor compared the different methods in terms of segmentation may be applied, such as clustering algorithms, visual
consumer segmentation. These issues will be addressed in the segmentation based on Principal Component Analysis (PCA) (Almli
present paper in the case of two rating and ranking approaches. et al., 2011) or fuzzy clustering (Johansen, Hersleth, & Næs, 2010;
Næs et al., 2010a; Westad, Hersleth, & Lea, 2004). It is also possi-
Self-explicated and conjoint approaches ble to induce segments in a latent class model (Hess, Ben-Akiva,
Gopinath, & Walker, 2011; Mueller, Lockshin, Saltman, &
Self-explicated approaches consist in testing consumers’ Blanford, 2010) or in a clustering around latent variables model
attitudes or preferences for product attributes by directly asking (Vigneau, Endrizzi, & Qannari, 2011; Vigneau, Qannari, Punter, &
about the attributes rather than presenting products. Such Knoops, 2001). Beyond the selection of a statistical approach, there
approaches are often seen in comparison to conjoint methods, are two main strategies to choose from when addressing clustering
which by using a complex design setup aim at collecting more reli- purposes: one may either create consumer groups of similar
able data than self-explicated measures. Among other, it is background such as gender, income, attitudes or purchase habits,
believed that conjoint methods increase the similarity to real or create consumer groups of similar product preferences. The first
choice situations and decrease the risk of collecting socially accept- strategy is sometimes called a priori segmentation (Næs et al.,
able answers (Sattler & Hensel-Börner, 2003). Sattler and Hensel- 2010a) and is based on splitting the consumer group into segments
Börner (2003), however, report that studies that compare conjoint according to consumer characteristics and analysing the group
and self-explicated measures generally conclude that their perfor- preferences separately or together in an ANOVA model. The second
mances are either equivalent, or different in favor of self-explicated strategy is based on analysing the actual preference, liking or
measures. It is therefore interesting to study how these methods purchase intent data to create segments, then relating segments
compare to each other when studying stated preferences for food to consumer characteristics a posteriori. In the present paper the
choices. second strategy will be used. To perform consumer segmentation
based on individual acceptance ratings, a multi-step approach
Data analysis introduced by Næs et al. (Næs, Lengard, Johansen, & Hersleth,
2010b; see also: Endrizzi, Menichelli, Johansen, Olsen, & Næs,
Acceptance rating tests generate (nearly) continuous data, 2011; Endrizzi, Gasperi, Rødbotten, & Næs, 2014) is applied. To
whereas preference ranking tests generate ordinal, discrete data. perform consumer segmentation in the case of preference ranking,
Accordingly, in conjoint experiments with rating scales the popula- a new approach is presented based on individual model estimates
tion effects from consumers’ evaluations are typically analysed by from mixed logit and inspired by the method in Næs et al. (2010b).
mixed model ANOVA (ANalysis Of VAriance), that is to say an In both cases, segmentation will be done based on visual interpre-
ANOVA model combining fixed and random effects and usually tation of PCA plots of the individual differences. The main advan-
assuming normal distributions for the random parts (Næs, tage of such an approach is that one can decide on which
Brockhoff, & Tomic, 2010a). In practice, ordinal measures can be segments or groups of consumers one is interested in studying.
approximated to continuous measures, such that ANOVA is also Another argument for such an approach is that using different
frequently used on ranking data even though this method is not automatic clustering methods can give quite different results,
designed for discrete data (Villanueva, Petenate, & Da Silva, 2000, and also results which are difficult to interpret in terms of samples
2005). One must, in particular, be aware of the fact that the ranks tested (see Endrizzi et al., 2014).
are highly dependent on each other in small studies and the
assumptions underlying standard ANOVA may be strongly vio- Objectives
lated. More appropriately, in the field of econometrics ranking data
and other choice-based data are routinely analysed by so-called The data presented in this paper are extracted from a large con-
discrete choice models. Discrete choice models aim at understand- joint experiment conducted in Norway in 2009 investigating the
ing the behavioural process that leads to a consumer’s choice effect of health information on consumers’ diet choices (Øvrum,
(Train, 2009). The approach consists in modelling Utility, that is Alfnes, Almli, & Rickertsen, 2012). In the present paper, only the
to say the net benefit a consumer obtains from selecting a specific control group of participants who did not receive health informa-
product in a choice situation. These models emerged in the 1970s tion are utilised. In particular, the study investigates consumer’s
and have undergone a rapid development from the original fixed willingness to buy full fat vs. low fat cheese and cheese of regular
coefficients models such as multinomial logit, to the highly general fat composition vs. new fat composition, which includes a higher
and flexible mixed logit, also called Random Parameter Logit unsaturated fat/saturated fat ratio. The factor corresponding to a
30 V.L. Almli et al. / Food Quality and Preference 39 (2015) 28–39

new, healthier fat composition is of major interest in this study and are somewhat overrepresented in the ranking group compared to
will guide the consumer segmentation. This innovation was not the rating group. The total sample (n = 219) compares to national
present yet on the Norwegian market at the time of the consumer census data for the targeted age group (30–70 years old) in terms
experiment. of gender composition and is slightly higher in mean age (Table 2).
The objective of this study is threefold: (i) present and compare
modelling strategies for studying population effects and preference Test protocol
heterogeneity in conjoint rating and ranking experiments, (ii) The same cheese pictures were used both in rating and ranking
investigate consumers’ stated preferences for various attributes conjoint experiments (Table 1). For all participants, the survey
in every day-use semi-hard cheese at population and segment started with a welcoming introduction and a brief presentation
levels and (iii) compare conjoint and self-explicated methods for of the three symbols used on the cheese packagings to ensure a
eliciting consumer acceptance. common interpretation of the conjoint factors. Then, for the rating
group eight successive screens presenting the eight cheeses were
Materials and methods shown in randomized balanced order. The consumers evaluated
their Willingness To Buy (WTB) the cheeses on 9-point scales
Consumer test anchored with ‘‘I would definitely not purchase’’ and ‘‘I would def-
initely purchase’’. For the ranking group, a ranking test was organ-
Cheese samples ised in seven successive screens. A first screen presented all eight
Eight pictures of generic every day-use semi-hard cheese pack- cheeses and participants were asked to click on the four items they
ages were generated according to a 241 fractional factorial design would most probably purchase. The second screen showed these
IV
with variations in fat content (full fat vs. low fat), fat composition four selected cheeses and participants were asked to indicate the
(regular vs. increased unsaturated fat/saturated fat ratio), sustain- item they would most probably purchase among the four. The third
able production (conventional vs. organic) and price (NOK 42 vs. and fourth screens showed the three (resp. two) remaining cheeses
NOK 58 per 500 g) as presented in Table 1. In this experimental and participants were asked to indicate the item they would most
design each two-way interaction is confounded with another one probably purchase among the three (resp. two). Then, the proce-
(Low fat * New fat + Organic * Price, New fat * Organic + Low fat * - dure was repeated on the four rejected cheeses from the original
Price and New fat * Price + Low fat * Organic) but not with main eight. In the following, these conjoint experiments will be referred
effects. to as ‘‘conjoint rating’’ and ‘‘conjoint ranking’’.
For each factor combination, the picture included the cheese’s Following the conjoint experiments, participants were ques-
price as well as symbols corresponding to factors organic produc- tioned about the importance of factors fat content, fat composition,
tion, low fat cheese and cheese with new fat composition (Fig. 1). organic production and price in self-explicated measures (Sattler &
By contrast, the absence of these symbols indicated full fat content, Hensel-Börner, 2003). They first rated each factor on a 5-point Likert
regular fat composition and conventional production process, scale anchored from ‘‘Very little importance’’ to ‘‘High importance’’,
respectively. All three symbols were present on the Norwegian then ranked the same factors from the most to the least important
market at the time of the experiment. In the following, reference one. In the following, these evaluations will be referred to as
to the cheese samples will refer to the constructed photographs ‘‘self-explicated rating’’ and ‘‘self-explicated ranking’’. These direct
of cheese packages with varying prices and symbols. measures of factor importance will be compared to the indirect
measures obtained through the conjoint experiments. Finally, the
participants filled in a questionnaire including behavioural and life-
Consumers style items, attitudinal items from the Food Choice Questionnaire
A sample of 219 Norwegian consumers across the country par- (Steptoe, Pollard, & Wardle, 1995) and socio-demographic items.
ticipated in a web-based experiment. They were selected on the
criteria that they eat semi-hard cheese at least once a week, are
Data analysis of conjoint rating
frequently responsible for food purchases for the household and
do not work in the food or marketing sectors. Participants were
Mixed model ANOVA
potentially rewarded by the draw of three universal gift coupons
A mixed model ANOVA was run to identify significant effects for
for a value of NOK 1000 (approx. € 125). In a first step, the study
the total group of consumers. This model includes low fat, new fat,
consisted in either a rating or a ranking conjoint test on the eight
organic, price and three interaction effects between conjoint factors
cheeses presented in Table 1. The assignment of participants to
as fixed factors, and consumer as random factor (see the confound-
one or the other test was done semi-randomly by the system,
ing pattern of the experimental design in section Cheese samples,
aiming at ensuring a balanced repartition according to gender,
above). In addition, random interaction effects between consumer
age, education and region of residence. Table 2 presents key
and the four conjoint factors and their interactions were included
socio-demographic indicators for the rating (n = 114) and ranking
to account for individual preferences. The model is written:
(n = 105) groups of consumers. The two groups present similar dis-
Y = mean + consumer effect + main effects for conjoint vari-
tributions in gender, age, household size and household income.
ables + 2-way interactions between conjoint variables + 2-way
Participants of university education and overweight participants
interactions between conjoint variables and consumer + 3-way
Table 1 interactions between consumer and 2-way interactions of conjoint
Fractional factorial design used in the conjoint experiments. variables + random noise.
Cheese Code Low fat New fat type Organic Price (NOK/500 g) More specifically,
1 1000 Yes No No 42 yijklmp ¼ l þ sm þ ai þ bj þ vk þ dl þ ðabÞij þ ðbvÞjk þ ðbdÞjl
2 1011 Yes No Yes 58
3 0001 No No No 58
þðsaÞmi þ ðsbÞmj þ ðsvÞmk þ ðsdÞml ð1Þ
4 0010 No No Yes 42 þðabsÞijm þ ðbvsÞjkm þ ðbdsÞjlm þ eijklmp
5 1101 Yes Yes No 58
6 1110 Yes Yes Yes 42 where l is the intercept, s is the consumer effect and a, b, v and d
7 0100 No Yes No 42
are the effects of factors low fat, new fat, organic and price. Further
8 0111 No Yes Yes 58
terms represent interactions and residuals (e). Note that this model
 V.L. Almli et al. / Food Quality and Preference 39 (2015) 28–39 31

Fig. 1. Picture of cheese sample 1110 (Table 1): low fat (keyhole symbol to the left), new fat type (LHL symbol in the middle), organically produced (Debio symbol to the right)
and low price (NOK 42).

Table 2
Selected socio-demographic characteristics of the consumer groups.

Rating group Ranking group Total sample National census data per 01.01.2011
n = 114 n = 105 n = 219
Gender (%)
Female 52.2 49.5 50.9 50.9%1
Male 47.8 50.5 49.1 49.1%1
Age (%)
30–39 20.3 23.8 22.0 26.4%
40–59 44.3 40.9 42.7 51.9%
60–70 35.4 35.3 35.3 21.7%
Mean in years (S.D.) 51.3 (11.2) 51.1 (12.3) 51.2 (11.8) 48.61 (n/a)
BMI (%)
<18.5 (underweight) 0.9 0 0.5 (n/a)
18.5–24.9 (normal weight) 47.8 38 43.1 (n/a)
25–30 (overweight) 34.5 44.8 39.4 (n/a)
>30–34.9 (obese) 16.8 17.1 17.0 (n/a)
Household size
Mean (S.D.) 2.5 (1.4) 2.5 (1.3) 2.5 (1.4) 2.22 (n/a)
Education (%)
Secondary school or lower 2.6 0.9 1.8
High school 31.9 24.8 28.4
University 65.5 74.3 69.7
Household income in kNOK (year) (S.D.) 640 (241) 670 (250) 655 (246) 617.13 (n/a)

Source of national census data: Statistics Norway, www.ssb.no.

1
Age group 30–70 years old specifically.
2
All age groups confounded.
3
Data from 2009 and all age groups confounded.

uses all available degrees of freedom for effects calculations and Individual preferences and consumer segmentation
will therefore give a random error equal to zero. This model is inter- First, a reduced mixed model ANOVA was run almost identical
preted in terms of mean acceptance in the total consumer sample. to the former model but without interaction effects between con-
The model was run in Minitab 16 (Minitab Inc.). sumer and conjoint factors, i.e., only the fixed effects and the main
32 V.L. Almli et al. / Food Quality and Preference 39 (2015) 28–39

consumer effect were retained. The residual vector e was rebuilt as where Vmjt is the explained part of Umjt in Eq. (2) and where the
a consumers  products (114  8) residual matrix. Note that the interactions follow the experimental design’s confounding pattern
model for each individual is saturated, leading to a residuals matrix presented above (see section Cheese samples). The mixed logit
with column sums and row sums equal to zero (Endrizzi et al., model used here assumes random parameters with normal distri-
2011). Then, this matrix was used to extract consumer segments. butions for all conjoint factors and two-way interactions. Thus, this
It was chosen to define segments visually, corresponding to the model provides estimates of the mean (bmean) and the standard
distribution of consumers along a relevant principal component deviation of the random conjoint parameters and interactions. Note
in PCA. These segments are directly interpretable with regard to that the mean coefficients for the population effects may be seen as
the products projected on the PCA loadings plot. Finally, the con- counterparts for the fixed factors in the mixed model ANOVA. Like-
sumer segments were characterised in terms of socio-demograph- wise, the individual effects (bm) correspond to the random interac-
ics, attitudes and self-explicated responses with the help of a tions between the conjoint factors and the consumer effect in the
Partial Least Squares Discriminant Analysis (PLS-DA) regression mixed model ANOVA. These individual parameters will be dis-
model relating the segments to the questionnaire. Multivariate cussed below. Further, the assumption of a random distribution
models were run in The Unscrambler X 10.1 (Camo Software AS). for price in this model accommodates the expectation that different
We refer to Almli et al. (2011), Endrizzi et al. (2011) and people prioritise price differently in comparison to other product
Hersleth, Lengard, Verbeke, Guerrero, and Næs (2011) for similar properties. This assumption leads to a number of positive individual
approaches to modelling and consumer segmentation from rat- coefficient estimates for price, suggesting a preference for the
ing-based conjoint analysis. higher price level relative to the lower price level for a number of
participants. In practice, these may be interpreted as price indiffer-
Data analysis of conjoint ranking ent consumers. The mixed logit models were run in Stata 11
(StataCorp LP) using the mixlogit add-on developed by Hole (2007).
Mixed logit
The ranking data were first reshaped in the form of choice sets
Individual preferences and consumer segmentation
following the pattern presented in Table 3. For eight products, this
First, the matrix of individual parameter estimates bm was
gives seven choice sets of decreasing sizes from eight to two items,
extracted from the mixed logit model (Eq. (2)). This matrix of indi-
leading to a total of 35 data rows per consumer. It is to be noted
vidual estimates is comparable to the residuals matrix from the
that in mixed logit, the seven choice sets per consumer are mod-
reduced mixed model ANOVA on the rating data in the sense that
elled as dependent observations, i.e., correspond to one consumer.
they both reflect individual variations from population effects.
This is an advantage over for example rank-ordered logit, which
Then, the bm matrix was submitted to a visual segmentation in
treats each decomposed choice set as an independent observation.
PCA. These segments are directly interpretable with regard to the
In the mixed logit model, the utility (i.e., the net benefit a
conjoint factors projected on the PCA loadings plot. Finally, the
consumer obtains from selecting a specific cheese) of cheese j for
consumer segments were characterised in terms of socio-
individual m in choice occasion t is written:
demographics, attitudes and self-explicated responses with the
U mjt ¼ b0m xmjt þ emjt ð2Þ help of a PLS-DA regression model relating the classes to the ques-
tionnaire, following the same procedure as for conjoint rating data.
where bm is a vector of individual-specific parameters accounting
for preference heterogeneity, xmjt is a vector of conjoint factors
(here: cheese attributes and interactions), and emjt is a random error Results and discussion
term which is assumed to be independent identically distributed
(i.i.d.) extreme value (Train, 2009). Further, it is assumed that the Population effects
b0 ms are random vectors representing the individuals while bmean
will be the random population mean, representing the mean of Main effects
the distribution of bm. In this way, both the individual effects and The ANOVA results studying population effects of factors low
the population average can be estimated. fat, new fat, organic and price in conjoint rating of pictorial
More specifically, the cheese utility model in the present case cheese-package stimuli are presented in Table 4. New fat, organic
may be written: and price present significant effects (p-values <0.01), while factor
low fat is not statistically significant at a 5% level. All effects are
V mjt ¼ b1m Low fatmjt þ b2m New fatmjt þ b3m Organicmjt þ b4m Pricemjt
estimated positive except price, that is to say that consumers on
þb5m ðLow fat  New fatÞmjt þ b6m ðNew fat  OrganicÞmjt average prefer new fat composition, organic production and lower
þb7m ðNew fat  PriceÞmjt price cheeses to regular fat composition, conventional production
ð3Þ and higher price cheeses (Fig. 2).
A mixed logit model as described in section Mixed logit was
used to investigate population effects from conjoint ranking.
Table 3 Table 5 reports the mean coefficients and standard deviations for
Reshaping ranking data of t products into choice sets for analysis with discrete choice each factor. In this model, price was coded as a 0/1 binary variable
models. Example for ranking order 4;2;6; . . . t  1;t. like the other factors in order to allow coefficients comparisons.
Choice set 1 Choice set 2 Choice set 3 ... Choice set t  1 Similarly to the rating group, consumers in the ranking group pre-
fer new fat, organic and lower price cheeses to regular fat, conven-
Sample Y Sample Y Sample Y ... Sample Y
tional production and higher price cheeses. Here again, factor low
1 0 1 0 1 0 ... t1 1 fat is not significant. Factor price shows the largest mean coeffi-
2 0 2 1 3 0 ... t 0
3 0 3 0 5 0 ...
cient, but the model also reveals a large consumer interest for attri-
4 1 5 0 6 1 ... bute new fat: consumers on average valued new fat nearly four
5 0 6 0 ... 0 ... times as much as low fat and twice as much as organic.
6 0 ... 0 t 0 Conclusively, population effects are consistent between the two
... 0 t 0
conjoint experiments, revealing in particular a large interest for
t 0
low price and new fat and a poor interest for low fat. Former
 V.L. Almli et al. / Food Quality and Preference 39 (2015) 28–39 33

Table 4
Mixed model ANOVA on conjoint rating data.

Sources of variation D.F. SS F-value p-Value

Low fat 1 8.685 3.54 0.063
New fat 1 58.510 21.05 0.000
Organic 1 14.001 7.35 0.008
Price 1 64.747 24.08 0.000
Low fat * New fat + Organic * Price 1 0.580 0.51 0.477
New fat * Organic + Low fat * Price 1 0.010 0.01 0.920
New fat * Price + Low fat * Organic 1 0.317 0.39 0.532
All consumer effects (main effect and interactions) 904 4046.149
Error 0
Total 911 4192.999

R-square: n/a.

Fig. 2. Main effects of the four factors in conjoint rating (original rating scale was from 1 to 9).

its healthier fat quality (reduced saturated fat) may provide health
Table 5 benefits.
Mixed logit model on conjoint ranking data.

Factors Coefficient z p-Value

Interaction effects
Mean None of the interaction effects are detected as statistically signif-
Low fat 0.185 1.52 0.127
icant in the mixed model ANOVA from conjoint rating (Table 4),
New fat 0.778 5.12 0.000
Organic 0.454 3.59 0.000 while one interaction is significant (New fat * Price + Low fat *
Price 1.600 6.53 0.000 Organic) and another one is nearly significant (Low fat * New
Low fat * New fat + Organic * Price 0.178 1.85 0.064 fat + Organic * Price) in the mixed logit model from conjoint rank-
New fat * Organic + New fat * Price 0.149 1.59 0.112 ing (Table 5). The significant interaction coefficient is, however,
New fat * Price + Low fat * Organic 0.376 3.84 0.000
smaller than the significant main effects coefficients. Unfortunately
Standard deviation the specific identification of the interactions at play is not possible
Low fat 0.775 4.09 0.000
New fat 1.088 5.58 0.000
because of the confounding pattern of the design. In order to under-
Organic 0.786 4.19 0.000 stand whether this difference in interaction sensitivity lies in the
Price 1.712 6.64 0.000 modelling methods or in the data sets, a mixed ANOVA using a con-
Low fat * New fat + Organic * Price 0.278 0.86 0.390 tinuous approximation of the eight product ranks and a mixed logit
New fat * Organic + New fat * Price 0.010 0.03 0.974
including parameter correlations instead of factor interactions were
New fat * Price + Low fat * Organic 0.006 0.04 0.971
run on the conjoint ranking data (Train, 2009). Both these models
Number of choice observations: 735. also detect significant interactions/factor combinations in the rank-
Number of consumers: 105.
ing data. All this indicates that the ranking data contains some
Log likelihood at convergence: 989.040.
interaction information that is not present in the rating data.

studies have shown that consumers are often not willing to com- Preference heterogeneity and consumer segmentation
promise on taste for health benefits (Tuorila & Cardello, 2002;
Verbeke, 2006). New-fat cheese may have come through as an New fat and Regular fat segments
attractive product to the consumers as its regular fat content In order to determine consumer segments based on individual
may give positive sensory expectations, while at the same time preference patterns in the conjoint rating and ranking groups,
34 V.L. Almli et al. / Food Quality and Preference 39 (2015) 28–39

PCA models were run on ANOVA residuals and mixed logit bm esti- however that this structure in PCA is clearer and shows higher cal-
mates, respectively, according to the method descriptions in sec- ibration (fitted) and cross-validation variances (Martens & Næs,
tion Materials and methods. 1989) in the case of ranking than rating results, with 85% of
The PCA bi-plot for conjoint rating includes consumers and explained variance restituted on the first two principal components
products, and conjoint factors were added on the plot to ease inter- for ranking data against 56% for rating data. Finally, for conjoint
pretation (Fig. 3a). The PCA bi-plot for conjoint ranking shows con- ranking PC5-PC7 span the variations of the three interactions, how-
sumers as well as main effects and interactions of conjoint factors ever these are negligible in comparison to the main effects.
(Fig. 3b). The results from these two PCAs are highly similar; in both Next, for each PCA model a visual consumer segmentation in
models, each conjoint factor spans one dimension from PC1 to PC4 two clusters was performed along PC2 on the scores plots, separat-
in the following order: price, new fat, organic and low fat. This order ing the consumers that are most favourable to new fat composition
matches the relative importance of the factors at a population level from those least favourable (Fig. 3a and b). Here it was chosen to
indicated in the ANOVA and mixed logit results above. Note perform a visual segmentation along PC2 rather than PC1 because

PCA bi−plot − Consumer segments from conjoint rating

(a) 1

REGULAR FAT
0.8

1000
0.6
0010
1011
0.4
Component 2 (21%)

0.2
0001

0 LOW PRICE HIGH PRICE

−0.2

−0.4 1110 1101

0111
−0.6 0100

−0.8
NEW FAT

−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Component 1 (35%)

(b) 1
PCA bi−plot − Consumer segments from conjoint ranking
New Fat
0.8

0.6

0.4
Component 2 (21%)

0.2 Low Fat

Organic
0
NewFxPrice NewFxOrganic Price
LowFxNewF
−0.2

−0.4

−0.6

−0.8

−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Component 1 (64%)

Fig. 3. PCA bi-plots on (a) ANOVA residuals from conjoint rating and (b) individual mixed logit parameter estimates from conjoint ranking. Consumers in the Regular fat
segment, consumers in the New fat segment.
 V.L. Almli et al. / Food Quality and Preference 39 (2015) 28–39 35

of the particular interest for factor new fat in this study. A visual Comparison of self-explicated and conjoint evaluations of factor
segmentation easily allows for flexibility in targeting the analysis importance
towards the objective of the study. Moreover there is no clear sep-
aration between the segments, indicating the strength of a visu- Fig. 5 (resp. 6) shows the results of self-explicated rating (resp.
ally-oriented approach. The consumer segments consist of 47 and ranking) evaluations presented per conjoint consumer group and
67 consumers for conjoint rating and of 59 and 46 consumers for per consumer segment. Self-explicated rating results are highly
conjoint ranking. In the following these segments are referred to consistent across conjoint conditions, showing the same patterns
as the ‘‘New fat’’ and ‘‘Regular fat’’ segments, respectively. of factor importance between the two New fat segments, between
the two Regular fat segments and between the two conjoint groups
(Fig. 5). Further, there is globally a good agreement between self-
Segments characteristics explicated rating and conjoint measures, corroborating the conclu-
To describe the consumer segments in terms of socio- sions of Sattler and Hensel-Börner (2003). On average, consumers
demographics, attitudinal characteristics and self-explicated in the New fat segments rated fat composition and fat content in
responses, identical approaches based on PLS-DA were used for top positions, while consumers in the Regular fat segments rated
conjoint rating and conjoint ranking data. In the PLS regressions, price and fat content in the first positions. This is logical with their
jack-knifing and uncertainty testing were used for variable selec- respective segment belongings. Note that the fact that fat content
tion and significance testing (Martens & Martens, 2000) and is highly rated in both segments may be due to the ambiguity of
cross-validation (CV) was run with 10 random segments. As the the self-explicated questions, which enquired about the impor-
questionnaire consisted of 46 items covering very different areas tance of fat content in general without specifying a low or high
of the consumer background (with possibly little relation between level of fat content. Fat content may be important both to consum-
them), a global PLS regression may have resulted in spurious vari- ers interested in low fat and to consumers interested in full-fat
able selections. To avoid this problem, several models were cheeses even though they belong to different segments.
attempted with different sets of predictor variables: (i) all ques- Self-explicated ranking results on the other hand are rather
tionnaire variables, (ii) socio-demographics variables only, (iii) inconsistent across conjoint conditions, showing different patterns
attitudinal variables only and (iv) self-explicated rating/ranking of factor importance between segments (Fig. 6). Some inconsisten-
evaluations only. In these models, category variables were re- cies can also be seen between self-explicated approaches by
coded as binary or ordinal variables. Finally, a summary model comparing Figs. 5 and 6. For example, in the New fat segment for
was built on the significant variables from these former models. conjoint ranking fat content is rated in first position in self-
The final PLS-DA models from conjoint rating (R2 = 0.23, explicated rating, but ranked in third position in self-explicated
RCV = 0.20) and conjoint ranking (R2 = 0.21, R2CV = 0.18) are presented
2
ranking. A possible explanation for these inconsistencies is that
in Fig. 4a and b. It should be mentioned that these R2 values might be self-explicated ranking is the only one of the four approaches in
somewhat overoptimistic since the models are based on variable the present study that did not enable ties between factors in the
selection. The results reveal that consumers in the New fat segment consumer test.
typically gave high ratings/low ranks in self-explicated measures for
the importance of fat type and the importance of fat content. In Method comparison discussion
addition, consumers in the New fat segment from conjoint ranking
typically gave a high rank (i.e., little importance) to factor price in Conjoint experimental setup and data analysis
self-explicated measures. These results are fully consistent with
these consumers’ belonging to the New fat segments. Further, these The same fractional factorial design was used in both the rating
results show a good correspondence between the two conjoint and ranking conjoint experiments, allowing a method comparison
approaches and self-explicated approaches. based on stated preference measures of the same eight cheeses.
Socio-demographic variables were not significant in submodels While orthogonal designs are state-of-the-art in the context of lin-
(i) and (ii) and do not appear in the final model. This highlights the ear models and still widely used in the context of stated choice
relevance of a segmentation approach based on common prefer- models, Ortúzar (2010) and Jaeger and Rose (2008) argue that
ences rather than common socio-demographic parameters, as the ‘‘orthogonality between attributes is not even a desired feature’’
latter may not always be pertinent. Regarding behavioural and in highly non-linear models such as mixed logit, and recommend
attitudinal characteristics, consumers in the New fat segments the use of so-called efficient designs. The selected samples may
from both conjoint approaches may be described as health- therefore not have been optimal for mixed logit modelling.
conscious. However, the PLS-DA for rating reveals two significant Further, multi-step approaches of equivalent complexity were
variables only: having a healthy diet and being very physically chosen for the modelling of conjoint rating and conjoint ranking.
active, whereas the PLS-DA for ranking reveals seven significant The mixed model ANOVA approach on rating data may appear
variables: having a healthy diet, importance to them that the food simpler in the sense that ANOVA is based on analysis of averages,
they eat on an ordinary day has a low fat content, is low in satu- which are intuitively appealing, and is a well-known, widely
rated fat, has few calories, helps them keep their weight, keeps spread modelling method in sensometrics. Mixed logit is neither
them healthy and is good for the skin. These attitudinal statements a standard tool in sensometrics nor in classical statistical software
may be related to the slight overrepresentation of overweight par- packages. Further, complex mixed logit models can require a large
ticipants in the ranking group. A possible explanation for the lower computation time due to the need for simulation algorithms
number of significant variables in PLS-DA from conjoint rating is (Ortúzar, 2010). However, computation time is seldom decisive
that these consumer segments may be less well-defined, due to a in the scope of a consumer experiment.
lower explained variance in PCA. Finally, by contrast to the New In this paper a visual segmentation approach was used as the
fat segments, the Regular fat segments include consumers that clustering algorithm that was originally attempted suggested
are less health-conscious, less physically active and more attracted clusters that did not show any interpretable trend in PCA. This
by regular fat composition and full fat content products as well as may be due to the fact that in this case there is not clear separation
by low prices. Conclusively, it seems that new-fat cheese appeals to between consumers. Segmenting consumers visually by help of PCA
existing consumers of low-fat cheese rather than attracts new con- and using the experimenter’s product and problem knowledge to
sumer groups to the healthy market. define relevant classes is a simple approach which can sometimes
36 V.L. Almli et al. / Food Quality and Preference 39 (2015) 28–39

PLS−DA correlation loadings − Conjoint rating

(a) 1

0.8 Fat composition best rank

0.6

0.4

Component 2 (23%, 2%)

0.2
NEW FAT
Fat type important
0

−0.2
Healthy diet
Physical activity
−0.4

Fat content important

−0.6

−0.8

−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Component 1 (32%, 23%)

(b) 1
PLS−DA correlation loadings − Conjoint ranking

0.8
Fat type best rank
0.6

0.4
Component 2 (12%, 5%)

0.2 NEW FAT

Fat type important
0
Fat content important

Good for skin

Few calories Keeps me healthy
−0.2 Healthy diet Helps me hold my weight
Low fat
−0.4 Low saturated fat
Price best rank
−0.6

−0.8

−1
−1 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 1
Component 1 (39%, 21%)

Fig. 4. Correlation loadings from PLS-DA models in (a) conjoint rating and (b) conjoint ranking.

be more sensible than standard algorithms (see also Endrizzi et al., whether conjoint ranking might be forcing an artificial structure
2014). in the data. Villanueva, Petenate, and Da Silva (2000 and 2005)
observed that ranking scales have a high discriminating power
Results consistency in different approaches on the condition that product differences are salient. In particular,
the ranking protocol consisted in first performing a partition of the
Conjoint experiments set of eight products into two groups. Thirty-four consumers out of
One of the results of this study is the overall equivalence of pop- 105 (32.4%) used the two levels of the price factor as a criterion for
ulation effects obtained in rating and ranking approaches, corrob- this dichotomy stage, leading to a high explained variance linked to
orating conclusions from Hein et al. (2008) and extending these price in PCA (64% explained variance on PC1, see Fig. 3b). This
toward picture stimuli in conjoint experiments. It should be noted, reflects the fact that price is an important factor of product choice
however, that the present results show a higher sensitivity to for these consumers. In addition, the numeric information for price
interaction effects in the ranking experiment than in the rating may have been cognitively easier to process than the symbols rep-
experiment, and a generally higher structure in ranking data than resenting qualitative factors (Rayner, 2009).
in rating data. Yet it is not known whether the stronger structure Further, the consumer segments derived from the rating/mixed
that is obtained better reflects true consumer preferences or ANOVA approach and from the ranking/mixed logit approach are
 V.L. Almli et al. / Food Quality and Preference 39 (2015) 28–39 37

Self-explicated rang Fat content

Fat composion
"How important were each of the following properes
Organic
when you performed your evaluaons?”
1: “Very lile important” - 5: “Very important” Price
4

Mean rang (scale 1-5)

3.5

2.5

1.5

1
CONJOINT CONJOINT CONJOINT CONJOINT CONJOINT CONJOINT
RATING, New RANKING, RATING, RANKING, RATING, All RANKING, All
fat (n=47) New fat Regular fat Regular fat consumers consumers
(n=59) (n=67) (n=46) (n=114) (n=105)
Conjoint groups and consumer segments

Fig. 5. Self-explicated rating of factors across conjoint groups and consumer segments.

Self-explicated ranking Fat content

"Rank the following cheese properes from the most (rank 1) Fat composion
to the least important one (rank 4)” Organic
Price
3.5

3
Mean rank

2.5

1.5

1
CONJOINT CONJOINT CONJOINT CONJOINT CONJOINT CONJOINT
RATING, New RANKING, RATING, RANKING, RATING, All RANKING, All
fat (n=47) New fat Regular fat Regular fat consumers consumers
(n=59) (n=67) (n=46) (n=114) (n=105)
Conjoint groups and consumer segments

Fig. 6. Self-explicated ranking of factors across conjoint groups and consumer segments.

similar in terms of self-explicated rating responses and attitudes, Further, as the evaluations obtained by self-explicated mea-
but here again the results from conjoint ranking show more struc- sures corroborate the conjoint results, one may wonder whether
ture and detect several additional significant characteristics to dis- a comprehensive conjoint experiment is necessary. Sattler and
tinguish between segments. Hensel-Börner (2003) reviewed 23 publications comparing self-
From a global perspective, this study validates two unrelated explicated to conjoint approaches and conclude that despite theo-
multi-step modelling approaches: one based on a mixed model retical advantages in conjoint experiments, their analysis ‘‘fails to
ANOVA and study of residuals from conjoint rating data, the other confirm the superiority of conjoint measurement’’. Nonetheless,
based on mixed logit and study of individual parameter estimates our study highlights three assets of conjoint analysis: firstly, infor-
from conjoint ranking data. Such multi-step approaches are chal- mation about attributes combinations is revealed. Secondly, in
lenging to validate by internal statistical validation. By separately conjoint analysis there is no possible ambiguity when interpreting
reaching the same conclusions, the two approaches serve as exter- preferred levels for important attributes. Thirdly, contrary to self-
nal validations for each other. explicated ranking, conjoint ranking always allows the possibility
of ties occurring between attributes – even if ties between prod-
Self-explicated measures ucts are not allowed.
The study of factor importance by self-explicated evaluations
revealed that self-explicated rating globally gives consistent results Respondents’ experience: time usage and monotony
with the conjoint experiments, while self-explicated ranking did A study of the respondents’ time usage reveals that the
not fully capture the same information. Possibly, self-explicated conjoint rating test was less time-consuming to perform than
ranking elicited more mental deliberation from the consumers than the conjoint ranking test, with averages of 83 s (median: 76 s,
self-explicated rating or conjoint experiments, which are monadic Standard Deviation: 37) against 127 s (median: 116 s, S.D.: 54),
tasks. In a series of preference experiments on Chinese ideograms, respectively, after removal of extreme time values in each group
paintings, jellybean flavours and apartments, Nordgren and (test time <10 s or >400 s). From a practical point of view, this dif-
Dijksterhuis (2009) found that deliberation leads to the inconsis- ference in time usage is unexpected as both tests required nearly
tent weighting of information, resulting in reduced preference the same number of screens (one fewer for the ranking test) and
consistency. Moreover, Lagerkvist (2013) compared attribute mouse-clicks (one more for the ranking test). Based on time
importance rankings for labelling of beef from two formats of usage, it seems therefore that the rating task is simpler for the
best–worst scaling (BWS) with those from direct ranking. It was consumers than the ranking task. This corroborates Hein et al.
found that direct ranking showed poorer individual choice predic- (2008), whom in their study comparing five acceptance and pref-
tions than BWS, and poorer transitivity of attribute importance. erence methods report that preference ranking was identified by
38 V.L. Almli et al. / Food Quality and Preference 39 (2015) 28–39

the consumers as ‘‘the least easy scale to use’’. A possible explana- Acknowledgements
tion is that ranking requires making many comparative decisions
between the cheeses and is thus more cognitively demanding This study was supported by TRUEFOOD (Traditional United
than rating, which is a monadic task. Ranking may force consum- Europe Food), an Integrated Project financed by the European
ers to establish a logical strategy while in rating consumers may Commission under the Sixth Framework Programme (Contract n°
rather answer by gut feeling. Finally, note that such time differ- FOOD-CT-2006-016264) and ‘‘Quality labelling and consumer
ences may possibly vanish or differ in acceptance tests involving responses in the Norwegian food sector’’ (Research Council of Nor-
tasting of products. way, Contract nr. 178200/110). The information in this document
Further, it is possible that some respondents got bored or even reflects only the authors’ views and the Community is not liable
annoyed during the conjoint rating experiment, as it consisted in for any use that may be made of the information contained therein.
a monotonous succession of nearly identical screens requiring
nearly identical tasks where only the picture of the cheese varied.
References
Whereas consumers in conjoint ranking saw from the first test
screen that eight cheeses were to be ranked, consumers in con- Almli, V. L., Næs, T., Sulmont-Rossé, C., Enderli, G., Issanchou, S., & Hersleth, M.
joint rating may have gone from screen to screen wondering (2011). Consumers’ acceptance of innovations in traditional cheese. A
when the test would be ending, thus loosing focus and generating comparative study in France and Norway. Appetite, 57, 110–120.
Barreiro-Hurlé, J., Colombo, S., & Cantos-Villar, E. (2008). Is there a market for
poorly structured data. An indication of this is the presence of functional wines? Consumer preferences and willingness to pay for resveratrol-
several consumers that did not fit well into the PCA model for enriched red wine. Food Quality and Preference, 19(4), 360–371. http://
conjoint rating and a generally poorer structure in the rating data dx.doi.org/10.1016/j.foodqual.2007.11.004.
Combris, P., Bazoche, P., Giraud-Heraud, E., & Issanchou, S. (2009). Food choices:
than in the ranking data. It may be advisable in the future to What do we learn from combining sensory and economic experiments? Food
inform consumers in a monadic (web-)experiment about the Quality and Preference, 20(8), 550–557. http://dx.doi.org/10.1016/
number of items that they will be evaluating. In contrast to con- j.foodqual.2009.05.003.
Endrizzi, I., Gasperi, F., Rødbotten, M., & Næs, T. (2014). Interpretation, validation
joint rating respondents, the respondents performing conjoint and segmentation of preference mapping models. Food Quality and Preference,
ranking may have remained better focused on the task throughout 32(Part C(0)), 198–209. http://dx.doi.org/10.1016/j.foodqual.2013.10.002.
the test as it consisted in the succession of varied screens requir- Endrizzi, I., Menichelli, E., Johansen, S. B., Olsen, N. V., & Næs, T. (2011). Handling of
individual differences in rating-based conjoint analysis. Food Quality and
ing varied tasks (‘‘select four out of eight cheeses’’, ‘‘select one out Preference, 22(3), 241–254. http://dx.doi.org/10.1016/j.foodqual.2010.10.005.
of three cheeses’’. . .). Finally, Hein et al. (2008) report that con- Grunert, K. G., Juhl, H. J., Esbjerg, L., Jensen, B. B., Bech-Larsen, T., Brunso, K., et al.
sumers ‘‘were more confident that they had provided accurate (2009). Comparing methods for measuring consumer willingness to pay for a
basic and an improved ready made soup product. Food Quality and Preference,
information’’ in preference ranking than in hedonic rating, proba-
20(8), 607–619. http://dx.doi.org/10.1016/j.foodqual.2009.07.006.
bly due to the simultaneous presentation of samples instead of a Hein, K. A., Jaeger, S. R., Carr, B. T., & Delahunty, C. M. (2008). Comparison of five
monadic one. common acceptance and preference methods. Food Quality and Preference, 19(7),
651–661.
Hersleth, M., Lengard, V., Verbeke, W., Guerrero, L., & Næs, T. (2011). Consumers’
Conclusion acceptance of innovations in dry-cured ham impact of reduced salt content,
prolonged aging time and new origin. Food Quality and Preference, 22(1), 31–41.
This study compared conjoint experiments and self-explicated http://dx.doi.org/10.1016/j.foodqual.2010.07.002.
Hess, S., Ben-Akiva, M., Gopinath, D., & Walker, J. Advantages of latent class over
measures based on rating and ranking approaches in consumer continuous mixture of Logit mo. (2011). <http://www.stephanehess.me.uk/
testing of cheese attributes. The data from rating and ranking con- papers/Hess_Ben-Akiva_Gopinath_Walker_May_2011.pdf>.
joint experiments were modelled with two parallel multi-step Hole, A. R. (2007). Fitting mixed logit models by using maximum simulated
likelihood. Stata Journal, 7, 388–401.
approaches respecting the different nature of the data. Thus, rating Jaeger, S. R., & Rose, J. M. (2008). Stated choice experimentation, contextual
data were analysed by a combination of mixed model ANOVA, PCA influences and food choice. A case study. Food Quality and Preference, 19(6),
and PLS-DA, while ranking data were analysed by a combination of 539–564. http://dx.doi.org/10.1016/j.foodqual.2008.02.005.
Johansen, S. B., Hersleth, M., & Næs, T. (2010). A new approach to product set
mixed logit, PCA and PLS-DA in a new approach. Findings show selection and segmentation in preference mapping. Food Quality and Preference,
that the two methods give similar conclusions both in terms of 21(2), 188–196. http://dx.doi.org/10.1016/j.foodqual.2009.05.007.
population effects and consumer segments. On average, consumers Kozak, M., & Cliff, M. A. (2013). Systematic comparison of hedonic ranking and
rating methods demonstrates few practical differences. Journal of Food Science,
favour cheese of new (healthier) fat composition, organic produc-
78(8), S1257–S1263. http://dx.doi.org/10.1111/1750-3841.12173.
tion and lower price to cheese of regular fat composition, conven- Lagerkvist, C. J. (2013). Consumer preferences for food labelling attributes:
tional production and higher price. The consumer segmentation Comparing direct ranking and best–worst scaling for measurement of
from conjoint ranking data reveals that consumers attracted by attribute importance, preference intensity and attribute dominance. Food
Quality and Preference, 29(2), 77–88. http://dx.doi.org/10.1016/
new fat composition are described as health-conscious consumers j.foodqual.2013.02.005.
who follow a healthy diet, consume low-fat and low-calorie prod- Martens, H., & Martens, M. (2000). Modified Jack-knife estimation of parameter
ucts and products that keep them healthy. The consumer segmen- uncertainty in bilinear modelling by partial least squares regression (PLSR).
Food Quality and Preference, 11(1–2), 5–16.
tation from conjoint rating data corroborates these results by Martens, H., & Næs, T. (1989). Multivariate calibration. Wiley, ISBN 0-471-90979-3.
indicating consumers who follow a healthy diet and are particu- Mueller, S., Lockshin, L., Saltman, Y., & Blanford, J. (2010). Message on a bottle:
larly physically active. It seems therefore that new-fat cheese The relative influence of wine back label information on wine choice. Food
Quality and Preference, 21(1), 22–32. http://dx.doi.org/10.1016/j.foodqual.
may especially appeal to already health-conscious consumers. 2009.07.004.
Seen from the respondents’ point of view, the conjoint ranking test Nordgren, L. F., & Dijksterhuis, A. P. (2009). The devil is in the deliberation: Thinking
is significantly more time consuming than the conjoint rating test too much reduces preference consistency. Journal of Consumer Research, 36(1),
39–46. http://dx.doi.org/10.1086/596306.
but may be perceived as less monotonous and generates more Næs, T., Brockhoff, P., & Tomic, O. (2010a). Statistics for sensory and consumer science.
structured data. Further, self-explicated ratings of the cheese attri- Chichester, UK: Wiley.
butes corroborate the conjoint approaches, while self-explicated Næs, T., Lengard, V., Johansen, S. B., & Hersleth, M. (2010b). Alternative methods for
combining design variables and consumer preference with information about
rankings differ from the three other approaches. Future research
attitudes and demographics in conjoint analysis. Food Quality and Preference,
may further investigate modelling of individual preferences in con- 21(4), 368–378. http://dx.doi.org/10.1016/j.foodqual.2009.09.004.
joint experiments, for example in choice-based conjoint. Finally, Ortúzar, J. D. D. (2010). Estimating individual preferences with flexible discrete-
hedonic and revealed preference studies may be conducted to bet- choice-models. Food Quality and Preference, 21(3), 262–269. http://dx.doi.org/
10.1016/j.foodqual.2009.09.006.
ter measure the potential of healthier semi-hard cheese on the Øvrum, A., Alfnes, F., Almli, V. L., & Rickertsen, K. (2012). Health information and
Norwegian market. diet choices: Results from a cheese experiment. Food Policy, 37, 520–529.
 V.L. Almli et al. / Food Quality and Preference 39 (2015) 28–39 39

Rayner, K. (2009). Eye movements and attention in reading, scene perception, and Verbeke, W. (2006). Functional foods: Consumer willingness to compromise on
visual search. Quarterly Journal of Experimental Psychology, 62(8), 1457–1506. taste for health? Food Quality and Preference, 17(1–2), 126–131. http://
http://dx.doi.org/10.1080/17470210902816461. dx.doi.org/10.1016/j.foodqual.2005.03.003.
Sattler, H., & Hensel-Börner, S. (2003). A comparison of conjoint measurement with Vigneau, E., Endrizzi, I., & Qannari, E. M. (2011). Finding and explaining clusters of
self-explicated approaches. In A. Gustafsson, A. Herrmann, & F. Huber (Eds.), consumers using the CLV approach. Food Quality and Preference, 22(8), 705–713.
Conjoint measurement: Methods and applications (pp. 147–159). Berlin: Springer. http://dx.doi.org/10.1016/j.foodqual.2011.01.004.
Sichtmann, C., & Stingel, S. (2007). Limit conjoint analysis and Vickrey auction as Vigneau, E., Qannari, E. M., Punter, P. H., & Knoops, S. (2001). Segmentation of a
methods to elicit consumers’ willingness-to-pay – An empirical comparison. panel of consumers using clustering of variables around latent directions of
European Journal of Marketing, 41, 1359–1374. http://dx.doi.org/10.1108/ preference. Food Quality and Preference, 12(5–7), 359–363. http://dx.doi.org/
03090560710821215. 10.1016/s0950-3293(01)00025-8.
Steptoe, A., Pollard, T. M., & Wardle, J. (1995). Development of a measure of the Villanueva, N. D. M., Petenate, A. J., & Da Silva, M. (2000). Performance of three
motives underlying the selection of food: The Food Choice Questionnaire. affective methods and diagnosis of the ANOVA model. Food Quality and
Appetite, 25(3), 267–284. http://dx.doi.org/10.1006/appe.1995.0061. Preference, 11(5), 363–370.
Train, K. (2009). Discrete choice methods with simulation (2nd ed.). Cambridge Villanueva, N. D. M., Petenate, A. J., & Da Silva, M. (2005). Performance of the hybrid
University Press, 2003. hedonic scale as compared to the traditional hedonic, self-adjusting and ranking
Tuorila, H., & Cardello, A. V. (2002). Consumer responses to an off-flavor in juice in scales. Food Quality and Preference, 16(8), 691–703. http://dx.doi.org/10.1016/
the presence of specific health claims. Food Quality and Preference, 13(7–8), j.foodqual.2005.03.013.
561–569. http://dx.doi.org/10.1016/s0950-3293(01)00076-3. PII: s0950- Westad, F., Hersleth, M., & Lea, P. (2004). Strategies for consumer segmentation with
3293(01)00076-3. applications on preference data. Food Quality and Preference, 15(7–8), 681–687.