You are on page 1of 22

A Weighted Ordinal Logistic Regression for Conjoint Analysis

and Kansei Engineering Experiments

Stefano Barone*, Alberto Lombardo*, Pietro Tarantino**

* University of Palermo
Department of Technology, Production and Managerial Engineering
Viale delle Scienze - 90128 Palermo (Italy)
stbarone@dtpm.unipa.it, lombardo@dtpm.unipa.it

** University of Naples "Federico II"
Department of Aeronautical Engineering
P.le V. Tecchio 80 - 80125 Napoli (Italy)
pietro.tarantino@unina.it
ABSTRACT
Customer need for emotional satisfaction is increasingly considered by product/service
designers. While several existing methods support the translation of customer requirements
into technical specifications, researchers are now working to develop methods aimed at
integrating affective aspects into product design. Kansei Engineering is a design philosophy
that considers customer perceptions and emotions through a multi-disciplinary approach.
Conjoint Analysis is a statistical tool that can be used for implementing a Kansei Engineering
study.
This article presents a new methodology for conducting a Kansei Engineering study in very
early development phases. This methodology uses two new procedures. The first one is aimed
at calculating attribute importance weights by using respondent choice time in controlled
interviews. The second procedure consists in the introduction of such weights in an ordinal
logistic regression model. By using the proposed methodology it is possible to better identify
the product/service attributes able to induce specific emotions and feelings in the customer.
An application of the method to the design of mobile phones is presented.
Keywords: Kansei Engineering, Conjoint Analysis, Attribute Rating, Ordinal
Logistic Regression, Importance Weights.

1 INTRODUCTION
In very competitive markets, identification of quality attributes of products and services
(hereinafter we’ll use the term ‘product’ for indicating either a physical good or a service) is
today a leading issue for successful organizations. In fact, once these quality attributes are
determined, it is possible to change the product development strategy earlier than competitors
and, therefore to favourably influence customer preferences and, consequently, their choices.
2
Several methodologies are aimed at translating customer needs and expectations in actual
product characteristics. Conjoint Analysis (CA) is among the earliest that try to integrate
psychology and statistics to give objective answers to such needs.
Today, customers are ever more demanding, expecting pleasure and fulfilment of their
emotions and psychological needs. This forces designers to put strong emphasis on capturing
and integrating these aspects into their product development strategy. This is the basis of
Kansei Engineering (KE), a new multidisciplinary approach developed in Japan over the past
few years. In such context, traditional CA can be integrated with KE to translate emotional
customers needs into actual product characteristics.
This work attempts to give a contribution to CA and KE literature, proposing a new
methodology useful in very early product development phases. This methodology can be
schematised as in Figure 1.

CA/KE
experiments
Noise
factors
Control
factors
output

CA/KE
experiments
Noise
factors
Control
factors
output
Factor
importance
weights
(a) (b)
Figure 1. Scheme of a standard CA/KE procedure (a) and the new proposed methodology (b).
In every experimental situation noise factors affect the output. Sometimes the effect of this
noise is so large that model fitting and estimation are weak. This is the case of most CA/KE
experiments where noise factors arise from several sources (Figure 1.a). In order to reduce
this effect and to improve the analysis, importance weights are applied to control factors.
Such importance weights are calculated for each experimental unit. In a CA/KE survey the
units are the respondents and the control factors are the product attributes. The procedure is
aimed at estimating how much such attributes are important for each respondent.
The new procedure for estimating attribute importance weights uses respondent choice time in
controlled interviews. The obtained weights are then used as correction coefficients in the
regression model. The interpretation of the so adjusted model allows experimenters to analyse
output data and, whereas they are less affected by noise factors, to get a more proper model
and more useful conclusions.
The article is composed of six sections. In Section 2 a brief introduction of CA and KE will
be given. The presence and the nature of noise factors in CA and KE experiments will be
discussed. In Section 3 the new procedure for estimating the attribute importance weights is
presented. The introduction of these weights in the ordinal logistic regression constitutes the
bulk of the work, as discussed in Section 4. A case study is presented in Section 5. In Section
6 conclusions and reflections for further work are made.
3
Table 1. An example of Conjoint Analysis design.
Attributes Customer satisfaction
Product profiles A B C Very Low Low Medium High Very High
1 - - -
2 + - -
3 - + -
4 + + -
5 - - +
6 + - +
7 - + +
8 + + +

2 CONJOINT ANALYSIS IN KANSEI ENGINEERING
In early stages of product development, an important issue is to identify the configuration of
product attributes that possibly best appeals customers. Therefore, designers need to
experiment several product profiles (combinations of product attribute levels) having major
impact on customer feelings and preferences. CA is a methodology developed in the 70’s to
this purpose [1]. Different approaches to CA have been developed over the years, so that, it is
now a family of techniques [2] [3] [4] [5].
Whatever technique the experimenter choose to conduct a CA, he/she has to decide how
many attributes shall compose product profiles. Since the need to consider as many attributes
as possible, without increasing respondent fatigue, economic designs are often used [6].
Once an experimental design is arranged, product profiles are presented to respondents for
their evaluation. An example is given in Table 1. The collected data are then analysed to
estimate respondent preferences.
Generally, a decompositive approach of the overall customer satisfaction is adopted by
estimating the so called part-worths, i.e. the effects associated with each attribute.
Therefore, through a statistical analysis of respondent data, it is possible to estimate the
preferences about product attributes and to identify the optimal product profile. Today, CA is
widely recognized as an important tool for modelling customer preferences.
2.1 CA/KE Experiments
Products affective and emotional properties (“kansei” in Japanese) have recently emerged as
important drivers for a successful marketing. In addition to classical methods for
understanding and integrating customer needs in product development, such as CA, Quality
Function Deployment [7] and Kano Model [8], new methodologies have been developed and
integrated into product design processes in order to measure the affective impact on
customers. These methodologies are part of the so called Emotional Design. Among these
methodologies, KE is finding a considerable interest of the academic as well as the industrial
research. It is used for analysing unexpressed and unconscious needs of customers and
translating them into emotion-related product specifications [9] [10].
In CA/KE experiments the respondent is asked to indicate his/her judgement of a product
profile in terms of kansei words instead of customer overall satisfaction.
4
Particular attention must be reserved to the way the selected product profiles are presented to
the respondent. Three possible strategies are:

S
1
: The experimenter builds physical prototypes, allowing respondents to interact with them.
S
2
: The experimenter builds virtual prototypes (digital mock-up), allowing respondents to
interact with them in a virtual environment.
S
3
: The experimenter uses products from those already existing in the market.

Building physical prototypes requires much time and resources. Hence, it is not suitable
especially in early phases of product development. Conversely, virtual prototypes are able to
simulate geometric characteristics and physical behaviours of the product saving time and
resources and obtaining other advantages, as for example early ergonomic evaluations [11].
Sometimes experimenters have not enough resources to build a prototype either physical or
virtual. Therefore, in this case real products, representative of the selected profiles, can be
chosen from the market and presented to respondents for their evaluation. Even if this
solution is the most economic and the easiest to realize, it introduces noise factors which can
heavily bias the analysis of results.
Generally speaking, noise factors affecting respondent evaluation can belong to two
categories.
The first category is composed by non-experimented product attributes influencing
respondent evaluation. We can call them “endogenous noise factors”.
The second category is composed by factors producing the so-called “halo effects” [12] [13],
biasing the customer perception of product attributes. The halo effect can be further
distinguished in “true halo” effect and “illusory halo” effect. True halo effect is a distortion of
respondent evaluation due to his/her incapacity to decompose a whole product into
components to be rated. Illusory halo effect is a distortion of respondent evaluation due to the
presence of context factors (e.g. the preference for a brand), which can uncontrollably affect
his/her judgement.
Endogenous noise and halo effect factors generate what we call “global noise”. A new
procedure has been conceived in order to reduce the effect of such global noise. It will be the
subject of Sections 3 and 4.
3 ESTIMATION OF ATTRIBUTE WEIGHTS IN CONTROLLED INTERVIEWS
A product attribute is considered important if a change in the customer perception of that
attribute determines a change in his/her attitude towards the product [14]. Different methods
for measuring the relative importance of product attributes have been developed. Such
methods can make use of direct questioning or indirect questioning [15]. In direct questioning
the respondent is asked to directly rate the product attributes. In indirect questioning a
respondent is not directly asked which product attribute is important. Among these
approaches, a multitude of methods have been proposed, see for example [16] [17] [18]. A
new procedure for estimating attribute rating is here proposed. It uses the choice time during
the ranking process of attributes.
The procedure is firstly described in its simplest form. Let’s imagine a respondent who is
asked to rank two attributes. We are interested to know the relative respondent “weight of
5
importance” of the two attributes. We assume that the relationship between the two relative
weights is a function of the respondent choice time, i.e. the time he/she takes to select the first
preferred attribute:


1
2
( )
w
f t
w
= (1)
where:
1
w is the relative weight of the first selected product attribute;
2
w is the relative weight of the second product attribute;
( ) f t is a generic function of the choice time t .
Furthermore, we assume that:

1
2
1 2
0 1
0 1
1
w
w
w w
≤ ≤
≤ ≤
+ =
(2)

The Preference Uncertainty theory [19] states that the more uncertain one is about the overall
value of an alternative, the slower is in assigning a value to the alternative. By extending this
concept to the case where respondents have to rank different product attributes, we assume
that:
• If the choice time tends to infinite, the respondent is absolutely undecided about the
order of importance between the product attributes. Therefore, these two product
attributes are theoretically considered equal and they have the same relative weight of
importance (
1 2
0, 5 w w = = ). In formulas:

1
2
lim 1
t
w
w
→∞
= (3)
• If the choice time tends to zero, the respondent considers the first selected attribute
absolutely more important than the second. The relative weight of importance of the
first selected attribute assumes its maximum value (
1
1 w = ), while the second attribute
has a weight equal to zero (
2
0 w = ). In formulas:

1
0
2
lim
t
w
w

= ∞ (4)
The function that better represents the previous conditions is:

1
2
1
1
w
w t
= + (5)
6
The relation (5) is not dimensionally homogeneous. To solve this problem we make this
reasoning. It is reasonable to consider that different respondents may have different times of
reaction to the same stimulus and therefore, different choice times. However, assuming that a
homogeneous sample is selected and experimental conditions are equal for each respondent,
we can define a reference time
*
t as the time a respondent takes to choose between two
product attributes of which the first selected is twice more important then the second one.
This time
*
t depends from the sample chosen for investigation, but in general is very low.
Introducing this reference time
*
t in (5) we obtain the dimensionless equation:

*
1
2
1
w t
w t
= + (6)
Once the choice time
c
t is measured from the respondent,
1
w and
2
w are determined by
resolving the system in two equations:

1
2
1 2
*
1
1
c
w t
w t
w w
¦
= +
¦
´
¦
+ =
¹
(7)
subject to (2). Then, the weights of importance are:

1
2
*
*
*
2
2
c
c
c
c
t t
w
t t
t
w
t t
+
=
+
=
+
(8)
Generally, the choice time coherently increases with the number of product attributes initially
presented to respondent. Therefore, the formulated model is valid for a limited number of
attributes. By extending the (7) to a generic number n of attributes, the weights
i
w
( 1, 2,..., ) i n = are calculated by resolving the system of 1 n + equations:

1
*
( )
1
1
1
i
c
i
i
i
n
i
w t
w t
w
+
=
¦
= +
¦
¦
´
¦
=
¦
¹

(9)
4 WEIGHTED ORDINAL LOGISTIC REGRESSION
In order to filter the effect of the noise factors affecting the results of CA/KE experiments
made accordingly to the strategy S
3
, a procedure is here proposed. The procedure is firstly
explained in the case of a simple linear regression model. Subsequently, the procedure is
extended to the Ordinal Logistic Regression, which is the most suitable for analysing the
results of CA/KE experiments.
7
4.1 Weighted linear regression
Let’s suppose that the response Y , given by n of respondents to stimuli characterized by a
single predictor x , can be modelled by a simple linear regression:

i i i
Y x α β ε = + + (10)
where 1, 2,..., i n = and the usual assumptions on the error term are made.
For simplicity let’s suppose that 3 n = . The regression parameters ˆ α and
ˆ
β are estimated by
the ordinary least squares method:

ˆ
ˆ y x α β = − (11)

3
1
3
2
1
( )
ˆ
( )
i i
i
i
i
y x x
x x
β
=
=

=



(12)
where
3
1
1
3
i
i
x x
=
=

and
3
1
1
3
i
i
y y
=
=

.
A qualitative picture of the linear regression is shown in Figure 2, where residuals from the
regression line are evidenced.
x
y
ˆ
ˆ y x α β = +
1
x
2
x
3
x
2
y
3
y
1
y
2
r
3
r
1
r

Figure 2. Qualitative picture of a regression line.
Let’s now suppose, that multiplicative coefficients, one for each respondent, are introduced in
the model:

' '
i i i i
Y x α β γ ε = + + (13)
If the coefficients are all equal, i.e.
1 2 3
γ γ γ γ = = = , then the estimates for
'
α and
'
β are:

' '
ˆ
ˆ y x α β = − (14)
8

3
' 1
3
2
1
( )
1
ˆ ˆ
( )
i i
i
i
i
y x x
x x
β β
γ
γ
=
=

= =



(15)
By posing:

1
i
i
w
γ = (16)
with
i
w the weight of importance, we obtain:

'
ˆ ˆ
w β β = (17)
Graphically, this situation is represented as a horizontal shift of the predictor points, while the
coefficient of determination
2
R and the significance of parameters remain unvaried (Figure
3).

y = 0,75x + 2,90
R
2
= 0,89
y = 0,23x + 2,90
R
2
= 0,89
0,00
1,00
2,00
3,00
4,00
5,00
6,00
0,00 2,00 4,00 6,00 8,00 10,00 12,00
x
y
1
2
3
0.3
0.3
0.3
w
w
w
=
=
=

Figure 3. Effect of introduction of equal respondent weights in linear regression model.
In reality, it is very rare the case where respondents express the same attribute weight. In
general the situation is:
1 2 3
w w w ≠ ≠ . In this situation the relation (17) is no more linear.
Parameters significance can change and model fitting can improve (Figure 4), up to a case
limit in which the model fitting is perfect (Figure 5), or can worsen (Figure 6).
9

y = 0,75x + 2,90
R
2
= 0,89
y = 0,58x + 1,64
R
2
= 0,96
0,00
1,00
2,00
3,00
4,00
5,00
6,00
0,00 1,00 2,00 3,00 4,00 5,00 6,00 7,00
x
y
1
2
3
0.3
0.4
0.5
w
w
w
=
=
=

Figure 4. An example of model fitting improvement due to the introduction of attribute weights in the
model.

y = 0,75x + 2,90
R
2
= 0,89
y = 1,00x + 0,01
R
2
= 1,00
0,00
1,00
2,00
3,00
4,00
5,00
6,00
0,00 1,00 2,00 3,00 4,00 5,00 6,00
x
y
1
2
3
0.286
0.425
0.6
w
w
w
=
=
=

Figure 5. A limit case of perfect model fitting due to the introduction of attribute weights in the model
10
y = 0,75x + 2,90
R
2
= 0,89
y = 0,18x + 3,31
R
2
= 0,72
0,00
1,00
2,00
3,00
4,00
5,00
6,00
0,00 2,00 4,00 6,00 8,00 10,00 12,00
x
y
1
2
3
0.35
0.4
0.3
w
w
w
=
=
=

Figure 6. An example of model fitting worsening due to the introduction of attribute weights in the model
4.2 Weighted Ordinal Logistic Regression model
When response data are collected by Likert scales, the ordinal logistic regression is a suitable
method to estimate the relationship between the response and the predictor variables (i.e.
product attributes). Logistic regression belongs to the class of general linear models with
logistic function as link function [22]. It is mainly used because of its simple interpretation of
involved parameters [20] [21].
Let’s suppose that the response variable can take only two values, 1 Y = (success) and
0 Y = and that there is only one predictor variable x . The logistic regression model can be
equivalently expressed by one of the following relations:

0 1
0 1
Pr( 1 ) ( )
1
x
x
e
Y x x
e
β β
β β
π
+
+
= = =
+
(18)

0 1
( )
1 ( )
x
x
e
x
β β
π
π
+
=

(19)

0 1
( )
log
1 ( )
x
x
x
π
β β
π
(
= +
(

¸ ¸
(20)
The three relations above are graphically presented in .Figure 7.
The relation (18) express the predicted probability of success, the relation (19) the odds of the
success and the relation (20) the logit of success.
11
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
x
P
r
o
b
a
b
ili t
y
o
f
s
u
c
c
e
s
s



(a) (b) (c)
Figure 7. Logistic Regression: (a) probability of success; (b) odds of success; (c) logit of success.
In case where the response variable is the respondents preference for product profiles, the
rating scale is ordinal, e.g. it goes from 1 to 5, with 5 being most satisfied. Logistic regression
has to be modified to calculate cumulative logits. They are based on cumulative probabilities
of the response category variable. In such case, the three most commonly used logistic models
are the adjacent-category model, the continuation-ratio model and the proportional odds
model [22]. For the scope of this work it is sufficient to present the baseline logit model in
which the category Y k = , for 1, 2,...., k K = is compared to the reference category 0 Y = :

0
0
( )
log
( )
k
k k
x
x
x
π
β β
π
(
= +
(
¸ ¸
(21)
The introduction of weights in the ordinal logistic model with a single predictor, resulted in
the introduction of these weights in the 2 n× X matrix containing the data for each
respondent:

1
1
2
2
1
1
1
n
n
x
w
x
w X
x
w
(
(
(
(
(
=
(
(
(
(
(
¸ ¸
(22)
By extending the relation (22) to the case with J predictors, the modified X matrix
becomes:
12

1 11
11 1
2 21
21 2
1
1
1
1
1
J
J
J
J
n nJ
n nJ
x x
w w
x x
w w X
x x
w w
(
(
(
(
(
=
(
(
(
(
(
¸ ¸
(23)
where
ik
w represents the weight of importance of respondent i ( 1, 2,..., ) i n = for attribute j
( 1, 2,..., ) j J =
A comparison with the linear regression case is possible. In both cases the introduction of
weights in the model corresponds to horizontal shifts of predictor variable values. In both
cases, the regression coefficients give the direction and the size of the effect of predictor
variable on the response variable. What differs is the interpretation of parameters in the two
regression models. In simple linear regression, the parameter β expresses how much the
response variable is expected to increase ( 0 β > ) or decrease ( 0 β < ) when the predictor
variable increases of one unit. In logistic regression the logit coefficient has a similar
meaning. It express how much the logarithm of the odds of success in response variable is
incremented when the predictor variable is incremented of one unit.
The procedure of Section 3 enable us to calculate the individual weights. Their introduction in
the CA/KE survey analysis, through the Ordinal Logistic Regression, allows us to reduce the
effect of the global noise factors. In fact, this is a way to enforce the interpretation of a
“signal” in an output in which it is mixed with a random noise. More details about theoretical
aspects on this topic are in preparation.
5 A CASE STUDY
In this work, the KE approach is applied to the design of a mobile phone. The massive
diffusion of this product allows designers to have a great deal of people opinions. In addition,
designers need to be careful to the customer tendencies and their feeling about the product.
In this study undergraduate student preferences were investigated. By looking at different
sources, 202 potential kansei words were initially identified. This set was too large to handle,
so it was reduced by using Affinity Diagram and Factor Analysis. The four chosen kansei
words have been: Appealing, Handling comfort, Stylish, Durable. By a similar procedure, 119
product properties initially found were reduced by Pareto diagram, leading to the six product
attributes listed in Table 2. For each attribute, two levels have been chosen.
13
Table 2. Description of the chosen product attributes and relative levels.
Levels
Attribute Description
0 1
A Integrated antenna No Yes
B Dimensions Small Very Small
C Internal memory Small Big
D USB port No Yes
E Music support No Yes
F VGA camera No Yes

Table 3. The experimental design and mobile phone models selected for the evaluation.
Product attributes Product attributes
A B C D E F
Concept
A B C D E F
Concept
1 0 1 0 0 1

1 0 0 0 1 0

0 1 0 0 0 1

1 1 1 1 0 0

1 1 0 1 1 1

0 0 0 1 0 0

0 0 1 1 1 1

0 1 1 0 1 0


The full factorial design would consist of 2
6
= 64 product profiles, which is an unaffordable
number for a conjoint experiment. In order to reduce this number, a fractional factorial design
(6 3)
2
III

has been selected consisting in 8 profiles. Such choice is allowed in case of absence or
negligibility of interactions of any order. The characteristics of the chosen attributes and the
explorative nature of this case study allow this choice.
By adopting the strategy S
3
(Section 2.1), eight products already existing in the market have
been chosen according to the experimental design. The experimental design and the selected
mobile phone models are shown in Table 3.
5.1 Data collection
A sample of 40 university students was contacted for the survey. Each respondent was
involved in a 2-phase interview. The first phase was aimed at assessing the individual
importance weights of the selected product attributes (Table 2). According to the method
presented in Section 3, each respondent was invited to fill in an electronic form (Appendix 1)
in which he/she stated the order of preference of the six attributes. Once a respondent had
14
selected an attribute, a computer programme calculated the time taken for the selection.
Accordingly, the individual list of preferences with the relative choice times was so obtained.
The choice times were used to determine the relative weights of importance, as described in
Section 3.
The weights obtained from the survey are graphically presented in Figure 8.
0,000
0,050
0,100
0,150
0,200
0,250
0,300
0,350
0,400
0 5 10 15 20 25 30 35 40
Respondent
W
e
i
g
h
t
Integrated antenna
Dimensions
Memory
Music Support
USB port
VGA camera

Figure 8. Attribute weights obtained from the survey.
In the second phase of the interview, the pictures of the eight concepts were shown to the
respondent (Appendix 2). The interviewer spent some time to explain how to fulfil the
questionnaire, and the meaning of each kansei word. The respondents were asked to give their
impression of each mobile phone concept by using a five-point Likert scale.
5.2 The relation Model
In order to quantify the relationships between respondent kansei and the selected product
attributes, the collected data have been analysed by the Ordinal Logistic Regression.
A model was estimated for each kansei word by the software MINITAB.
The method of analysis was applied twice. The first time, the weights of importance were not
taken into consideration. The second time, the weights were introduced. In Figure 9 and
Figure 10 an example of software output, for the response variable “Appealing”, is presented
for both cases.
Firstly, we can assess whether the observed data are consistent with the model to fit. Output in
Figure 9 and Figure 10 report Pearson chi-square and Deviance chi-square statistics. The
distribution of these statistics – under the null-hypothesis that the fitted model is adequate – is
a
2
χ -distribution with n − (k + 1) degrees of freedom, with n number of observations and k
number of predictors. Details of these tests are available in [22]-[24].
15

Figure 9. Ordinal Logistic Regression for the kansei word “Appealing” (without weights)

Figure 10. Ordinal Logistic Regression for the kansei word “Appealing” (with weights).
16
It is important to underline that these are badness-of-fit tests, i.e. the failure to reject the null
hypothesis supports the adequacy of the regression logistic model. The significance values are
reported in the columns labelled by P.
For the model without weights (Figure 9) we can see that both P-values (0.080, 0.034) are
low, giving some concerns about model fitting. Nevertheless, we now comment the logistic
regression output of Figure 9.
In order to assess the overall significance of regression, the test statistic based on the
likelihood ratio test is:

( ) ( )
2 ln ln 2 ln ln
null k null perfect k perfect
G Deviance Deviance L L L L = − = − − − − (24)
where the likelihood under a perfect fitted models is 1. Under the hypothesis that all logistic
regression coefficients are equal to zero, the G statistic is distributed as a
2
χ with k degrees
of freedom. In the model the P-value of the G-test is less than the chosen significance level
0.05, so we can conclude that at least one predictor is related to response variable Appealing.
Other elements of the output are reported in the ‘Lgistic Regression Table’: regression
coefficients and their standard deviation, P-value of the Wald test for the predictor
significance, odds ratios and relative confidence intervals (logistic regression coefficients and
odds ratios
ˆ
i
ϑ are in relation
ˆ
ˆ i
i
e
β
ϑ = ).
Under the null hypothesis 0
i
β = , the statistic

( )
ˆ
Z
ˆ
i
i
SE
β
β
= (25)

follows a standard Normal distribution.
Confidence interval around the odds ratio

( ) ( )
{ }
2 1
1
ˆ ˆ ˆ ˆ
exp exp
i i i i
z SE z SE
α α
β β ϑ β β

( (
− ≤ ≤ −
¸ ¸ ¸ ¸
(26)
is another way to assess the significance of predictors. If this interval contains 1, the
corresponding predictor is not significant.
Ignoring the concerns about model fitting, we can say that the product attributes having an
impact over the kansei word “Appealing” are: A = integrated antenna, B = dimensions and F
= VGA camera. For all of them the p-value is 0.000 and the confidence interval does not
contain the value of 1. The logit coefficients show that the attribute B = dimensions has an
impact on the dependent variable which is double than the other two.
By introducing the weights obtained by the procedure proposed in Section 3, the results of the
analysis change. They are presented in Figure 10. By looking at the Pearson's and Deviance
test significance values, we could remove our concerns about model fitting.
However, these statistics are based on the differences between observed and fitted data, where
fitted data depend on the estimated probability of each covariate pattern [22]. When data are
17
sparse, i.e. the number of covariate pattern is large, the p-values from the
2
χ distributions are
no longer accurate.
To overcome the problem related to data sparsity, a possible solution consists in grouping
data. The used coding merges data according to the Table 4 below:
Table 4. Importance weights grouping.
0.000→0.045 0
0.046→0.091 1
0.092→0.183 2
0.184→0.229 3
0.230→0.321 4
0.322→0.367 5

By means of this grouping, the goodness-of-fit test becomes more reliable. The output of the
ordinal logistic regression is presented in Figure 11.


Figure 11. Ordinal Logistic Regression output for the kansei word “Appealing” (with grouped weights)
In comparison with the output in Figure 9, the product attribute D (USB port) has now an
influence on the kansei word “Appealing”. In fact its p-value is less than 0.05 and the
confidence interval does not include 1.
18
Table 5. Relationships between product attributes and kansei words.
Model Fitting
Integrated
Antenna
Dimensions
Internal
memory
USB port
Music
Support
VGA
Camera

OLR without weights
Appealing No ** *** ***
Handling
comfort
Yes ** * *
Stylish Yes ** ** ** *
Durable No * ** * *
OLR with weights
Appealing Yes * *** * **
Handling
comfort
Yes * * *
Stylish Yes * * * *
Durable Yes *

* Weak relation
** Moderately strong relation
*** Strong relation

In conclusion, introducing the individual weights in the logistic model clarifies the
interpretation of model parameters, and consequently it can help designers to choose the right
product development strategy.
A qualitative synthesis of the analysis and a comparison of results with and without
introduction of weights are presented in Table 5.
The situation illustrated above is just an example. In fact the introduction of weights could
even lead to a worse fitting and interpretation of the model. Therefore, it is important that the
analyst always compares the results of the new proposed procedure with the results of the
standard procedure before drawing final conclusions.
6 CONCLUSIONS
The need to fully understand and interpret the wishes of customers has led researchers to
develop several methodologies aimed at bringing the “voice of customer” into the
development process. In this context Conjoint Analysis (CA) and Kansei Engineering (KE)
are methods by means of which it is possible to incorporate customer emotions and
perceptions into the product development process.
In this work three innovations for carrying out CA/KE experiments are proposed.
1. Determination of attribute rating on the basis of choice time. It allows a quick and
economic way to asses the respondent preferences for product attributes. Moreover, it
is a very efficient and objective method for obtaining preference measurements based
on metric scales. A further work is ongoing to discuss the theoretical framework of
these aspects.
2. Introduction of weights to regressor levels for each statistical unit, in order to improve
model fitting and interpretation. Further work is ongoing to show the theoretical bases.
In the context of this article, this procedure allows us to incorporate customer
19
individual preferences in the CA/KE model, depurating it from the action of global
noise factors.
3. Integration of the previous procedures for a weighted ordinal logistic regression.
Through the presented case study we show how the introduction of customer
individual weights in ordinal logistic regression model can help the experimenter to
identify the right relationships between product design solutions and customer
feelings/impressions, and therefore, to obtain substantial competitive advantages.
ACKNOWLEDGEMENTS
This work was financially supported by the Italian Ministry of University funded project
PRIN 2005 “Progettazione statistica dell’innovazione continua di prodotto”, and by the EU
FP6 funded project ENGAGE. The authors who mutually acknowledge an equal contribution,
wish to thank the University of Linkoping, Division of Quality Technology and Management
where part of this work was developed.
REFERENCES
[1] Green, P.E. and Rao, V.R. (1971) Conjoint Measurement for Quantifying
Judgmental Data. Journal of Marketing Research, 8, 355-63
[2] Cattin, P. and Wittink, D.R. (1982) Commercial Use of Conjoint Analysis: A
Survey. Journal of Marketing, 46(Summer), 44-53.
[3] Green, P.E. and Srinivasan, V. (1978) Conjoint Analysis in Consumer Research:
Issues and Outlook. Journal of Consumer Research 5(2), 103-123.
[4] Green, P.E. and Srinivasan, V. (1990) Conjoint Analysis in Marketing: New
Developments with Implications for Research and Practice. Journal of Marketing
(October), 3-19.
[5] Green, P.E et al.(1991) Adaptive Conjoint Analysis: Some Caveats and Suggestions.
Journal of Marketing Research, 28, 215-222.
[6] Bergman, B. et al. (1999) Conjoint Analysis: a useful tool in the design process.
Total Quality Management, 10 (3), 327-343.
[7] Cohen, L. (1995) Quality Function Deployment- How to Make QFD Work for You.
Addison-Wesley Publishing Company
[8] Center for Quality Management (2003) Kano’s Methods for Understanding
Customer-defined Quality. The Center for Quality Management Journal, 2(4), 3-36
[9] Nagamachi, M. (1995a) Kansei Engineering: A new ergonomic consumer-oriented
technology for product development. International Journal of Industrial
Ergonomics,15, 3-11.
[10] Schütte, S. and Eklund, J.( 2005) “Design of rocker switches for work-vehicles- an
application of Kansei Engineering”. Applied Ergonomics 36: 557-567.
[11] Barone, S. and Lanzotti, A. (2002) Quality Engineering approach to improve
comfort of a new vehicle in virtual environment. Proceedings of the American
Statistical Association.
20
[12] Lance, C.E. and Woehr, D.J. (1986) Statistical Control of Halo: Clarification From
two Cognitive Models of the Performance Appraisal Process. Journal of Applied
Psychology 71(4), 679-685.
[13] Murphy, K.R. et al. (1993) Nature and Consequence of Halo Error: A Critical
Analysis. Journal of Applied Psychology ,78(2), 218-225.
[14] Jaccard, J. et al. (1986) Assessing Attribute Importance: A Comparison of Six
Methods. Journal of Consumer Research, 12 (March), 463-68.
[15] Alpert, M. I. (1971) Identification of Determinant Attributes: A comparison of
Methods. Journal of Marketing Research, 8 (May), 184-91.
[16] Heeler, R. M. et al. (1979) Attribute Importance: Contrasting Measurements.
Journal of Marketing Research, 16 (February), 60-3.
[17] Kohli, R. (1988) Assessing Attribute significance in Conjoint Analysis:
Nonparametric Test and Empirical Validation. Journal of Marketing Research, 15
(May), 123-33.
[18] Barone, S. and Lombardo, A. (2004) Service Quality Design Through a Smart Use
of Conjoint Analysis. The Asian Journal on Quality, 5(1), 34-42.
[19] Fischer, G. W. et al. (2000) Attribute Conflict and Preference Uncertainty: Effects
on Judgment Time and Error. Management Science, 46(1), 88-103.
[20] Agresti, A. (2002) Categorical Data Analysis. John Wiley & Sons. New York, 2
nd

Edition
[21] Lawson, C. and Montgomery, D.C. (2006) Logistic Regression Analysis of
Customer Satisfaction Data. Quality and Reliability Engineering International. In
press
[22] McCullagh, P. and Nelder, J.A. (2000) Generalized Linear Models. Chapman &
Hall. 2
nd
Edition
[23] Hosmer, D.W. and Lemeshow, S. (2000) Applied Logistic Regression. John Wiley &
Sons. New York, 2
nd
Edition
[24] Cohen, J. and Cohen, P. (1983) Applied Multiple Regression/Correlation Analysis
for Behavioural Sciences. Lawrence Erlbaum Associates. 2
nd
Edition.
21
APPENDIX 1
Electronic Form used for attribute weight estimation


Respondent n°1 weights
Dimensions
VGA camera
Internal memory
USB port
Music support
Integrated antenna
22
APPENDIX 2
An example of the second phase of interview





How do you consider the mobile phone above according to the following kansei words?

Kansei word
Very
Low
Low Medium High Very
High
Appealing
Handling comfort
Stylish
Durable