-4
An Investigation of Preferences for Product
Appearance: a Case Study of Residential Solar
Panels
AR4?Vms
by
MASSACHUSETTS INSTITUTE
OF TECHNOLOCY
Qifang Bao
AUG 15 2014
B.S., Tsinghua University (2012)
LIBRARIES
Submitted to the Department of Mechanical Engineering
in partial fulfillment of the requirements for the degree of
Master of Science in Mechanical Engineering
at the
MASSACHUSETTS INSTITUTE OF TECHNOLOGY
June 2014
@ Massachusetts Institute of Technology 2014. All rights reserved.
Signature redacted
Author .................................
Department of Mechanical Engineering
May 18, 2014
CePrtified b
Signature redacted,
y.
................
Maria C. Yang
Associate Professor of Mechanical Engineering & Engineering Systems
Thpsis Supervisor
Signature redacted'
Accepted by ...............
..........
David E. Hardt
Professor of Mechanical Engineering
Chairman, Department Committee on Graduate Students
2
An Investigation of Preferences for Product Appearance: a
Case Study of Residential Solar Panels
by
Qifang Bao
Submitted to the Department of Mechanical Engineering
on May- 18, 2014, in partial fulfillment of the
requirements for the degree of
Master of Science in Mechanical Engineering
Abstract
The importance of the styling and appearance of consumer products is widely understood. This paper evaluates the appearance of a technology-oriented product, the
residential solar panel, using a quantitative approach known as visual conjoint analysis. The goal of this research was to determine the design attributes that make a solar
panel more visually appealing, and to understand how different visual representations
may reveal consumer preference differently. Approximately 200 survey respondents
were shown two kinds of images of solar panels, one of a standalone panel and the
other of a panel installed on a roof. The preferences for product appearance of solar
panels were determined on two levels: one from an individual perspective and another from an aggregated point of view. By analyzing the preference on an individual
level, the study further explores how presenting a solar panel in its context of use
can influence the consistency of consumer preferences. The results show a significant
shift of preferences when first showing the non-contextualized image and then showing the contextualized image. Such preference inconsistency provides insights with
which to inform the process of user-needs revealing. By analyzing the preference on
an aggregated level, attributes such as color, shape and pattern are determined to be
more important than the frame style on influencing peoples preference for solar panel
appearance.
Thesis Supervisor: Maria C. Yang
Title: Associate Professor of Mechanical Engineering & Engineering Systems
3
4
Acknowledgments
I wish to thank, first and foremost, my advisor Professor Maria Yang for her guidance
and support. Maria, thank you for the hours and hours of time you've spent with
me in discussing and brainstorming. Thank you for helping me editing the papers
and taking the trouble to correct every single mistake in the writing. And thank you
so much for being supportive of my attempts to explore all research areas that I'm
interested in. You are an awesome advisor!
My special appreciation goes to fellow Ideationers: Jim, Bo, Jesse, Anders, Geoff,
Maxime, Jasmine, Mike, Francisco, Lisa, Justin, Heidi and Catherine. Thank you for
your friendship, encouragement and help. The numerous fruitful discussions with you
are really enjoyable and it's you that make my graduate life such a great experience.
I would like to thank Cummins-Tsinghua Women Fellowship to support my first
year and provide me with internship opportunities to gain some valuable industry
experience. And thank KFUPM for supporting the second year of my graduate study.
To my loving Mom and Dad, thank you for being such supportive parents. To
my friends, it is great to have you all. Special thanks goes to my dearest roommates
Di, Lei and Yi - it's you that make me convinced, with your own passion for science
and engineering, that it's not a bad idea to keep going with research life. I'll never
forget all the midnight conversations about lithium-ion battery, nonlinear control, or
photon and phonon.
5
6
Contents
1 Introduction
13
2 Literature Review
17
2.1
Product Appearance ...........................
17
2.2
Conjoint Analysis .............................
18
2.3
User Preference Inconsistency . . . . . . . . . . . . . . . . . . . . . .
20
2.4
Research Gap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
3 Data Collection
23
3.1
Attribute Identification . . . . . . . . . . . . . . . . . . . . . . . . . .
23
3.2
Image Creation . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . .
24
3.3
Survey Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
25
3.4
Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
28
4 Individual Preference Model
31
4.1
Research Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . .
31
4.2
Data Analysis Methods . . . . . . . . . . . . . . . . . . . . . . . . . .
32
4.3
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
33
4.3.1
Bradley-Terri-Luce (BTL) Model . . . . . . . . . . . . . . . .
33
4.3.2
Preference Inconsistency Across Different Visual Representations 35
4.3.3
Directly Stated Preference Under Contextualized Condition
.
38
4.3.4
Changing of 'Most Preferred' Attribute . . . . . . . . . . . . .
39
4.4
Discussion and Conclusions
. . . . . . . . . . . . . . . . . . . . . . .
7
40
43
5 Aggregated Preference Model
5.1
Research Questions ............................
43
5.2
Conditional Logit Model . ......................
43
5.3
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
44
5.3.1
Full Factory Model . . . . . . . . . . . . . . . . . . . . . . . .
44
5.3.2
Reduced Model . . . . . . . . . . . . . . . . . . . . . . . . . .
46
5.3.3
Aesthetic Levels . . . . . . . . . . . . . . . . . . . . . . . . . .
47
. . . . . . . . . . . . . . . . . . . . . . .
47
5.4
Discussion and Conclusions
51
6 Summary
6.1
Discussion and Conclusion . . . . . . . . . .. . . . . . . . . . . . . .
51
6.2
Limitation of Current Work . ..
. . . . . . . . . . . . . . . . . . . .
52
6.3 Future Work. . . . . . . . . . .. . . . . . . . . . . . . . . . . . . . .
53
8
List of Figures
3-1 Examples of the created images.
. . . . . . . . . . . . . . . . . . . .
25
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
3-3 Example survey question, Part II . . . . . . . . . . . . . . . . . . . .
27
3-4
Example survey question, Part IV . . . . . . . . . . . . . . . . . . . .
28
3-5
Example survey question, Part V . . . . . . . . . . . . . . . . . . . .
29
4-1
Data Analysis
33
4-2
Preference model from one participant
. . . . . . . . . . . . . . . . .
34
4-3
Summary of individual preference model . . . . . . . . . . . . . . . .
35
4-4
Summary of most preferred level of attributes
38
3-2
Survey structure
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9
. . . . . . . . . . . . .
10
List of Tables
3.1
Levels of attributes influencing the appearance of solar panels
. . . .
24
4.1
Prediction accuracy for two types of images . . . . . . . . . . . . . .
36
4.2
Prediction accuracy comparison . . . . . . . . . . . . . . . . . . . . .
36
4.3
Preference matching rate . . . . . . . . . . . . . . . . . . . . . . . . .
37
4.4
Color of the roof and first two preferred color in five design questions
39
4.5
Changing of 'Most Preferred' attributes from conjoint analysis revealed
to directly stated . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
5.1
Coefficients of the full model . . . . . . . . . . . . . . . . . . . . . . .
45
5.2
Levels of attributes of the reduced model . . . . . . . . . . . . . . . .
46
5.3
Coefficients of the reduced model . . . . . . . . . . . . . . . . . . . .
46
5.4
Examples of solar panels with different aesthetic level . .... . . . . .
49
11
12
Chapter 1
Introduction
Understanding user needs is critical to the formulation of comprehensive design requirements and the future success of a product. User needs can touch on many aspects
of a product, including price, functional performance, sustainability, brand, as well
as product appearance. The aesthetic value of a product can be defined as the pleasure derived from seeing the product, without consideration of utility [1]. For some
technology-push products that are developed around technological innovation rather
than user-needs or market-needs [2], appearance is usually considered after performance or price. However, as a product rises on the S-curve of technology diffusion,
that is, the technology becomes more mature and more familiar to the consumer,
product appearance may play a larger role in influencing customers buying decisions.
To obtain meaningful and accurate information from users, conjoint analysis has
been widely used as a cost-efficient tool to capture preferences for products. Traditional conjoint has proved to be successful in capturing descriptive features [3].
Visual conjoint analysis allows users to express preferences for visual features of a
product such as form, size and color [4]. Studies have demonstrated the ability to accurately capture consumer preferences of this method by simplifying symbolic graphs
to represent the appearance and identity of a product, such as using Bezier curves to
represent the outline of an automobile [4].
Many strategies exist for obtaining and estimating consumer preferences, from
qualitative ethnographic approaches to quantitative computational strategies. How13
ever, inconsistency is an innate feature of individual preference due to the case-by-case
preference construction process [5], thus it is often difficult to make the 'best' decision
on product design [6]. This is also true for preferences for product appearance. In
some sense, preferences for product appearance can be more easily biased, since it
deals with a primarily visceral response [7]. Understanding these inconsistencies will
provide product designers and engineers insights on the decision making process of
customers
[5], thus helping them better understand consumer preferences
for product
appearance.
In this.thesis, the preferences for appearance of residential solar panels were examined. Much research has been conducted on technologies to improve engineering
performance, particularly efficiency and manufacturability of the panels, but less attention has been paid to panel styling. Typically, elements such as panel color, size,
and frame style are by-products of engineering performance considerations. Darker
solar panels are generally more efficient, while manufacturing methods and ease of
installation issues govern the size and shape of a panel, as well as their rigidity, weight
and texture. However, a solar panel's appearance can dramatically change the visual
appeal of a home and thus influence whether a homeowner chooses to invest in a
system and further adoption of the technology.
There are two main research questions explored in this study:
1. What design attributes make a solar panel more visually appealing than the
others?
Products can be characterized by various attributes such as performance (efficiency etc.) and physical properties (dimensions, weight etc.). Here we'd like to
study which attributes and what values of them lend themselves to visual appeal.
2. Do different visual representations reveal consumer preferences for product appearance differently?
Product may be visually presented to a consumer in a myriad of ways, and the
particular format may influence the consumers' opinion. Here we explore how the
introduction of product context might play a role in user perception.
14
Chapter 2 starts the thesis with a literature review of product appearance, conjoint analysis method and preference inconsistency. In Chapter 3, the design of an
online survey was introduced in detail and a summary of the survey participation was
presented. In Chapter 4, the survey results were analyzed on an individual level. The
Bradley-Terri-Luce (BTL) equation was used to create individual preference models,
and the preference inconsistency between product images with/without context was
analyzed. Further in Chapter 5, a conditional logit model was built to estimate the
user preference for the solar panel appearances on an aggregated level. The results
showed that some of the attributes were of higher impact on users preferences than
the others. Finally, Chapter 6 serves as a summary of the study and discussed future
works.
15
16
Chapter 2
Literature Review
2.1
Product Appearance
Human interfaces and industrial design are often regarded as key marketing elements
of a product [8]. The appearance of a product is partly determined by the functions
it performs, and is partly designed to convey performance and emotional information
to users [9]. It serves its own 'Aesthetic Function' [10] and also plays an important
role in defining the product-person relationship [11].
Industrial design plays an important role in technological innovation. This was
supported by direct evidence in Moody's study, where designers from nine sciencebased companies were interviewed [12]. In the marketing of function driven products,
technology sophistication is the key of product competence. Thus the designers may
want to draw on the product form to emphasize the underlying technology. Also,
physical constrains exist which the product form must connect to and designers should
explore in the realistically feasible subset of product form [13].
A core practice of the field of industrial design is product styling, developing
forms for a product that are both appropriate for the design and attractive. At the
intersection of industrial design and engineering, some research has been conducted
on product appearance for cases such as vehicles and mobile phones [14, 15]. Other
research has been conducted on how style and fidelity of preliminary design representations can influence users response to product concepts [16], and has found that users
17
believe realistic, finished drawings to be more appealing. User interaction designers
have also considered the topic for software design, and found that designs that are
perceived as more attractive are considered better, whether or not they actually are
more effective [17].
The appearance of residential solar panels, which are traditionally considered engineering performance driven products rather than styling driven, has not received
as much research attention. In a survey of 193 solar panel installers in California,
Chen, et al. found that installers were generally the stakeholders who make the primary decision on what solar panels will be made available to the homeowners [18, 19].
When asked to list the top five attributes considered when recommending a panel to
a homeowner, the survey found that the aesthetic of the panels was the second highest priority attribute voted after cost. In the same study, installers were presented
images on a blank (no context) background of four different kinds of panels, varying
the color of frame and the surface texture. Results of the survey suggested that, an
even surface with dark-colored frame is the most preferable solar panel style.
2.2
Conjoint Analysis
In previous works of user preference for solar panels, methods of eliciting revealed
preference (preference extracted from real marketing data) and stated preference
(preference explicated by consumers over hypothetical alternatives) were conducted
independently and the results are compared to each other [18, 19]. Ranking and
rating were used as the main methods of stated preference elicitation. In this study,
choice based conjoint analysis is used as the main preference elicitation method. It
belongs to the category of stated preference eliciting methods, however different from
directed stating preference eliciting methods in many ways.
When using conjoint analysis, several attributes of a product are chosen, each of
which can be varied in discrete levels. A controlled set of potential products is created
by experimental design, each of which contains a combination of the attributes [20].
Full-factorial design is a choice for product profile generation when the number of
18
product feature is limited. However, when the number of product features is large,
fractional-factorial experimental designs can be used to capture consumer preferences,
substantially reducing the number of questions [21].
After the set is prepared, users are asked to rate, rank or choose among the
product profiles. By their rating, ranking or choices, utility models can be built and
predictions can be made about the users' future preferences. Choice based conjoint
presents several products with different combinations of attributes to users and asks
them to choose the one he/she likes the most [22]. Compared to the ranking or
rating conjoint method, choice based conjoint has the advantage of more realistically
mimicking customers purchasing behavior.
Conjoint analysis is used to address the additive effect of a set of individual vaiables on consumer preference. Despite its disadvantage that it only considers limited
number of product attributes and cannot reveal any latent user need, it is still widely
used in market research to determine how people value different features that make
up an individual product or service [23].
Conjoint analysis has been traditionally used on descriptive features of products.
Visual conjoint expands the use of the method by allowing judgment on a products
appearance. In Kelly, et al.'s work, two attributes were varied to generate different
shape of a cola bottle [24]. In Swamy, et al.'s work, four-control-point Bezier curves
were defined to generate the shape of a vehicle headlight [25]. In Orsborn, et al.'s
study, seven attributes were determined to fully describe the outline of an SUV [26].
These 2D visual representations are further combined with functional attributes of
products to understand consumer preference from other perspectives such as perceived
environmental friendliness or the emotion and reasoning behind the choices [27, 28].
Experimental conjoint methodology was proposed by Tovares, et al. [3]. Virtual
reality was combined with conjoint analysis to create a vivid use environment. This
method allowed users to judge not only how a product looked, but also how it would
feel to use as a product.
19
2.3
User Preference Inconsistency
Preference inconsistency is an unavoidable issue when collecting preference data from
users. Inconsistency can be a result of using different preference elicitation methods, testing preference within various contexts, or interpret preference using different
models.
There have been many studies conducted in understanding the preference inconsistency. MacDonald, et al. compared the preference elicitation results of three widely
used engineering/market design methods and distinguished the internal and external
inconsistencies [5]. Horsky, et al. demonstrated that the weights of attributes are not
always correlated between stated preference and revealed preference. They also found
that the relative importance of intangible attributes (prestige) are sometimes overstated and that of tangible attributes (price) are frequently understated [29]. Reid, et
al. studied the preference inconsistency across forms of vehicals. The results showed
that the opinions for products are inconsistent between computer sketch / FSV (3D
file system visualizer) silhouette and realistic rendering, though the inferences are
consistent [30]. To analyze the underlying reason and to reveal latent variables that
have not been addressed in explicit preference models are the main goals of these
studies. For product designers and engineers, better understanding the user preference inconsistency will help them gain insights on the decision making process of
customers, thus further make better design decision.
Different solutions have been proposed to deal with this preference inconsistency.
To make up for the deficiency of traditional random utility model, which ignores
latent psychological factors such as perception and attitude, or latent segmentation
of population, Ben-Akiva et al. dedicated themselves to develop advance models
such as hybrid choice model to include comprehensive decision related factors [31].
In Kulok and Lewiss study, a systematic method, preference consistency check, was
proposed to help decision makers expressing consistent preference. [6].To deal with
inconsistency caused by different visual representations, Reid et al. proposed to use
pilot studies pretesting a variety of visual representation forms and use one or more
20
forms to improve the robustness of the preference test [30].
2.4
Research Gap
Visual conjoint analysis is a relatively new approach for studying product appearance.
Only a limited variety of products have been studied using this method. In addition,
it has previously focused on line-drawing visual representations. This work seeks to
build on this approach in a number of ways. First, this work extends the variety of
products that have been studied by exploring a technology product. Second, here we
consider photorealistic contextualized and non-contextualized images other than line
drawings.
21
22
Chapter 3
Data Collection
This study involves presenting users with an online survey of images of solar panels
with varying appearance attributes. The details of the creation of the survey and
images, survey administration, and participants demography follow below.
3.1
Attribute Identification
Datasheets of 265 different models of residential solar panels from 37 different brands
were collected as part of a previous study [191. These panel types were successfully sold
on the open California market from 2007-2011, and were drawn from the California
Solar Initiative's database [32].
In this current study, analysis of the product images included in the datasheets
identified four key attributes of panels' appearance:
1. Color: The color of solar cells
2. Shape: The shape of the corners of solar cells
3. Pattern: The pattern of the front contact wires of solar cells
4. Frame: The style of the frame of solar panel modules
In the experiment, each attribute was varied on 3 4 different levels, which can be
found in Table 3.1.
23
Table 3.1: Levels of attributes influencing the appearance of solar panels
Attributes Num. of
Levels
Level 1
Level 2
Level 3
Level 4
Green
Color
4
Black
Blue
Red
Shape
3
Big rounded
corner cell
Small rounded
corner cell
Square cell
Pattern
3
Two main contact
No patternwiecottwrs
Frame
3
Silver frame
Three main
wires
contact wires
Black frame
No frame
In fact, the attributes identified are not only related to the appearance of solar
panels, but are also connected with engineering criteria such as performance and price.
For example, the color of the panel is directly determined by the solar cell coating
material. The shape of a solar cell is related to the usage rate of silicon wafer, the
fabrication materials, and the power generation efficiency per area of the solar panel.
For the sake of simplicity, here we do not consider the variation of any criteria except
those relating to aesthetics, holding all other attributes constant.
3.2
Image Creation
Two different kinds of images of solar panels were created specifically for this survey:
* Non-contextualized: Images of solar panels from a front view on a blank
background.
* Contextualized: Images of solar panels installed on houses of different styles
and with differnt roof colors. The panels may be shown at an angle, depending
on the roof style.
Images were created using Adobe Photoshop, and all levels of the attributes were
controlled to minimize possible bias introduced by factors such as image resolution.
24
For the contextualized images, the combination of panels and roofs were chosen randomly, without specific match of styles between the two. Figure 3-1 shows examples
of the two kinds of images. Panels in the two images have the same combination of
attributes: blue, big rounded corner cell, two main contact wires and no frame.
Non-contextualized
Contextualized
Figure 3-1: Examples of the created images.
3.3
Survey Design
A six-part survey was designed. Figure 3-2 is a summary of the survey structure.
A pilot study was conducted with 8 participants and their feedback on overall
content, wording, and survey length were considered in refining the survey. It took
an average time of 15 min to complete. The sequence of different parts of the survey
was designed that directly stated questions come after pairwise choice questions, to
reduce potential bias of preference as the respondents progressed from one section to
the next. Also, specific questions are introduced as cross checks to see if respondents
answer same questions the same way twice.
Question examples of Part II, Part IV and Part V of the survey can be found in
Figure 3-3 to Figure 3-5.
The first 27 questions in part II would be used to create the preference model. The
remaining 5 questions in part II were the first holdout set of questions used to cross
25
validate the prediction accuracy of the model. The 5 questions of the contextualized
images in part IV were the second cross validating set.
These two question sets
correspond with 5 randomly picked questions in the 27-question set, which means
that the extra 5+5 questions have the same solar panel pairs as 5 previously asked
questions.
Part I
Introduction: A general introduction of the study and an introduction
of the four attributes of solar panels.
Part II
Preference for Non-contextulized Panel Images: A 32-question
pairwise choices questionnaire. Each question asks the participants to
choose a more preferred solar panel from two options.
Respondents
are asked to use the panel's appearance as the only criteria to consider
when comparing the two panels. Non-contextualized images are used.
Part III
Direct Stated Attributes Preference Without Context: A direct
statement survey asking participants to choose their most preferred
level for each of the four attributes.
Part IV
Preference for Contextulized Panel Images: A five-question pairwise choices questionnaire.
Each question asks the participants to
choose a preferred solar panel between two options.
Contextualized
images are used.
Part V
Direct Stated Attributes Preference Within Context: A fivequestion solar panel design questionnaire.
Each question presents a
differnt house, and asks respondents to make a combination of the four
attributes to create a solar panel that looks nice on the roof of that
house.
Part VI
Demography: A questionnaire asking basic demographic information
such as gender, age, geographical location and type of residence.
Figure 3-2: Survey structure
This repetition of questions should not be obvious to respondents if the total
number of questions is sufficiently large. Thus there is only limited risk that par26
ticipants memorize their answers and make the same choice within same solar panel
pairs intentionally. We can assume that every time a participant answers a question,
whether he/she has seen it before or not, he/she will make the choice only based on
their current judgment of the product appearance, not their memory of answers to
previous questions and an obligation to answer consistently.
Directly stated preferences for each attribute were elicited in part III and part V of
the survey. The difference between these two parts is that part III asks participants'
preference for solar panel attributes without any context; part V asks the attributes
preference within the context of a given house image.
Later, the results would be
compared to the conjoint revealed preference.
A
B
Only considering the appearance, which of the solar panels would you prefer?
e A
o B
o None of the two
Figure 3-3: Example survey question, Part II
27
A
Which solar panel looks better on the roof?
B
sA
eB
e None of the two
Figure 3-4: Example survey question, Part IV
3.4
Participants
The survey was created using Qualtrics and distributed through Amazon Mechanical Turk, a crowdsourcing Internet marketplace where Human Intelligence Tasks are
posted and completed by MTurk Workers. The respondents of the survey were confined to adults in the US. Informed consents from the participants were obtained at
the beginning of the survey. Each participant was compensated with $1.50, approximately $6.00 per hour.
A total of 227 participants responded to the survey. Two criteria were used to
screen the data:
1. Response time: Surveys completed in too little time (less than 8 min) were
considered to be low quality, and therefore were rejected;
2. Correct answer: In part II, an objective question (which also shows two solar
panel images but instead of asking participants to choose the one that looks
good, instructing the participants to choose the option 'None of the two') was
set , and those that failed to choose the option as instructed were considered
28
Image that this is the house that you're going to install the solar panels on. What
combination of attributes for the solar panel you would like to choose?
Color
Shape
Pattern
Frame
" Black
" Big rounded corner
" No pattern
" Silver frame
" Blue
" Small rounded corner
main
conducting
" Red
" Square
" Two
wire
" Three
wire
main
conducting
" Green
" Black frame
" No frame
Figure 3-5: Example survey question, Part V
low quality thus were rejected.
33 of the responses were rejected, leaving 194 responses to the survey being used
in the final data analysis.
Among the valid data, the demographic information of the participants is as the
following:
110 male, 84 female, age ranging from 20 to more than 60, all living in
United States.
29
30
Chapter 4
Individual Preference Model
4.1
Research Questions
First, all the data was analyzed on an individual level. Preference models were built
for each participants, and the models were used to predict their preference choices.
The main purpose of building preference models on an individual level is to assess
how sensitive a customer might be to the particular way a panel is presented. As
one might expect, solar panel datasheets generally showcase a panel on a blank background, without any visual context. However, residential solar panels are installed on
roofs of widely varying color, shape, materials and slope. And the homeowners may
not see how a panel looks on their own roof until after it is installed.
This study poses three questions about preferences for a solar panel's appearance:
1. Are preferences for a solar panels appearance the same when a panel is presented
in images within/without a context?
2. Are preferences consistent when the attributes of product appearance are evaluated all together (conjoint analysis) compared with evaluated separately (directly stated)?
3. If the preferences for product appearance are not consistent, then what are the
main factors that influence preferences?
31
4.2
Data Analysis Methods
The conjoint analysis data was analyzed using the Bradley-Terri-Luce (BTL) equation, which is often applied to pairwise comparison data [4]:
P(xi,) =
c(4.1)
Wij
Here xij is the jth level of attribute i. The probability that attribute xij will be
chosen, P(x,,), is given by dividing wijc, the number of times attribute level x 3 was
selected in the conjoint survey, by we,, the total number of times attribute level xij,
was presented in survey.
(4.2)
ni = fi(xi) = P(Xi,)
The part-worth utility of each attribute is estimated by the chosen probability.
In Equation 4.2, ni is the estimation of part-worth utility of attribute i.
fi
is the
mapping from level of attributes to the part-worth utility. The level with the highest
utility was noted as the most preferred level of the attribute.
The total part-worth utility of a product is given by the sum of that of each
attribute:
U=En
(4.3)
A larger part-worth utility represents a higher preference. When using the model
to predict future choice, the product with the higher estimated part-worth utility will
be chosen.
Figure 4-1 shows a flow chart that explains the process of model creation, prediction, and comparisons of preference consistency.
As can be seen in the flow chart, the 27 questions with non-contextualized images
are used to create the preference model. Then the model is used to predict the customers preferences for the extra 5 questions with non-contextualized images, as well
as the 5 questions with contextualized images. The prediction results are compared
32
Created from non contextualized image test data
Model
Data Collect
Prediction
sd
theic
Predict customer's
choices using
icd by2 w
a theutilityprt-rc
o
worth
Survey of
preferences (for
creating model)
Survey of
preferences (for
aomnaring with
Compare preferred levels
elicited in different ways
Non-contextualized
contextualie
DiractySad
image
to the
choices
real participantslcmade
n for t 5+ usin oshwtepeito
Compare the aesthetic
preference elicited for 2
different kind of images
to hei
co
preference model
Compare actual choices
with model predictions
Figure 4-1:2-uIose.Ticmpronaorval
Data Analysis
reodiing qusInaige
to the real choices participants made for the 5+5 questions to show the prediction
accuracy of the model and to evaluate the p
er tyence consistency.
In addition, the answers to the extra 5 questions with non-contextualized images
and the answers to the 5 questions with contextualized images are compared directly
to their corresponding questions in the 27-question set. This comparison also reveals
information about participants preference inconsistency.
Later the most preferred levels revealed by conjoint analysis are compared to the
responses to part III of the survey, where participants give answers to their most
preferred level for each of the four attributes directly. The most preferred levels are
also compared to the design outcome of the survey part V.
4.3
4.3.1
Results
Bradley-Terri-Luce (BTL) Model
By calculating the part-worth utility of each attribute level, the preference models
were built for each participant.
Figure 4-2 is an example from one participant.
33
The preference models are de-
scribed by blue dots representing equation 4.1. The most preferred levels are marked
by red bars.
Preference for color
5O.S
Preference for shape
0.5
0.4
0.4
0.3
0.3
0.2
0.2
0
0
Black
Blue
Red
Green
Big Corner
Preference for pattern
Small Corner
Square
Preference for frame
0.5
0.5
- 0.4
- -
- -..
- -
- -
- -
0.4
I 0.3
0.3
a
.51 0.2
0.1
0
0
No Pattern
Two Wires
Three Wires
Silver
Black
No Frame
Figure 4-2: Preference model from one participant
Figure 4-3 is a summary of the individual preference models.
Later, the models are used to predict participants choices on other questions using
equation 4.2.
Using the preference model in Figure 4-2 as an example, when comparing:
" A black panel with big rounded corner, no pattern cell, and a silver frame with
" A blue panel with small rounded corner, two main contact wires cell with a
black frame
the estimated part-worth utility of the two panels are given as below:
U1
=iColor
-
+
tSh ape
+
UPattern
0.38 + 0.44 + 0.48 + 0.41
= 1.71
34
+
fFrame
(4.4)
Color Preferences
Shape Preferences
50.00%
50.00%
40.00%
30.00%
20.00%
10.00%
0.00%
40.00%
10.00%
0.00%
-
30.00%
20.00%
Black
Red
Blue
Means 35.70% 27.26% 18.39% 18.65%
SD
15.07% 11.04%
Big
Green
8.04%
Means
9.63%
SD
Pattern Preferences
corner
Sur
34.45%
37.11%
28.44%
11.18%
7.68%
13.28%
Frame Preferences
50.00%
40.00%
30.00%
20.00%
10.00%
0.00%
-
50.00%
40.00%
30.00%
20.00%
10.00%
0.00%
Small 3.1 284
corner
No
Pattern
Two
wires
Three
wires
Means
39.22%
30.68%
30.10%
SD
10.31%
7.81%
8.38%
Silver
Black
No
Frame
Means
30.36%
34.51%
35.13%
SD
6.84%
6.98%
6.26%
Figure 4-3: Summary of individual preference model
U2
---
lorO
Shape +
UPattern + UFrame
= 0.13 + 0.37 + 0.15 + 0.33
-
Since U 1
>
U2 , the
(4.5)
0.98
model predicts that this participant will prefer the first panel
over the second one. If he/she does choose panel ii) over panel i), we claim that the
model makes the right prediction.
4.3.2
Preference Inconsistency Across Different Visual Representations
The last 5 questions in Part II (with non-contextualized images) and the 5 questions in
Part VI (with contextualized images) of the survey are used for checking the accuracy
of the model.
Prediction accuracy is the number of right predictions made over the number of
total predictions made by a preference model. Every participant has his/her own preference model and his/her own prediction accuracy value for both non-contextualized
35
image questions and contextualized image questions.
Table 4.1 is a summary of 194 individual prediction accuracies.
An average of
70.8% prediction accuracy of non-contextualized images suggests a relatively good
prediction ability of the model when the image style doesn't change (note that the
preference models were built based on participants' answers to the questions with
non-contextualized images). However when the style of the images changes to contextualized image, the model does not work as well any more.
Table 4.1: Prediction accuracy for two types of images
Average
St Dev.
ANOVA
Non-contextualized images
70.80%
22.30%
F=57.44
Contextualized images
53.00%
24.00%
p = 2.61 x 10-13
The small p-value for one-way ANOVA indicates that there is a significant difference between the prediction accuracy for questions with different styles of images.
The prediction accuracy for contextualized image questions is much lower than that
for non-contextualized image questions.
For some people, their own preference models predict contextualized image questions more accurately than predict contextualized image problems for them. For some
people, it is the opposite. For the others, their preference models predict both kinds of
questions with same accuracy. Table 4.2 is a summary of the number of participants
in these three categories. The results show that, more'than half of the participants
have higher prediction accuracy for non-contextualized image questions.
Table 4.2: Prediction accuracy comparison
Non-contextualized >
Contextualized
Non-contextualized
Contextualized
119
53
=
Non-contextualized <
Contextualized
22
In Table 4.3, the answers for the five non-contextualized image questions and the
36
Table 4.3: Preference matching rate
Average
St Dev.
ANOVA
Non-contextualized images
83.6%
18.8%
F=80.39
Contextualized images
64.4%
23.2%
p = 1.34 x 10-17
five contextualized image questions are compared directly to his/her previous answers
to the corresponding questions.
The results show the inconsistency of preference. When respondents answered a
question the second time with the same style of image (non-contextualized), there
was 83% chance they would state the same preferences that they did before. This
random inconsistency has been observed before and the reason of the inconsistency
was concluded as stochastic choice [33].
However, when respondents answered a question the second time in response to a
different style of images (the contextualized images), only 64% chance they would still
prefer the same solar panel. The small p-value of ANOVA indicates that the difference
between the matching rates is significant. Thus the inconsistency of preference caused
by changing image style is not random, but systematic.
These together show a discrepancy of users response when first using non- contextualized images and then using contextualized images in the visual conjoint. In other
words, when comparing images of non-contextualized panels, a respondent might
prefer a green frameless panel to a black silver-frame, for example. However, when
comparing the same two panels presented in the context of a roof, the respondent is
very likely to change the preference and prefer the latter to the former, suggesting
that the roof is able to change the respondents preference.
37
Directly Stated Preference Under Contextualized Con-
4.3.3
dition
Figure 4-4 is a summary of the answers to the five solar panel design questions. The
most preferred levels of attributes are also presented in the graph as a reference. It
can be seen that people's choices for the most preferred level of each attribute changes
under different circumstances.
L100.00%
Most preferred Shape
Most preferred Color
90.00%
80.00%
70.00%
60.00%
50.00%
40.00%
30.00%
20.00%
10.00%
0.00%
70,00%
ODesign I
A
- 3
-
--
50.00%
Design 3
4000D%
Design 4
30.00%
Green
Red
0
4
Design 5
Conjoint analysis
0.00%
Big Corner
stated
Directly
00Des ign
10.00%
nomConjoint analysis
Blue
Design I
ODesign 2
20,00%
Design 5
Black
.-
60.00%
Design 2
Smal Corner
Square
atllirectly statement
Most preferred Frame
Most preferred Pattern
70.00%
70,00%
-Design
60.00%
1
-m"
Design 3
anDesign
4
30.00%
in
2000%
20.00%
- ---
40 00%
-- ---- - -
Design 5
Design 1
Design 2
50M0%
40.00%
30.00%
-
60.00%
Design
-------__
-- -__ - 2-
5000%
410
-Co
Design 3
Design 4
""Design
5
10.00%
10.00%
-4-Conjoint analysis
BNFm rConjoint
0.00%
0.00%
No Pattern
Two Wires
Three Wires
Silver
-101Directly stated
Black
No Frame
analysis
4HOmDirectly statement
Figure 4-4: Summary of most preferred level of attributes
As for shape, Square is the most preferred in four design questions out of five and
Big Rounded Corner is always the least preferred.
These are different from either
conjoint analysis or directly stated answer. The reason for the differences could be
that the shape of the tiles on the roof influences peoples preference for the shape of
the cells on the panel.
For pattern, No Pattern is the most preferred in all the five design questions,
suggesting that a majority of the respondents find an even surface of the solar panels
more attractive than patterned surface of the panels.
For frame, No Frame is the most preferred for three design questions and Black
Frame is the most preferred for the other two design questions. It seems that when
the roof of the house has a dark color (gray), customers prefer Black Frame the most,
38
and when the roof has a light color (red, green or blue), No Frame is considered to
be the best match. Silver Frame is always the least preferred one.
Preferences for color changed the most under different design situations. The color
of the roofs in the images provided is likely the main cause for the preference change
as can be seen in Table 4.4. It was found that respondents tend to choose the color
of the solar panel to match the roof: blue panels for blue roofs, red panels for red
roofs, etc. In addition, Black is always a second preferred choice if not the first.
Table 4.4: Color of the roof and first two preferred color in five design questions
Design 1
Design 2
Design 3
Design 4
Design 5
Color of the roof
Grey (with
blue eaves)
Grey (with
white eaves)
Red
Green
Blue
First preferred
color
Blue
Black
Red
Green
Blue
Second preferred
color
Black
Blue/Red
Black
Black
Black
4.3.4
Changing of 'Most Preferred' Attribute
In Figure 4-4 the most preferred attributes revealed by conjoint analysis seem to match
the directly stated most preferred attributes well. However, looking into details we
can see that there is a high rate of chage for the most preferred attributes.
Table 4.5 is a summary of the changes of most preferred attributes. The digit
in row X colum Y represents the number of people whose most preferred attribute
revealed by conjoint is X, and stated directly is Y. The bolded digits represent the
number of people, of whom the most preferred attributes are consistent. Digits in
gray cells represent the number of people whose most preferred attributes changed.
We can see that the most preferred color revealed in the two ways generally consistent with each other. For pattern and frame however, though the summations of
the most preferred attribute on each single level are similar, the inconsistency is high.
Take pattern as an example: there are 36 people whose conjoint revealed favorite
39
pattern is No Pattern. However their directly stated favorite pattern is Two Contact
Wires. It is almost 1/3 of the population whose conjoint revealed favorite pattern is
No Pattern.
4.4
Discussion and Conclusions
The results of this study show the existence of inconsistency when revealing appearance preference for solar panel from consumers. To answer the questions we asked in
the beginning of this chapter:
1. Are preferences for a solar panels appearance the same when a panel is presented
in images within/without a context?
No. When the style of product representation of the solar panel changes from a
non-contextualized image with a blank background to contextualized image within
realistic context, the appearance preference shifts. We can see this from the significant
difference of prediction accuracies as well as the significant difference of preference
matching rates (Table 4.1 to Table 4.3).
2. Are preferences consistent when the attributes of product appearance are evaluated all together (conjoint analysis) compared with evaluated separately (directly stated)?
No. The preference model reveals the appearance preference for solar panel when
all the attributes are presented together within an integrated image. The directly
stated preference represents the appearance preference considering the attributes independently. In this study, the preference for color is pretty consistent. However the
preference for shape, pattern and frame are not consistent when evaluated in different
ways. Table 4.5 shows the changing of most preferred attributes revealed in these two
ways. However which one of these two preference-revealing methods is better is open
to question and is out of the scope of this study.
40
3. If the preference for product appearance is not consistent, then what are the
main factors that influence the preference?
In this study the context in which the product is presented plays an important
role in bias the appearance preference. This can be seen from respondents choices
of color in part V of the survey (Table 4.4), when they are asked to choose the best
combination of the attributes to create solar panels to match the roof of house given.
It appears that, a majority of respondents consider the consistency of the color of
solar panel and the color of roof represents a better looking. Thus the preference for
the appearance of the solar panel is not only a question about the product itself, but
also a question about what the environment in which the product is used looks like.
Further we could argue that, the closer the visual representation looks to the
final product and the closer the testing environment looks to the real using context,
the more realistic the revealed appearance preference will be. Which product looks
better can be a subtle question to ask, since insignificant factors such as the intensity
of lighting could change people's opinion. Thus, when revealing consumers' preference
for product appearance, it is important to make the visual representation look vivid,
not only the product itself, but also the using context. In contrast, if the visual
representation of the product is different from what it really looks like, designers may
receive biased information about customer preference.
The preference inconsistency could also be explained from a cognitive perspective
[34]. When revealing preference using conjoint analysis, participants are shown a
combination of the attributes as a whole product. Thus they can have a visual
impression of what the product looks like with all the attributes. When deciding
which solar panel looks better, they actually don't have to distinguish the individual
attributes, but make decisions based on the general impression received from the
images. However, when revealing preferences using direct statement, a list of attribute
levels is showed and participants need to consider them independently. How a product
looks with all the attributes is largely dependent on the participants' imagination.
The difference between the ways customers evaluate the product could be the reason
why preference for its appearance is changed.
41
Table 4.5: Changing of 'Most Preferred' attributes from conjoint analysis revealed to
directly stated
Directly stated
Color
Conjoint
Analysis
Black
Blue
Red
Green
Sum
Black
95
5
1
1
102
Blue
9
40
1
2
52
Red
2
2
4
1
9
Green
11
8
2
10
31
Sum
117
55
8
14
194
Directly stated
Shape
Conjoint
Analysis
Big C
Small C
Square
Sum
Big C
29
38
22
89
Small C
10
34
29
73
Square
0
9
23
32
Sum
39
81
74
194
Directly stated
Pattern
Conjoint
Analysis
None
Two
Three
Sum
None
62
36
15
113
Two
22
18
8
48
Three
8
12
13
33
Sum
92
66
36
194
Directly stated
rame
Conjoint
Analysis
Silver
Black
None
Sum
Silver
14
20
7
41
Black
12
50
24
86
None
14
39
14
67
Sum
40
109
45
194
42
Chapter 5
Aggregated Preference Model
5.1
Research Questions
In addition to individual preference model, an aggregated preference model was built
for the solar panel appearance.
This model could be an aesthetic representation
of the population. The purpose of building this model is to identify the most important attributes that influence the appearance of a solar panel, reduce indifferent
attributes/levels of attributes, and give out aesthetic levels of solar panels.
There are two questions we would like to answer by building aggregated preference
model in this chapter:
1. Generally, what levels of attributes help a solar panel to gain visual appealingness and what levels of attributes may impair the visual appealingness of a solar
panel?
2. What attributes/levels of attributes dominate the appearance preference and
what attributes/level of attributes are tend to be ignored?
5.2
Conditional Logit Model
Analyze conjoint data on an aggregated level is a usual practice of marketing research,
since it answers questions such as how attributes of a product are welcomed to the
43
market, what is the segmentation of the market etc., which are directly related to
the business success of a product. Logit and probit discrete choice models are widely
used in conjoint analysis. Conditional logit model was chosen here according to the
pairwise choice questions used in the survey.
Compared to multinomial logit model, conditional logit model is more appropriate
here because the utility of choice among alternatives is modeled as a function of the
characteristics of the alternatives, rather than the characteristics of the individual
making the choice [35].
The conditional logit model:
Ui = fz + ei
(5.1)
Where Ur is the utility of the customer choosing alternative i in the choice set n;
,8 is a set of parameters giving the effects of variables on probabilities, which are
estimated statistically;
zn is a vector of observed variables relating to alternative i in the choice set n
that depends on attributes of the alternative;
em is an iid random variable.
The probability of alternative i being chosen in the choice set n is given by:
P
exp(pzi)
_1=V
(5.2)
J=1eXP(#Zni)
An underlying assumption of this model is that, there is no correlation among
different attributes. Software R was used to estimate the conditional logit model.
5.3
5.3.1
Results
Full Factory Model
Since the levels of design attributes are nominal, a dummy variable 1 was used to
indicate which level was used for each attribute. The dummy variables for the other
44
levels were set to 0. For the choice None of the two, dummy variables for all attribute
levels were set to 0.
One level of each design attribute was held out for reference, and their coefficients
Ps were set to 0. For the following model, level Green of attribute color, level Square
of attribute shape, level Three Main Contact Wires of attribute pattern, and level
No frame of attribute frame are held out as reference.
A dummy variable Constant (Cons) is added to the model to distinguish the null
choice None of the two. If a solar panel was chosen, either A or B, the variable Cons
is 1; is the choice None of the two was chosen, the variable Cons is 0. Table 5.1 shows
the result of the coefficient estimation. P-value larger than 0.05 indicates that the
coefficient of this level is not significantly different from that of the reference level,
which is zero. Significant p-values are marked out by *.
Table 5.1: Coefficients of the full model
Attributes
#
exp(#)
Standard Error
z-value
p-value
Cons
1.2933
3.645
0.1015
12.74
O.Oe+00*
CBlack
1.1235
3.076
0.0635
17.681
0.Oe+00*
C..Blue
0.6873
1.988
0.0638
10.779
0.Oe+00*
C..Red
-0.124
0.883
0.0868
-1.429
1.50e-01
S.Large
0.2979
1.347
0.0565
5.276
1.3e-07*
S-Small
0.3015
1.352
0.0493
6.117
9.5e-10*
P.Even
0.6465
1.909
0.0536
12.068
0.Oe+00*
P.Two
0.1047
1.11
0.0579
1.808
7.10e-02
F.Silver
-0.0273
0.973
0.0552
-0.495
6.20e-01
F-Black
0.0813
1.085
0.0522
1.557
1.20e-01
Likelihood ratio test=3430 on 10 df, p=O n= 15714, number of events= 5238
We can see that, the coefficient of color Red is not significant, which means that
theres no significant difference of preference for color Red to color Green. Similarly,
theres no significant difference of preference for pattern Two Main Contact Wires to
45
pattern Three Main Contact Wires; and there's no significant difference of preference
among all the frame styles.
5.3.2
Reduced Model
A new model was built by reducing the number of variables and combining the attribute levels which receive no significantly different preferences. The attributes and
their levels are listed out in Table 5.2. The levels marked by * will be held out as
reference in the model. Table 5.3 shows the estimated coefficients of the reduced
model.
Table 5.2: Levels of attributes of the reduced model
Level 1
Level 2
Color
3
Black
Blue
Shape
2
Square
With corner*
Pattern
2
Even
Uneven*
Level 3
Others
*
Attributes Num. of
Levels
Table 5.3: Coefficients of the reduced model
Attributes
0
exp(#)
Standard Error
z-value
p-value
Cons
1.591
4.907
0.0732
21.73
0.00e+00
C.Black
1.188
3.28
0.0497
23.92
0.00e+00
CBlue
0.745
2.106
0.0541
13.77
0.00e+00
S..Square
-0.271
0.762
0.0435
-6.24
4.50e-10
P..Even
0.606
1.833
0.0443
13.66
0.00e+00
Likelihood ratio test=3419 on 5 df, p=0 n= 15714, number of events= 5238
The positive coefficient of Cons (P = 1.591, p-value = 0) shows that participants
usually chose one solar panel out of the two choices, instead of indicating they dislike
any one of them. This preference is significant.
46
As for color, the combination of Red and Green were held out as reference. Black
was indicated to be the most preferred one. And Blue was second preferred.
As for shape, theres no significant difference between Small Corner and Large
Corner, thus they were merge into one level as With Corner and was held out as
reference. Generally they were more preferred than the Square shape, which has no
corner.
As for pattern, solar panels with even surfaces (No Conduct Wires) were shown to
be much more preferred than those with uneven surfaces (Two/Three Main Conduct
Wires).
As for frame, theres no significant difference between any levels. Thus this attributes is eliminated from the final model.
In addition, the demography information was explored to be included in the preference model. However none of them were proved to be significant.
5.3.3
Aesthetic Levels
Using the reduced preference model, the estimated utility Ur = fz3r of a solar panel
varies in the rage -0.271 to 1.794. Six designs of solar panels were chosen to be
representatives of three aesthetic levels: high, medium and low (see Table 5.4).
5.4
Discussion and Conclusions
By build conditional logit model, we can answer the two questions brought up in the
beginning of this chapter:
1. Generally, what levels of attributes help a solar panel to gain visual appealingness and what levels of attributes may impair the visual appealingness of a solar
panel?
As for color, it seems that dark colors, such as black and blue, will make a solar
panel looks more appealing; light colors, such as red and green, will make a solar panel
less appealing. This is interesting since black and blue are colors normally used for
47
solar panels on market. Whether this preference is biased by the prevailing products
on market could be interesting to explore.
As for shape, solar cells with rounded corner make the solar panel more appealing
to people. The corners are outcomes of the fact that monocrystalline solar cells are
fabricated from round wafers and four edges are cut out from each wafer creating a
square like shape with four round corners. Thus instead of a intentional design, the
rounded corners are actually a byproduct of manufacturing constrains. However, the
result still indicate that, solar panel with geometry patterns is more welcomed than
that without any geometry pattern, which could be a guideline for design of future
solar panels.
As for pattern, solar panels with even surface are generally more preferred than
those with contact wires on the surface. From a technical perspective, the contact
wires cast shadows on solar cells thus reduce the panel efficiency. Therefore, to make
the solar cells free from front side contact wire will satisfy both aesthetic tendency
and efficiency requirement.
2. What attributes/levels of attributes dominate the appearance preference and
what attributes/level of attributes axe tend to be ignored?
In the scope of this study, the frame style of solar panels tends to be ignored
by participants when stating their preference for panel appearance. All other three
attributes are showed to be important.
However it is possible that, some frame styles not discussed in this study can
dominant the appearance preference. In addition, it is also possible that there exists
design attributes significantly influencing the appearance preference for solar panels,
however are not addressed in this study. Those attributes may worth further exploring
as to optimize the appearance design of a solar panel, which is out of the scope of
this study.
48
Table 5.4: Examples of solar panels with different aesthetic level
Aesthetic Attributes combination
Estimated
Image
level
Low
Medium
High
Color
Shape
Pattern
Others
Square
Uneven
-0.271
Others
With
corner
Uneven
0
Blue
With
corner
Uneven
0.745
Black
Square
Uneven
0.917
Black
Square
Even
1.523
With
Even
1.794
Black
corner
49
utility
U
50
Chapter 6
Summary
6.1
Discussion and Conclusion
This thesis investigates the consumers preference for the appearance of technology
product with a case study of solar panels. In this work, four main attributes that
influence the appearance of a residential solar panel are identified. Visual choice based
conjoint analysis is used as a main method to reveal the preference model.
The Bradley-Tem-Luce (BTL) model helps us to analyze people's preference on
an individual level. And the preference revealed from non-contextualized images of
solar panel with a blank background were compare to the preference revealed from
contextualized images of solar panel within a realistic background. Inconsistencies of
product appearance preference arise when the style of images presenting the product
changes. The context in which the product is presented is believed to be the main
factor that bias consumers preference. It is not necessarily true that one type of
representation is inherently better than another, but if the goal is to increase solar
panel adoption, the suggestion of this result is that panels that are designed to match
home roofs are strongly preferred, regardless of the standalone styling of the panel.
Other preference inconsistency observed in this study is the inconsistency between
conjoint revealed preference and directly stated preference. This inconsistency could
be caused by the difference between the appearance value construction processes of
different preference elicitation methods of conjoint analysis and directly stated.
51
The conditional logit model enabled us to build the solar panel appearance model
on an aggregated level. Some attributes and levels of attributes that related to solar
panel appearance were found to be having significant influence in the preference model
while the others plays minor role in the model. Therefore, the insignificant factors
were eliminated and a reduced model was built. Using the reduced model, a set of
examples of solar panels with different estimated aesthetic value was established.
Here we give out conclusions for the two questions asked at the very beginning of
this thesis:
1. What design attributes make a solar panel more visually appealing than the
others?
2. Do different visual representations reveal consumer preferences for product appearance differently?
On one hand, the study gives some explicit conclusion on what kind of solar
panel has more appearance attractiveness to users: darker solar panels with rounded
corner and non-texturized surfaced cells are the most appealing. When considering
the appearance of solar panels on a house, the one that has similar color of the roof
is more preferred.
On the other hand, the result of the study has implications for the way stakeholders and companies represent their products to users and the broader world. The
existence of preference inconsistencies suggests designers to be cautious when studying
customers preference for product appearance. And the reasons behind the existence
of inconsistencies give designers insights into how to address the real customer preference that will lead to higher satisfaction after the design of product completed.
6.2
Limitation of Current Work
In this study, the existence of preference inconsistencies to product appearance is
demonstrated by significant statistical data. However, the reasons behind them are
52
more of a qualitative analysis than quantitative experimental study. Thus designing
an experiment to further proving the arguments made in this work is important.
In addition, the attributes and levels of attribute studied in the work is not necessarily sufficient, thus the solar panels with combination of different attributes do
not necessarily cover the whole design space. Explore the design space further could
end up with solar panels with very different appearance (integrating solar cells with
tiles on roof, for example) thus gives users a totally different experience of how this
technology product looks like.
Whats more, only two visual representatives are explored in this study: noncontextualized image and contextualized image.
Different from usually consumer
product, solar panel is basically a 2-D object (with large length and width but little
thickness). This makes user perceive its appearance differently. Thus the usual visual representatives, such as sketching and FSV silhouette, may not be able to apply.
However there could be other visual representatives, such as Illustrator generated pattern vs. realistic rendering image, that could influence the perception of appearance
of the product.
Last but not the least, only one kind of technology product is studied in this paper.
Expanding this study to other technology products, even to other consumer products
with more complicated constitution of aesthetic, should give out results with more
general value.
6.3
Future Work
The next step of this study will be investigating the tradeoff of preference among
multi-attributes of a technology product
the efficiency, reliability and price, in ad-
dition to the product appearance. Thus the weight of product appearance in this
tradeoff can be quantified.
In addition, this work could be expanded from a stated preference elicitation to
a revealed preference elicitation: using real market data and analyzing the tradeoffs
between technical factors and appearance factors. This could give out information
53
of how people make their choice within a realistic context. Also, this could provide
reference to the stated preference results.
54
Bibliography
[1] Marielle E. H. Creusen and Jan P. L. Schoormans. The Different Roles of Product
Appearance in Consumer Choice*. Journal of Product Innovation Management,
22(1):63-81, January 2005.
[2] Rolf A. Faste. Perceiving needs. Society of Automotive Engineers, 1987.
[3] Noah Tovares, Jonathan Cagan, and Peter Boatwright. Capturing Consumer
Preference Through Experiential Conjoint Analysis. In ASME 2013 International
Design Engineering Technical Conferences and Computers and Information in
Engineering Conference, pages 1-13, 2013.
[4] Seth Orsborn, Jonathan Cagan, and Peter Boatwright. Quantifying Aesthetic Form Preference in a Utility Function. Journal of Mechanical Design,
131(6):061001, 2009.
[5] Erin F. MacDonald, Richard Gonzalez, and Panos Y. Papalambros. Preference
Inconsistency in Multidisciplinary Design Decision Making. Journal of Mechanical Design, 131(3):031009, 2009.
[6] Michael Kulok and Kemper Lewis. Preference Consistency In Multiattribute
Decision Making. In ASME 2005 International Design Engineering Technical
Conferences and Computers and Information in Engineering Conference. American Society of Mechanical Engineers, pages 1-10, 2005.
[7] Gary Bamossy, Debra L Scamoon, and Marilyn Johnston. A preliminary investigation of the reliability and validity of an aesthetic judgement test. Advances
in consumer..., 10(1):685-691, 1983.
[8] RW Veryzer. The place of product design and aesthetics in consumer research.
Advances in consumer research, 22(1), 1995.
[9] Peter H. Bloch. Seeking the Ideal Form: Product Design and Consumer Response. Journal of Marketing, 59(3):16, July 1995.
[10] Nathan Crilly. The roles that artefacts play: technical, social and aesthetic
functions. Design Studies, 31(4):311-344, July 2010.
55
[111 Nathan Crilly, James Moultrie, and P.John Clarkson. Seeing things: consumer
response to the visual domain in product design. Design Studies, 25(6):547-577,
November 2004.
[12] S Moody. The role of industrial design in technological innovation. Design
Studies, 1(6):329-339, 1980.
[13] Nathan Crilly, James Moultrie, and P. John Clarkson. Shaping things: intended
consumer response and the other determinants of product form. Design Studies,
30(3):224-254, May 2009.
[14] Seth Orsborn, Jonathan Cagan, Richard Pawlicki, and Randall C. Smith. Creating cross-over vehicles: Defining and combining vehicle classes using shape
grammars. Ai Edam, 20(03):217-246, June 2006.
[15] Dan Nathan-Roberts. Using Interactive Genetic Algorithms to Support Aesthetic
Ergonomic Design. Phd thesis, The University of Michigan, 2012.
[16] Bryan Macomber and Maria Yang. The role of sketch finish and style in user
responses to early stage design concepts. In ASME 2011 InternationalDesign
Engineering Technical Conferences and Computers and Information in Engineering Conference, pages 567-576, 2011.
[17] M. Walker, L. Takayama, and J. a. Landay. High-Fidelity or Low-Fidelity, Paper
or Computer? Choosing Attributes when Testing Web Prototypes. Proceedings
of the Human Factors and Ergonomics Society Annual Meeting, 46(5):661-665,
September 2002.
[18] Heidi Q. Chen, Tomonori Honda, and Maria C. Yang. An Approach for Revealed
Consumer Preferences for Technology Products: A Case Study of Residential
Solar Panels. In ASME 2012 InternationalDesign Engineering Technical Conferences and Computers and Information in Engineering Conference. American
Society of MechanicalEngineers, pages 379-390. Asme, August 2012.
[19] Heidi Q Chen, Tomonori Honda, and Maria C Yang. Approaches for identifying
consumer preferences for the design of technology products: a case study of
residential solar panels. Journal of Mechanical Design, 135(6):061007, 2013.
[20] Warren F Kuhfeld. Conjoint Analysis. SAS Technical Papers, MR201OH:681801, 2010.
[21] Hideo Aizaki and Kazushi Nishimura. Design and Analysis of Choice Experiments Using R: A Brief Introduction. Agricultural Information Research,
17(2):86-94, 2008.
[22] Warren F Kuhfeld. Discrete Choice. SAS Technical Papers, MR-2010F:285-663,
2010.
56
[23] PE Green. On the design of choice experiments involving multifactor alternatives.
Journal of Consumer Research, 1(2):61-68, 1974.
[24] Jarod Kelly and PY Papalambros. Use of shape preference information in product
design. InternationalConference on Engineering Design, ICED, (August):1-11,
2007.
[25] Surya Swamy, Seth Orsborn, Jeremy Michalek, and Jonathan Cagan. Measurement of headlight form preference using a choice based conjoint analysis. In
ASME 2007 InternationalDesign Engineering Technical Conferences and Computers and Information in Engineering Conference, pages 197-206, 2007.
[26] Seth Orsborn, Jonathan Cagan, and Peter Boatwright. Quantifying Aesthetic Form Preference in a Utility Function. Journal of Mechanical Design,
131(6):061001, 2009.
[27] Tahira Reid, Richard Gonzalez, and Panos Papalambros. A methodology for
quantifying the perceived environmental friendliness of vehicle silhouettes in engineering design. ASME 2009 InternationalDesign Engineering Technical Conferences and Computers and Information in Engineering Conference. American
Society of Mechanical Engineers, 2009.
[28] Brian Sylcott, Jonathan Cagan, and Golnaz Tabibnia. Understanding Consumer
Tradeoffs Between Form and Function Through Metaconjoint and Cognitive Neuroscience Analyses. Journal of MechanicalDesign, 135(10):101002, August 2013.
[29] Dan Horsky, Paul Nelson, and SS Posavac. Stating preference for the ethereal but
choosing the concrete: how the tangibility of attributes affects attribute weighting in value elicitation and choice. Journal of Consumer Psychology, 14(1):132140, 2004.
[30] Tahira N. Reid, Erin F. MacDonald, and Ping Du. Impact of Product Design Representation on Customer Judgment. Journal of Mechanical Design,
135(9):091008, July 2013.
[31] M Ben-Akiva, D McFadden, K Train, and J Walker. Hybrid Choice Models:
Progress and Challenges. Marketing Letters, pages 163-175, 2002.
[32] California solar Initiative. Current CSI data, 2011.
[33] John D Hey. Do Rational People Make Mistakes. In Game Theory, Experience,
Rationality, pages 55-66. 1998.
[34] Gianluca Consoli. A Cognitive Theory of the Aesthetic Experience. Contemporary Aesthetics, 10, 2012.
[35] S D Hoffman and G J Duncan. Multinomial and conditional logit discrete-choice
models in demography. Demography, 25(3):415-27, August 1988.
57