Chapter 5.

Conjoint analysis
1. Objectives
 Conjoint analysis is to understand how respondents develop preferences for any types of objects
(products, services or ideas), assuming that
o Consumers evaluate the value of an object (real or hypothetical) by combining the separate
amounts of value provided by each attribute
o Consumers can best provide their estimates of preference by judging objects by
combinations of attributes

 Trade-offs analysis
o To understand the trade-offs of different attributes

Example

3 important features for a new golf ball

 Average driving distance
 Average ball life
 Price
Average driving distance Average ball life Price
300 yards 54 holes $10
250 yards 36 holes $15
200 yards 18 holes $20

 Market’s ideal ball is

o 300 yards, 54 holes and $10

 Manufacturer’s ideal ball is

o 200 yards, 18 holes and $20

 Where is the most viable product?

 How does the market make a trade-off among different features?

2. Study design

 Utility
o A subjective judgment of preference unique to each individual
o Encompasses all features of the object and such is a measure of an individual’s overall
preference
o Assumed to be based on the value placed on each of the levels of the attributes
o Expressed by a relationship reflecting the manner in which the utility is formulated for any
combination of attributes

 Factors
o To define utility, a researcher must describe the object in terms of both its attributes and all
relevant values for each attribute
o Factor
 A specific attribute or characteristic of the product or service
 A researcher must identify all of the important attributes that could affect the preference
and thus utility
 Actionable

1
 o Level
 Possible values for a factor
 Describe an object in terms of its levels on the set of factors
 Range reflected in market place (somewhat greater than that in marketplace)
 Non-linear relation: 3 levels

 Stimulus
o Treatment
o A combination of different levels of different factors
o Describe an object according to a specific plan

Example

A company, HBAT, is to develop a new cleaner. After discussions with sales representatives and
focus groups, management decides 3 important attributes.
Factor Levels
Ingredients Phosphate-free Phosphate-based
Form Liquid Powder
Brand name HBAT Generic brand

 3 factors
o Each with 2 levels

 A hypothetical cleaning product can be constructed by selecting one level of each attribute
o 8 = (2 x 2 x 2) combinations (stimuli) can be formed
o HBAT phosphate-free powder
o Generic phosphate-based liquid
o Generic phosphate-free liquid
o …

 Steps
o Constructs a set of real or hypothetical objects by combining selected levels of each attribute
 Results in a design which is the set of stimuli presented to the respondent
o These stimuli are presented to respondents
 Respondent provide only their overall evaluation
 Respondents need not to tell the researcher anything else
 How important an individual attribute is to them
 How well the object performs on any specific attribute
o The influence of each attribute and each value of each attribute on the utility judgment of a
respondent can be determined from the respondents’ overall ratings

3. Models
 Additive model
o Utility represent the total worth or overall preference of an object
o Sum of what the product parts are worth (part-worth)

Utility of product of level i for factor 1, level j of factor 2, …, level n of factor m

= (Total worth)i,j,…,n
= Part-worth of level i for factor 1 + Part-worth of level j for factor 2 + … + Part-worth of level n
for factor m

2
Example

 Preference structure of the industrial cleanser represented as based on adding up the 3 factors
o Utility = brand effect + ingredient effect + form effect

 For the preference for HBAT phosphate-free powder

o Utility = Part-worth of HBAT brand + Part-worth of phosphate-free cleaning ingredient +
Part-worth of powder form

 Interactive model
o The utilities of certain combinations of levels are more or less than just the sum of the part-
worth

4. Data collection
 Trade-off method
o Compares attributes two at a time by ranking or rating all combinations of levels
o All possible combinations of attributes are used
o For N factors, number of trade-off matrices = N (N – 1)/2

Example

Consider 4 factors for a product

 Price, brand name, form and color
o Totally 4 (4 – 1)/2 = 6 comparisons
 Price vs brand name, price vs form, price vs color, brand name vs form, brand name vs
color and form vs color
o Each trade-off matrix has r × c cells where r and c are the number levels of the factors of
concerned.

 Rank the combinations

o 1 is the most preferred and largest value is the least preferred

 Price vs Brand name

Price
$1.19 $1.39 $1.49
Brand name Generic 3 6 9
KX-19 2 5 8
Clean-all 1 4 7

 Brand name vs form

Form
Powder Liquid
Brand name Generic 6 5
KX-19 4 3
Clean-all 2 1

3
 Full-profile method
o Each stimulus is described separately
o Perceived realism
o Reduce the number of comparisons through the use of fractional factorial designs
o Respondent evaluates all possible stimuli or a smaller number of stimuli

Example

Ask respondent rank or give a rating on the profiles

 When price has 3 levels, brand name has 3 levels, form has 2 levels and color has 2 levels, there
are 3 × 3 × 2 × 2 = 36 stimuli

 Profile 1
o Rate 7
Brand name: KX-19
Price: $1.19
Form: Powder
Color brightener: Yes

 Profile 2
o Rate 10
Brand name: Clean-all
Price: $1.19
Form: Liquid
Color brightener: No

 Pairwise comparison method

o Comparison of two profiles
o Respondent most often using a rating scale to indicate strength of preference for one profile
over the other
o Profile does not contain all the attributes
o Respondent evaluates all possible stimuli/comparisons or a smaller number

Example

 Comparison 1
o Rate 5 for preference for 1 over 2
Profile 1 vs Profile 2
Brand name: KX-19 Brand name: Generic
Price: $1.19 Price: $1.49
Form: Powder Form: Liquid

 Comparison 2
o Rate 7 for preference for 2 over 1
Profile 1 vs Profile 2
Brand name: KX-19 Brand name: Clean-all
Price: $1.39 Price: $1.19
Color brightener: Yes Color brightener: No

4
 Subsets of stimuli
o Factional factorial design
 Full profile method
 Orthogonal design
 No correlation among levels of an attribute
 Balanced design
 Each level in a factor appears the same number of times

o Cyclical design
 Trade-off method

 Measures
o Rank-ordering or rating

Example

Consider rating all 8 possible full profiles in a 8-point scale

 2 respondents are considered

Stimulus Form Ingr. Brand Resp. 1 Resp. 2

1 Liquid P-F HBAT 8 8
2 Liquid P-F Generic 7 7
3 Liquid P-B HBAT 4 6
4 Liquid P-B Generic 3 5
5 Powder P-F HBAT 6 2
6 Powder P-F Generic 5 4
7 Powder P-B HBAT 2 1
8 Powder P-B Generic 1 3

5. Estimation of part-worth
 Rank-order measures
o Monotonic ANOVA model (MONOANOVA)
o Modified form of ANOVA specified for ordinal data
 Rating or metric measure
o Multiple regression
o It is an ANOVA when all factors are categorical

Example

 For metric measure (rate), employ the ANOVA

o Part-worth is the deviation of the level average from the overall average for full factorial
design.
 Overall rate = (1 + 2 + … + 8 )/8 = 4.5
 For respondent 1
o For liquid form
 Rates across stimuli are 8, 7, 4, 3
 Average rate = (8 + 7 + 4 + 3)/4 = 5.5
 Part-worth = level average rate – overall average = 5.5 – 4.5 = 1

5
 Respondent 1 Respondent 2
Rates Average Part-worth Rates Average Part-worth
Form
Liquid 8,7,4,3 5.5 +1.0 8,7,6,5 6.5 +2.0
Powder 6,5,2,1 3.5 -1.0 4,3,2,1 2.5 -2.0
Ingredients
P-F 8,7,6,5 6.5 +2.0 8,7,4,2 5.25 +0.75
P-B 4,3,2,1 2.5 -2.0 6,5,3,1 3.75 -0.75
Brand
HBAT 8,6,4,2 5.0 +0.5 8,6,2,1 4.25 -0.25
Generic 7,5,3,1 4.0 -0.5 7,5,4,3 4.75 +0.25

Example
 There are 2 levels of Form and Price is a continuous factor.
Form Price Rate
Liquid 1 8
Liquid 2 7
Liquid 4 4
Powder 1 6
Powder 3 2

 In general, particularly when there are continuous factors, apply the linear regression model.
𝑌 𝛼 𝑈𝐹 𝑈𝑃
1 1 1
⎡1 1 2⎤ 𝛼
⎢ ⎥
𝐘 ⎢1 1 4⎥ 𝑈
⎢1 1 1⎥ 𝑈
⎣1 1 3⎦
o F1 = 1 for liquid and -1 for powder, U1 = part-worth of liquid, U2 = -U1 = part-worth of
powder and (U3 P) = part-worth of price P.

 The design matrix is given as

1 1 1
⎡1 1 2⎤
⎢ ⎥
𝐗 ⎢1 1 4⎥
⎢1 1 1⎥
⎣1 1 3⎦
 Then the estimate of the part-worth is
𝛼
𝑈 𝐗 𝐗 𝐗 𝐘
𝑈
𝛼 8.525
𝑈 1.425
𝑈 1.55
 We have
Part-worth
Form
Liquid 1.425
Powder -1.425

6
 Price
1 -1.55
2 -3.10
3 -4.65
4 -6.20

6. Interpretation
 Examination of the part-worth estimates for each factor
o Part-worth estimates are typically scaled so that the higher the part-worth, the more impact it
has on overall utility
o Part-worth values can be plotted graphically to identify patterns

 Assessing the relative importance of attributes

o Assess the relative importance of each factor
o The most important factor is the factor with the greatest range of part-worths

Example

Importance of the 3 factors

 For respondent 1
o Range of form factor = 1 – (1) = 2
o Range of ingredient factor = 2 – (2) = 4
o Range of brand factor = 0.5 – (0.5) = 1
o Sum of all ranges = 2 + 4 + 1 = 7
o Importance of form factor = 2/7 = 28.6%

Respondent 1 Respondent 2
Range Importance Range Importance
Form 2 28.6% 4 66.7%
Ingredients 4 57.1% 1.5 25%
Brand 1 14.3% 0.5 8.3%
Total 7 6

 Ingredients is the most important factor to respondent 1

 Form is the most important factor to respondent 2

7. Managerial applications
 Segmentation
o Group respondents with similar part-worths or importance values to identify segments
o Apply cluster analysis to the part-worth estimates or the importance scores

 Conjoint simulators
o Predict the share of preferences
o Choice simulators

o Specify the scenarios

 Researcher provides the set of stimuli representing the objects available in the market
scenario being examined

7
 o Stimulate choices
 Part-worths for each individual are used to predict the choices across the stimuli in each
scenario
o Calculate share of preferences
 Predict preference for each individual and then calculate share of preferences for each
stimulus by aggregating the individual choices
 Maximum utility (first choice) model
 Assumes the respondent chooses the stimulus with the highest predicted utility score
 Most often used
 Preference probability model
 Overall share of preference is measured by summing the preference probability
across all respondents
 Bradford-Terry-Luce (BTL) model
share x j   U x j   U  xi 
i

 Not often used

 Logit model
share x j   Exp U x j   ExpU x 
i
i

 Sometimes used

Example
 Respodnent 1
o Utility for liquid, phosphate-free and HBAT = 4.5 + 1 + 2 + 0.5 = 8
o Ranking the predicted utilities and the actual rates for all 8 stimuli
Part-worth Ranking
Stimulus Form Ingr Brand Form Ingr Brand Utility Est. Act.
1 Liquid P-F HBAT 1 2 0.5 8 8 8
2 Liquid P-F Generic 1 2 -0.5 7 7 7
3 Liquid P-B HBAT 1 -2 0.5 4 4 4
4 Liquid P-B Generic 1 -2 -0.5 3 3 3
5 Powder P-F HBAT -1 2 0.5 6 6 6
6 Powder P-F Generic -1 2 -0.5 5 5 5
7 Powder P-B HBAT -1 -2 0.5 2 2 2
8 Powder P-B Generic -1 -2 -0.5 1 1 1
o Estimated part-worth predict the preference order perfectly
o Preference was successfully represented in the part-worth estimates
o Respondent made choices consistent with the preference structure
 Respondent 2
Part-worth Ranking
Stimulus Form Ingr Brand Form Ingr Brand Utility Est. Act.
1 Liquid P-F HBAT 2 0.75 -0.25 7.0 7 8
2 Liquid P-F Generic 2 0.75 0.25 7.5 8 7
3 Liquid P-B HBAT 2 -0.75 -0.25 5.5 5 6
4 Liquid P-B Generic 2 -0.75 0.25 6.0 6 5
5 Powder P-F HBAT -2 0.75 -0.25 3.0 3 2
6 Powder P-F Generic -2 0.75 0.25 3.5 4 4
7 Powder P-B HBAT -2 -0.75 -0.25 1.5 1 1
8 Powder P-B Generic -2 -0.75 0.25 2.0 2 3
o Inconsistency in rankings prohibits a full representation of the preference structure
o Interaction effect

8
Example

HBAT use conjoint results to simulate choices among 3 possible products. Product 1 and 2 are
existing products while Product 3 is new
 Product 1: Liquid, phosphate-based and HBAT brand
 Product 2: Powder, phosphate-free and Generic brand
 Product 3: Powder, phosphate-free and HBAT brand

 Utilities
Utility BTL Logit
Product Form Ingr Brand R1 R2 R1 R2 R1 R2
1 Liquid P-B HBAT 4.0 5.5 27% 46% 9% 82%
2 Powder P-F Generic 5.0 3.5 33% 29% 24% 11%
3 Powder P-F HBAT 6.0 3.0 40% 25% 67% 7%

 Market share
o Maximum utility model
 Respondent 1 selects product 3
 Respondent 2 selects product 1
 Product 1 has 50% market share
 Product 2 has 0% market share
 Product 3 has 50% market share

o BTL probabilistic model

 Respondent 1, market share of product 1 = 4 / (4 + 5 + 6) = 27%
 Respondent 2, market share of product 2 = 3.5 / (5.5 + 3.5 + 3.0) = 29%
 Averages of the market shares
Product Form Ingr Brand Share
1 Liquid P-B HBAT 36%
2 Powder P-F Generic 31%
3 Powder P-F HBAT 33%
o Logit model
 Respondent 1, market share of product 1 = Exp(4) / (Exp(4) + Exp(5) + Exp(6)) = 9%
 Respondent 2, market share of product 2 = Exp(3.5) / (Exp(5.5) + Exp(3.5) + Exp(3.0))
= 11%
 Averages of the market shares
Product Form Ingr Brand Share
1 Liquid P-B HBAT 46%
2 Powder P-F Generic 18%
3 Powder P-F HBAT 37%

 Profitability analysis
o If the cost of each feature is known, the cost of each product can be combined with the
expected market share and sales volume to predict its viability
o Identify combinations of attributes that would be profitable even with smaller market shares

9
Example (eg1)

 In developing a new industrial cleanser, a company wants to understand the needs and
preferences of its industrial customers. Focus group research established specific levels for each
attribute that were deemed actionable and communicable through a small-scale pretest and
evaluation study.

Factor Levels
Color Red, blue, yellow
Form Cubic, cylinder, spherical
Scent Yes, no

 Preferences from 2 respondents are collected. Totally, there are 18 profiles. Full-profile method
is employed to obtain the respondent evaluations. The respondent rank 18 for the most preferred
product while 1 for least preferred product.

 Apply the preference-based conjoint analysis to study the data.

 Simulate choices among 3 possible products and find the market share.
o Product 1: blue, cubic and with scent
o Product 2: Red, cylinder and with scent
o Product 3: Red, Sphere and without scent

Solution
 Conjoint analysis for preference data
 Monotonic ANOVA

 Fitness
> summary([Link]$lm)
Response r1 :
Residual standard error: 0.5761 on 12 degrees of freedom
Multiple R-squared: 0.9918, Adjusted R-squared: 0.9884
F-statistic: 289.6 on 5 and 12 DF, p-value: 4.473e-12

Response r2 :
Residual standard error: 0.3954 on 12 degrees of freedom
Multiple R-squared: 0.9961, Adjusted R-squared: 0.9945
F-statistic: 617.3 on 5 and 12 DF, p-value: 4.922e-14
o For respondent A, R2 = 0.9918 which is close to 1 and hence the model fits well.
o For respondent B, R2 = 0.9961 which is close to 1 and hence the model fits well.

 Utilities and importance

> summary([Link]$lm)
Response r1 :
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 9.500e+00 1.358e-01 69.967 < 2e-16 ***
color1 -1.374e-05 1.920e-01 0.000 1
color2 -3.062e+00 1.920e-01 -15.944 1.93e-09 ***
form1 -6.123e+00 1.920e-01 -31.887 5.71e-13 ***
form2 3.062e+00 1.920e-01 15.944 1.93e-09 ***
scent1 -1.304e+00 1.358e-01 -9.606 5.52e-07 ***

10
Response r2 :
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 9.5000 0.0932 101.929 < 2e-16 ***
color1 5.0207 0.1318 38.091 6.90e-14 ***
color2 -3.3596 0.1318 -25.489 8.09e-12 ***
form1 3.7155 0.1318 28.189 2.46e-12 ***
form2 -0.2539 0.1318 -1.926 0.0781 .
scent1 -2.2605 0.0932 -24.254 1.45e-11 ***
> contrasts(clean$form)
[,1] [,2]
cubic 1 0
cylinder 0 1
spherical -1 -1
> contrasts(clean$color)
[,1] [,2]
red 1 0
blue 0 1
yellow -1 -1
> contrasts(clean$scent)
[,1]
yes 1
no -1
> [Link]$part
r1 r2
red -1.373984e-05 5.0207547
blue -3.061516e+00 -3.3596183
yellow 3.061529e+00 -1.6611365
cubic -6.123014e+00 3.7155523
cylinder 3.061507e+00 -0.2538641
spherical 3.061507e+00 -3.4616882
yes -1.304291e+00 -2.2605185
no 1.304291e+00 2.2605185
> [Link]$imp
r1 r2
color 0.3417613 0.4173773
form 0.5126393 0.3574563
scent 0.1455995 0.2251664
o For respondent A
 Part-worth of spherical (reference level) = – (-6.123+3.062) =3.062
 Part-worth of yellow = – (0 – 3.062) = 3.062
 Part-worth of no scent = – (-1.304) = 1.304
 Range of color part-worth = 3.062 – (–3.062) = 6.123
 Range of form part-worth = 9.185
 Range of scent part-worth = 2.609
 Total range = 17.916
 Importance of form = 9.185/17.916 = 51.26%
 Importance of color = 6.123/17.916 = 34.18%
 Importance of scent = 2.609/17.916 = 14.56%

 Form is the most important factor among the 3 attributes which account for 51.26% of
the utility
 The second important factor is color.
 Yellow is the most preferred color.

11
  Cylinder and spherical are the most preferred form.
 No scent is preferred.

 The optimal profile is yellow, cylinder and no scent.

o For respondent B
 Color is the most important factor (41.74%).
 Form is the second most important factor (35.75%).
 Red, cubic and no scent is the most preferred.

 The predicted utilities for the respondents

o For respondent A
 For the product with attributes red, cubic and scent, the predicted utility is
9.5 + 0.0 – 6.123 – 1.304 = 2.073
o For respondent B
 For the product with attributes yellow, cylinder and no scent, the predicted utility is
9.5 – 1.661 – 0.254 + 2.261 = 9.846
o Manager can search down from the list the profile that has high utility and is feasible to
make within company resources and offers an acceptable profit
> cbind(profile,pred=predict([Link]$lm,newdata=profile))
color form scent pred.r1 pred.r2
1 red cubic yes 2.0726815 15.975788
2 red cubic no 4.6812634 20.496826
3 red cylinder yes 11.2572022 12.006372
4 red cylinder no 13.8657841 16.527409
5 red spherical yes 11.2572022 8.798548
6 red spherical no 13.8657841 13.319585
7 blue cubic yes -0.9888204 7.595415
8 blue cubic no 1.6197615 12.116453
9 blue cylinder yes 8.1957003 3.625999
10 blue cylinder no 10.8042822 8.147036
11 blue spherical yes 8.1957003 0.418175
12 blue spherical no 10.8042822 4.939212
13 yellow cubic yes 5.1342246 9.293897
14 yellow cubic no 7.7428065 13.814934
15 yellow cylinder yes 14.3187453 5.324481
16 yellow cylinder no 16.9273272 9.845518
17 yellow spherical yes 14.3187453 2.116657
18 yellow spherical no 16.9273272 6.637694

 Choice simulator
o The predicted utilities of the 3 products are
> cbind(csimp,pred=predict([Link]$lm,newdata=csimp))
color form scent pred.r1 pred.r2
19 blue cubic yes -0.9888204 7.595415
20 red cylinder yes 11.2572022 12.006372
21 red spherical no 13.8657841 13.319585
o Both respondents A and B choose product 3
o By maximum utility model, the market shares are
> maxconj(conjanal=[Link],simp=csimp)
color form scent share
19 blue cubic yes 0
20 red cylinder yes 0
21 red spherical no 1

12
 o BTL model (average of the individual predictions)
> btlconj(conjanal=[Link],simp=csimp)
color form scent share
19 blue cubic yes 0.09487101
20 red cylinder yes 0.41557048
21 red spherical no 0.48955851

o Logit model (average of the individual predictions)

> logitconj(conjanal=[Link],simp=csimp)
color form scent share
19 blue cubic yes 0.001283773
20 red cylinder yes 0.139996849
21 red spherical no 0.858719378

o Different market shares obtained by different methods

o Try metric ANOVA.

Example (eg2)

 A study of how students evaluate sneakers.

 Qualitative research identified 3 attributes: the sole, the upper and the price.
o Each defined in terms of 3 levels

Sole 1 Rubber
2 Polyurethane
3 Plastic
Upper 1 Leather
2 Canvas
3 Nylon
Price 1 $30
2 $60
3 $90

 Full-profile approach is used.

o Given 3 attributes, defined in 3 levels each, a total of 27 profiles can be constructed.
o To reduce the respondent evaluation task, a fractional factorial design was employed and a
set of 9 profiles was constructed.
 Respondents were required to provide preference ratings for the sneakers described by the 9
profiles. These ratings were obtained using a 9-point Likert scale (1 = not preferred, 9 = greatly
preferred).

 Study the segmentation of the market based on how the customers evaluate the importance of
the attributes.
Solution
 Metric scale is considered based on rate.
o ANOVA model is employed

13
 For customer 1
o Price is the most important attribute
> [Link]$`Response COL4`
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 4.2222 0.1111 38.000 0.000692 ***
COL11 0.7778 0.1571 4.950 0.038476 *
COL12 -0.5556 0.1571 -3.536 0.071523 .
COL21 0.1111 0.1571 0.707 0.552786
COL22 0.4444 0.1571 2.828 0.105573
COL31 0.7778 0.1571 4.950 0.038476 *
COL32 0.4444 0.1571 2.828 0.105573
> [Link]$part[,1]
rubber polyurethane plastic leather
0.7777778 -0.5555556 -0.2222222 0.1111111
canvas nylon 30 60
0.4444444 -0.5555556 0.7777778 0.4444444
90
-1.2222222
> [Link]$imp[,1]
COL1 COL2 COL3
0.3076923 0.2307692 0.4615385

 Cluster analysis is applied to all customers based on the importance of the attributes.
o Ward’s method
o The standardized importance.
o Plot of the distances and dendrogram

100
100
height

Height

60

50
0 20

0
0 20 40 60
stage
 3 clusters are suggested

 Consider a 3-cluster solution by K-means method

> fit1
K-means clustering with 3 clusters of sizes 21, 27, 12

Cluster means:
COL1 COL2 COL3
1 -1.0422240 0.06859145 0.96887371
2 0.3956916 0.56576154 -0.78042756
3 0.9335859 -1.39299850 0.06043303

14
o Cluster 1 : Sole is the most important attribute.
o Cluster 2 : Price is the most important attribute.
o Cluster 3 : Upper is the most important attribute.

15