Selecting a Technique

• Dependency
Univariate and { Dependent (criterion) variables

M lti i t Data
Multivariate D t and
d iindependent
d d t ((predictor)
di t )
variables are present
Analysis. • Interdependency
{ Variables are interrelated without
designating some dependent and
others independent

Dependency Techniques Interdependency Techniques

„Multiple regression. „Factor analysis

„Cluster analysis
y
„Conjoint analysis.
„Multidimensional Scaling
(MDS)
„Discriminant analysis.

1
 Products/Services are Composed of
Features/Attributes

Conjoint Analysis „ Credit Card:

Brand + Interest Rate + Annual Fee + Credit Limit

„ On-Line Brokerage:

Brand + Fee + Speed of Transaction + Reliability of

Transaction + Research/Charting Options

Breaking the Problem Down

How to Learn What Customers Want?

„ Ask Direct Questions about preference:

„ If we learn how buyers value the { What brand do you prefer?

components of a product, we are in a { What Interest Rate would you like?
{ What Annual Fee would you like?
better position to design those that
{ What Credit Limit would you like?
improve profitability
„ Answers often trivial and unenlightening (e.g.
respondents prefer low fees to high fees,
higher credit limits to low credit limits)

2
How to Learn What Is Important?
Stated Importances

„ Ask Direct Questions about importance „ Importance Ratings often have low discrimination:

{ How important is it that you get the <<brand, Average Importance Ratings
i t
interest
t rate,
t annuall fee,
f credit
dit limit>>
li it that
th t you
want? Brand 6.7

Interest Rate 7.2

Annual Fee 8.1

Credit Limit 7.5

0 5 10

Stated Importance Conjoint Analysis

„ Answers often have low discrimination, with „The objective is to decompose a set of overall
most answers falling in “very important” responses to designed stimuli so that the utility
categories of each stimulus attribute can be inferred from
the respondent's overall evaluations of the
„ Answers sometimes useful for segmenting
stimuli.
market, but still not as actionable as could be

3
 History

„ CONsidered JOINTly „ Luce and Tukey (1964)

„ To Conjoin „ Green and Rao (1971) – Full profile

„ Richard Johnson (1971) – Paired
„ Backdoor DECOMPOSITIONAL comparison
approach to estimate preference „ McFadden (1974) – Discrete choice
„ Green and Wind (1975, HBR) – Full
profile

Creating Conjoint Design

„ Green and Srinivasan (1978) –

LINMAP
„ Steve Herman (1982 / 83) Bretton
Bretton-
„ Attribute
Clark Software „ Level
„ Richard Johnson (1985) – Sawtooth „ Stimuli / Concept
Software (ACA)

4
 Assumption:
„ Self –explicated [Not conjoint] Utility = f ( product attributes ) Sounds funny!;
f(.) is an linear additive form (heroic assumption!);
„ Full profile No interactions between different attributes;
Consumers are able to assign `scales' to their
utilities.
tiliti
„ Paired comparison
Method of Scaling:
Metric: (preference) rating scales for each product
„ ACA attribute combination;
Non-metric: a ranking scale for each combination.
„ CBC

Different Perspectives, Different Goals

Demand Side of Equation

„ Typical market research role is to focus first

on demand side of the equation
„ Buyers want all of the most desirable features at
lowest possible price
„ After figuring out what buyers want, next
assess whether it can be built/provided in a
„ Sellers want to maximize profits by: cost- effective manner
1) minimizing costs of providing features
2) providing products that offer greater overall value
than the competition

5
 What’s So Good about Conjoint?
How Does Conjoint Analysis Work? (cont)

„ We vary the product features (independent variables) to

build many (usually 12 or more) product concepts
„ When respondents are forced to make difficult
„ We ask respondents
p to rate/rank those p
product concepts
p tradeoffs we learn what they truly value
tradeoffs,
(dependent variable)

„ Based on the respondents’ evaluations of the product

concepts, we figure out how much unique value (utility)
each of the features added

„ (Regress dependent variable on independent variables;

betas equal part worth utilities.)

Rules for Formulating

First Step: Create Attribute List attribute Levels
„ Attributes assumed to be independent
(Brand, Speed, Color, Price, etc.)
„ Levels are assumed to be mutually exclusive
„ Each attribute has varying degrees, or
“levels” Attribute: Add-on features

{ Brand: Coke, Pepsi, Sprite level 1: Sunroof

{ Speed: 5 pages per minute, 10 pages per minute level 2: GPS System
{ Color: Red, Blue, Green, Black level 3: Video Screen
„ Each level is assumed to be mutually exclusive of
{ If define levels in this way, you cannot determine
the others (a product has one and only one level of
the value of providing two or three of these
that attribute)
features at the same time

6
 Rules for Formulating Rules for Formulating
Attribute Levels Attribute Levels

„ Levels should have concrete/unambiguous „ Don’t include too many levels for any one
meaning attribute
{ The usual number is about 3 to 5 levels per attribute
“Very
Very expensive
expensive” vs.
vs “Costs
Costs $575”
$575 { The temptation (for example) is to include many
many, many
levels of price, so we can estimate people’s preferences for
“Weight: 5 to 7 kilos” vs. “Weight 6 kilos” each
{ But, you spread your precious observations across more
parameters to be estimated, resulting in noisier (less
{ One description leaves meaning up to individual precise) measurement of ALL price levels
interpretation, while the other does not { Better approach usually is to interpolate between fewer
more precisely measured levels for “not asked about” prices
in NPD

Rules for Formulating

Rules for Formulating Attribute Levels
Attribute Levels
„ Make sure levels from your attributes can combine
„ Whenever possible, try to balance the number of
freely with one another without resulting in utterly
levels across attributes
impossible combinations (very unlikely combinations
are still okay)
„ There is a well-known bias in conjoint analysis called
the “Number
Number of Levels Effect”
Effect { Resist temptation to make attribute prohibitions (prohibiting
levels from one attribute from occurring with levels from
{ Holding all else constant, attributes defined on more levels other attributes)!
than others will be biased upwards in importance { Respondents can imagine many possibilities (and evaluate
{ For example, price defined as ($10, $12, $14, $16, $18, them consistently) that the study commissioner doesn’t plan
$20) will receive higher relative importance than when to/can’t offer. By avoiding prohibitions, we usually improve
defined as ($10, $15, $20) even though the same range the estimates of the combinations that we will actually focus
was measured on.
{ The Number of Levels effect holds for quantitative (e.g. { But, for advanced analysts, some prohibitions are OK, and
price, speed) and categorical (e.g. brand, color) attributes even helpful

7
Approach of Obtaining Scales:
2. multiple-factor procedure: respondents are
1. two-at-a-time procedure: respondents are shown a combination of the levels of all
shown two attributes at a time, each attribute attributes.
is of several levels.
Example:
E
Example:
l b brand
d vs. tread
t d life,
lif Card 1
Brand: Sears
Sears Goodyear Goodrich Tread Life: 40K miles
Price: $50
40K Sidewall: Black
50K
Each respondent is asked to reveal his preferences
60K
by assigning a utility level (or a rank) to each
attribute combination.

18 Combinations for the estimation

Card No. Brand Tread Life Price Sidewall

Multiple-Factor Evaluation 1 Sears 40K $50 White
2 Sears 50K $60 White
3 Sears 60K $70 Black
Tire: 4 Goodyear 40K $60 Black
Brand: Sears, Goodyear, Goodrich; 5 Goodyear 50K $70 White

Tread Life: 40K, 50K, 60K miles; 6 Goodyear 60K $50 White
7 Goodrich 40K $70 White
Price: $50 $60,
$50, $60 $70; 8 Goodrich 50K $50 Black
Sidewall: white, black. 9 Goodrich 60K $60 White
10 Sears 40K $70 Black

Total possible combinations: 3x3x3x2=54, but 11 Sears 50K $50 White

only 18 combinations are needed:

12 Sears 60K $60 White
13 Goodyear 40K $50 White
14 Goodyear 50K $60 Black
15 Goodyear 60K $70 White
16 Goodrich 40K $60 White
17 Goodrich 50K $70 White
18 Goodrich 60K $50 Black

8
Metric Method
α: the mean utility value (score) for the
combination of the levels, Sears, 40K, $50,
Suppose a respondent gives a set of rating scores and white wall;
(18 numbers), βgy: the mean difference between Goodyear
3.5, 5.4, …. , 8.7, …. , 5.6, 9.2. and Sears;
βgr: the mean difference between Goodrich and
Run a linear regression with dummy variables: Sears;
β50: the mean difference between 50K and 40K;
Score = α + β gy GY + β gr GR + β 50 K 50 + β 60 K 60 β60: the mean difference between 60K and 40K;
+ β p 6 P 60 + β p 7 P 70 + β bk BLK + ε βp6: the mean difference between $60 and $50;
βp7: the mean difference between $70 and $50;
where GY, GR, K50, K60, P60, P70 and BLK are βbk: the mean difference between Black and
dummy variables and White sidewall.

Part-Worth
Part-Worths
the marginal utility contributed by a specific
attribute level 1.2

Suppose βgy = 0.8 and βgr = - 0.3 (estimated from 1

th above
the b regression).
i ) BBy normalizing
li i βgr to
t be
b 0,
0 08
0.8
Utility

we have a plot of part-worths as: 0.6

0.4

Goodrich Sears Goodyear 0.2

Part-Worths 0 0.3 1.1 0

Goodrich Sears Goodyear

9
 Example 1: suppose that John evaluates the
Run a dummy variable regression with the following
18 combinations and provide the following preference ratings;
Input;
Card No. Brand Tread Life Price Sidewall Pref ratings Ratings GY GR K50 K60 P60 P70 BLK
1 Sears 40K $50 White 6.5 6.5 0 0 0 0 0 0 0
2 Sears 50K $60 White 7.1 7.1 0 0 1 0 1 0 0
3 Sears 60K $70 Black 4.5 4.5 0 0 0 1 0 1 1
4 Goodyear 40K $60 Black 5.1 5.1 1 0 0 0 1 0 1
5 Goodyear 50K $70 White 4.9 4.9 1 0 1 0 0 1 0
6 Goodyear 60K $50 White 9.8 9.8 1 0 0 1 0 0 0
7 Goodrich 40K $70 White 21
2.1 21
2.1 0 1 0 0 0 1 0
8 Goodrich 50K $50 Black 8.0 8 0 1 1 0 0 0 1
9 Goodrich 60K $60 White 8.8 8.8 0 1 0 1 1 0 0
10 Sears 40K $70 Black 4.0 4 0 0 0 0 0 1 1
11 Sears 50K $50 White 7.5 7.5 0 0 1 0 0 0 0
12 Sears 60K $60 White 7.8 7.8 0 0 0 1 1 0 0
13 Goodyear 40K $50 White 6.0 6 1 0 0 0 0 0 0
14 Goodyear 50K $60 Black 9.0 9 1 0 1 0 1 0 1
15 Goodyear 60K $70 White 6.9 6.9 1 0 0 1 0 1 0
16 Goodrich 40K $60 White 5.8 5.8 0 1 0 0 1 0 0
17 Goodrich 50K $70 White 5.5 5.5 0 1 1 0 0 1 0
18 Goodrich 60K $50 Black 9.4 9.4 0 1 0 1 0 0 1

Model Summary

Adjusted Std. Error of

Validity of the Model

Model R R Square R Square the Estimate
1 .945a .893 .818 .8790
a. Predictors: (Constant), BLK, P70, K60, GR, P60, K50,
GY
ANOVAb

Model
Sum of
Squares df Mean Square F Sig.
Internal examination: use these coefficients to
1 Regression
Residual
64.284
7.726
7
10
9.183
.773
11.887 .000a recalculated the mean score of each
Total 72.009 17 combination. Plot the actual and the
a. Predictors: (Constant), BLK, P70, K60, GR, P60, K50, GY
b Dependent Variable: Ratings
b.
estimated scores,, respectively.
p y Examine the
Coefficientsa
discrepancy (such as rank-order).
Unstandardized Standardized
Coefficients Coefficients
Model
1 (Constant)
B Std. Error Beta t Sig. External examination: use these coefficients to
5.792 .567 10.208 .000
GY .717 .507 .169 1.412 .188 estimate the remaining combinations that are
not used in the estimation process. Examine
GR .367 .507 .086 .723 .487
K50 2.083 .507 .491 4.105 .002
K60
P60
2.950
-.600
.507
.507
.695
-.141
5.813
-1.182
.000
.264
the discrepancy.
P70 -3.217 .507 -.758 -6.339 .000
BLK .108 .439 .026 .247 .810
a. Dependent Variable: Ratings

10
 Rating Score = α + β BR BR + β BM BM + βWBWB + βWCWC
Example 2: Color TV set + β13 S13 + β19 S19 + β R R + ε

where BR, BM, WB, WC, S13, S19, and R are dummy
Attributes: variables and
α: the valuation of the base model: a Sony TV set with
Brand: Sony, RCA, and Magnavox; warranty package A and 10 inch screen, but without a
remote control;
Warranty: Package A: labor and parts for the
βBR: the mean difference between RCA and Sony y TV sets;;
first 9 months,
months βBM: the mean difference between Magnavox and Sony TV
Package B: free exchange during the sets;
first 6 months, βWB: the mean difference between warranty packages B and
A;
Package C: parts only for the first 18 βWC: the mean difference between warranty packages C and
months; A;
Screen Size: 10, 13, and 19 inches; β13: the mean difference between screen size 13 and 10
inches;
Remote Control: Yes or No. β19: the mean difference between screen size 19 and 10
inches;
βR: the mean difference between with a remote controller
and without a controller.

18 Combinations for estimation

Card No. Brand Warranty Screen Remote Rating

1 Sony Package A 10 inch Yes 58
2 Sony Package B 13 inch Yes 75
3 Sony Package C 19 inch No 92
4 RCA Package A 13 inch No 48 3 Combinations for validation
5 RCA Package B 19 inch Yes 90
6 RCA Package C 10 inch Yes 55 Card No. Brand Warranty Screen Remote Rating

7 Magnavox Package A 19 inch Yes 88 19 Sony Package A 10 inch No 50

8 Magnavox Package B 10 inch No 45 20 RCA g A

Package 10 inch No 38

9 Magnavox Package C 13 inch Yes 65 21 Magnavox Package A 10 inch No 38

10 Sony Package A 19 inch No 95

11 Sony Package B 10 inch Yes 60
12 Sony Package C 13 inch Yes 70
13 RCA Package A 10 inch Yes 52
14 RCA Package B 13 inch No 50
15 RCA Package C 19 inch Yes 85
16 Magnavox Package A 13 inch Yes 68
17 Magnavox Package B 19 inch Yes 90
18 Magnavox Package C 10 inch No 35

11
 Model Summary
Run a dummy variable regression with the first
18 observations, Adjusted Std. Error of
Model R R Square R Square the Estimate
Rating BR BM WB WC S13 S19 R 1 .984a .968 .945 4.387
58 0 0 0 0 0 0 1 a. Predictors: (Constant), R, S19, WC, BM, S13, WB, BR

75 0 0 1 0 1 0 1 ANOVAb

92 0 0 0 1 0 1 0 Sum of
Model Squares df Mean Square F Sig.
48 1 0 0 0 1 0 0
1 Regression 5762.000 7 823.143 42.761 .000a
90 1 0 1 0 0 1 1 Residual 192.500 10 19.250
55 1 0 0 1 0 0 1 Total 5954.500 17
a. Predictors: (Constant), R, S19, WC, BM, S13, WB, BR
88 0 1 0 0 0 1 1
b. Dependent Variable: Ratings
45 0 1 1 0 0 0 0
65 0 1 0 1 1 0 1 Coefficientsa

95 0 0 0 0 0 1 0 Unstandardized Standardized
Coefficients Coefficients
60 0 0 1 0 0 0 1
Model B Std. Error Beta t Sig.
70 0 0 0 1 1 0 1 1 (Constant) 51.333 3.102 16.546 .000
BR -11.667 2.533 -.302 -4.606 .001
52 1 0 0 0 0 0 1
BM -9.833 2.533 -.255 -3.882 .003
50 1 0 1 0 1 0 0 WB .167 2.533 .004 .066 .949
85 1 0 0 1 0 1 1 WC -1.167 2.533 -.030 -.461 .655
S13 11.833 2.533 .307 4.671 .001
68 0 1 0 0 1 0 1
S19 39.167 2.533 1.015 15.462 .000
90 0 1 1 0 0 1 1 R 10.500 2.194 .272 4.786 .001
35 0 1 0 1 0 0 0 a. Dependent Variable: Ratings

Using the estimated coefficients, we predict the preference ratings,

Rating BR BM WB WC S13 S19 R (intercept) Pred. Rating
Coefficients -11.667 -9.8333 0.16667 -1.1667 11.8333 39.1667 10.5 51.33333

58 0 0 0 0 0 0 1 61.83
75 0 0 1 0 1 0 1 73.83
92 0 0 0 1 0 1 0 89.33
48 1 0 0 0 1 0 0 51.50 Observed vs. Predicted Ratings
90 1 0 1 0 0 1 1 89.50
55 1 0 0 1 0 0 1 49.00
88 0 1 0 0 0 1 1 91.17
100
45 0 1 1 0 0 0 0 41.67 90
65 0 1 0 1 1 0 1 62 67
62.67
95 0 0 0 0 0 1 0 90.50 80
Rating

60 0 0 1 0 0 0 1 62.00
70 Predicted Rating
70 0 0 0 1 1 0 1 72.50
52 1 0 0 0 0 0 1 50.17 60 Rating
50 1 0 1 0 1 0 0 51.67
50
85 1 0 0 1 0 1 1 88.17
68 0 1 0 0 1 0 1 63.83 40
90 0 1 1 0 0 1 1 91.33
35 0 1 0 1 0 0 0 40.33
30
Validation 1 3 5 7 9 11 13 15 17
50 0 0 0 0 0 0 0 51.33
Observations
38 1 0 0 0 0 0 0 39.67
38 0 1 0 0 0 0 0 41.50

12