Conjoint Analysis
What is conjoint analysis:
A general class of procedures for measuring, analyzing and predicting consumers responses to:
1. new products
2. new features of existing products
Decomposes consumers’ preferences for any product into utilities for each option of each feature or attribute.
Utilities (part-worths) can be combined to predict preferences for any product that can be developed using those features or attributes
Can determine optimal product concept
Can identify segments that value a product concept highly
Conjoint Analysis
Statistics & Terms
• Part-worth functions - utility of each level of each attribute
• Relative importance weights - which attributes are important in choice of a brand, etc.
• Attribute levels - feature options available in product
• Full profiles - all attributes and levels included in product profiles
• Pairwise tables - two attributes with all levels evaluated at a time
• Full factorial design - all combinations of all attributes levels included in profiles
• Fractional factorial designs - subset of full factorial design
• Orthogonal arrays - fractional design to capture only main effects
Conjoint Analysis
Steps in conjoint analysis
Formulate the problem
Construct the stimuli
Determine form of input data
Select a conjoint procedure
Interpret results
Assess reliability & validity
Conjoint Analysis
Formulate the problem
1. Select attributes - salient in influencing preference & choice;
-- not similar across products
2. Select attribute levels - ranges beyond availability in market, but believable
Attribute & Level Selection.
Assume:
A. Product = Bundle of Attribute levels
B. Utility of Product = f (Utility of Attribute levels)
C. Highest Utility = pr (Purchases) highest
D. Exhaustive set of attributes & levels
E. No redundancy
F. High External Validity re: Buying Action.
Conjoint Analysis
Construct the stimuli
Two Methods
1. Paired Comparison - two-factor (attribute/feature) evaluation
2. Full Profile - multiple-factor (attribute/feature) evaluations
1. Paired - Comparison
• Easier for respondents
• Requires more evaluations than full profile
• Evaluations may be unrealistic
Conjoint Analysis
Tradeoff Matrix - Pairwise Comparisons
RANK 1 TO 9 FOR EACH MATRIX
Course Course Course
Grading 1 2 3
1 1 4 3
2 9 2 5
3 8 7 6
Conjoint Analysis
Tradeoff Matrix - Pairwise Comparisons
RANK 1 TO 9 FOR EACH MATRIX
Command Command Command
Grading 1 2 3
1 4 3 2
2 5 1 6
3 8 7 9
Course Course Course
Grading 1 2 3
1 9 3 2
2 8 1 4
3 5 6 7
etc.
Conjoint Analysis
2. Full Profile Methods
A. Full Factorial *** Multiple Factors
determines both main & interaction effect
(recall ANOVA, multiple regression)
B. Fractional Factorial -
Balanced Design = Independence of Attributes
i.e. Attribute Levels are Uncorrelated
Conjoint Analysis
2. Full Profile Methods
Fractional factorial
1. Randomized Block - homogeneous groups are independent
Group 1 E (R) O1 X O3
e.g., Family C (R) O2 O4
Group 2 E (R) X O5
Individual C (R) O6
2. Orthogonal arrays
• substitute new factor for interaction presumed negligible
• main effects only - uncorrelated with other main effects
Conjoint Analysis
2. Full Profile Methods
Latin Square (example: Computer use in schools)
Operating system: (5)Win95 (8)Win98 (M)MAC
Software for: 1 Course 2-courses 3-courses
Monitor: 15” 17” 19”
Full Factorial Latin Square
5 8 M 5 8 M
1 5,7,9 5,7,9 5,7,9 1 5 7 9
2 5,7,9 5,7,9 5,7,9 2 7 9 5
3 5,7,9 5,7,9 5,7,9 3 9 5 7
Conjoint Analysis
2. Full Profile Methods
Latin Square
Example 2
COURSE REQUIREMENTS
GRADING Heavy Moderate Light
Generous Good Fair Poor
Moderate Fair Poor Good
Strict Poor Good Fair
Within cell - COMMAND OF SUBJECT
Conjoint Analysis
Determine form of input data
1. Nonmetric - rank order profiles / pairs
2. Metric - rate on a scale
a. Form of scale
– e.g., Likert (1 = not preferred; 7 = greatly preferred)
b. Dependent variable
i. Preference
ii. Intent to buy (e.g., probability of purchase)
iii. Choice
iv. Actual purchase
v. Attitude (e.g., Liking Scale)
Conjoint Analysis
Select a conjoint procedure
Goal: Decompose overall responses into utilities of various attribute levels.
Model:
where: U(X) = overall utility of brand
= part worth of jth level of ith attribute
ki = number levels of attribute I
m = number of attributes
a = part-worth or utility of level of attribute
x = level of attribute or feature
Conjoint Analysis
Select a conjoint procedure
Technique: Regress Rank or Rate on Attribute Levels
Tools: OLS regression, LINMAP, MONANOVA, Logit
OLS regression:
Independent variables (IV) are Dummy Variables
==> number of levels - 1 (the referent level)
Dependent variables (DV) are ratings or rankings
• If ratings, use directly as DV
• If rankings, convert to 0 or 1 by making paired comparisons between brands
Conjoint Analysis
Interpret results
* Can predict preference for any attribute level combination
* Graph part worths for each attribute to illustrate results
* Determining importance of attributes
Self-explicated (reported)
• Self-rated
– 5-point scale (1 = of no importance to 5 = extremely important)
– snake plot to visualize
• Anchored measures
– 10 points to most important
– 0 to 10 points to all other attributes
• Constant-sum
– 0 to 100 points divide among attributes
Conjoint Analysis
Interpret results
* Determining importance of attributes
Problems with self-explicated (reported) importance
• self-rated scales don’t force tradeoffs or hierarchies among attributes
• anchored scales are more difficult to administer
• constant sum have independence from irrelevant needs (IIN) problem: more needs = fewer points allocated per need
• everyone wants best at lowest price
• report desirability better than importance (regress towards mean)
• importance depends on range of attribute values, e.g., price cut-off.
Conjoint Analysis
Interpret results
* Revealed importance of attribute = Difference between highest and lowest attribute level utilities.
• Importance of an attribute = [Max (aij) - Min (aij)] for each “i”
• To determine importance relative to other attributes, normalize importance
that is:
Conjoint Analysis
Assess reliability and validity
• Goodness-of-fit: R2 in dummy variable regression
• Reliability: Correlate response to stimuli used early and late in the study
• Internal validity: Predict values in a holdout sample
• Stability: Compare subsamples in an aggregate-level analysis
Conjoint Analysis
Using Conjoint Analysis
• Determining relative importance of attributes in consumer choice
• Estimating market share of brands that differ in attribute levels
• Determining composition of the most preferred brand
• Segmenting the market based on similarity of preferences for attribute levels
Conjoint Analysis
Using Conjoint Analysis
a. As a market / choice simulator:
– compare estimated & known brand market shares.
– highest utility = highest preference
– aggregate preferences to form market share estimates
– prior vs. post aggregation.
- average responses - estimate individual part-worths
- run conjoint on averages - cluster (analysis) on basis of
part-worth values
– Logit analysis ===> greater detail, but, ratio data needed
Conjoint Analysis
Using Conjoint Analysis
b. Independence of attributes -- handling dependence
i. interaction terms
ii. multiplicative (not additive) model.
c. Perceptual Mapping / Joint Space vs. Conjoint Analysis
– Conjoint stimuli are attribute levels not products or brands
– Conjoint Compliments * PM/JS Analysis
• PM/JS link perception to preference of product benefits
• Conjoint links features to preference and perception
– Engineering needs => attribute level importance
Conjoint Analysis
Using Conjoint Analysis
d. Limitations
• assumes important attributes can be identified
– may not consider brand name or image variables
• assumes consumers evaluate choice alternatives based on these attributes
• assumes consumers make tradeoffs (compensatory model)
• tradeoff model may not respresent choice process (non-compensatory models?
• data collection can be difficult and complex