Business Intelligence & Data Mining-2

8/10/2019 Business Intelligence & Data Mining-2

1/52

Conjoint Analysis


2/52

Different Perspectives, Different Goals

Buyers want all of the most desirable features atlowest possible price

Sellers want to maximize profits by:

1) minimizing costs of providing features2) providing products that offer greater overall value than thecompetition


3/52

Demand Side of Equation

Typical market research role is to focus first ondemand side of the equation

After figuring out what buyers want, next assesswhether it can be built/provided in a cost-effective manner


4/52

Products/Services are Composed of

Features/Attributes Credit Card:

Brand + Interest Rate + Annual Fee + Credit Limit

On-Line Brokerage:

Brand + Fee + Speed of Transaction + Reliability of

Transaction + Research/Charting Options


5/52

Breaking the Problem Down

If we can find out how buyers value the

components of a product, we will be in a better position to design those that improve profitability


6/52

How to find out What Customers Want?

Ask Direct Questions about preference:

What brand do you prefer? What Interest Rate would you like? What Annual Fee would you like? What Credit Limit would you like?

Answers often trivial / obvious and unenlightening

(e.g. respondents prefer low fees to high fees, highcredit limits to low credit limits)


7/52


8/52

Stated Importances

Importance Ratings often have low discrimination:

Average Importance Ratings

7.5

8.1

7.2

6.7

0 5 10

Credit Limit

Annual Fee

Interest Rate

Brand


9/52

Stated Importances

Answers often have low discrimination, with mostanswers falling in very important categories

Answers sometimes useful for segmenting themarket, but still not as actionable as could be


10/52

What is Conjoint Analysis?

Research technique developed in early 70s

Measures how buyers value components of a

product/service bundle

Dictionary definition-- Conjoint: Joined together,combined.

Marketers term -- Features CONsidered JOINTly


11/52

How Does Conjoint Analysis Work?

We vary the product features (independent variables) to buildmany (usually 12 or more) product concepts (Bundle of features)

We ask respondents to rate/rank those product concepts(dependent variable)

Based on the respondents evaluations of the product concepts,we figure out how much unique value (utility) each of thefeatures added

Regress dependent variable on independent variables (the betasequal part worth utilities.)


12/52

Whats So Good about Conjoint?

More realistic questions:

Would you prefer . . .

210 Horsepower or 140 Horsepower 17 MPG 28 MPG

If choose left, you prefer Power. If choose right, you prefer Fuel Economy

Rather than ask directly whether you prefer Power overFuel Economy, we present realistic tradeoff scenarios andinfer preferences from your product choices


13/52

Tradeoffs

When respondents are forced to make difficult

tradeoffs, we learn what they truly value


14/52

First Step: Create Attribute List

Attributes assumed to be independent (Brand,Speed, Color, Price, etc.)

Each attribute has varying degrees, or levels

Brand: Coke, Pepsi, Sprite Speed: 5 pages per minute, 10 pages per minute Color: Red, Blue, Green, Black

Each level is assumed to be mutually exclusive of theothers (a product has one and only one level among the possible levels of that attribute)


15/52

Rules for FormulatingAttribute Levels

Levels are assumed to be mutually exclusive

Attribute: Add-on features

level 1: Sunroof level 2: GPS Systemlevel 3: Video Screen

If define levels in this way, you cannot determine thevalue of providing two or three of these features at thesame time


16/52


Levels should have concrete/unambiguousmeaning

Very expensive vs. Costs $575

Weight: 5 to 7 kilos vs. Weight 6 kilos

One description leaves meaning up to individualinterpretation, while the other does not


17/52


Dont include too many levels for any oneattribute

The usual number is about 3 to 5 levels per attribute The temptation is to include many, many levels of price, so we can

estimate peoples preferences for each But, you spread your precious observations across more

parameters to be estimated, resulting in noisier (less precise)measurement of ALL price levels

Better approach usually is to interpolate between fewer more precisely measured levels for not asked about prices


18/52


Whenever possible, try to balance the number of levelsacross attributes

There is a well-known bias in conjoint analysis called theNumber of Levels Effect

Holding all else constant, attributes defined on more levels thanothers will be biased upwards in importance

For example, price defined as ($10, $12, $14, $16, $18, $20) willreceive higher relative importance than when defined as ($10, $15,

$20) even though the same range was measured The Number of Levels effect holds for quantitative (e.g. price,

speed) and categorical (e.g. brand, color) attributes


19/52


Make sure levels from your attributes can combine freelywith one another without resulting in utterly impossiblecombinations (very unlikely combinations are OK)

Resist temptation to make attribute prohibitions (prohibiting levelsfrom one attribute from occurring with levels from otherattributes)!

Respondents can imagine many possibilities (and evaluate themconsistently) though the organization conducting the study doesnt

plan to/cant offer some of the combinations. By avoiding prohibitions, we usually improve the estimates of the combinationsthat we will actually focus on.


20/52

Conjoint Analysis Output

Utilities (part worths)

Importances

Market simulations


21/52

Conjoint Utilities (Part Worths)

Numeric values that reflect how desirable differentfeatures are:

Feature & Level UtilityVanilla 2.5

Chocolate 1.8

25 5.335 3.250 1.4

The higher the utility, the more desirable for the buyer


22/52

Conjoint Importances

Measure of how much influence each attribute has on buyers choices

Best minus worst level of each attribute, percentaged:

Vanilla - Chocolate (2.5 - 1.8) = 0.7 15.2%25 - 50 (5.3 - 1.4) = 3.9 84.8%

----- --------Totals: 4.6 100.0%


23/52

Market Simulations

Make competitive market scenarios and predict which products respondents would choose

Accumulate (aggregate) respondent predictions to makeShares of Preference (some refer to them as marketshares)


24/52

Market Simulation Example

Predict market shares for 35 Vanilla cone vs. 25Chocolate cone for Respondent #1:

Vanilla (2.5) + 35 (3.2) = 5.7Chocolate (1.8) + 25 (5.3) = 7.1

Respondent #1 chooses 25 Chocolate cone!

Repeat for rest of the respondents. . .


25/52

Market Simulation Results

Predict responses for 500 respondents, and we might seeshares of preference like:

65% of respondents prefer the 25 Chocolate cone

35%

65%

Vanilla @ 35

Chocolate @ 25


26/52

Conjoint Market Simulation Assumptions

All attributes that affect buyer choices in the real worldhave been accounted for

Equal availability (distribution)

Respondents are aware of all products

Long-range equilibrium (equal time on market)

Equal effectiveness of sales force

No out-of-stock conditions


27/52

Shares of Preference Dont Always Match

Actual Market Shares Conjoint simulator assumptions usually dont hold true in

the real world

But this doesnt mean that conjoint simulators are notvaluable!

Simulators turn esoteric utilities into concrete shares

Conjoint simulators predict respondents interest in products/services assuming a level playing field


28/52

Value of Conjoint SimulatorsSome Examples

Lets you play what-if games to investigate the value ofmodifications to an existing product

Lets you estimate how to design new products to maximize buyer interest at low manufacturing cost

Lets you investigate product line extensions: do wecannibalize our own share or take mostly fromcompetitors?

Lets you estimate demand curves, and cross-elasticitycurves

Can provide an important input into demand forecastingmodels (for various price levels)


29/52

Three Main Flavors of ConjointAnalysis

Traditional Full-Profile Conjoint

Adaptive Conjoint Analysis (ACA)

Choice-Based Conjoint (CBC), also known asDiscrete Choice Modeling (DCM)


30/52

Strengths of Traditional Conjoint

Good for both product design and pricing issues

Can be administered on paper or computer/internet

Shows products in full-profile, which many arguemimics real-world

Can be used even with very small sample sizes


31/52

Weaknesses of Traditional Full-Profile

Conjoint Limited ability to study large number of attributes

(more than about six)

Limited ability to measure interactions and otherhigher-order effects (cross-effects)


32/52

Traditional Conjoint(Six Attributes)

Using a 100-pt scale where 0 means definitely

would NOT and 100 means definitely WOULD

How likely are you to purchase

1997 Honda AccordAutomatic transmission

No antilock brakesDriver and passenger airbag

Blue exterior/Black interior $18,900

Your Answer:___________


33/52

Six Attributes: Challenging

Respondents find six attributes in full-profilechallenging

Need to read a lot of information to evaluate each card Each respondent typically needs to evaluate around 24-

36 cards

T diti l C j i t C d S t M th d


34/52

Traditional Conjoint: Card-Sort Method(15 Attributes)

Using a 100-pt scale where 0 means definitely would

NOT and 100 means definitely WOULD

How likely are you to purchase

1997 Honda AccordAutomatic transmission

No antilock brakesDriver and passenger airbagBlue exterior/Black interior

50,000 mile warrantyLeather seats

optional trim package3-year loan

5.9% APR financingCD-player

No cruise controlPower windows/locksRemote alarm system

$18,900

Your Answer:___ ________


35/52

15 Attributes: Near Impossible

Faced with so much reading, respondents are forcedto simplify (focus on just the top few attributes inimportance)

To get good individual-level results, respondentsneed to evaluate around 60-90 cards


36/52

Adaptive Conjoint Analysis

Developed in 80s by Rich Johnson, Sawtooth Software

Devised as way to study more attributes than was prudent

with traditional full-profile conjoint

Adapts to the respondent, focusing on most importantattributes and most relevant levels

Shows only a few attributes at a time (partial profile) ratherthan all attributes at a time (full-profile)


37/52

Steps in ACA Survey (1)

Self-Explicated Priors Section Preference Ratings for the levels of any attributes that

we do not know ahead of time (e.g. brand, color).


38/52


Self-Explicated Priors Section Show best and worst levels of each attribute, and ask

respondents how important the difference is.


39/52


Conjoint Pairs / trade-offs (show only two tofive attributes at a time)


40/52


41/52

Adaptive Conjoint Analysis Example

Sample ACA survey:http://www.sawtoothsoftware.com/support/technical-support/104-support/sample-surveys/77-aca-sample-study


42/52

Strengths of ACA

Ability to measure many attributes, withoutwearing out respondent

Respondents find interview more interesting and

engaging

Efficient interview: high ratio of informationgained per respondent effort

Can be used even with very small sample sizes


43/52

Weaknesses of ACA

Partial-profile presentation less realistic than realworld Respondents may not be able to assume attributes not

shown are held constant

Often not good at pricing research Tends to understate importance of price

Must be computer-administered (PC or Web)


44/52

Choice-Based Conjoint (CBC)

Became popular starting in early 90s

Respondents are shown sets of cards and asked to

choose which one they would buy

Can include None of the above response

Ch i B d C j i Q i


45/52

Choice-Based Conjoint Question


46/52

Strengths of CBC

Questions closely mimic what buyers do in real world:choose from available products

Can investigate interactions, alternative-specific effects

Can include None alternative

Paper or Computer/Web based interviews possible

k f


47/52

Weaknesses of CBC

Usually requires larger sample sizes than with Full Profileor ACA

Tasks are more complex, so respondents can process fewerattributes (CBC recommended


48/52

The basic conjoint analysis model can be represented by the

following formula:

where

U(X) = overall utility of an alternative= the part-worth contribution or utility associated with

the j th level ( j, j = 1, 2, . . . k i) of the i th attribute(i , i = 1, 2, . . . m)

x jj = 1 if the j th level of the i th attribute is present= 0 otherwise

k i = number of levels of attribute im = number of attributes

Conjoint Analysis Model

xij j

ij

m

i

k X U

i

==

=11

)(

ij

Importance of Attributes


49/52

The importance of an attribute, I i is defined in terms of the

range of the part-worths, across the levels of thatattribute:

The attribute's importance is normalized to ascertain itsimportance relative to other attributes, W i :

So that

i

=

=m

ii

ii

I I W

1

11

==

m

iiW

Importance of Attributes

H h M d l W k


50/52

How the Model Works

Conjoint analysis is run as a regression model

Dependent variable Ratings or rankings

Aim of the model estimation: is to estimate the part-worth of every attribute level.

Methods: ANOVA, Dummy Variable Regression,Logit model and Probit model.

The regression coefficients provide the part utility ofeach level of attributes.

From that we can find range of utility and relativeimportance of attributes.


51/52

Assessing reliability and validity

Goodness-of-fit of the model (if dummy-variable regression R Square)

Test-retest reliability repeat interview

Holdout/ validation internal validity

ld d


52/52

Holdout Cards

Rated by subjects but are not included in the conjointanalysis for coefficient / utilities estimations, or

Generated from another random plan, (not the maineffects experimental plan) without any duplication.

Conjoint computes correlations between the observedand predicted rank orders for these cards as a checkon the validity of the utilities.

It can be expected that holdouts will always havelower correlation .

Business Intelligence & Data Mining-2

Documents