segmentation frankwyman slides

7/28/2019 Segmentation FrankWyman Slides

1/52

Best Segmentation Practices andBest Segmentation Practices andTargeting Procedures that ProvideTargeting Procedures that Provide

the most Clientthe most Client -- Actionable Strategy Actionable Strategy

Frank Wyman, Ph.D.Frank Wyman, Ph.D.Director of Advanced AnalyticsDirector of Advanced AnalyticsM/A/R/CM/A/R/C ResearchResearch


2/52

Segmentation DefinedSegmentation Defined

Segmentation : T he dividing of a marketscustomers into subgroups in a way that optimizesthe firms ability to profit from the fact thatcustomers have different needs, priorities, andeconomic levers.


3/52

Keep in mind the end goal of

enhancing profitability , as this canhelp increase the actionability of the segmentation. At each step

ask how can these results helpimprove profits?

Best Practice #1Best Practice #1Best Practice #1


4/52

Product Positioning ResearchProduct Positioning Research

Value AddsValue Adds Key DriversKey Drivers

Chewy

Sweet

Rich

Chocolaty

Hearty

Crunchy

Peanutty

Long lasting

Inexpensive

Salty

Low YieldLow Yield Entry TicketsEntry Tickets

D e r i v e d

I m p o r t a n c e

Low

High

Performance

Well AboveCompetitors

Slightly Above

Slightly Below

Well BelowCompetitors

Low HighStated Importance


5/52

Conduct product positioningresearch around the same timeas segmentation to improve the

interpretability and enhance theactionability of the segmentation.



6/52

Three basic approachesThree basic approachesto segmentationto segmentation

Analytic AnalyticSegmentationSegmentation

InterdependenceInterdependence

(clustering)(clustering)

DependenceDependence

(CHAID)(CHAID)

NonNon -- Analytic Analytic(convenience)(convenience)SegmentationSegmentation

SegmentationSegmentation


7/52






(CHAID)(CHAID)




8/52

Choose to take an analyticapproach to segmentation.



9/52






(CHAID)(CHAID)




10/52


Strat egy => In ter dep end en ce (clu ste ring )

Ta ctics = > Dep en den ce (CH AID)

C id i i d idi bC id i i d idi b


11/52

Considerations in deciding between anConsiderations in deciding between aninterdependence or dependenceinterdependence or dependence

approach, or bothapproach, or bothIs a primary objective to know the entire market (what are theneeds of customers, how many needs groups exist, and how largeis each)?Is a primary objective to know how to target those who will buymy offering/product/service?

How old is the market?To what degree are customer needs being met?How old is your product (existing, new launch, line extension)?How crowded is the market?How much room for differentiation is there?How dynamic is the market (how much has the market changedsince your last segmentation)?


12/52


13/52

A general taxonomy of analytical A general taxonomy of analyticalsegmentation methodssegmentation methods

Analytic Segmentation Analytic Segmentation

InterdependenceInterdependence DependenceDependence

ClusteringClustering QQ--FactorFactor Analysis Analysis TreeingTreeing

NeuralNeuralNetwork Network

HierarchicalHierarchical K K --meansmeans 22--stepstep Latent ClassLatent Class

C&RTC&RT CHAIDCHAID QuestQuest


14/52

A general taxonomy of analytical A general taxonomy of analyticalsegmentation methodssegmentation methods

Analytic Segmentation Analytic Segmentation

InterdependenceInterdependence DependenceDependence

ClusteringClusteringDistancesDistances

QQ--FactorFactor Analysis Analysis Treeing

Treeing NeuralNeuralNetwork Network

HierarchicalHierarchical K K --meansmeans 22--stepstep Latent ClassLatent Class

C&RTC&RT CHAIDCHAID QuestQuest


15/52


Friends don t let f riends use Q-f ac tor ana lysis and Neura l netf or se gmentation!


16/52


17/52


Best all-around approaches arek-means clustering (forinterdependence) and CHAID(for dependence).

Obj i f h 2 l iObj i f h 2 l i


18/52

Objectives of the 2 analyticObjectives of the 2 analyticapproaches to segmentationapproaches to segmentation

Interdependence (clustering) What are the existing camps of needs among consumers?

1. How many segments (natural camps) are there?2. How is each defined w.r.t needs/beliefs/behaviors?3. What is the size of each segment?4. What are the markers of each segment?

5. Given their needs and my products features, which segments aregood targets ?

Dependence (CHAID) Who will buy my product and who will not?

1. How many segments are there with distinctly different propensitiestoward buying my product?

2. What are the key drivers that define the segments who will buy (andnot buy) my product?

3. How do I best reach those most likely to buy my offering?4. Given propensities and sizes, which segments are good targets?

D C id i fD t C id ti f


19/52

Data Considerations forData Considerations forSegmentation in GeneralSegmentation in General

Sample size: 400-600 minimum; 1000+ typical. Figurethat a 5% segment can hold promise and that you do not want to make inferencesfrom subgroup (segment) sample sizes of less than 30 thus, 30/5% = 600

minimum.

Representative (random) sample; no over-sampling.


20/52


For analytic segmentation,gather a large (600+),random sample.

Q ti i C id tiQ ti i C id ti


21/52

Questionnaire ConsiderationsQuestionnaire Considerationsfor Segmentation in Generalfor Segmentation in General

Include (towards end) lots of demographics, as wellas media and channel use itemsInclude at least some competitive product usage andperceived performance items


22/52

Best Practice #9Best Practice #9Best Practice #9Questionnaire should include lots of demographics, media use, and channeluse items so that you can discern howto best reach target segments. Also,including at least some items regardingcompetitive product use and perceivedperformance is a good idea.

Data Considerations forData Considerations for


23/52

Data Considerations forData Considerations forClusteringClustering

Preceded by good qualitative work (focus groups) in order tohave solid knowledge of all the needs dimensions currently ineffect in the marketShorter scale (5-point) better than longer (11-point) since,

Battery may/should be long Helps avoid differential scale use

Place distinct anchors (labels) on each point of scale to helpavoid differential scale use , e.g.,

1 totally agree, 2 somewhat agree, 3 neutral, 4 somewhat disagree, 5totally disagree

Needs statements/questions should themselves be extreme. A types are better than B:

A. I like my candy to be extremely crunchy. I absolutely love chocolate.B. I like my candy to be crunchy. I like chocolate.


24/52


Precede inter de pen de nce(clus tering ) seg menta tion w ithfresh qu alitat ive re sea rch(f ocu s g roup s).


25/52

Best Practice #11Best Practice #11Best Practice #11Slay d ifferen tial sc ale use !

Avoid th is evil b ias by design ing a ne eds batte ry base d on a short (1 -5 agr ee -disag ree ) sc ale with c lea rlydiffer entiate d anch ors on every pointand it ems w orded in the extre me.


26/52


Consider us ing con joint - or discrete-

choic e-ba sed ut ilities a s the basis o f cluste ring.


27/52

Data Cleansing for ClusteringData Cleansing for Clustering

Biggest problem in clustering is differential scale use

(response style bias).Test for differential scale use via clustering 2segments. If, across attributes, the 2 profiles are

strongly correlated and different only in general levelof response, then a differential scale use problemexists in the data and needs to be fixed.


28/52


29/52


1

2

3

4

5

Crunchy Sweet Chocolatey Peanutty Rich Chewy

N e e

d L e v e

l

Segment 1 Segment2


30/52

Best Practice #13Best Practice #13Best Practice #13Before clustering in earnest, alwaysfirst test for response style(differential scale use) bias byexamining the parallel-ness of profiles for 2-3 clusters.


31/52


Semi and full ipsatization (re-centering and re-dispersing)Outliers (ok leave them in)Missing data presents a problem so fill in. Usemean of other attributes for the respondent adjustedby a sample-wide factor for that attribute. Or if onlya few missing data points then just throw out casesvia listwise deletion.Will need to standardize (e.g. Z scores) if basisvariables arise from different scales (why I stronglyurge just one single same-scaled needs battery)


32/52


33/52


Whe n c luster ing data , o utliers ar e O K (leave em alone) missings a re no t (f ill em in).

Precede Clustering Analysis withPrecede Clustering Analysis with


34/52

Precede Clustering Analysis withPrecede Clustering Analysis withFactor AnalysisFactor Analysis

Original item Factor 1 Factor 2 Factor 3 Factor 4

Crunchy +

Sweet +

Chocolaty +

Peanutty +

Rich +

Chewy +


35/52


Cluster on ly uniqu e d imens ions; facto r a na lyze o rigina l item s t oge t t he un ique dimen sions .

D A l i i Cl iD t A l i i Cl t i


36/52

Data Analysis in ClusteringData Analysis in Clustering

1. Determine the best number of cluster segments: Via k-means, examine the decomposition of segment sizes in higher

and higher order solutions (2-12); crosstab prior-run results withnext higher-order solution. Stop once the larger segments stabilizeand/or new segments pull from too many prior segments and/orsome segments begin to substantially grow in size, then a too manysegments point has been reached.

Use hierarchical (dendrograms), latent class, and two-step clusteranalysis to help determine final answer for # of segments.

Client input.

2. Obtain final cluster solution via two runs of k-means: Use first pass results (centroids) as starting points for 2 nd , final k-

means run.

3. ANOVA of basis variable means across segments, whileopportunistic, gives a basic sense of how muchdifferentiation exists between final segments.


37/52


In clustering, use numerous

algorithms to solidify answer to Howmany segments? Once the numberis decided, obtain final cluster

segments using a two-step k-meansprocess.

D t C id ti f CHAIDData Considerations for CHAID


38/52

Data Considerations for CHAIDData Considerations for CHAID

Stated purchase intent

Exaggeration-corrected, derived purchase intent(Assessor )Conjoint/choice derived

May be something different than purchase intent: Retention Advertising response Targetability = difference in predicted usage and actual

usage

The long list of demographics, media use, andchannel use become potential drivers.


39/52


In CHAID, the dependent variableshould be as reliable and accurateas possible, usually better than a

simple stated intention item.

No special data cleansingNo special data cleansing


40/52

p gp gneeded for CHAIDneeded for CHAID

No need for factor analysis

No need for ipsatizing or standardizingOutliers usually ok

Missing data ok (as it is treated as diff value)


41/52


In CHAID, there is no real needfor heavy data cleansing.

Data Analysis in CHAIDData Analysis in CHAID


42/52

Data Analysis in CHAIDData Analysis in CHAID

Usually want to look at perhaps 3 or 4 differentlevels of solutions

Just the demographics Just the media and channel use items Some mid level (e.g., demographics and media/channel

use) Kitchen sink = everything in the questionnaire!

Typical CHAID SettingsTypical CHAID Settings


43/52

Typical CHAID SettingsTypical CHAID Settings

Drivers must be defined correctly as to level of measurement since will be treated differentlySmallest child node at 5% of N, parent at twice childGrow trees usually to around 4-5 branch levels

Alpha=.05Turn Bonferroni off


44/52


In CHAID, set the minimum size of child nodes at 5% of total N, andparent nodes at twice child node

size.


45/52


In CHAID, grow trees to a depthof about 4-5 branches with alphaset at .05 with no Bonferroni

adjustment.

CHAID Requires HumanCHAID Requires Human


46/52

Judgment OverlayJudgment Overlay

Merging

Redefining splitsCutting whole branches

Eliminating some drivers


47/52

Presentation of InterdependencePresentation of Interdependence( l )


48/52

(Cluster) Segments(Cluster) Segments

1

2

3

4

5

s a l t y

c h e w y

c r u n

c h y / p

e a n u t t y

c h o c o l a t

e y

l o n g l a s t i n

g

s w e e t / r i c h

i n e x p e n s i v

e

h e a r t y

N e e d

L e v e

l

segment 1 (40%)segment 2 (20%)

segment 3 (20%)

segment 4 (12%)

segment 5 (8%)

Presentation of DependencePresentation of Dependence( )(CHAID) S


49/52

(CHAID) Segments(CHAID) Segments

Segment Industry Company Size Position

SegmentSize

AveragePropensity

1 Real Estate ---N/A--- ---N/A--- 4% 96%

2Services other than RealEstate 0-50 employees ---N/A--- 22% 90%

3Services other than RealEstate 50+ employees Marketing Executive 8% 82%

4Services other than RealEstate 50+ employees

Not a MarketingExecutive 18% 70%

5 Non-Service Industries ---N/A--- Director 18% 60%

6 Non-Service Industries ---N/A--- Higher than director 4% 38%

7 Non-Service Industries ---N/A--- Lower than Director 26% 14%

Sum = 100%


50/52


The best way to present clustersegments is with profile linecharts; the best way to present

CHAID segments is with tables.

For more information:For more information:


51/52

For more information:For more information:

Frank WymanFrank WymanDirector of Advanced AnalyticsDirector of Advanced Analytics

M/A/R/CM/A/R/C ResearchResearch(864) 938(864) 938 --02820282

[email protected]@marcresearch.com


52/52

Copyright 2005 M/A/R/C Research

segmentation frankwyman slides

Documents