how to avoid some of the pitfalls when deploying legacy targeting models david dipple - adroit...
DESCRIPTION
TRANSCRIPT
How to avoid some of the pitfalls when deploying legacy targeting models
David Dipple
Introduction
Fellow of Royal Statistical Society Worked with Not For Profit and Charity Clients for over
25 years Recognised as an expert data modeller Trained numerous analysts and fundraisers in the use
of analysis in fundraising Worked with charities in UK and mainland Europe
David Dipple
Words from the wise
An approximate answer to the right question is worth a great deal more than the precise answer to the wrong question.
-The first golden rule to applied mathematics
The formulation of a problem is often more essential than its solution which may be merely a matter of mathematical or mental skill.
•A. Einstein
Processes
The analysis process?
Question
Answer
Gubbins
Process Flows
The only point where there is interaction is at the start – no time is allocated for re-visiting the question
The Analysis Processes
Results workshop
Answer
Marketing Brief
Question Initial Brief Analysis Brief
Analysis reqs
Initial Analysis
Initial results
Final analysis
Needs, wants and requirements
Marketing
Analysis
Forget complex relationships – simplicity is your friend Analysis follows the 80/20 rule◦ 80% of the analysis can be done in 20% of the time.◦ The last 20% takes 80% of the time
Looking for the perfect answer?
Who?
HealthNature
Humanity
Environment
Wildlife
Disability
Cancer & Medical Research
Religion
Inequality
3rd Word & Overseas
Animal Welfare
Binary Clustering: Charity Sector
Our Target?
Or
Traditionally many legacy campaign have been designed and devised around a message they are not shaped around supporters needs and requirements
To fully tap the legacy potential of the base a more supporter lead strategy would match supporter interests and propensity to legacy message
Campaign vrs Supporter Lead
Method◦ Mail◦ Phone◦ Event◦ Online
The halo effect
Getting the Message Across
Data
The Data Triangle
Questionnaires, Interests & Beliefs
Lifestage, Age, GenderGeodems
Segmentation
Recency, Frequency,Value, Forms of help.
AttitudinalDemographic
Behavioural
Supporter Information
Donor &Demographic
DetailsDatabaseDerived
Donations
Communications
Attitu
dinal
LTVsRFVsScores
Media codesResponsesMethod
InterestsLifestyleCause
NameAddressGenderAgeIncome
Payment TypeAmountDate
Donor Information
Donor & Demographic
Details
DatabaseDerivedDonations
CommunicationsAttitudinal
Geo-Dems are great for cold and certain aspects of warm targeting
For small population analysis they tend to be less useful◦ For one model that I created by using a geo-dem it added
0.5% to the power of the model Take care with including or excluding people based on
their geo-dem coding
Geo-Dems
Do they mean me? I think that they do!
Academic Centres, Students and Young Professionals
Retired - Low income - Aged in the City Suburbs
Acorn Description
PersonicxDescription
People tend to be interested in people◦ But why are they interested?◦What aspects of your cause excites them?◦What motivates them to give you money?
Interests
What data do we currently have?◦What is its quality
What data would we like to have?◦What barriers are there to getting it?
So what we do not not know?
Modelling & Analysis
But what type of model?◦ Legacy◦ Pledger◦ Legacy & Pledger◦ Residuary/Pecuniary
The past determines the future◦ Lifetime Model◦ Time Limited Model◦ Something Else
But We Need a Propensity Model!
SPSS Excel FastStats MapInfo & MapPoint My own software
My Analysis Toolkit
Modelling techniques◦Binary Logistic◦Discriminant◦Multinomial Logistic◦CHAID◦Proxy
Legacy Models
Type of Data◦Number of Relationships◦Supporter Lifetime◦Number of Gifts◦Age of Supporter◦Gift Aider
Time is not our friend!
Important Legacy Factors
Gender Response Age ResponseMale 8% Young 12%Female 10% Old 12%
PopulationResponse: 10%
Gender: MaleResponse: 8%
Gender: FemaleResponse: 12%
Age: YoungResponse: 15%
Age: OldResponse: 5%
Age: YoungResponse: 10%
Age: OldResponse: 16%
Beware of False Relationships
Understanding models: Basic outputClassification Tablec
776 134 85.3 908940 153597 85.5
173 725 80.7 83 272 76.6
83.0 85.5
Observed0
1
Legator
Overall Percentage
Step 10 1
Legator PercentageCorrect
Selected Casesa
0 1
Legator PercentageCorrect
Unselected Casesb
Predicted
Selected cases sel_var EQ 1a.
Unselected cases sel_var NE 1b.
The cut value is .500c.
Multiple ways of understanding if a model has worked. Most of the output can be ignored by non statisticians and the key – The key is finding what needs to be communicated to marketers and in what form. used to determine power.
Tom Smith
Testing Down the Model
High ScoreSelected
Supporters
Model Score
Even with a small population outcome models – test down the model to reduce the Tom Smith effect.
Building legacy models has so far been carried out by building statistical propensity models. These need previous results to determine what will happen.
But if there are no previous results you can’t build a model or can you?
No Legacy Info – don’t worry!
The factors that increase propensity to make a pledge or leave a legacy are fairly well know – as we saw earlier
Create binary flags for each of the data items given earlier and then add them all up. The higher the result, the more likely to make a pledge (and it works).
Proxy Models
Analysis of a legacy campaign tends to be point based, That is how many responded to being contacted
To truly understand the effect of legacy campaigning the relationship over time needs to be examined, including the effect on non legacy messages – that is the full supporter journey
Longitudinal Analysis
Going Forward
The Future?
Warehouse
Model
Message 1 Message 2 Message 3 Message 4
No Contact (at this point…)
Single model that determines both who should be contacted and with what message.
The biggest barrier to producing efficient models is lack of data – especially demographic and attitudinal data
Understand what the data is saying and then use an appropriate model - There is no one perfect solution
There is no certainty in modelling – models are built from past behaviour and if you change what you are doing it can take a while for the data to catch up
Examine the whole supporter journey to understand the full relationship
Define the question and the answer will be much easier – remember a model is not a panacea
Conclusions
Thank You For Listening