Breakfast seminar: The Business Value of Survival Analysis
Evi Nagler Methodologist - European Renal Best Practice Renal Unit, Ghent University Hospital Veerle Liébaut Consultant – 4C Consulting Wannes Rosius, Client Technical Professional - IBM SPSS
01 Introduction
Agenda
02 Survival analysis: origin and possible application
03 IBM SPSS modeler and Survival Analysis
04 Closing remarks
05 Q&A
4C Consulting | Our Customers
3
Marketing Excellence
Customer Experience Management
Customer Insight Management
Experience Identity | Customer Journeys | Moments of Truth | Cross-channel | Unique Customer View | CRM Roadmap | Cultural Change
About 4C Consulting | Our service portfolio
Sales Excellence Service Excellence
• Marketing Maturity Assessment
• Campaign Management & Automation
• Campaign Management Outsourcing
• Marketing Resource Management
• SFA Management & Automation
• Sales Portfolio Management
• Sales Middle Office
• Training & Coaching
• Customer Service Automation
• Self Service Strategy & Management
• Complaints Handling
Performance Management | Data Quality | Data Mining I Segmentation | Scoring | Profiling | Forecasting | Predictive Analytics
4
5
BI / CI Services
Customers
How to optimize your
service to answer their
needs
Who are they / What do they want
How to communicate
with them
DATA
01 Introduction
Agenda
02 Survival analysis: origin and possible application
03 IBM SPSS modeler and Survival Analysis
04 Closing remarks
05 Q&A
1 Introductory example
Attention:
• Preferably use 4CC library images (see 4CPedia or ask Priskilla)
• If not: make sure that image resolution is sufficiently high! No blurry pictures
• Text bar should be moved for optimal
position in picture
• Text color is white or black in function of
contrast needs
• Font = calibri, 44pt, regular
• Alignment: default = left
The origin
How to treat?
Metastatic cancer
Classic treatment
30% alive
New treatment
50% alive AFTER 12 MONTHS
9
Metastatic cancer
Classic treatment
20% alive
New treatment
21% alive
How to treat?
AFTER 16 MONTHS
10
Attention:
• Preferably use 4CC library images (see 4CPedia or ask Priskilla)
• If not: make sure that image resolution is sufficiently high! No blurry pictures
• Text bar should be moved for optimal
position in picture
• Text color is white or black in function of
contrast needs
• Font = calibri, 44pt, regular
• Alignment: default = left
Time is Crucial
Interesting, but how can we use this?
Attention:
• Preferably use 4CC library images (see 4CPedia or ask Priskilla)
• If not: make sure that image resolution is sufficiently high! No blurry pictures
• Text bar should be moved for optimal
position in picture
• Text color is white or black in function of
contrast needs
• Font = calibri, 44pt, regular
• Alignment: default = left
An Application
Attention:
• Preferably use 4CC library images (see 4CPedia or ask Priskilla)
• If not: make sure that image resolution is sufficiently high! No blurry pictures
• Text bar should be moved for optimal
position in picture
• Text color is white or black in function of
contrast needs
• Font = calibri, 44pt, regular
• Alignment: default = left
Customer Churn
How to treat?
Churn
Classic marketing program
30% stays
New marketing program
50% stays AFTER 12 MONTHS
15
Churn
Classic marketing program
20% stays
New marketing program
21% stays
How to treat?
AFTER 16 MONTHS
16
Attention:
• Preferably use 4CC library images (see 4CPedia or ask Priskilla)
• If not: make sure that image resolution is sufficiently high! No blurry pictures
• Text bar should be moved for optimal
position in picture
• Text color is white or black in function of
contrast needs
• Font = calibri, 44pt, regular
• Alignment: default = left
Time is Money
2 The idea
Survival curves | Customer Example
New marketing program
Classic marketing program
Event=churn
Time (months)
Survival probability
19
Survival curves | Traditional Example
New treatment
Classic treatment
Event=death
Time (months)
Survival probability
Median survival time: 9.6 versus 8 months
Douillard JY et al. J Clin Oncol 2010; 28 (31): 4697-4705
20
Added value | Entire Sample
21
Start of study End of study
Time (months)
=event occurs
=enter the study
2 4 6 8 10 12 0
Added value | Entire Sample
22
Start of study End of study
Time (months)
=event occurs
=enter the study
2 4 6 8 10 12 0
Added value | Entire Sample
23
Start of study End of study
Time (months)
=event occurs
=enter the study
2 4 6 8 10 12 0
Added value | Entire Sample
24
Time (months)
=event occurs
2 4 6 8 10 12 0
Time in study =censored
An individual censored at time t should have the same survival
chance as all subject who survive up to time t
Condition | Non-informative censoring
25
Hospital A Hospital B
Condition | Non-informative censoring
26
Hospital A Hospital B
Time (months)
2 4 6 8 10 12 0 Time (months)
2 4 6 8 10 12 0
Condition | Non-informative censoring
27
Hospital A Hospital B
Condition | Non-informative censoring
28
Hospital A Hospital B
Time (months)
2 4 6 8 10 12 0 Time (months)
2 4 6 8 10 12 0
Condition | Non-informative censoring
29
Compare 2 loyalty programs
For who: valuable customers (gold status)
30
Condition | Non-informative censoring
Condition | Non-informative censoring
31
Bank A Bank B
Condition | Non-informative censoring
32
33
How is your data collected?
Which customers are included?
Comparing survival curves
Treatment A
Treatment B
Survival probability
34
Comparing survival curves
New marketing program Classic marketing program
Event=churn
Time (months)
Survival probability
35
Randomised trial All patients Treatment A
Treatment B
Follow-up
Follow-up
Compare results
RANDOM
36
Observational study All patients Treatment A
Treatment B
Follow-up
Follow-up
Compare results
CHOICE
37
Observational study All patients Campaign A
Campaign B
Follow-up
Follow-up
Compare results
CHOICE
Business setting
We need to adjust for
confounders
38
3 Modelling
39
0%
20%
40%
60%
80%
100%
0 1 2 3 4 5 6 7 8 9 10 11 12
Definitions
S(t)=Survival curve F(t)=Cumulative Incidence
Time (months)
40
Definitions
41
0%
5%
10%
15%
20%
25%
30%
0 1 2 3 4 5 6 7 8 9 10 11
Incidence Hazard
Definitions
Time Survival Curve Cumulative incidence Incidence Hazard
0 100% 0% 20% 20%
1 80% 20% 20% 25%
2 60% 40% 10% 17%
3 50% 50%
42
Definitions
Time Survival Curve Cumulative incidence Incidence Hazard
0 100% 0% 20% 20%
1 80% 20% 20% 25%
2 60% 40% 10% 17%
3 50% 50%
43
Definitions
44
0%
5%
10%
15%
20%
25%
30%
0 1 2 3 4 5 6 7 8 9 10 11
Incidence Hazard
Cox proportional hazards model
Most common used model for survival data (*)
Flexible choice of covariates
Fairly easy to model
Standard software exists
Well developed elegant mathematical theory
Few distributional assumptions
Non informative censoring
Proportional hazards
Independence
45
(*)Goetghebeur E and Van Rompaye B. Survival analysis edition 2011
Cox proportional hazard model
𝜆 𝑡, 𝒁 = 𝜆0 𝑡 𝑒𝑥𝑝 𝛽1𝑍1 + 𝛽2𝑍2+…+𝛽𝑝𝑍𝑝
𝜆0=baseline hazard
𝑍1, 𝑍2,… , 𝑍𝑝= covariates
46
Cox proportional hazard model
𝜆 𝑡, 𝒁 = 𝜆0 𝑡 𝑒𝑥𝑝(𝛽1𝑍1)
𝜆0=baseline hazard
𝑍1 = 0 = 𝑛𝑜 𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡
1 = 𝑡𝑟𝑒𝑎𝑡𝑚𝑒𝑛𝑡
𝛽1=-0.7 exp(𝛽1)=0.5
47
Take home messages
Classic regression ignores time – time is crucial
Solution: survival analysis
Advantages
Use of entire sample Instantaneous risk estimation
Conditions
Non informative censoring Proportional hazards Independence
48
01 Introduction
Agenda
02 Survival analysis: origin and possible application
03 IBM SPSS modeler and Survival Analysis
04 Closing remarks
05 Q&A
Questions?
Let’s have coffee first
50