2. Very powerful: tells you not only who but also when
Who is more likely to die, kill, get cured, go bankrupt,
attrite, drop spend, catch a cold, etc etc and when:
sort of like astrology
The who part can also be answered by other methods:
logistic regression, and a host of segmentation
techniques comes to mind
The when part is the most attractive part about
survival (or astrology). When is a person likely to
attrite, when is he likely to die, or experience an
accident (or get a job, get married, go abroad:
astrology)
3. One bad thing about survival analysis. It isnt as clear about
the when as an astrologer. For every event you can think
of, a survival analysis model will give you a host of
probabilities. But you’ll have to do the interpretation part
so:
A doctor asks when should i give the treatment to this flu
patient?
An actuary asks what premium should I set for this guy who
wants accident insurance?
A telecom company asks, who out of my prepaid customers is
likely to walk out and when?
What do you tell them? They dont wanto look at you host
of probabilities. They are paying you to do that
4. 3 outputs from survival analysis we would consider
The chances of surviving till particular period
▪ Prob that an accident victim would not die 3 days from today
The expected lifetime of an individual or a group
▪ The period when the survival probability of the indv becomes 50%
▪ The time when 50% of the group would have had the accident in 6
years
The chances of the event occurring in a particular interval
given that it has not occurred till before that
▪ Given that a telecom customer has not attrited till the third month,
what are his chances of attriting between periods 3 and 6
5. The most detailed information that you can
provide is a table of expected time to event
for every possible individual profile and all
events of interest
6. event sneezing throat pain fatigue diarhoaea death
age
0-20
30
21-40
41-60
6
61-80
81-100
This means 50% of the patients in the 61-80 age group
are likely to have throat pain in 6 days from the
inception of the disease compared to 21-40 age group
people.
So the doc better treat the older people first
7. event death
age
0-20
30
21-40
41-60
10
61-80
81-100
This means 50% of the patients in the 61-80 age group are
likely to die in 10 years compared to the 21-40 age group
people, half of whom are likely to live for 30 more years
So the actuary would charge higher premium for selling life
insurance to the seniors since their chances of dying are
much higher and sooner
8. event attrition
spend
0-20
21-40
41-60
10
61-80
30
81-100
2 of the highest spend groups have dramatically different median survival times
This means 50% of the customers in the 61-80 spend group are likely to stop
renewing in 10 months compared to the 81-100 spend group people, half of
whom are likely to continue renewing for 30 more months
So the marketer would provide more freebies like free roaming, lower call
charges, free ringtones, etc to the 61-80 group and before 10 months. For the 81-
100 group, it can wait for 15-20 more months
Once the treatment is given, it can be measured whether it was effective in
extending the tenure
9. The treatment to be given to prevent or takle
advantage of the event is out of the scope of survival
analysis
Survival analysis can be used to measure the effect of
these treatments too, once they are given
Survival analysis models would typically produce a
much more detailed profiling rather than only one
variable like age, spend etc.
Be careful where the expected lifetime is way into the
future, the assumption is that whatever was
happening in the modelling period would continue;
maybe true for medical data, but usually untrue for all
others