Survival Analysis
Decision Systems Group
Brigham and Womens Hospital
Harvard-MIT Division of Health Sciences and Technology
HST.951J: Medical Decision Support
Outline
Basic concepts & distributions
Survival, hazard
Parametric models
Non-parametric models
Simple models
Life-table
Product-limit
Multivariate models
Cox proportional hazard
Neural nets
What we are trying to do
Predict survival
Variable 1 Variable 2 days (or probability of survival)
Case 1
0.7 -0.2 8
Case 2
0.6 0.5 4
-0.6 0.1 2 and evaluate
0 -0.9 3
performance on
new cases
-0.4 0.4 2
-0.8 0.6 3 and determine
0.5 -0.7 4 which variables are
important
Using these
Survival function
Probability that an individual survives at least t
S(t) = P(T > t)
By definition, S(0) = 1 and S()=0
Estimated by (# survivors at t / total patients)
100
Survival
(%)
time
Unconditional failure rate
Probability density function (of death)
f(t) = lim t-> 0 P(individual dies (t,t+t))/ t
f(t) always non-negative
Area below density is 1
Estimated by
# patients dying in the interval/(total patients*interval_width)
Same as # patients dying per unit interval/total
Some other definitions
Just like S(t) is cumulative survival, F(t) is
cumulative death probability
S(t) = 1 F(t)
f(t) = - S(t)
Conditional failure rate
AKA Hazard function
h(t) = lim t-> 0 P(individual aged t dies (t,t+t))/ t
h(t) is instantaneous failure rate
Estimated by
# patients dying in the interval/(survivors at t *interval_width)
So can be estimated by
# patients dying per unit interval/survivors at t
h(t) = f(t)/S(t) h(t) = -S(t)/S(t) = -d log S(t)/dt
Parametric estimation
Example: Exponential
f(t) = e-t
S(t) = e-t
h(t) =
S(t) h(t)
1
t t
Weibull distribution
Generalization of the
exponential
h(t)
For , > 0
f(t) = (t) e-1 -t
S(t) = e -t =2 t
h(t) = (t) -1
S(t) h(t)
=1
t t
Non-Parametric estimation
Product-Limit (Kaplan-Meier)
S(ti) = (nj - dj )/ nj
S(t)
dj is the number of deaths in interval j
1 nj is the number of individuals at risk
Product is from time interval 1 to j
One interval per death time
2 4 5 8
Kaplan-Meier
Example
Deaths: 10, 37, 40, 80, 91,143, 164, 188, 188, 190,
192, 206,
Life-Tables
AKA actuarial method
S(ti) = (nj - dj )/ nj
dj is the number of deaths in interval j
nj is the number of individuals at risk
Product is from time interval 1 to j
Pre-defined intervals j are independent of death times
Kaplan-Meier
S(t) S(t)
1 1
2 2 1 2 3
Life-Table
hazard
survival
density
Simple models
Multiple strata
Multivariate models
Several strata, each defined by a set of
variable values
Could potentially go as far as one
stratum per case?
Can it do prediction for individuals?
Cox Proportional Hazards
Regression model
Can give estimate of hazard for a
particular individual relative to baseline
hazard at a particular point in time
Baseline hazard can be estimated by,
for example, by using survival from the
Kaplan-Meier method
Proportional Hazards
i = exi
where is baseline hazard and xi is covariate for patient
Cox proportional hazards
hi(t) = h0(t) e xi
Survival
exi
Si(t) = [S0(t)]
Cox Proportional Hazards
hi(t) = h0(t) e xi
New likelihood function is how we extimate
From the set of individuals at risk at time j (Rj), the probability of
picking exactly the one who died is
h0(t) e xi
m h0(t) e xm
Then likelihood function to maximize to all j is
L() = (e xi / m e xm )
Important details
Survival curves cant cross if hazards are
proportional
There is a common baseline h0, but we dont need to
know it to estimate the coefficients
We dont need to know the shape of hazard function
Cox model is commonly used to interpret importance
of covariates (amenable to variable selection
methods)
It is the most popular multivariate model for survival
Testing the proportionality assumption is difficult and
hardly ever done
Estimating survival for a
patient using the Cox model
Need to estimate the baseline
Can use parametric or non-parametric model
to estimate the baseline
Can then create a continuous survival curve
estimate for a patient
Baseline survival can be, for example:
the survival for a case in which all covariates are
set to their means
Kaplan-Meier estimate for all cases
What if the proportionality
assumption is not OK?
Survival curves may
cross
Other multivariate
models can be built
Survival at certain time
points are modeled and
combined100 A
Survival
(%)
B
time
0
Single-point models
Logistic regression
Neural nets
age
gender
blood pressure
CHD in ta
cholesterol
smoking
weight
Problems
Survival
(%)
Dependency nonmonotonic
between intervals is
curve
not modeled (no
links between
networks)
2 4
Nonmonotonic
1 3 5 6
Year
curves may appear S(1)=0.9 S(2)=0.6 S(3)=0.4 S(4)=0.3 S(5)=0.5 S(6)=0.3
How to evaluate?
patients
followed
for >1 year >2 years >3 years >4 years >5 years >6 years
input nodes: patient data
output nodes: probability of survival in a given time point
Accounting for dependencies
Survival (%)
Link networks
in some way to
account for monotonic
curve
dependencies
0 1 2 3 4 5
Year S(4)=0.2
S(3)=0.3
Output from lower network serves as
input for upper network.
Summary
Kaplan-Meier for simple descriptive
analysis
Cox Proportional for multivariate prediction
if survival curves dont cross
Other methods for multivariate survival
exist: logistic regression, neural nets,
CART, etc.
Censoring
Non-censored
Non-censored
Type I
Type II
Left
Study begin Study end
