Module 2: NP Models Flashcards
Type 1 Right Censoring definition
Event (e.g. failure such as death) is observed only if it occurs
prior to some prescribed time CR (right censoring time).
Type 1 Right Censoring example
Life Insurance PH surrenders their policy.
Type 2 Right censoring
Observations continues until a predetermined number (r) of events have occured.
Left sensoring
Death/Event occurs before the observation starts.
So we only know that T is less than the left censor point CL
Interval Censoring
T is only known to fall within an interval.
Truncation
Events within a certain period are observed.
Otherwise no infomation available.
Example of truncation
Insurance claim where deductible is not reached.
Random bounds is when
The censoring/truncation point is subject to randomness.
Ci is the
Time when the ith observation is censored.
Its a RV.
Non-informative censoring is when
it gives no information about lifetimes
Example of non-informative censoring
The end of an investigation period.
Condition for non-informative censoring
Independance of all T’s and C’s
Example of informative censoring
Those in better health are more likely to withdraw from life insurance.
Why do truncation likelihoods have a quotient
They are conditional on actually being able to observe the occurance in the first place.
N represents
population of N lives.
m represents
the amount of deaths we observe
(N-m deaths are right censored)
t12k
represent…
The ordered times of death
(Multiple can occur at the same time)
dj represents
Number of deaths that occured at time tj
cj represents
Number of lives that are right censored between [tj , tj+1)
tj1 , tj2 etc are the…
Times that obersvations are censored in the period [tj,tj+1)
nj represents
The number of lives alive at risk time ti- (just before time ti)
Discrete hazard function eqn
λj = Pr [T = tj |T ≥ tj ]
Discrete hazard function implies what about the survival function?
S(t) = 1 − F(t) = Π j:tj≤t (1 − λj)
Non-informative sensoring means
Time to sensoring is indepedent to time of death. Ie censoring time tells us nothing.
λ is
d/n for a certain j
S(t) for the KM estimator is
Π (1- d/n)
(product from j=1 to t)
Where is S(t) KM poorly defined?
beyond tk
we could either set; they all die, or all survive.
In practice find a middle ground.
Variance of KM
(1-F(t))2 * Σ d / n(n-d)
summed from j=1 to t
How to find CI of S(t) KM
S(t) +- Zquantile * sqrt(Var [S(t)])
NA estimator S(t) =
exp( -Σd/n)
sum from j=1 to t
NA estimator and KM estimator related via….
Also which is bigger ?
Taylor series approximation. See notes.
KM < NA for all t
How do we compare survival functions…
use H0 : h1(t)=h2(t)
then see if we can reject it.
Ie looking for a test statistic to be small enough to accept. <0.05
Trick with data when comparing survial funcs;
must amalgamate the data essentially then compare the sample against the group.