Why Conditional Survival Needs Its Own Calculation
A Kaplan-Meier estimate at time \(t\) describes estimated survival from the study’s starting point through \(t\). A different question is: among subjects who have remained event-free to time \(s\), what proportion are estimated to remain event-free through time \(t\)? This is conditional survival between two times. It is useful when a decision or prognosis is made at a later point in follow-up, rather than at the original study start.
The calculation is compact, but the time endpoints matter. In particular, a Kaplan-Meier estimate at \(s\) includes any failure at \(s\). Dividing by that estimate therefore conditions on being event-free after failures at \(s\), consistent with the condition \(T>s\). The event factors included in the conditional estimate are those strictly after \(s\) and no later than \(t\).
How the Ratio Cancels Earlier Survival
Suppose observed failures occur at distinct times \(t_1,t_2,\ldots\). At failure time \(t_j\), let \(Y_j\) be the number at risk just before that time and \(d_j\) the number of failures. The Kaplan-Meier estimate is the product of the observed conditional survival factors:
For \(s\le t\), the product at \(t\) contains all the factors through \(s\), followed by the factors for failures in \((s,t]\). Dividing by the product at \(s\) cancels the shared earlier factors. This proves the interval-factor form:
This cancellation is also consistent with the probability identity. Since \(T>t\) implies \(T>s\) when \(s\le t\), the event \(\{T>t\}\) is a subset of \(\{T>s\}\). Thus, when \(P(T>s)>0\),
The estimated ratio is meaningful under the usual Kaplan-Meier setting: subjects are followed from a common time origin, event times are observed or right-censored, and censoring is non-informative for the event time (commonly expressed as independence of event and censoring times). If censoring is related to event risk in a way not accounted for, the risk sets may not represent those still event-free, and the Kaplan-Meier interpretation can fail.
Use the estimate at \(t\) divided by the estimate at \(s\). The standard condition is \(T>s\), not \(T\ge s\).
Confirm \(\widehat S_{\rm KM}(s)>0\). If it is zero, the ratio is undefined and the data do not support this conditional estimate from that denominator.
Equivalently, multiply factors for event times satisfying \(s<t_j\le t\). A failure at \(t\) is included; a failure at \(s\) is not.
Worked Conditional-Survival Calculations
Worked Example: Conditional Survival Across Two Failure Times
A hypothetical device study has six subjects at risk just before time 2, with one failure at time 2. There are five at risk just before time 4, with two failures then, and three at risk just before time 6, with one failure then. Estimate survival from time \(s=2\) through \(t=6\), conditional on being event-free at time 2.
Method. Calculate the Kaplan-Meier estimates at both endpoints and take their ratio. The failures at times 4 and 6 are in the interval; the failure at time 2 is not, because the condition is \(T>2\).
At time 2, the survival factor is \(1-1/6=5/6\), so
At time 4, the factor is \(1-2/5=3/5\). At time 6, the factor is \(1-1/3=2/3\). Therefore,
The endpoint ratio is
As a check, multiplying only the interval factors gives \((3/5)(2/3)=2/5\), the same result. The estimated conditional survival is 0.4000, or 40.00%. The corresponding estimated conditional probability of failure by time 6, given \(T>2\), is \(1-0.4000=0.6000\).
Common mistake avoided: Dividing by survival just before time 2 would include the failure at time 2 in the interval calculation. That would answer a different question, about subjects event-free immediately before time 2.
Worked Example: Tied Censoring and Endpoint Failures
In a hypothetical follow-up study, the risk set immediately before time 1 has \(Y_1=8\), with \(d_1=2\) failures and one censoring at that time. At time 3, \(Y_3=5\), with \(d_3=1\) failure and one censoring. At time 5, \(Y_5=3\), with \(d_5=1\) failure. Estimate conditional survival from time 1 to time 5, given event-free status at time 1.
Under the usual tied-time convention, the risk set at a failure time is counted just before that time; censorings recorded at that time do not remove subjects from the risk set used for that failure factor. The time-1 factor is \(1-2/8=6/8=3/4\), so
The conditional interval includes failures strictly after 1, namely the failures at 3 and 5. Their factors are \(1-1/5=4/5\) and \(1-1/3=2/3\). Thus the interval-factor estimate is
to four decimal places. To check by the endpoint ratio, first calculate
Then
Both methods agree. The failure at time 5 is included because \(t=5\); the failure at time 1 is excluded because \(s=1\) and the condition is \(T>1\). The tied censorings affect subsequent risk sets, but do not change the failure factors just calculated.
Worked Example: Comparing Conditional and Overall Survival
A hypothetical maintenance study records failures at time 2 with \(Y_2=10,d_2=2\), at time 4 with \(Y_4=7,d_4=1\), and at time 7 with \(Y_7=5,d_7=2\). Censoring between failure times accounts for the reductions in the later risk sets. Calculate (a) conditional survival from time 4 to time 7 and (b) conditional survival from time 2 to time 7.
(a) From 4 to 7. The only failure time strictly after 4 and no later than 7 is time 7. Its factor is \(1-2/5=3/5\). Thus conditional survival is
For an endpoint check, \(\widehat S_{\rm KM}(4)=(1-2/10)(1-1/7)=(4/5)(6/7)=24/35\). Also,
Therefore, \((72/175)/(24/35)=(72/175)(35/24)=3/5\), as required.
(b) From 2 to 7. Both failures after time 2, at 4 and 7, contribute. Their factors are \(6/7\) and \(3/5\), so
The time-2 failure factor is not included. As a check, \(\widehat S_{\rm KM}(2)=4/5\), and \[ \frac{72/175}{4/5} =\frac{72}{175}\cdot\frac{5}{4} =\frac{18}{35}. \] The conditional estimate from time 2 is lower than the estimate from time 4 because the interval from 2 to 7 includes the additional failure drop at time 4. This is an interval comparison, not a claim that the subjects at the two conditioning times are the same observed individuals.
Common mistake avoided: Using only the factor at time 7 for part (b) would give 0.6000, which answers part (a), not part (b). Always list the event times in \((s,t]\) before selecting factors.
Endpoint Conventions and Interpretation
The Kaplan-Meier curve is a right-continuous step function: at an observed failure time, its value has already dropped. Consequently, \(\widehat S_{\rm KM}(s)\) includes failures at \(s\), while the interval product for conditional survival begins strictly after \(s\). If the question instead conditions on being event-free immediately before \(s\), the denominator would need to be the left limit \(\widehat S_{\rm KM}(s-)\), and a failure at \(s\) would be included. These are distinct conditioning statements; do not switch between them silently.
At the right endpoint, a failure at \(t\) is included in the estimate of survival through \(t\), because survival is defined using \(T>t\). A censoring at \(t\) is not a failure and does not produce a Kaplan-Meier drop. The censoring convention for tied event times must match the risk-set definition being used.
A positive denominator is essential. If \(\widehat S_{\rm KM}(s)=0\), the ratio is undefined. If the curve has no observed failure drops between \(s\) and \(t\), the product over that interval is 1, so the estimated conditional survival is 1. This is a property of the fitted step curve, not proof that the true event risk is zero. In particular, if follow-up has ended or risk sets have become sparse, a flat curve may contain little information about later survival.
The ratio is an estimate of a marginal conditional survival probability for the population represented by the study, under the censoring assumptions. It is not automatically an individual prediction or a covariate-adjusted estimate. The earlier material on Kaplan-Meier survival and censoring supplies the risk-set estimates; the essential additional step here is to align the conditioning event and interval endpoints with the ratio.
Common Exam Traps
- Including a failure at \(s\). For \(P(T>t\mid T>s)\), use factors with \(s<t_j\le t\). A failure at \(s\) is already reflected in \(\widehat S_{\rm KM}(s)\).
- Omitting a failure at \(t\). A failure at the right endpoint is included, because the product runs through \(t\), not strictly before \(t\).
- Using the wrong denominator. The standard ratio is \(\widehat S_{\rm KM}(t)/\widehat S_{\rm KM}(s)\), not the reverse. The result should be between 0 and 1 when \(s\le t\) and the estimates are valid.
- Ignoring the positive-denominator condition. Division by a zero Kaplan-Meier estimate is undefined; it is not a conditional survival estimate of zero.
- Treating a censoring mark as a failure drop. Censoring changes later risk sets but does not itself multiply the curve by a failure factor.
Check Your Understanding
For each problem, use the stated endpoint convention: conditional survival from \(s\) to \(t\) means \(P(T>t\mid T>s)\). Give probabilities to four decimal places where appropriate.
- A study reports \(\widehat S_{\rm KM}(3)=0.80\) and \(\widehat S_{\rm KM}(8)=0.56\). Find the estimated conditional survival from 3 to 8 and the corresponding conditional failure probability by 8. (Answer: survival \(0.7000\); failure probability \(0.3000\).)
- Failures occur at time 2 with \(Y=9,d=3\), at time 5 with \(Y=6,d=2\), and at time 8 with \(Y=4,d=1\). Find the conditional survival from time 2 to time 8. Verify using the product of interval factors. (Answer: \((4/6)(3/4)=0.5000\).)
- At time 1, \(Y=10,d=2\); at time 4, \(Y=7,d=2\); and at time 6, \(Y=4,d=1\). A censoring is also recorded at time 4, using the convention that the risk set is measured immediately before failures and tied censorings. Find conditional survival from time 1 through time 6. (Answer: \((5/7)(3/4)=15/28=0.5357\).)
- At time \(s=3\), there is a failure and the Kaplan-Meier estimate drops from \(0.75\) just before time 3 to \(0.60\) at time 3. The estimate at time 7 is \(0.30\). Find survival from time 3 to time 7 given \(T>3\), and explain why the time-3 failure is excluded. (Answer: \(0.30/0.60=0.5000\); the value at time 3 already includes the failure at 3.)
- A fitted Kaplan-Meier curve has \(\widehat S_{\rm KM}(4)=0\) and \(\widehat S_{\rm KM}(9)=0\). Can the ratio formula provide conditional survival from time 4 to 9? State the denominator issue. (Answer: No; the denominator is zero, so the ratio is undefined.)