Why Some Kaplan–Meier Quantiles Are Not Estimable
A survival quantile is located by finding where the estimated survival curve reaches a specified level. But observed follow-up may end while the curve is still above that level. In that case, the data do not show the requested quantile: reporting a later time would require extrapolation beyond what the Kaplan–Meier estimate establishes.
This issue commonly arises when many observations are right-censored before enough failures occur. Censoring reduces the risk set used at later event times, but censoring by itself does not make the estimated survival curve step downward. The key question is therefore not simply how many observations were censored, but whether the observed Kaplan–Meier curve reaches the survival threshold for the requested quantile.
Here \(p\) is expressed as a proportion: the 40th percentile has \(p=0.40\), so its survival threshold is \(1-0.40=0.60\). “Reaches” includes equality. Thus, a curve value of exactly 0.60 is sufficient to estimate the 40th percentile; a value that remains above 0.60 is not.
The convention uses the first time the estimated survival is at or below the threshold. The Kaplan–Meier curve is a step function, so the estimate is reported at an observed event time, not at an interpolated time between steps. Right-censoring contributes to the risk sets but does not itself create a downward step.
How to Decide Whether the Curve Reaches the Threshold
At each distinct failure time \(t_j\), let \(n_j\) be the number at risk just before that time and \(d_j\) the number of failures at that time. The product-limit estimate updates as \(\hat S_{\mathrm{KM}}(t_j)=\hat S_{\mathrm{KM}}(t_{j-1})(1-d_j/n_j)\), starting at 1. Censorings between failure times reduce the later risk sets. To assess estimability, compare the successive survival estimates with the requested threshold and find the first observed crossing, if one exists.
The censoring assumption is the standard one for interpreting the Kaplan–Meier estimate: censoring is independent of the event time, at least conditional on any relevant groups or covariates used in the analysis. No parametric event-time distribution is assumed. The examples below use hypothetical equipment lifetimes; interest and mortality assumptions do not apply.
For a \(p\)-quantile, calculate \(1-p\). Keep the percentile and its complementary survival probability distinct.
Use the risk set immediately before each failure time, updating it for prior failures and censorings.
The quantile is estimated at the first observed failure time with \(\hat S_{\mathrm{KM}}\leq 1-p\).
State that the quantile is not estimable from the observed data. Do not invent a time beyond the observed curve.
Worked Example: A Quantile Reached Exactly at the Threshold
In a hypothetical study of 12 identical cooling units, three units fail at time 1, four are censored at time 2, one fails at time 4, and the remaining four are censored at time 6. Determine whether the 40th percentile of failure time is estimable, and give its estimate if it is. Assume independent censoring and no parametric lifetime model.
Method. Convert the requested percentile to a survival threshold: \(1-0.40=0.60\).
At time 1, all 12 units are at risk and three fail. The Kaplan–Meier estimate becomes \(\hat S_{\mathrm{KM}}(1)=1(1-3/12)=9/12=0.75\).
The four units censored at time 2 do not cause a survival step. They leave \(12-3-4=5\) units at risk immediately before time 4. One of those five fails at time 4, so \(\hat S_{\mathrm{KM}}(4)=0.75(1-1/5)=0.75(4/5)=0.60\).
The estimate equals the required threshold at time 4. Therefore, the 40th percentile is estimable and its Kaplan–Meier estimate is time 4. The four censorings at time 6 do not change this conclusion because the threshold was already reached.
Check. The successive estimates are 1 before the first failure, 0.75 at time 1, and 0.60 at time 4. Time 4 is the first observed time at or below 0.60. The common mistake avoided here is requiring the curve to fall strictly below the threshold; equality is enough.
Worked Example: The Final Observed Estimate Remains Above the Threshold
A hypothetical portfolio of ten monitored components has two failures at time 1, two censorings at time 2, one failure at time 4, and five censorings at time 5. Is the 40th percentile of failure time estimable? Assume independent censoring and no parametric lifetime model.
Method. The 40th-percentile survival threshold is \(1-0.40=0.60\). At time 1, the risk set is 10 and two components fail: \(\hat S_{\mathrm{KM}}(1)=1(1-2/10)=8/10=0.80\).
The two censorings at time 2 reduce the risk set from 8 to 6 but do not change the survival estimate. Thus, six components are at risk just before time 4. One fails then, giving \(\hat S_{\mathrm{KM}}(4)=0.80(1-1/6)=0.80(5/6)=4/5(5/6)=2/3\approx0.6667\).
The five components censored at time 5 are the remaining at-risk components. They do not create another downward step. The final observed estimate is \(2/3\), which is greater than \(0.60\). Consequently, the curve never reaches the 40th-percentile threshold, and the 40th percentile is not estimable from the observed data.
Check. At each observed failure time, the estimate is above 0.60: it is 0.80 at time 1 and approximately 0.6667 at time 4. The curve stays at that last value after time 4; the censorings do not imply an unobserved further drop. This example also illustrates why one must compare the computed estimate with the threshold rather than infer estimability from the fact that failures continued after the first event.
Worked Example: Crossing After Censoring Shrinks the Risk Set
In a hypothetical study of 15 pump seals, four fail at time 2, six are censored at time 3, two fail at time 5, and the remaining three are censored at time 7. Determine whether the median failure time is estimable. Assume independent censoring and no parametric lifetime model.
Method. The median corresponds to \(p=0.50\), so its survival threshold is \(1-0.50=0.50\).
At time 2, 15 seals are at risk and four fail: \(\hat S_{\mathrm{KM}}(2)=1(1-4/15)=11/15\approx0.7333\).
Six seals are then censored at time 3, leaving \(15-4-6=5\) at risk just before time 5. Two fail at time 5, so \(\hat S_{\mathrm{KM}}(5)=(11/15)(1-2/5)=(11/15)(3/5)=33/75=11/25=0.44\).
The estimate is above 0.50 at time 2 and below 0.50 at time 5. Therefore, the first crossing occurs at time 5, and the estimated median failure time is 5. The three censorings at time 7 do not alter the estimate.
Check. The time-5 risk set is five, not eleven or fifteen: both the four earlier failures and six censorings have been removed. The resulting estimate, 0.44, is at or below 0.50, confirming the crossing.
Why the Non-Estimability Rule Holds
For an event time \(T\), the survival function is \(S(t)=P(T>t)\), and the distribution function is \(F(t)=P(T\leq t)=1-S(t)\). A \(p\)-quantile is the earliest time at which the distribution function reaches \(p\), using the usual generalized-inverse convention. Replacing \(S\) by its Kaplan–Meier estimate gives \(\hat F_{\mathrm{KM}}(t)=1-\hat S_{\mathrm{KM}}(t)\).
This equivalence explains both the threshold and the first-crossing rule. If no observed step satisfies the final inequality, then the estimated distribution function has not reached \(p\) in the observed data. The Kaplan–Meier curve is constant between observed failure times, and censorings do not create downward survival steps. Thus, if the last observed failure leaves the estimate above \(1-p\), subsequent censorings alone cannot make the estimated quantile appear.
A flat segment after the final failure is not evidence that survival has fallen further; it records that the estimate has no additional observed failure step. The data may be consistent with future failures, but estimating when enough additional failures would occur requires a model or extrapolation beyond the empirical Kaplan–Meier estimate. For this question, the appropriate conclusion is “not estimable from the observed data,” not a guessed time.
Common Exam Traps and Reporting
- Comparing with the percentile instead of its complement. The 40th percentile is found at survival 0.60, not survival 0.40. Always calculate \(1-p\) first.
- Calling equality a failure to cross. If the estimate is exactly the threshold, the quantile is estimable at that event time.
- Treating censoring as a failure. A censoring reduces later risk sets but does not directly lower \(\hat S_{\mathrm{KM}}\).
- Using the wrong risk set. Each event-time factor uses the number at risk immediately before that time, after prior failures and censorings have been accounted for.
- Extrapolating a flat curve. A final estimate above the threshold means the quantile was not reached in observed follow-up, even if every remaining individual was eventually censored.
- Reporting only “not estimable.” A clear answer identifies the threshold and the final relevant estimate, then states why no observed step reaches it.
A concise, defensible report gives the requested percentile, its survival threshold, the last observed Kaplan–Meier value relevant to that threshold, and either the first event time that crosses it or the statement that no crossing occurred. This makes the role of censoring and the limits of follow-up explicit.
Check Your Understanding
For each hypothetical study, use the Kaplan–Meier steps to decide whether the requested quantile is estimable. Assume independent censoring and no parametric lifetime model.
- Eight devices have two failures at time 1, three censorings at time 2, and one failure at time 4; the remaining two are censored later. Is the median estimable? (Answer: Yes, at time 4. The estimate is \(6/8=0.75\) at time 1, then \(0.75(2/3)=0.50\) at time 4, exactly the median threshold.)
- Ten sensors have two failures at time 1, three censorings at time 2, one failure at time 3, and four later censorings. Is the 40th percentile estimable? (Answer: No. The threshold is 0.60; \(\hat S(1)=0.80\) and \(\hat S(3)=0.80(4/5)=0.64\), so no observed step reaches 0.60.)
- Twenty valves have five failures at time 2, ten censorings at time 3, and two failures at time 5; the remaining three are censored later. When is the median estimable? (Answer: At time 5. The estimates are \(15/20=0.75\) at time 2 and \(0.75(3/5)=0.45\) at time 5, crossing the threshold 0.50.)
- Four units fail successively at times 1, 2, 3, and 4, with no censoring before those failures. Is the 75th percentile estimable, and when? (Answer: Yes, at time 3. The Kaplan–Meier estimates are 0.75 at time 1, 0.50 at time 2, and 0.25 at time 3; the threshold is \(1-0.75=0.25\).)