Censoring and Informative Dropout · Tutorial 229 of 501
Restricted Mean Survival Time Under Informative Censoring
Learn to integrate an IPCW-adjusted survival curve through a fixed horizon and quantify the resulting uncertainty without discarding covariance over time.
Advanced
12 min read
What You'll Learn
Define restricted mean survival time through a fixed horizon and relate it to the survival curve.
Calculate RMST exactly from a stepwise IPCW-weighted survival estimate.
Derive the RMST influence contribution by integrating the survival-curve influence function.
State the iid and remainder conditions needed for a central limit theorem and variance calculation.
Account for covariance across times and for uncertainty from estimating censoring weights.
Recognize positivity, model, and numerical-integration problems that can undermine the estimate.
From a Weighted Survival Curve to Restricted Mean Survival Time
A survival curve summarizes survival at each follow-up time, while a restricted mean gives one quantity that accumulates survival over a chosen period. The restriction matters: the target is the mean survival time up to a prespecified horizon \(\tau\), not an extrapolation of mean lifetime beyond the observed period. Under informative censoring, the survival curve used for this calculation must account for the censoring process.
Definition: Let \(T\) be a nonnegative event time, \(S(t)=P(T>t)\) its marginal survival function, and \(\tau>0\) a fixed horizon. The restricted mean survival time through \(\tau\) is
$$
\mu(\tau)=E\{\min(T,\tau)\}=\int_0^\tau S(t)\,dt.
$$
The identity follows by writing \(\min(T,\tau)=\int_0^\tau I(T>t)\,dt\) and taking expectations. Interchanging expectation and integration is valid because the indicator is nonnegative and bounded, so Tonelli's theorem applies. The horizon must be selected as part of the estimand; changing \(\tau\) changes the quantity being estimated.
For continuous-time follow-up, an IPCW weighted product-limit estimator is built from weighted risk and event counts. Write \(X_i=\min(T_i,C_i)\), where \(C_i\) is the censoring time, and let \(\delta_i=I(T_i\le C_i)\). If \(W_i(t)\) is the inverse probability of remaining uncensored through \(t\), conditional on the measured history used in the censoring model, define
$$
Y_w(t)=\sum_{i=1}^n W_i(t)I(X_i\ge t),\qquad
dN_w(t)=\sum_{i=1}^n W_i(t)I(X_i=t,\delta_i=1).
$$
The weighted product-limit curve is
$$
\widehat S_w(t)=\prod_{u\le t}\left\{1-\frac{dN_w(u)}{Y_w(u)}\right\},
\qquad
\widehat\mu(\tau)=\int_0^\tau \widehat S_w(t)\,dt.
$$
This construction relies on the same censoring assumptions used for IPCW survival estimation: conditional independent censoring given the relevant observed history, positivity of remaining uncensored through \(\tau\), and a correctly specified censoring model or a consistent alternative for estimating the required weights. The weighted survival curve must consistently estimate the marginal survival curve on \([0,\tau]\). The curve's uncertainty also depends on how the weights are estimated; treating fitted weights as fixed is not generally justified.
Calculating the Area Under a Step Curve
A product-limit survival estimate is constant between event times. Its integral is therefore a sum of rectangle areas, with no approximation required when all jumps and the horizon are included. Censoring changes the weighted risk sets and hence the estimated curve; it is not itself a survival-curve drop.
Worked Example: Integrating a Weighted Product-Limit Curve
Suppose the weighted event-to-risk ratios at event times \(1,3,4\) are \(0.10,0.20,0.25\), respectively, with horizon \(\tau=5\). The survival curve starts at 1. Calculate its value on each interval and the RMST.
At time 1, the product-limit factor is \(1-0.10=0.90\). Thus the curve is 1 on \([0,1)\) and 0.90 on \([1,3)\). At time 3, the factor is \(1-0.20=0.80\), so the curve becomes \(0.90(0.80)=0.72\) on \([3,4)\). At time 4, the factor is \(1-0.25=0.75\), so it becomes \(0.72(0.75)=0.54\) on \([4,5]\).
The estimated restricted mean is 4.06 time units. This is an area under the estimated survival curve, not the event time of an individual and not an estimate of unrestricted mean survival.
If the curve is represented on a time grid, integrate each interval using its appropriate curve value. Trapezoidal integration is not the exact calculation for a step curve if it replaces the constant interval value with a line between endpoints. For a smooth fitted curve, numerical quadrature may be needed, and its error should be small relative to the statistical uncertainty.
How Curve Uncertainty Propagates Through the Integral
The key uncertainty calculation is an integration of the survival-curve influence contribution. State the regularity conditions explicitly: subjects are iid; for each \(t\in[0,\tau]\), the weighted survival estimator has an asymptotic linear representation with mean-zero subject-level contributions \(\phi_i(t)\); and its integrated remainder is negligible at the root-\(n\) scale. In particular, assume
$$
\widehat S_w(t)-S(t)=\frac{1}{n}\sum_{i=1}^n\phi_i(t)+r_n(t),
\qquad
\int_0^\tau |r_n(t)|\,dt=o_p(n^{-1/2}),
$$
where the \(\phi_i\) are iid copies of a measurable, mean-zero process \(\phi(t)\), and \(E[\{\int_0^\tau|\phi(t)|\,dt\}^2]<\infty\). The contributions must represent the estimator actually used, including fitted censoring weights when those weights are estimated.
Theorem: Under the conditions above, define
$$
\psi_i(\tau)=\int_0^\tau\phi_i(t)\,dt.
$$
Then
$$
\widehat\mu(\tau)-\mu(\tau)
=\frac{1}{n}\sum_{i=1}^n\psi_i(\tau)+o_p(n^{-1/2}),
\qquad
\sqrt{n}\{\widehat\mu(\tau)-\mu(\tau)\}
\xrightarrow{d}N\bigl(0,\operatorname{Var}\{\psi_i(\tau)\}\bigr).
$$
Consequently, the first-order variance of \(\widehat\mu(\tau)\) is \(\operatorname{Var}\{\psi_i(\tau)\}/n\).
To prove the representation, integrate the curve expansion over \([0,\tau]\). The sum has finitely many terms, so it can be interchanged with the integral:
The moment condition ensures that \(\psi_i(\tau)\) has finite second moment: \(|\psi_i(\tau)|\le\int_0^\tau|\phi_i(t)|\,dt\). Its mean is zero by integrability and \(E\{\phi_i(t)\}=0\). Since the subjects are iid, the ordinary iid central limit theorem applies to their scalar contributions \(\psi_i(\tau)\). Combining that limit with the negligible remainder gives the stated convergence. The iid assumption is substantive: mean zero and finite variance alone do not make dependent contributions obey this central limit theorem. \(\square\)
The corresponding variance can also be written as a double integral. The assumed second-moment bound permits the required interchange of expectation and integration:
This expression shows why uncertainty at different times cannot generally be treated as independent. In practice, if \(\widehat\psi_i(\tau)\) estimates each subject's integrated influence contribution, a plug-in standard error is
$$
\widehat{\operatorname{SE}}\{\widehat\mu(\tau)\}
=\sqrt{\frac{1}{n}\left\{\frac{1}{n-1}
\sum_{i=1}^n\bigl(\widehat\psi_i(\tau)-\overline{\widehat\psi}(\tau)\bigr)^2\right\}},
\qquad
\overline{\widehat\psi}(\tau)=\frac{1}{n}\sum_{i=1}^n\widehat\psi_i(\tau).
$$
This plug-in calculation requires consistent influence estimates and appropriate regularity for their empirical variance. A curve-level standard error at one time point is not enough to recover the RMST standard error: the dependence of the curve estimates across the whole interval matters.
Worked Uncertainty Calculations
Worked Example: Integrating Estimated Influence Contributions
Suppose \(\tau=2\), and an influence calculation for the weighted curve gives the following subject-level contributions, constant within each unit-length interval. The four subjects are treated as a small illustrative sample.
Subject \(i\)
\(\phi_i(t)\) on \([0,1)\)
\(\phi_i(t)\) on \([1,2]\)
1
\(-0.4\)
\(0.0\)
2
\(-0.1\)
\(0.0\)
3
\(0.2\)
\(0.0\)
4
\(0.3\)
\(0.0\)
Integrating across the two intervals gives \(\widehat\psi_i(\tau)=\int_0^2\phi_i(t)\,dt=(-0.4,-0.1,0.2,0.3)\). Their mean is zero. The empirical variance with denominator \(n\), used here as the plug-in estimate of the per-subject variance, is
Thus the plug-in estimated variance of the RMST estimator is \(0.075/4=0.01875\), and its standard error is \(\sqrt{0.01875}\approx0.137\). This calculation uses the integrated contributions, rather than averaging pointwise standard errors.
Worked Example: Why Covariance Across Time Matters
Suppose the curve's influence contribution equals a random value \(A\) on \([0,2)\) and a random value \(B\) on \([2,3]\). Let \(P\{(A,B)=(1,1)\}=P\{(A,B)=(-1,-1)\}=1/2\). Then both components have mean zero, variance 1, and covariance 1.
The integrated contribution is \(\psi=2A+B\). It equals 3 with probability \(1/2\) and \(-3\) with probability \(1/2\), so \(E(\psi)=0\) and \(\operatorname{Var}(\psi)=9\). Equivalently,
Ignoring covariance would give 5 instead of 9. This example illustrates why an uncertainty calculation must preserve the joint contribution over time, including dependence induced by shared subjects and estimated censoring weights.
Worked Example: Including Censoring-Model Uncertainty
For illustration, suppose the integrated influence contribution is decomposed into an outcome-curve component \(U_i\) and a censoring-weight estimation component \(V_i\). Let the iid subject-level pairs be \((U_i,V_i)=(1,2)\) or \((-1,-2)\), each with probability \(1/2\). Both components have mean zero. Their variances are 1 and 4, and their covariance is 2.
The full contribution is \(\psi_i=U_i+V_i\), which equals 3 or \(-3\) with equal probability. Therefore its variance is 9. Direct calculation confirms the decomposition:
If the censoring model were estimated but its contribution were omitted, the variance calculation would keep only \(\operatorname{Var}(U_i)=1\). That is not the variance of the estimator that was actually fitted. Influence decompositions vary by estimator, but uncertainty from estimating weights must be represented when it affects the first-order limit.
Diagnostics, Alternatives, and Limitations
An RMST estimate is only as credible as the weighted survival curve and its uncertainty calculation. Before interpreting it, check that the chosen horizon is supported by follow-up and that estimated probabilities of remaining uncensored do not approach zero so closely that weights become extreme. Also inspect whether the censoring model includes relevant observed history and whether weighted risk sets become sparse near \(\tau\). These checks can reveal instability, but no diagnostic can prove that conditional independent censoring or the censoring model is correct.
A common mistake is to calculate the RMST standard error by integrating pointwise standard errors, or by treating curve estimates at different times as independent. Neither operation generally gives the variance of an integral. The correct first-order variance integrates the covariance function over pairs of times, or equivalently computes the variance of each subject's integrated influence contribution. Another mistake is to use a variance formula for fixed weights when the censoring weights were estimated; that can omit a first-order source of uncertainty.
Compared with reporting survival at a single time, RMST uses the entire curve up to \(\tau\) and remains interpretable in time units. Compared with a parametric survival model, the weighted product-limit approach avoids specifying a full parametric event-time distribution, but it still relies on censoring assumptions, adequate positivity, and valid weight estimation. Choosing a longer horizon may include more area but can also expose weak follow-up support and unstable weights. The estimand and its support should therefore be considered together.
1
Fix the target horizon. Specify \(\tau\) and define the target as \(\int_0^\tau S(t)\,dt\).
2
Estimate the adjusted curve. Construct the IPCW survival estimate under the censoring assumptions and positivity conditions.
3
Integrate the curve. Sum interval areas for a step curve, or use sufficiently accurate numerical integration for a smooth curve.
4
Propagate uncertainty. Integrate subject-level influence contributions and estimate their variance, retaining covariance over time and weight-estimation effects.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
Why is \(\mu(\tau)\) an area under the survival curve, and what does the horizon restrict?
A step curve is 1 on \([0,1)\), 0.8 on \([1,2)\), and 0.5 on \([2,4]\). Calculate its RMST through \(\tau=4\).
What assumptions about the subject-level contributions justify the iid central limit theorem in the uncertainty result?
Why is integrating pointwise variances without covariance generally incorrect?
What uncertainty component may be missed if estimated censoring weights are treated as fixed?
Solutions. (1) Since \(\min(T,\tau)=\int_0^\tau I(T>t)\,dt\), taking expectations gives \(\int_0^\tau P(T>t)\,dt\). The horizon limits the accumulated survival time to \(\tau\). (2) The area is \(1(1-0)+0.8(2-1)+0.5(4-2)=1+0.8+1=2.8\) time units. (3) The contributions must be iid, mean zero, and have finite variance, along with the negligible integrated remainder condition; iid sampling supplies the central limit theorem used here. (4) The variance of an integral includes covariance between contributions at different times, as the double-integral formula shows. (5) The first-order contribution from estimating the censoring model and its weights may be omitted.