Why Censoring Adjustment Is Needed
A prediction for an event by a fixed horizon can be evaluated by comparing the predicted risk with whether the event actually occurred by that time. But in a validation cohort, a person censored before the horizon may have unknown event status: the event might have happened later, or not at all by the horizon. Simply dropping such people can distort the evaluation if the subjects whose outcomes remain observable differ systematically from those whose outcomes are censored.
Inverse-probability-of-censoring weighting (IPCW) addresses this by giving greater weight to observable subjects who represent similar subjects whose status was not observed. The method can estimate a range of horizon-specific losses. It does not make unknown individual statuses known, nor does it repair a misspecified prediction model; it adjusts the evaluation for incomplete outcome ascertainment under stated assumptions.
If \(\Delta=1\) and \(Y\leq\tau\), the person is a known case: \(D=1\), and the event time is observed. If \(Y>\tau\), the person is a known control: \(D=0\), because follow-up establishes that no event occurred by \(\tau\). If censoring occurs before \(\tau\) without an earlier event, the horizon status is unknown. These definitions assume a consistent convention for event and censoring ties; for simplicity, the examples below treat event and censoring times as continuous, so ties have probability zero.
Assumptions and the IPCW Formula
Let \(G(t\mid Z)=P(C\geq t\mid Z)\) denote the conditional probability of remaining uncensored through time \(t\). The conditioning variables \(Z\) must include the variables determining the prediction, so \(r(Z)\) is fixed given \(Z\). They must also be sufficient for the censoring assumption below to be reasonable.
For a known case, the chance its status is observed, conditional on its event time and \(Z\), is \(G(T\mid Z)\). For a known control, status is observed if the person remains uncensored through \(\tau\), with probability \(G(\tau\mid Z)\). Thus cases receive the reciprocal of the first probability and controls the reciprocal of the second. Censored subjects with unknown status contribute zero directly; the increased weights on observed subjects compensate in expectation.
Here \(n\) is the full validation-cohort size and \(\widehat G\) is the estimated censoring survival function. The case indicator selects observed events by \(\tau\); the control indicator selects subjects known to be event-free at \(\tau\). Using \(n\), rather than only the number with known status, is part of this inverse-probability average. The loss may be any specified measurable loss for binary status and a risk prediction for which the target expectation is well-defined and finite; the weighting argument does not otherwise depend on a particular choice.
Why the Weights Recover the Target Loss
The justification is a conditional expectation calculation. First consider a person whose event time satisfies \(T\leq\tau\). Given \(T\) and \(Z\), the prediction and case loss are fixed. Conditional independent censoring gives \(P(C\geq T\mid T,Z)=G(T\mid Z)\). Therefore:
For a person with \(T>\tau\), the status is a control. The indicator that the subject remains observable through \(\tau\) has conditional probability \(G(\tau\mid Z)\), so the corresponding calculation is:
Under the continuous-time convention, \(P(C>\tau\mid Z)=G(\tau\mid Z)\). Combining the event and event-free cases and then averaging over \(T\) and \(Z\) shows that the population expectation of the weighted contribution equals \(E\{L(D,r(Z))\}\). This is the target mean loss. In a finite cohort, the displayed estimator is an estimate of that expectation; its realized value need not equal the complete-data cohort loss exactly.
Worked Example: Calculating an IPCW Mean Loss
Suppose a hypothetical validation cohort has four people and horizon \(\tau=2\) years. Use absolute error \(L(D,r)=|D-r|\). The censoring survival probabilities shown are assumed known for this illustration.
Person 1: An event is observed at \(Y=0.5\) years, with \(r=0.20\) and \(G(0.5\mid Z)=0.90\). The loss is \(|1-0.20|=0.80\), so the weighted contribution is \(0.80/0.90=0.8889\).
Person 2: Follow-up extends beyond 2 years without an event, with \(r=0.30\) and \(G(2\mid Z)=0.65\). The loss is \(|0-0.30|=0.30\), giving \(0.30/0.65=0.4615\).
Person 3: An event is observed at \(Y=1.5\) years, with \(r=0.60\) and \(G(1.5\mid Z)=0.70\). The loss is \(|1-0.60|=0.40\), and the weighted contribution is \(0.40/0.70=0.5714\).
Person 4: Censoring occurs at 1.2 years without an earlier event. The status at 2 years is unknown, so this person's direct contribution is zero.
The cohort IPCW estimate is \((0.8889+0.4615+0.5714+0)/4=0.4805\). It estimates the validation-population mean absolute error at 2 years under the assumptions above; the censored person's unknown status has not been classified as either an event or a non-event.
Worked Example: The Role of Each Reciprocal Probability
Consider a hypothetical subject with \(T=1\) year, horizon \(\tau=3\) years, and case loss \(L\{1,r(Z)\}=0.60\). Suppose \(G(1\mid Z)=0.80\). Among subjects with these same event time and prediction inputs, the probability of observing the event before censoring is 0.80. If observed, the weighted loss is \(0.60/0.80=0.75\); if not observed, its direct contribution is zero. The conditional expected contribution is \(0.80(0.75)+0.20(0)=0.60\), which recovers the case loss.
For a control with loss 0.25 and \(G(3\mid Z)=0.50\), the corresponding expected contribution is \(0.50(0.25/0.50)+0.50(0)=0.25\). This calculation shows why weights are reciprocals of the relevant observation probabilities, not arbitrary adjustments based on the overall percentage censored.
Prediction Inputs Must Be Accounted For
A crucial detail is that conditional independent censoring must hold given variables sufficient to make the prediction fixed. If a prediction depends on a covariate associated with censoring, weighting only by a marginal censoring probability can fail, even when event time and censoring time are independent without conditioning on that covariate. The weighted sample may then overrepresent some prediction values and underrepresent others.
Worked Example: Why Marginal Censoring Weights Can Fail
Let \(X\) be 0 or 1 with equal probability. Suppose every subject is event-free beyond \(\tau\), so \(D=0\), and the prediction is \(r(X)=X\). The control loss is absolute error: it is 0 when \(X=0\) and 1 when \(X=1\). The target mean loss is therefore \((0+1)/2=0.50\).
Suppose subjects with \(X=0\) are always followed beyond \(\tau\), while subjects with \(X=1\) are followed beyond \(\tau\) with probability 0.25. Since \(T>\tau\) for everyone, \(D=0\) for everyone, while censoring depends on \(X\). The marginal probability of being observed through \(\tau\) is \((1+0.25)/2=0.625\). Applying the reciprocal marginal weight gives expected contribution \(0.5(0)+0.5\{0.25(1/0.625)\}=0.20\), not the target 0.50.
Conditioning the censoring probability on \(X\) instead gives \(G(\tau\mid X=1)=0.25\). The \(X=1\) expected weighted contribution is then \(0.25(1/0.25)=1\); the \(X=0\) contribution is zero. Averaging over \(X\) recovers \((0+1)/2=0.50\). In practice, \(Z\) must include the prediction inputs and any further covariates needed to support the conditional independence assumption.
Diagnostics, Alternatives, and Limitations
The weights can reveal when the evaluation is fragile. Since \(1/\widehat G(t\mid Z)\) grows as the estimated probability of remaining uncensored shrinks, a few subjects may dominate the estimate. Examine the distribution of weights, the smallest estimated censoring probabilities in the relevant time and covariate ranges, and whether high weights occur for influential predictions or outcomes. Also check whether follow-up supports the chosen horizon: if nearly no comparable subjects remain observable through \(\tau\), the data provide little information about the target loss there.
A very small or zero \(G(t\mid Z)\) violates positivity. No finite weight can represent subjects with those covariate values and times if their outcomes are never observable. Weight truncation can reduce variance from extreme weights, but changes the estimator and may introduce bias; it is not a substitute for diagnosing poor support. Uncertainty in an estimated censoring model also affects uncertainty in the final metric and should be reflected in inference.
| Approach | What it does | Main concern |
|---|---|---|
| Complete-case evaluation | Computes loss only for subjects with known status at the horizon. | Can target a selected observed subset rather than the full validation population. |
| IPCW evaluation | Weights known cases and controls by their inverse probabilities of being observed. | Requires conditional independent censoring, positivity, and adequate estimation of \(G\). |
| Unweighted evaluation of observed subjects | Uses known statuses without adjusting for differing observation probabilities. | Does not generally correct selection associated with censoring or prediction inputs. |
IPCW is useful because it makes its adjustment explicit and can target a stated population mean loss. Its validity is not guaranteed by the algebra alone: if censoring depends on unobserved information related to event time even after conditioning on \(Z\), or if the censoring model is badly estimated, the weighted result can remain biased. Nor can IPCW identify performance beyond the region where observation probabilities are positive. The censoring model must be chosen for the prediction metric and validation setting; its practical specification requires careful attention to the data available.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions.
- At horizon \(\tau\), what status is known for a subject whose event is observed at \(Y\leq\tau\)? What status is known when \(Y>\tau\)?
- A known case has loss 0.30 and estimated \(G(Y\mid Z)=0.60\). What is its IPCW contribution?
- Why is a subject censored before \(\tau\) without an earlier event not treated as a control?
- What does positivity require, and what practical warning arises when estimated censoring survival is close to zero?
- Explain why a marginal censoring weight can be inadequate when the prediction depends on a covariate associated with censoring.
Solutions. (1) The subject is a case, \(D=1\); if \(Y>\tau\), follow-up establishes no event by the horizon, so \(D=0\). (2) The contribution is \(0.30/0.60=0.50\). (3) The subject's event status at \(\tau\) remains unknown; censoring before the horizon does not establish that the subject stayed event-free through \(\tau\). (4) The relevant \(G(t\mid Z)\) must be positive for the target times and covariate values. Values near zero create very large weights and unstable estimates; values equal to zero make the target unsupported by observed data. (5) The observed sample can overrepresent some prediction values, and the loss can vary with those values. Conditioning the censoring model on prediction inputs, under conditional independent censoring, is needed to recover the full-population mean loss.