Why the Censoring Model Matters
A prediction metric at a fixed horizon depends on whether each person experienced the event by that horizon. When follow-up is censored first, that person's horizon status is unknown. Inverse-probability-of-censoring weighting (IPCW) compensates for this incomplete observation by giving more weight to outcomes that were less likely to be observed. The weights are only appropriate, however, if the model for censoring matches the assumptions used to construct them.
The central choice is whether to model censoring marginally, using the same censoring distribution for everyone, or conditionally, allowing that distribution to vary with observed covariates. A conditional model can handle covariate-related censoring, but it is not automatically better: it requires appropriate covariates, adequate data, and a correctly specified model. The goal is not to make the censoring model elaborate; it is to make its assumptions credible for the target validation population and metric.
Set-Up and Assumptions
Let \(T\) denote event time, \(C\) censoring time, \(Y=\min(T,C)\) observed follow-up time, and \(\Delta=I(T\le C)\) the event indicator. At horizon \(t\), the event status is \(D_t=I(T\le t)\). Let \(Z\) be a set of covariates observed for every validation subject, and let \(q(t,Z)\) be that subject's predicted event risk by \(t\). For example, a Brier loss uses \(\ell(D_t,q)=(D_t-q)^2\). The target is the average loss in the validation population, \(E\{\ell(D_t,q(t,Z))\}\).
A conditional censoring model supports IPCW when censoring is conditionally independent of the event time given the modeled covariates: \(T\perp C\mid Z\). A marginal censoring model requires the stronger marginal independence condition \(T\perp C\), unless the target and metric have some special structure that makes the marginal weighting valid. Conditional independence does not generally imply marginal independence: if both event risk and censoring vary with \(Z\), averaging over \(Z\) can make \(T\) and \(C\) associated.
Both approaches also require positivity over the relevant follow-up range: the censoring survival probability used as a denominator must be greater than zero for covariate patterns and times contributing to the target metric. In practice, probabilities near zero produce unstable, very large weights even before positivity is literally violated. The covariates in \(Z\) must be available for the validation subjects and sufficiently capture the dependence needed for the chosen independence assumption. No observed-data diagnostic can prove that conditional independence holds; it is an assumption supported by design knowledge and scientific reasoning.
Which Observations Receive Which Weights?
For a loss defined at horizon \(t\), an event observed by \(t\) contributes its event-status loss even if censoring would otherwise have happened before \(t\). Its censoring weight is based on the event time, not on the horizon. By contrast, an observed event-free status at \(t\) requires being followed and uncensored through \(t\). With a conditional censoring model, the IPCW estimate of the target average loss is
A subject censored before \(t\) without an observed event contributes neither term. The marginal version replaces both conditional denominators by \(\widehat G(Y_i-)\) and \(\widehat G(t)\), respectively. These are not interchangeable weights: using \(\widehat G(t\mid Z_i)\) for an event observed at \(Y_i<t\) incorrectly treats that event as if it needed to remain uncensored until \(t\).
Worked Example: Marginal Weights Can Bias a Horizon Brier Score
Consider a hypothetical validation population with two equally frequent covariate groups, \(Z=0\) and \(Z=1\). Every event by horizon \(t\) occurs at time \(s<t\); otherwise the subject remains event-free past \(t\). Let the event risks be \(0.20\) and \(0.80\) in the two groups. Suppose censoring is independent of event status within each group, with \(G(s\mid 0)=0.90\), \(G(t\mid 0)=0.80\), \(G(s\mid 1)=0.50\), and \(G(t\mid 1)=0.30\). The predicted risk is \(q=0.50\) for everyone. These assumptions make \(D_t\perp C\mid Z\), but the different event risks and censoring distributions mean marginal independence is not justified by the setup.
The true population Brier score is \(0.25\): since \(q=0.50\), either event status has squared error \((1-0.50)^2=(0-0.50)^2=0.25\). The marginal censoring probabilities are \(G(s)=\{0.90+0.50\}/2=0.70\) and \(G(t)=\{0.80+0.30\}/2=0.55\). Under these marginal weights, the expected weighted contribution within group 0 is
The first term accounts for events observed at \(s\), and the second for event-free outcomes observed through \(t\). For group 1, the corresponding contribution is
Averaging the two equally frequent groups gives \((0.3552+0.1701)/2=0.2627\), rather than the target \(0.25\). With the correct conditional censoring probabilities, each group's event and event-free terms sum to \(0.25\), so their average is also \(0.25\). The example shows the source of bias: marginal weights do not correct the different observation probabilities within covariate groups.
Choosing and Checking a Censoring Model
A marginal model is reasonable when censoring is plausibly independent of the event time in the target population and a common censoring distribution is adequate. It is parsimonious and can be estimated with a reverse Kaplan–Meier approach. A covariate-dependent model is appropriate when conditional independence is plausible given measured variables that explain censoring and are relevant to the relationship between censoring and event status. It may be estimated with stratified reverse Kaplan–Meier curves, a regression model for the censoring distribution, or a discrete-time model for censoring. The model must estimate survival probabilities at both observed event times and the evaluation horizon.
| Choice | When it is defensible | Main risk |
|---|---|---|
| Marginal \(G(u)\) | Marginal event–censoring independence is plausible, or a common censoring distribution is otherwise justified for the target. | Covariate-related censoring can leave residual association between censoring and event status. |
| Conditional \(G(u\mid Z)\) | Censoring may vary with measured \(Z\), and conditional independence given \(Z\) is plausible. | Omitted covariates, a wrong functional form, or sparse covariate patterns can make weights biased or unstable. |
| Highly flexible conditional model | There is enough data to estimate censoring probabilities reliably across relevant covariate patterns. | Overfitting and near-zero estimated probabilities can make weighted metrics erratic. |
For a fitted conditional model, inspect whether its estimated censoring survival probabilities reproduce observed censoring patterns over time, both overall and across important covariate groups. For a regression model, assess functional forms and time dependence rather than assuming a single covariate effect is adequate at every time. Check the distribution of estimated weights, the smallest estimated probabilities, and whether a few subjects dominate the weighted metric. Compare model-based probabilities with empirical censoring estimates in well-populated strata when possible. Such checks assess fit and stability; they cannot establish the required independence assumption.
Worked Example: Correct Conditional Weights Recover the Target Metric
Use the two-group hypothetical setup above, now with group-specific predictions \(q(t,0)=0.20\) and \(q(t,1)=0.80\). Suppose a constructed validation sample has 100 people in each group. In group 0, 18 events are observed at \(s\), and 64 people are observed event-free through \(t\). These are the expected counts under event risk \(0.20\) and censoring probabilities \(G(s\mid0)=0.90\), \(G(t\mid0)=0.80\). The event weight is \(1/0.90\), and the event-free weight is \(1/0.80\). Its weighted Brier contribution sum is
In group 1, 40 events are observed at \(s\) and 6 people are observed event-free through \(t\), consistent with event risk \(0.80\), \(G(s\mid1)=0.50\), and \(G(t\mid1)=0.30\). The weighted contribution sum is
The estimate divides the combined sum by all 200 validation subjects: \((16+16)/200=0.1600\). This agrees with the target Brier score because each group's prediction equals its event risk: in either group, the expected squared error is \(p(1-p)^2+(1-p)p^2=p(1-p)=0.16\), for \(p=0.20\) or \(p=0.80\). The counts are illustrative expected counts, not claims about a real study.
Worked Example: A Misspecified Conditional Model Still Biases the Metric
Keep the same event risks, censoring distributions, and group-specific predictions as in the preceding example. Suppose instead that a fitted conditional model uses \(\widehat G(s\mid0)=0.80\), \(\widehat G(t\mid0)=0.70\), \(\widehat G(s\mid1)=0.60\), and \(\widehat G(t\mid1)=0.40\). These probabilities do not equal the true censoring probabilities. The expected weighted loss in group 0, using the true observation probabilities in the numerators and the fitted probabilities in the denominators, is
For group 1, the expected contribution is
The resulting population metric is \((0.1806+0.1227)/2=0.1516\), not the true \(0.1600\). Including covariates in a censoring model does not guarantee validity: the estimated conditional probabilities must also be adequate.
Common Pitfalls and Limitations
- Using the horizon weight for an earlier event. An event observed at \(Y<t\) needs the probability of remaining uncensored until \(Y\), not until \(t\). Use the event-time denominator for event contributions and the horizon denominator for event-free contributions.
- Assuming marginal and conditional independence are equivalent. They are different assumptions. Dependence through \(Z\) can invalidate marginal weights even when conditional independence holds.
- Adding predictors without checking support. More covariates can improve the plausibility of conditional independence but can also create sparse strata and extreme weights. Positivity and weight stability remain essential.
- Treating fit checks as proof of independent censoring. Observed censoring patterns can reveal poor model fit, but unmeasured causes of both censoring and event status may remain. The independence assumption needs substantive justification.
- Assuming a correct censoring model fixes every validation problem. IPCW addresses incomplete event-status observation under its assumptions. It does not make the risk predictions correct, remove sampling error, or change the target population.
Weight truncation can reduce variability from extreme weights, but it changes the estimator and can introduce bias; it is not a substitute for examining positivity or reconsidering the target. Likewise, a more flexible censoring model may reduce functional-form error while increasing estimation variability. The choice should balance credible conditional independence, model adequacy, available sample size, and weight stability.
Check Your Understanding
Answer each question, then compare with the solutions below.
- For an event observed at time \(Y<t\), which censoring probability belongs in its IPCW denominator, and why?
- Conditional independence \(T\perp C\mid Z\) holds, but censoring and event risk both vary with \(Z\). Is a marginal censoring model automatically valid?
- In a group with event risk \(0.30\), suppose \(G(s\mid Z)=0.80\) for events at \(s<t\), and \(G(t\mid Z)=0.50\). What is the expected correctly weighted event contribution to an event-risk estimate, and what is the weighted event-free contribution to a status total?
- Name two diagnostics for a conditional censoring model and one important assumption those diagnostics cannot verify.
Solutions. (1) Use \(G(Y-\mid Z)\), because the event is observed if censoring has not occurred before its event time; it need not remain uncensored until \(t\). (2) No. Dependence induced by \(Z\) can make \(T\) and \(C\) dependent marginally, so marginal weights are not generally justified. (3) The expected weighted event contribution is \(0.30\times0.80/0.80=0.30\). The event-free status contribution is \(0.70\times0.50/0.50=0.70\). Together they recover total status probability \(1\). (4) Examples include comparing estimated and empirical censoring survival across covariate groups, checking time-dependent model fit, and inspecting estimated weight extremes. These diagnostics cannot verify that censoring is independent of event time conditional on the measured covariates.