Why Augment Censoring Weights?
Inverse probability of censoring weighting (IPCW) gives more representation to people whose observed histories resemble those of people who were censored. It can correct selection from censoring, but a few very large weights can make an estimate unstable. Augmentation adds predictions from models for future event outcomes. These predictions can improve precision and, under stated conditions, protect consistency if one of the two nuisance models is wrong.
The construction below targets the probability of experiencing a specified event type by a fixed horizon. Time is divided into intervals, and competing events are treated as terminal outcomes for this endpoint. The same logic applies to other fixed-horizon summaries when their interval-specific outcome contributions are defined appropriately. It is not a general correction for arbitrary informative censoring: the assumptions about censoring and the histories used in both models matter.
Write \(D_{kj}=1\) when the event of interest occurs in interval \(j\), and \(N_j=1\) when no event of any type occurs in that interval. A competing event has \(D_{kj}=0\) and \(N_j=0\). For fitted outcome regressions \(Q_j(H_j)\), define the interval pseudo-outcome \(Z_j=D_{kj}+N_jQ_{j+1}(H_{j+1})\), with \(Q_{K+1}=0\). When interval \(j\) is observed, \(Z_j\) is calculable: an event of interest contributes 1, a competing event contributes 0, and remaining event-free carries forward the predicted risk from the next interval.
The sequential outcome regression is correct when \(Q_j(H_j)=E[Z_j\mid H_j,\text{event-free at the start of }j]\) at every interval. This is a sequence of conditional predictions, not merely a model for the event risk from baseline. The censoring model is correct when the fitted \(G_j\) equals the conditional probability of remaining observable through interval \(j\), given the recorded at-risk history.
Estimator and Conditions
Let \(\widehat Q_j\) and \(\widehat G_j\) denote fitted models. For each person, the augmented contribution starts with the predicted baseline risk and adds weighted residuals from intervals whose outcomes are observed. If an event occurs, later intervals do not contribute; if censoring occurs, no unobserved interval is treated as known.
Here the sum is over intervals in which person \(i\) is still event-free and at risk; terms after an event are zero by convention. In \(Z_{ij}\), a no-event interval carries forward \(\widehat Q_{i,j+1}\). The estimator is an average of augmented contributions, not necessarily an average of values confined to 0 and 1. Finite-sample estimates can fall outside that range.
The interpretation requires a well-defined outcome and follow-up horizon, correctly recorded histories, and conditional independent censoring: given the recorded history at each interval, censoring must not additionally select people according to their future event outcomes. There must also be positivity: each relevant at-risk history must have a nonzero probability of remaining observable through the intervals needed to identify the risk. In estimation, very small fitted \(\widehat G_{ij}\) can make contributions unstable even when the theoretical probability is positive.
Why Either Model Can Protect Consistency
The two protections are often called double robustness. They are alternatives, not a claim that arbitrary misspecification is harmless. A short derivation shows what each condition contributes.
Conditional on a complete event history, \(E[I_j/G_j]=1\), because \(G_j\) is the probability that the history through interval \(j\) is observed. Thus the expected weighted residual sum equals the expected residual sum along the complete event path.
At a no-event interval, \(Z_j-\widehat Q_j=\widehat Q_{j+1}-\widehat Q_j\). At the terminal event interval, the residual is the terminal outcome minus \(\widehat Q_j\). If no event occurs by \(K\), the last carried-forward value is \(Q_{K+1}=0\). Therefore, along a fully observed path, \(\widehat Q_1+\sum_j(Z_j-\widehat Q_j)=Y\), whether or not the regressions are correct.
At each at-risk history, \(E[Z_j-Q_j\mid H_j]=0\). Under conditional independent censoring, the observation indicator has conditional mean \(G_j\). Even if a fitted weight is wrong but is a function of the recorded history, the conditional expected weighted residual is \((G_j/\widehat G_j)E[Z_j-Q_j\mid H_j]=0\). Averaging leaves the correctly predicted baseline risk.
Consequently, under the identification and positivity conditions, the estimator targets \(E[Y]\) if the censoring model is correct or if all the sequential outcome regressions are correct. If both are wrong, consistency is not guaranteed. Model fitting and uncertainty estimation must also account for estimated nuisance functions; for complex models, sample splitting or cross-fitting can reduce problems from fitting and evaluating the same flexible model on the same observations.
Worked Calculations
Worked Example: One-Interval Robustness to a Wrong Weight Model
Suppose a hypothetical population is evenly divided between histories \(H=0\) and \(H=1\). The event risks are 0.20 and 0.60, respectively. In this one-interval illustration, the outcome is observed with probabilities 0.50 and 0.80. Assume censoring is conditionally independent of the interval outcome given \(H\). The target risk is \(0.5(0.20)+0.5(0.60)=0.40\).
First use correct outcome predictions \(Q(0)=0.20\) and \(Q(1)=0.60\), but fitted observation probabilities \(\widehat G(0)=0.25\) and \(\widehat G(1)=0.40\). Within either history group, the expected augmented contribution is \(Q+(G/\widehat G)E[Y-Q\mid H]\). For \(H=0\), this is \(0.20+(0.50/0.25)(0.20-0.20)=0.20\). For \(H=1\), it is \(0.60+(0.80/0.40)(0.60-0.60)=0.60\). Averaging gives 0.40. The wrong weights do not change this population expectation because the outcome regression is correct.
This calculation demonstrates an expectation, not a promise that a finite-sample estimate will equal 0.40. Observed samples still vary, and estimated nuisance functions add uncertainty.
Worked Example: Correct Censoring Model With a Wrong Prediction
Use the same hypothetical population, but now take wrong predictions \(Q(0)=0.10\) and \(Q(1)=0.90\), while the observation probabilities are correctly specified. For \(H=0\), the expected contribution is \(0.10+(0.50/0.50)(0.20-0.10)=0.20\). For \(H=1\), it is \(0.90+(0.80/0.80)(0.60-0.90)=0.60\). The population average is again \(0.5(0.20)+0.5(0.60)=0.40\).
Here the correct weights restore the average residual within each history group. This one-interval example is the terminal step of the sequential construction: the same argument applies interval by interval when the censoring model correctly describes observability through each at-risk interval.
Worked Example: A Two-Interval Contribution With a Competing Event
Consider a hypothetical person with two intervals, fitted baseline risk \(\widehat Q_1=0.30\), and fitted next-interval risk \(\widehat Q_2=0.20\). Suppose the cumulative fitted observation probabilities are \(\widehat G_1=0.80\) and \(\widehat G_2=0.60\). These are illustrative fitted values, not estimates from a real study.
If the event of interest occurs in interval 1, then \(Z_1=1\), and the contribution is \(0.30+(1/0.80)(1-0.30)=0.30+0.875=1.175\). The value exceeds 1, which is possible for an individual augmented contribution.
If instead the person remains event-free in interval 1 and experiences the event of interest in interval 2, then \(Z_1=0.20\), \(Z_2=1\). The contribution is \(0.30+(1/0.80)(0.20-0.30)+(1/0.60)(1-0.20)\) \(=0.30-0.125+1.333=1.508\), to three decimal places.
If a competing event occurs in interval 1, then \(Z_1=0\) and there is no interval-2 term. The contribution is \(0.30+(1/0.80)(0-0.30)= -0.075\). These values are not individual risks; averaging contributions estimates the population risk. In particular, a competing event is not treated as censoring or as an event of interest.
Diagnostics, Alternatives, and Limitations
Augmentation does not remove the need to inspect the censoring process and fitted models. Check the estimated probabilities of remaining observed across intervals and histories, the distribution of inverse weights, and whether influential observations dominate the estimate. A small \(\widehat G_j\) produces a large multiplier \(1/\widehat G_j\); truncation can reduce variance but may introduce bias and does not repair a lack of positivity. Review whether histories used in the censoring model capture plausible predictors of loss to follow-up and whether the outcome regressions represent event risks across the relevant histories.
A useful comparison is with the two simpler approaches:
| Approach | Strength | Main vulnerability |
|---|---|---|
| IPCW only | Uses a censoring model to correct observed event histories. | Can be unstable with extreme weights and depends on a correct censoring model. |
| Outcome regression | Predicts the target risk from recorded histories. | Depends on correctly specifying the sequential event-outcome regressions and may extrapolate into poorly observed histories. |
| Augmented IPCW | Combines weighted residuals with outcome predictions; either nuisance model can suffice under the stated conditions. | Does not protect against both models being wrong, positivity failure, or censoring that depends on unrecorded future outcomes. |
The assumptions are substantive, not just computational. If censoring remains associated with future event risk after conditioning on recorded history, neither fitted model automatically identifies the target risk. Incorrectly measured covariates, sparse histories, and misspecified time-varying predictors can undermine both models. Augmentation also does not make an estimate causal; it estimates the stated population event risk under the observation assumptions.
Common Pitfall
A common mistake is to assume that adding a regression term makes the estimate robust to any censoring problem. It does not. The protection requires either a correct sequential censoring model or correct sequential outcome regressions, along with conditional independent censoring and positivity. Another error is to interpret an individual augmented contribution as a probability. As the worked calculations show, contributions can be negative or greater than 1; the target is their population average.
Check Your Understanding
Answer each question, then compare with the solutions below.
- For an endpoint distinguishing a particular cause by a fixed horizon, how should an event of a competing cause contribute to the outcome?
- In the two-history example, what is the target population risk?
- Why can a correct outcome regression protect the estimator when the fitted censoring probabilities are wrong?
- Name two conditions, beyond fitting one of the nuisance models correctly, required for the stated consistency result.
Solutions. (1) A competing event contributes 0 to the indicator for the specified cause and ends the at-risk event process; it is not treated as censoring. (2) The groups have equal population shares, so the risk is \(0.5(0.20)+0.5(0.60)=0.40\). (3) Correct sequential regressions make each residual have conditional mean zero given its recorded history. Under conditional independent censoring, weighting that residual by a history-dependent factor still gives conditional expectation zero. (4) Examples include conditional independent censoring given recorded history and positivity of remaining observable; a well-defined outcome and adequate measurement of the histories are also required.