Independent Censoring In Aalen-Johansen Estimation

Why Censoring Conditions Matter

Aalen–Johansen estimation builds transition estimates from people who are still observed and at risk in each state. Its risk sets are useful only if the observed people provide appropriate information about the transitions that would occur in the target population. When censoring selectively removes people with different future transition patterns, the remaining risk sets can give a misleading picture, even if the product-integral calculations are performed correctly.

A second condition is easy to overlook. Independent censoring can justify estimating state-specific transition intensities from risk sets, but it does not by itself make those intensities sufficient to recover every target transition probability. For the usual pooled Aalen–Johansen transition-probability interpretation, the process must also be Markov, or a suitable alternative model must justify that interpretation.

This is a conditional statement, not necessarily the claim that \(C\) is unrelated to everything about a participant. Censoring may depend on observed covariates or the current state. The issue is whether, among people comparable on the information used in the analysis, those censored would have had the same future transition distribution as those remaining under observation. If censoring depends on a recorded covariate associated with transitions, an unadjusted pooled risk set may not be representative even when censoring is independent conditional on that covariate.

What the Risk-Set Calculation Assumes

For states \(j\) and \(k\), let \(Y_j(t)\) be the number observed and at risk in state \(j\) immediately before time \(t\), and let \(dN_{jk}(t)\) be the number of observed transitions from \(j\) to \(k\) at that time. At a transition time with \(Y_j(t)>0\), the Aalen–Johansen estimate of the cause-specific transition-intensity increment is

The diagonal increment is \(d\widehat A_{jj}(t)=-\sum_{k\ne j}d\widehat A_{jk}(t)\). The product integral combines these increments in time order to form an estimated transition matrix. The required risk-set condition is that the people still observed in state \(j\) at time \(t\) provide valid information about the transition intensity of the target people in that state at that time. There must also be adequate observation and a positive chance of remaining uncensored over the time range needed; if everyone at risk is censored before a transition can be observed, that transition intensity is not identified from the data.

For a precise transition-probability interpretation, suppose \(X(t)\) is a time-inhomogeneous Markov process: given its current state and time, its future does not additionally depend on its earlier state history. Write its transition intensity as \(q_{jk}(t)\). For small \(h>0\), the Markov condition gives

The intensity depends on \(j\) and \(t\), not on unmodeled features of the path that led to state \(j\). Under this condition, independent censoring for the relevant risk sets, consistent recording of transitions and censoring, and adequate follow-up, the pooled risk-set estimates can estimate the transition intensities used to construct the target transition-probability matrix. If transition intensities depend on earlier history, risk-set estimates can still be calculated, but pooling people in the same current state can combine different histories. The resulting product integral should not automatically be described as a consistent estimate of the target conditional transition-probability matrix.

Worked Example: A Risk-Set Increment

Why the Markov Condition Matters

The product integral propagates state probabilities using transition intensities indexed by current state and time. The Markov condition makes this propagation appropriate: once the current state and time are known, the transition rate does not need to distinguish how an individual arrived there. Without that condition, people in the same state at the same time may have different transition rates because of their prior paths or unmeasured characteristics. The observed risk-set rate is then an average over the histories represented among those still at risk. That average may change as the composition of the risk set changes, and it need not serve as a common transition intensity for every origin-state cohort in the target matrix.

A short justification clarifies the role of censoring. At a fixed time, a risk-set ratio compares observed transitions with observed person-time in the state. Under independent censoring, remaining observed does not select people with different future transition behavior after conditioning on the information defining that risk set. With sufficient data and positivity, the ratio can therefore estimate the relevant transition intensity. But this argument does not establish that a single intensity \(q_{jk}(t)\) applies across different histories in state \(j\). That further conclusion comes from the Markov assumption, or from a correctly specified alternative that accounts for the history differences.

Worked Example: Future-Related Censoring Can Bias an Estimate

Addressing Conditional Censoring and Checking Assumptions

When censoring depends on measured covariates, an analysis can sometimes condition on those covariates, stratify the transition estimates, or use inverse-probability-of-censoring weights. Such approaches require that censoring be independent of future transitions after conditioning on the modeled information, that the censoring model be adequately specified, and that there be a positive probability of remaining uncensored across relevant covariate patterns and times. A weighted or adjusted estimate is not automatically valid merely because a model was fitted.

Useful diagnostics include comparing censoring patterns across observed states, times, and measured risk factors; checking whether few or no people remain at risk in important groups; and examining sensitivity to plausible censoring models. Such checks can reveal sparse support or measured patterns that merit adjustment. They cannot prove independence from unobserved future transitions. Administrative end of follow-up may be plausibly independent by design, but loss to follow-up related to worsening health, treatment changes, or impending transitions deserves particular scrutiny.

Common Mistake and Limitations

Wrong reasoning: “Censoring is independent, so the Aalen–Johansen estimator consistently estimates the transition-probability matrix.” Correction: Independent censoring supports valid risk-set information, conditional on the relevant risk-set information. The usual pooled transition-probability interpretation also needs the Markov condition, or another justified model that accounts for history dependence. State-specific risk-set rates alone do not establish that condition.

The method also cannot recover transition behavior beyond times with adequate observation without additional assumptions. Dependence on unmeasured prognostic factors is especially difficult: observed data may not distinguish informative censoring from a genuinely different transition pattern. Alternative analyses can include covariate-adjusted transition models, stratified estimates, or censoring-weighted procedures, but each targets a particular population and relies on its own modeling and positivity conditions. These methods do not remove uncertainty about unobserved causes of censoring.