Informative Censoring and Event-Dependent Dropout

Why Censoring Can Change an Exposure–Event Result

Time-to-event analyses use information from people who remain under observation: those still at risk contribute to the risk set, and observed events contribute event-time information. This works when censoring is noninformative for the event process, conditional on the covariates and exposure history in the model. But patients may withdraw, discontinue follow-up, or become unobservable for reasons related to their event risk. If that relationship is not accounted for, the observed risk sets can become a selected subset of the original population.

In PK/PD studies, event-related dropout is plausible when worsening symptoms, toxicity, treatment discontinuation, or impending clinical deterioration affects follow-up. An exposure–hazard estimate can then be distorted: high-risk patients may leave observation disproportionately at a particular exposure, making the remaining patients appear safer. The direction and size of bias depend on how dropout relates to exposure and event risk; they are not guaranteed to be toward zero.

The condition is model-relative. Censoring can be associated with event risk marginally but become conditionally independent after accounting for measured predictors, such as prior symptoms or concentration history. Conversely, adjusting for exposure alone may be insufficient if an unmeasured disease-severity factor influences both dropout and the event. Administrative end of study is often plausibly independent conditional on design and enrollment time; early withdrawal requires more scrutiny.

Under conditionally independent right censoring, an event-time likelihood can use the event-process contribution \( [\lambda_T(Y\mid H)]^\delta \exp[-\Lambda_T(Y\mid H)]\), where \(\lambda_T\) is the event hazard and \(\Lambda_T\) its cumulative hazard. The censoring process can be omitted from estimation of the event model when it is conditionally independent and its parameters are distinct. If censoring depends on unmodeled event risk, that factorization does not justify ignoring dropout.

How Selective Dropout Distorts Exposure Effects

An exposure coefficient in a hazard model describes how the event hazard varies with exposure, conditional on the model and the observed risk sets. Informative dropout changes which people remain in those risk sets. If high-exposure patients with high underlying event risk are more likely to disappear before an event is recorded, the observed high-exposure risk set becomes depleted of high-risk patients. The fitted exposure effect can be too small, absent, or even reversed. If low-risk patients preferentially drop out, the distortion can go in the opposite direction.

No single observed-data pattern proves that censoring is informative. Useful clues include different dropout rates by exposure, concentration, prior event-risk markers, or treatment response; dropout following worsening symptoms or toxicity; and disagreement between results under plausible censoring assumptions. Reasons and timing of withdrawal should be summarized alongside the event data. A lack of association between recorded covariates and dropout is not proof of independence, because an unmeasured common cause may remain.

The key limitation is fundamental: after dropout, the unobserved event time is unknown. The observed data alone generally cannot establish whether those patients would have had higher or lower event risk. Remedies therefore rely on assumptions, measured predictors, or a model for the joint processes, and should be accompanied by sensitivity analyses when the assumptions are uncertain.

Inverse-Probability-of-Censoring Weighting

Inverse-probability-of-censoring weighting (IPCW) gives greater weight to observed patients who resemble people likely to have dropped out. At a given time, let \(R(t)=1\) indicate that a patient remains uncensored and observed. Let \(\pi(t\mid H)\) be the conditional probability of remaining uncensored through that time, given the recorded history \(H\). A basic weight is \(w(t)=1/\pi(t\mid H)\). For time-varying histories, the probability of remaining uncensored is commonly built from conditional probabilities over successive intervals.

The basic justification is an inverse-probability identity. Let \(Z\) be a future event outcome and suppose \(R\) is the indicator that it is observed. If \(\pi(H)=P(R=1\mid H)\), and \(R\) is independent of \(Z\) conditional on \(H\), then:

The first equality uses the definition of conditional expectation. The second uses conditional independence, and the third uses \(E[R\mid H]=\pi(H)\). Averaging over histories therefore recovers the target mean under the stated assumptions. For survival data, weights are applied to event-risk-set contributions or estimating equations over time; the formula is not a claim that a single weighted endpoint analysis replaces the event-time model.

In practice, censoring probabilities are estimated, often using a regression model for remaining observed at each interval. Stabilized weights can reduce variability by using a numerator probability conditional on a smaller set of baseline variables and a denominator conditional on the fuller history. Very large weights indicate that some histories have little chance of remaining observed, threatening positivity and producing unstable estimates. Weight distributions, overlap in predicted probabilities, and the sensitivity of estimates to reasonable weight truncation should be reported. Truncation can improve precision but changes the weighting and may reintroduce bias. Uncertainty calculations should account for estimated weights, for example through an appropriate bootstrap or robust variance procedure.

Joint Models for Event and Dropout

A joint model specifies both the event process and the dropout process, allowing them to share latent predictors. This is useful when unobserved heterogeneity is a plausible source of dependence, but it trades the weighting assumptions for stronger model assumptions. Let \(b_i\) be a patient-specific latent effect. One illustrative specification is:

Here \(E_i\) is an exposure measure, \(\lambda_{T,0}\) and \(\lambda_{C,0}\) are baseline event and dropout hazards, and \(\alpha\) controls how the shared latent effect relates to dropout. Conditional on \(b_i\) and included covariates, the two processes are commonly assumed independent. The marginal likelihood integrates over the assumed distribution of \(b_i\); for an observed event at time \(y\), a contribution has the form \(\int \lambda_{T,i}(y\mid b)\exp[-\Lambda_{T,i}(y\mid b)-\Lambda_{C,i}(y\mid b)]f(b)\,db\), under the stated conditional-independence model and with dropout-free observation up to the event.

Choosing and Checking an Approach

A practical analysis begins by describing censoring timing and reasons by exposure and by measured predictors of event risk. The event model and its covariate history should be specified before choosing a correction. If measured history can plausibly explain dropout and positivity is adequate, IPCW may be reasonable. If unmeasured shared risk is a central concern, a joint model can be considered alongside sensitivity analyses. Comparing estimates across defensible approaches is informative, but agreement does not prove that censoring is noninformative.

Common pitfall: “The exposure groups have similar overall dropout percentages, so censoring cannot bias the exposure effect.” Equal percentages do not guarantee equal selection: dropout may concentrate among high-risk patients in one exposure group and low-risk patients in another, or depend on predictors not included in the event model. The correction is to assess dropout conditional on exposure and relevant history, examine reasons and timing, and state the assumptions behind any weighting or joint model. Neither method can recover missing event information without assumptions.