Choosing Covariates for Censoring Models

Why Covariate Choice Matters

A censoring model describes which individuals remain under observation over time, conditional on measured information. Its covariates matter because censoring may depend on characteristics that also predict when an event occurs. If those characteristics are ignored, the observed event histories can be systematically unrepresentative of the histories that would have been observed with complete follow-up. On the other hand, adding every available variable is not automatically safer: some variables are measured too late, are recorded only for selected participants, or can make the intended analysis harder to interpret.

The goal is not to find variables that predict censoring as accurately as possible in isolation. It is to specify a defensible set of measured predictors that makes the censoring assumption plausible for the analysis, while respecting the timing and meaning of those variables. This tutorial concerns that selection task, not the estimation of censoring survival probabilities.

Definition: Let \(T\) be the time to the event of interest and \(C\) the time to censoring. The observed time is \(\tilde T=\min(T,C)\), and \(\Delta=I(T\le C)\) indicates whether the event was observed before censoring. Let \(Z\) denote baseline covariates and \(H(t)\) the measured covariate history available just before time \(t\). Conditional independent censoring means that, given the relevant measured covariates or history, the censoring process contains no further information about the event process. In a baseline formulation this is written \(T\perp C\mid Z\); with evolving information, the corresponding condition is stated given \(H(t)\) at each time.

In competing-risks and multi-state data, a transition to another event state is an event, not censoring. Covariate selection for a censoring model concerns loss of observation or another censoring mechanism, not variables chosen to reclassify a competing event as censoring. The earlier material on marginal and conditional independent censoring gives the distinction between the unconditional and conditional assumptions.

A Practical Basis for Selecting Predictors

Start with the data collection and follow-up process, then ask which measured variables could explain both who is censored and whose event history differs. Relevant candidates commonly include disease severity, prior events or health status, treatment, site, calendar time, access to care, and recorded reasons for withdrawal. A variable can matter even if it is not a strong predictor in a particular sample: selection should reflect the design and plausible mechanisms, not only a test of association.

For a baseline conditional model, consider baseline variables known before follow-up begins. For a time-varying model, consider information available at the time it would be used. A value measured after a person has left observation cannot ordinarily serve as their predictor of censoring at that earlier time. The earlier material on time-varying covariates in censoring models addresses the timing and updating of such histories.

It is useful to distinguish three roles. A variable may predict censoring; it may predict the event outcome; or it may plausibly help explain why censoring and the event process are associated. A predictor of censoring that is unrelated to the event process can still be relevant to modeling censoring. An outcome predictor alone does not necessarily need to enter the censoring model, but it deserves consideration when it is related to censoring or when conditioning on it helps make the conditional independence assumption credible. These are substantive judgments, not universal inclusion rules.

Define the censoring process.
Specify what ends observation and when. Separate censoring from event transitions and record censoring dates and reasons where possible.

Map plausible predictors.
Use study design and subject-matter knowledge to identify variables related to censoring, the event process, or both.

Check timing and data quality.
Include only information available at the relevant prediction time; examine missingness, coding, and category counts.

Assess whether the set is adequate and estimable.
Ask whether important shared predictors are represented and whether the data support the proposed detail. Document unresolved uncertainty rather than treating a fitted model as proof of independence.

Why Omitting a Shared Predictor Can Matter

A simple calculation shows why conditional independence does not generally imply marginal independence. Suppose \(Z\) has two categories and, within each category, censoring by a fixed time is independent of having had the event by another fixed time. Write \(q_z=P(C\le c\mid Z=z)\) and \(r_z=P(T\le t\mid Z=z)\). Then the joint probability is a weighted average of the within-category products, while the product of the marginal probabilities multiplies the two weighted averages. These need not be equal.

P(C\le c,T\le t)=\sum_z P(Z=z)q_zr_z, \qquad P(C\le c)P(T\le t)= \left\{\sum_zP(Z=z)q_z\right\} \left\{\sum_zP(Z=z)r_z\right\}.

If \(Z\) predicts both censoring and the event, omitting it can therefore leave censoring and event occurrence associated in the pooled data. This is a justification for considering shared predictors; it is not a guarantee that adding a variable will make the conditional independence assumption true. Unmeasured predictors, incorrect timing, and model misspecification can still matter.

Worked Example: A Shared Predictor of Censoring and Event Risk

Consider a hypothetical population in which half the participants have mild disease and half have severe disease. By a specified follow-up horizon, suppose censoring probabilities are 0.10 and 0.30, respectively, and the event probabilities under complete ascertainment are 0.20 and 0.50. Assume censoring and event occurrence are independent within each severity group. Compare the conditional and marginal calculations.

Within the mild group, the joint probability of censoring and event occurrence is \(0.10(0.20)=0.020\). Within the severe group it is \(0.30(0.50)=0.150\). Averaging across the equally sized groups gives \(P(C\le c,T\le t)=0.5(0.020)+0.5(0.150)=0.085\).

The marginal censoring probability is \(0.5(0.10)+0.5(0.30)=0.20\), and the marginal event probability is \(0.5(0.20)+0.5(0.50)=0.35\). Their product is \(0.20(0.35)=0.070\), not 0.085. Thus the two processes are independent within severity groups but associated marginally in this hypothetical setting. Disease severity is a strong candidate for the conditional censoring model because it predicts both processes. The calculation illustrates a reason to include it; it does not establish that severity is the only needed variable in a real analysis.

Sparsity, Overfitting, and Inappropriate Adjustment

Covariates may be conceptually important but poorly supported by the observed data. A rare category with few participants, few censoring events, or little follow-up can make category-specific estimates unstable. A large collection of finely divided predictors can also produce an unwieldy model with weak support in some combinations of covariates. The remedy is not automatically to delete a variable known to be important. First examine coding errors and whether categories can be combined in a substantively meaningful way; where appropriate, consider a simpler functional form or a method that regularizes estimation. Report the simplification and its rationale.

Avoid conditioning on variables that are unavailable at the relevant time or are only observed among people who remain under observation. Also consider the role of post-baseline variables. Such variables may be consequences of prior treatment or health, and their adjustment can change the question being addressed. In particular, a variable that is a common effect of an exposure and a cause of the event can induce an association between those causes when conditioned on. A variable’s predictive value alone does not determine whether adjustment is appropriate.

Worked Example: A Sparse Category Is Not a License to Ignore It

A hypothetical study has 240 participants at three sites. Site A has 120 participants and 24 censored observations; Site B has 115 participants and 23 censored observations; Site C has 5 participants and 0 censored observations. The corresponding crude censoring proportions are \(24/120=0.200\), \(23/115=0.200\), and \(0/5=0\).

The zero in Site C does not show that its censoring probability is truly zero: only five participants were observed there. A site effect may still be important if sites differ in follow-up practice, so dropping Site C merely because its sample proportion is zero could omit relevant structure. At the same time, a separate effect for Site C may be estimated very imprecisely.

The analyst should verify the site coding and investigate whether Sites A, B, and C have substantively comparable censoring procedures. If A and B have the same relevant process and can be combined without obscuring meaningful differences, combining them may simplify the representation while retaining Site C as a separate category. If they cannot be combined, an appropriate estimation strategy may be needed, and uncertainty should be made explicit. These counts diagnose sparse support; they do not by themselves justify either pooling or exclusion.

Worked Example: Why a Predictive Variable May Be an Inappropriate Adjustment

Suppose a hypothetical study compares treatment \(A\) with usual care. A post-baseline measure \(W\) is affected by treatment and by an underlying health factor \(U\); \(U\) also predicts the event. The diagram is \(A\rightarrow W\leftarrow U\rightarrow T\). Assume \(U\) also predicts loss to follow-up. The measure \(W\) is recorded only at a later visit, which some participants do not reach.

Although \(W\) may predict censoring among participants with a recorded value, it is a common effect of \(A\) and \(U\). Conditioning on \(W\) can associate \(A\) with \(U\), even if treatment assignment was initially independent of \(U\). Since \(U\) predicts the event, this induced association can complicate an exposure-outcome comparison. In addition, \(W\) is unavailable for participants censored before the visit, so it cannot simply be used as a baseline predictor for their earlier censoring.

The correction is not to ignore all post-baseline information. Instead, define the analysis and its target carefully, respect when each measurement becomes available, and assess whether conditioning on that history is justified. A time-updated value can be considered for censoring after it is measured if the model and analysis support that use; it cannot supply information retroactively. If the needed information is unmeasured, selecting a different observed covariate does not guarantee that censoring is conditionally independent.

Diagnostics and Alternatives

Diagnostics can reveal weaknesses in the chosen covariate set, but no diagnostic can prove conditional independent censoring. Compare censoring patterns over follow-up across important covariate groups and review recorded reasons, dates, missingness, and changes in follow-up practice. Look for categories with few participants or censoring observations, unusual combinations of covariates, and abrupt changes by site or calendar period. Check whether continuous predictors have plausible functional forms rather than assuming every relationship is linear. These checks assess data support and model plausibility; they do not rule out unmeasured determinants.

The main alternatives serve different purposes. A marginal censoring model uses no covariates and is simplest, but it is inadequate when censoring depends on measured characteristics relevant to the event process. A stratified approach permits different censoring patterns across a small number of well-supported groups but becomes impractical with many cross-classified categories. A regression model can represent several predictors and functional forms, but depends on its specification and the available data. Automated selection based only on p-values can exclude scientifically important predictors, select chance associations, and behave poorly with sparse data. None of these choices removes the need for a defensible conditional independence assumption.

Key takeaway: Choose censoring-model predictors using the censoring mechanism, the event process, measurement timing, and data support. Include plausible shared predictors when justified, scrutinize sparse categories and post-baseline variables, and describe the conditional independence assumption as an assumption—not a result established by model fit.

Common Mistakes

Keeping only statistically significant predictors. A large p-value does not show that a variable is irrelevant to the censoring assumption. Use scientific and design knowledge as well as observed patterns.
Adding every available variable. More predictors can create sparse combinations, use information measured too late, or condition on variables that alter the analysis question.
Treating a zero censoring count as a zero censoring probability. Sparse data can yield extreme proportions by chance and provide little information about a category-specific rate.
Claiming that conditional independence has been proved. Diagnostics can reveal problems, but measured data cannot rule out unmeasured predictors or misspecification.
Confusing a competing event with censoring. A transition to a competing event state is an outcome, not a loss of observation.

Check Your Understanding

Answer each question, then compare your reasoning with the brief solutions.

A baseline variable predicts both withdrawal and event risk. Give one reason to consider it in a conditional censoring model.
In a category of four participants, none is censored. What can and cannot be concluded from this count?
A measure is collected at a visit six months after baseline and predicts later withdrawal. Can its value be used to model censoring before that visit?
Why can a variable that predicts censoring still be inappropriate to adjust for in a particular analysis?
Does a well-fitting censoring model establish conditional independent censoring? Explain briefly.

Solutions. (1) It may account for a shared source of dependence between censoring and the event process, making conditional independence more plausible. (2) The observed proportion is zero, but four participants provide too little information to conclude the underlying censoring probability is zero. (3) No. The value is not available before it is measured; it may be considered for later times only if the analysis and model appropriately use the time-updated history. (4) It may be measured after baseline, observed only for selected participants, or be a common effect whose adjustment changes associations relevant to the target analysis. (5) No. Fit and diagnostics can identify some shortcomings, but cannot prove the assumption or exclude unmeasured predictors.