Copula-Graphic Estimator in Survival Analysis

• The paper introduces the copula-graphic estimator as a copula-based generalization of Kaplan–Meier, enabling survival analysis under dependent censoring.
• It establishes an explicit dependence model using primarily the Clayton copula to reconstruct marginal survival and support two-sample permutation tests and survival tree splits.
• Empirical studies show that the estimator maintains type I error control and robustness against copula misspecification, though it may struggle with unbalanced sample sizes.

Searching arXiv for the named paper and closely related work on copula-graphic estimators in survival analysis. The copula-graphic estimator (CGE) is a nonparametric estimator of the survival function for right-censored data in which the event time $T$ and censoring time $C$ may be dependent. In the formulation emphasized in "Permutation Tests Based on the Copula-Graphic Estimator and Their Use for Survival Tree Construction" (Baur et al., 28 Jul 2025), the estimator replaces the standard independent-censoring assumption of Kaplan–Meier by an explicit copula model for the joint survival of $T$ and $C$. Within that framework, the CGE serves both as an estimator of marginal survival and as the basis of a two-sample permutation test and a recursive partitioning algorithm for survival trees under dependent censoring.

1. Statistical setting and motivation

The CGE is defined for right-censored survival data with observed variables

$X_i=\min(T_i,C_i), \qquad \Delta_i=I(T_i\le C_i),$

so that the observed sample is $\{(x_i,\delta_i)\}_{i=1}^n$. Standard procedures, especially Kaplan–Meier, assume independent censoring, $T \perp C$. The paper explicitly notes that when dropout depends on prognosis this assumption is often violated and Kaplan–Meier can be badly biased (Baur et al., 28 Jul 2025).

To relax independent censoring, the joint survival of event and censoring times is modeled through a bivariate copula $\mathcal C$ acting on the marginal survival functions

$S_T(t)=P(T>t)=1-F(t), \qquad S_C(s)=P(C>s)=1-G(s),$

with

$$P(T>t,C>s)=\mathcal C\bigl(S_T(t),S_C(s)\bigr), \qquad t,s>0.$$

This formulation separates the marginal survival distributions from the dependence structure. In the paper, the dependence model is restricted to Archimedean copulas, which provide a convenient representation through a decreasing convex generator $C$0 satisfying $C$1 (Baur et al., 28 Jul 2025).

The resulting perspective places the CGE within survival analysis with dependent, or informative, censoring. The paper situates it alongside inverse probability of censoring weighting, joint models, and frailty models, but distinguishes copula-based methods by their explicit specification of dependence between event and censoring while keeping the margins conceptually separate. This suggests that the CGE is best understood not as a minor modification of Kaplan–Meier, but as a copula-driven reconstruction of $C$2 from censored observations under an assumed dependence model.

2. Copula formulation and estimator construction

For an Archimedean copula,

$C$3

where $C$4 is decreasing and convex, and

$C$5

The paper focuses primarily on the Clayton copula, with parameter $C$6 and generator

$C$7

for which Kendall’s $C$8 satisfies

$C$9

The limit $T$0 corresponds to independence, and within the Archimedean framework this is represented by the independence generator $T$1 (Baur et al., 28 Jul 2025).

The central empirical quantity is the observed survivor of $T$2,

$T$3

which is the empirical estimator of $T$4, since $T$5 if and only if both $T$6 and $T$7. The CGE enforces the copula relationship at the empirical level through

$T$8

Rivest and Wells derive an explicit expression for $T$9 under an Archimedean copula. In the notation used by the paper, the CGE of the survival function is

$C$0

This estimator satisfies $C$1, is right-continuous, piecewise constant, and decreases with negative jumps only at event times (Baur et al., 28 Jul 2025).

Under the independence copula, with generator $C$2, the expression simplifies to the Kaplan–Meier estimator. The paper therefore characterizes the CGE as a copula-adjusted generalization of Kaplan–Meier: the dependence specification controls how censorings affect the step sizes. A frequent misconception is to regard the CGE as merely another graphical survival smoother; in the present formulation, its defining feature is instead the imposition of a copula model linking event and censoring times.

3. Identifiability, copula choice, and dependence specification

A central practical issue emphasized in the paper is that $C$3 and $C$4 are not identifiable from bivariate censored data without additional assumptions, echoing the classic nonidentifiability problem for competing risks (Baur et al., 28 Jul 2025). The copula assumption provides identifiability, but only once the copula family and its parameter have been specified. This is a key methodological constraint rather than a secondary implementation detail.

The paper adopts a fixed copula family, namely Clayton, chosen a priori. The stated justifications are its ability to model both positive and negative dependence, a simple $C$5 mapping, successful use in previous survival-copula work, and computational convenience through a simple generator and vectorized implementation. For robustness studies, however, data are also generated from a Frank copula. The paper notes that the Frank copula is used in simulations for data generation to study misspecification, but not for estimation.

Rather than estimating $C$6 directly from the data, the paper treats Kendall’s $C$7 as a sensitivity parameter. In simulations it evaluates $C$8; in the real-data example it considers values from $C$9 to $X_i=\min(T_i,C_i), \qquad \Delta_i=I(T_i\le C_i),$0 (Baur et al., 28 Jul 2025). Prior work by Zheng and Klein and by others is reported to suggest that the CGE is relatively robust to copula misspecification provided the strength of dependence is roughly correct. The simulations in the paper are described as confirming that type I error does not degrade under misspecified copulas, specifically Clayton versus Frank.

The practical implication is that the CGE is neither fully nonparametric nor fully parametric in the usual sense. Its marginal estimator is nonparametric, but its inferential validity depends on a fixed dependence model. A plausible implication is that sensitivity analysis over $X_i=\min(T_i,C_i), \qquad \Delta_i=I(T_i\le C_i),$1 is not optional in many applications but part of the substantive interpretation of results.

4. Two-sample testing based on group-specific CGEs

The paper considers a two-sample problem in which group $X_i=\min(T_i,C_i), \qquad \Delta_i=I(T_i\le C_i),$2 has event times $X_i=\min(T_i,C_i), \qquad \Delta_i=I(T_i\le C_i),$3, censoring times $X_i=\min(T_i,C_i), \qquad \Delta_i=I(T_i\le C_i),$4, and observed pairs $X_i=\min(T_i,C_i), \qquad \Delta_i=I(T_i\le C_i),$5. The null hypothesis is

$X_i=\min(T_i,C_i), \qquad \Delta_i=I(T_i\le C_i),$6

against the alternative that the survival functions differ at some $X_i=\min(T_i,C_i), \qquad \Delta_i=I(T_i\le C_i),$7. A crucial assumption is equality of censoring distributions across groups, $X_i=\min(T_i,C_i), \qquad \Delta_i=I(T_i\le C_i),$8, which renders the pooled observations exchangeable under $X_i=\min(T_i,C_i), \qquad \Delta_i=I(T_i\le C_i),$9 and makes a permutation test available (Baur et al., 28 Jul 2025).

For each group, the method computes a group-specific copula-graphic estimator $\{(x_i,\delta_i)\}_{i=1}^n$0 using the same copula family and assumed dependence parameter. The test statistic is a global distance between the two estimated survival curves: the integrated absolute distance of the group-specific CGEs, normalized by the common observation span $\{(x_i,\delta_i)\}_{i=1}^n$1. The integration range is $\{(x_i,\delta_i)\}_{i=1}^n$2, where

$\{(x_i,\delta_i)\}_{i=1}^n$3

Because the CGE is stepwise, the integral is computed numerically over the grid of observed times. The paper also defines a signed variant $\{(x_i,\delta_i)\}_{i=1}^n$4, without the absolute value, to determine direction when terminal nodes are ordered in the tree (Baur et al., 28 Jul 2025).

The permutation procedure computes the observed statistic, repeatedly permutes the pooled $\{(x_i,\delta_i)\}_{i=1}^n$5 records between groups, recalculates the statistic, and forms the standard Monte Carlo randomization $\{(x_i,\delta_i)\}_{i=1}^n$6-value with approximately $\{(x_i,\delta_i)\}_{i=1}^n$7 permutations in practice. With all $\{(x_i,\delta_i)\}_{i=1}^n$8 permutations and the exact critical region, the procedure is an exact level-$\{(x_i,\delta_i)\}_{i=1}^n$9 test under the null. Under independence, where the CGE equals Kaplan–Meier, the statistic reduces to the integrated absolute difference of Kaplan–Meier curves as in Moradian et al. (2019).

This test differs conceptually from the log-rank test. The log-rank test targets equality of hazards through weighted differences between observed and expected events under independent censoring. The CGE-based procedure instead measures a functional distance between survival curves and is intended to remain meaningful when censoring is dependent, provided the copula model captures that dependence and the censoring marginals are equal (Baur et al., 28 Jul 2025).

5. Embedding the estimator in survival trees

The main methodological contribution of the paper is to use the CGE-based permutation test as the splitting rule in a survival tree. The tree follows classical binary recursive partitioning: for each covariate $T \perp C$0 and feasible cutoff $T \perp C$1, the node is split into $T \perp C$2 and $T \perp C$3, and the two resulting child groups are compared by the CGE-based permutation test (Baur et al., 28 Jul 2025).

At a node with data $T \perp C$4, the algorithm considers all covariates and candidate cutoffs, computes a permutation $T \perp C$5-value $T \perp C$6 for each candidate split, and selects the split with minimum $T \perp C$7-value,

$T \perp C$8

If the minimum $T \perp C$9-value is below a pre-specified threshold $\mathcal C$0, for example $\mathcal C$1, the split is accepted. The signed statistic $\mathcal C$2 is then used to order the child nodes so that the better-survival group is placed on the left. If the threshold is not met, or if the node size is too small, such as $\mathcal C$3, the node becomes terminal.

This construction yields a tree in which the splitting criterion is a global test of survival difference under dependent censoring rather than a log-rank comparison. Traditional survival trees use log-rank tests at candidate splits and Kaplan–Meier estimates in terminal nodes, thus inheriting the independent-censoring assumption. In contrast, the CGE-based tree uses a copula-based survival estimator at each node and permits a specified dependence between survival and censoring through the copula parameter (Baur et al., 28 Jul 2025).

If censoring is independent, the CGE with $\mathcal C$4 reduces to Kaplan–Meier, so the resulting tree becomes close in spirit to a standard survival tree, albeit with an integrated-distance statistic rather than a log-rank score. If censoring is dependent, the paper argues that the log-rank framework is no longer valid, whereas the CGE-based tree remains interpretable under the copula specification. This suggests that the estimator functions not only as a terminal-node summary but as the core inferential engine of the partitioning procedure.

6. Empirical performance, applications, and limitations

The paper reports two simulation studies. The first evaluates the CGE-based permutation test under exponential baseline survival with dependence generated by Clayton or Frank copulas, multiple censoring rates, several sample sizes, and multiple assumed dependence levels (Baur et al., 28 Jul 2025). The reported findings are that all CGE-based tests maintain type I error close to $\mathcal C$5, with 90% of type I error estimates in $\mathcal C$6 and median approximately $\mathcal C$7. No systematic inflation is observed under Frank-generated data, which the paper interprets as robustness to copula misspecification. Power rises with $\mathcal C$8, and tests with smaller assumed dependence $\mathcal C$9 generally have higher power; the version with $S_T(t)=P(T>t)=1-F(t), \qquad S_C(s)=P(C>s)=1-G(s),$0, essentially Kaplan–Meier, often performs best. Under independent censoring, log-rank tends to be slightly more powerful, while under certain heteroscedastic normal designs the CGE tests are often more powerful, especially at higher variance ratio and higher censoring.

A notable limitation appears in unbalanced sample sizes. Because the statistic integrates only up to $S_T(t)=P(T>t)=1-F(t), \qquad S_C(s)=P(C>s)=1-G(s),$1, it can ignore tail separation when the smaller group has shorter survival and correspondingly smaller maximal follow-up. In such settings, the paper reports that the CGE tests can fail to detect differences identified by log-rank, producing very low power for some $S_T(t)=P(T>t)=1-F(t), \qquad S_C(s)=P(C>s)=1-G(s),$2 combinations (Baur et al., 28 Jul 2025).

The second simulation study evaluates tree performance using $S_T(t)=P(T>t)=1-F(t), \qquad S_C(s)=P(C>s)=1-G(s),$3 covariates generated in correlated blocks, with informative and noise variables, and censoring induced through a Clayton copula with $S_T(t)=P(T>t)=1-F(t), \qquad S_C(s)=P(C>s)=1-G(s),$4. Performance is assessed by selection ability, Harrell’s $S_T(t)=P(T>t)=1-F(t), \qquad S_C(s)=P(C>s)=1-G(s),$5-index, and Integrated Brier Score. The paper reports that log-rank trees produce more terminal nodes than CGE trees; for example, at $S_T(t)=P(T>t)=1-F(t), \qquad S_C(s)=P(C>s)=1-G(s),$6 with low censoring, a CGE tree with $S_T(t)=P(T>t)=1-F(t), \qquad S_C(s)=P(C>s)=1-G(s),$7 has about $S_T(t)=P(T>t)=1-F(t), \qquad S_C(s)=P(C>s)=1-G(s),$8 terminal nodes, whereas the log-rank tree has about $S_T(t)=P(T>t)=1-F(t), \qquad S_C(s)=P(C>s)=1-G(s),$9. CGE trees achieve precision above $$P(T>t,C>s)=\mathcal C\bigl(S_T(t),S_C(s)\bigr), \qquad t,s>0.$$0 in all settings and systematically higher precision than log-rank trees, while Harrell’s $$P(T>t,C>s)=\mathcal C\bigl(S_T(t),S_C(s)\bigr), \qquad t,s>0.$$1 is above $$P(T>t,C>s)=\mathcal C\bigl(S_T(t),S_C(s)\bigr), \qquad t,s>0.$$2 for all methods and is slightly higher or comparable for log-rank. Integrated Brier Score worsens with higher censoring for all methods; log-rank trees often have slightly higher IBS, though the differences are described as not dramatic.

In the Mayo PBC application, after preprocessing to remove nonrandomized subjects and missing covariates, the authors fit trees across a grid of assumed $$P(T>t,C>s)=\mathcal C\bigl(S_T(t),S_C(s)\bigr), \qquad t,s>0.$$3 values and compare them with a log-rank tree using 5-fold cross-validation (Baur et al., 28 Jul 2025). The log-rank tree attains the highest average $$P(T>t,C>s)=\mathcal C\bigl(S_T(t),S_C(s)\bigr), \qquad t,s>0.$$4-index, approximately $$P(T>t,C>s)=\mathcal C\bigl(S_T(t),S_C(s)\bigr), \qquad t,s>0.$$5, compared with about $$P(T>t,C>s)=\mathcal C\bigl(S_T(t),S_C(s)\bigr), \qquad t,s>0.$$6 to $$P(T>t,C>s)=\mathcal C\bigl(S_T(t),S_C(s)\bigr), \qquad t,s>0.$$7 for the CGE trees. Some CGE trees have better IBS than the log-rank tree, and no single $$P(T>t,C>s)=\mathcal C\bigl(S_T(t),S_C(s)\bigr), \qquad t,s>0.$$8 dominates. With a stricter split threshold $$P(T>t,C>s)=\mathcal C\bigl(S_T(t),S_C(s)\bigr), \qquad t,s>0.$$9, the resulting CGE trees are of moderate size and identify age, hepatomegaly, and serum bilirubin as important variables, while repeated splitting on bilirubin is noted as a possible sign of overfitting.

Several limitations are explicit. Equal censoring distributions across groups is a strong assumption and underlies the permutation test. The choice of $C$00 is not estimated but treated as a sensitivity parameter, creating a trade-off between stronger adjustment for dependent censoring and possible loss of power under weak or misspecified dependence. Computation is substantial because permutations are nested within recursive partitioning, even though the CGE itself is fast and the implementation is fully vectorized in R. The paper further notes that neither pruning nor amalgamation is implemented beyond the $C$01-value stopping rule, and that forests would likely improve prediction while reducing interpretability and increasing computation (Baur et al., 28 Jul 2025).

Within the broader literature, the paper links the CGE to the work of Zheng and Klein, Rivest and Wells, Emura and coauthors, Lo and Wilke, and Huang and Zhang. In that context, the estimator can be viewed as a copula-based generalization of Kaplan–Meier that supports not only marginal survival estimation under dependent censoring but also hypothesis testing and interpretable tree construction when the dependence between event and censoring times is explicitly modeled.