Right-Censored Log-Likelihood: Methods & Applications

• Right-censored log-likelihood is a statistical tool that combines density and survival contributions to handle censored event-time data.
• It enables likelihood-based estimation that yields consistent and asymptotically normal estimators under the independent censoring assumption.
• The framework extends to various models including regression, cure, and copula-based analyses, facilitating rigorous inference and model selection.

A right-censored log-likelihood is the objective function used in parametric, semiparametric, and nonparametric inference for event-time data when the response variable is susceptible to right-censoring—i.e., observation is limited by a censoring mechanism so that only a lower bound on the actual event time is available for some subjects. The log-likelihood accommodates both observed events and censored subjects by synthesizing density and survival contributions, weighted according to the observed status indicator. This construction undergirds likelihood-based estimation for censored survival regression, discrete lifetime models, clustered time-to-event analysis, and recent modal regression frameworks with censoring. The right-censored log-likelihood is foundational to maximum likelihood estimation (MLE), information-theoretic inference, EM-based semiparametrics, and model selection criteria in censoring contexts.

1. Notation and General Formulation

Let $X_i$ denote the true event time for subject $i$, $C_i$ the right-censoring time, and $T_i = \min(X_i, C_i)$ the observed time. The event indicator is $\delta_i = \mathbf{1}\{X_i \leq C_i\}$, with $\delta_i=1$ for an uncensored event and $\delta_i=0$ for a censored observation. Let $f(x;\theta)$ be the model's density and $S(x;\theta) = 1 - F(x;\theta)$ the survival function, where $F$ is the cumulative distribution. Under independent censoring, the likelihood for $i$0 i.i.d. observations is

$i$1

and the right-censored log-likelihood is

$i$2

This form is applicable for both parametric and nonparametric modeling, and the above reduces to the standard uncensored log-likelihood when all $i$3 (Zhao et al., 2024, Matthews et al., 10 Apr 2025, Galarza et al., 7 Mar 2026).

2. Theoretical Properties and Interpretation

The right-censored log-likelihood provides the basis for likelihood-based estimation. Under regularity and independent censoring, maximizing $i$4 yields estimators that are consistent and asymptotically normal, and the usual Fisher information and large-sample theory apply (Zhao et al., 2024, Matthews et al., 10 Apr 2025). The censoring indicator $i$5 switches between the contribution of the observed density and the survival probability, replacing each missing density contribution by a survival-term contribution (Zhao et al., 2024). For a parametric model, score equations follow from differentiating $i$6 with respect to $i$7, and the observed information matrix is the negative second derivative.

A plausible implication is that, by construction, the censored likelihood efficiently utilizes partial information when the failure time is unobserved, and inference remains valid provided the censoring mechanism is independent.

3. Extensions to Regression, Discrete and Nonparametric Models

Regression

In regression with right-censoring, a covariate vector $i$8 is introduced, and the conditional density $i$9 specifies the model. The right-censored regression log-likelihood is

$C_i$0

In the absence of left-truncation, only right-censoring modifies the standard log-likelihood (Matthews et al., 10 Apr 2025). For models such as accelerated-failure-time regression, the density and survival have closed forms in terms of the error distribution, which may be parametric or semiparametric.

Discrete Time and Cure Models

Right-censored log-likelihoods for discrete failure times (e.g., the discrete Bilal distribution) follow the identical structure, replacing $C_i$1 by the pmf and $C_i$2 by the discrete survival function. For mixture (cure) models, an extra mass at $C_i$3 is accommodated: the likelihood contributions for censored cases account for the sum of the cure fraction and the standard survivor function (Freitas et al., 2021, Duembgen et al., 2013).

Nonparametric and Log-Concave Density Estimation

In nonparametric settings with log-concave density constraints, the log-likelihood accommodates a subprobability density $C_i$4 and an atomic mass at infinity $C_i$5. The observed-data log-likelihood is

$C_i$6

The maximization occurs over all concave $C_i$7 suitably normalized, and an EM algorithm is often implemented for numerical optimization (Duembgen et al., 2013). Consistency and uniqueness of the nonparametric MLE under mild conditions are established for this setting.

4. Multivariate and Copula-Based Right-Censored Log-Likelihoods

In clustered survival data settings, the right-censored log-likelihood extends to parametric copula models for the dependence structure. For bivariate clusters, the contribution to the log-likelihood depends on the censoring indicators and involves partial derivatives of the copula function and its density. Specifically, for cluster $C_i$8 with margins transformed to $C_i$9, the log-likelihood combines log-copula density and log-copula partial derivatives according to censoring patterns (Geerdens et al., 2016): $T_i = \min(X_i, C_i)$0 Here, all $T_i = \min(X_i, C_i)$1 (for general cluster size $T_i = \min(X_i, C_i)$2) combinations of censoring status are represented with appropriate contributions. Marginal and joint estimation is often performed via local (kernel-weighted) maximization for conditional models.

5. Right-Censored Log-Likelihood in Modal Regression

Recent parametric modal regression frameworks explicitly accommodate right-censoring by reparameterizing densities in terms of the conditional mode, for instance in the Gamma and Weibull families. The censored log-likelihood is maximized as a function of regression parameters linked directly to the mode, and asymptotic inference is based on the observed Fisher information (Galarza et al., 7 Mar 2026): $T_i = \min(X_i, C_i)$3 where $T_i = \min(X_i, C_i)$4 is the mode as a function of covariates, and parameters $T_i = \min(X_i, C_i)$5 are analytically re-expressed in terms of $T_i = \min(X_i, C_i)$6 and dispersion.

This methodology allows for direct conditional mode modeling and provides inference tools (score equations, observed/expected information) compatible with right-censored data.

6. Practical Applications and Computational Aspects

The right-censored log-likelihood is central to survival analysis in biomedical cohort studies, actuarial science (insurance loss modeling), reliability theory, and clustered time-to-event inference. Maximization is generally performed via Newton–Raphson or EM algorithm, depending on model complexity and parametric versus nonparametric context (Freitas et al., 2021, Duembgen et al., 2013). Penalized likelihood approaches employ the right-censored likelihood as a criterion, and model selection can utilize likelihood-ratio-based statistics tailored to censoring (Zhao et al., 2024, Geerdens et al., 2016).

In the context of large datasets or complex models (e.g., cure models, copulas, or high-dimensional covariates), specialized numerical routines—such as active-set algorithms for log-concave densities or kernel-weighted local likelihood maximization in conditional copula models—are standard (Duembgen et al., 2013, Geerdens et al., 2016).

Summary Table: Canonical Right-Censored Log-Likelihood Forms

Data Type / Model	Log-Likelihood Expression	Reference
Univariate parametric/semiparam.	$T_i = \min(X_i, C_i)$7	(Zhao et al., 2024)
Regression	$T_i = \min(X_i, C_i)$8	(Matthews et al., 10 Apr 2025)
Discrete/cure models	$T_i = \min(X_i, C_i)$9	(Freitas et al., 2021)
Nonparametric log-concave	$\delta_i = \mathbf{1}\{X_i \leq C_i\}$0	(Duembgen et al., 2013)
Copula (clustered)	$\delta_i = \mathbf{1}\{X_i \leq C_i\}$1 log-densities/partial-derivatives of the copula, selected by censoring indicators	(Geerdens et al., 2016)
Modal regression	$\delta_i = \mathbf{1}\{X_i \leq C_i\}$2	(Galarza et al., 7 Mar 2026)

7. Model Assumptions, Mis-specification, and Extensions

A key theoretical requirement is that censoring is non-informative (independent of the event time given covariates and parameters) (Zhao et al., 2024, Matthews et al., 10 Apr 2025). Failure of this assumption invalidates the standard log-likelihood formulation and necessitates explicit modeling of the joint censoring-event mechanism or alternative inferential strategies. Extensions exist for left-truncated and interval-censored data (the right-censoring log-likelihood is retrieved as a special case of more general truncation/censoring likelihoods when the truncation time is set to $\delta_i = \mathbf{1}\{X_i \leq C_i\}$3 or elimination of interval endpoints) (Matthews et al., 10 Apr 2025, Duembgen et al., 2013).

This suggests that methodological developments for right-censored log-likelihood remain active in research, especially as frameworks for covariate-dependent, clustered, or high-dimensional models proliferate, and as censored data arises in new application domains.