Censoring Unbiased Transformations

• Censoring Unbiased Transformations are mathematical mappings and algorithmic procedures that recode incomplete data to yield unbiased statistical inferences under right-censoring and truncation.
• They are widely applied in survival analysis, causal effect estimation, fairness correction, and regression with imperfect labels using approaches like IPCW, pseudo-observations, and doubly robust methods.
• These transformations enable consistent and efficient estimation by adjusting observed outcomes through weighting, augmentation, or recoding to match true counterfactual targets.

Censoring unbiased transformations are mathematical mappings, estimation frameworks, and algorithmic procedures constructed to ensure that statistical inference remains unbiased in the presence of censoring, typically right-censoring, truncation, or coarsened data. These transformations underlie modern approaches for survival analysis, causal effect estimation, fairness under censorship, and regression with imperfect labels. Foundationally, they correct for missingness (often not-at-random) by recoding, weighting, or otherwise transforming the data so that standard modeling tools remain consistent, efficient, and diagnostic-valid, even when direct observation of the response is incomplete.

1. Foundational Principles and Definitions

Censoring unbiased transformations (CUTs) are constructed so that the transformed outcomes possess conditional expectations equal to the unobserved target functionals (such as survival probabilities, cumulative incidence, or functional of a full path) under minimal and explicit assumptions. In the prototypical right-censored failure time setting, for observed $(T, C, \Delta)$ with event time $T$, censoring time $C$, and $\Delta = 1\{T \leq C\}$, a transformation $Y^*$ is said to be "censoring-unbiased" if, for any covariates $X$,

$$\mathbb{E}\left[Y^* \mid X\right] = \mathbb{E}[Y \mid X],$$

where $Y$ is the (not-always-observed) outcome of inferential interest (Xu et al., 2024, Sandqvist, 2024).

CUTs generalize classical techniques such as inverse-probability of censoring weighting (IPCW), Buckley-James (BJ) pseudo-observations, and doubly robust augmentation schemes, and can be constructed for both pointwise (events, survival), functional (restricted mean), and dynamic (functional of a trajectory) targets (Sandqvist, 2024). In modern implementations, they serve as "pseudo-outcomes" for plug-in estimation using ordinary regression, boosting, or modern machine learning algorithms, with robust finite-sample and oracle properties.

2. Classes of Censoring Unbiased Transformations

Several classes of censoring unbiased transformations have been established, each tied to a particular inferential target and censoring structure.

2.1. IPCW and Pseudo-observations

For estimation of conditional survival probabilities or restricted mean survival time, classical IPCW-type transformations include:

• IPCW-1: $$Y^{S,\mathrm{IPCW1}}(t; G) = I(\widetilde{T} > t) / G(t \mid X)$$, where $G$ is the time-varying conditional survival probability for censoring; $\widetilde{T} = \min(T, C)$ (Xu et al., 2024).
• IPCW-2: $$Y^{S,\mathrm{IPCW2}}(t; G) = \Delta I(\widetilde{T} > t) / G(\widetilde{T}^-\mid X)$$ (pseudo-observation form).

2.2. Buckley-James-Type and Doubly Robust Cut

• Buckley-James (BJ) type: $Y^{S,\mathrm{BJ}}(t; S)$ is derived using survival function plug-ins and censored data augmentation (Xu et al., 2024).
• Doubly robust (AIPCW) CUT: $Y^{S,\mathrm{AIPCW}}$ augments the IPCW with a correction term to attain double robustness, i.e., unbiasedness if either the survival or censoring model is correctly specified (Xu et al., 2024, Sandqvist, 2024).

2.3. Extensions

Analogous transformations are available for:

• Competing risks (cause-specific, subdistribution functionals): CUTs extend to cumulative incidence function estimation via cause-specific or Aalen-Johansen-based constructions (Xu et al., 2024).
• Trajectory and general functionals: The doubly robust construction can handle path-dependent or cumulative functionals, not just the first event time, by re-expressing the estimating equations in counting process notation (Sandqvist, 2024).

These transformations can be summarized in the table below:

Transformation Type	Key Formula	Unbiasedness Guarantee
IPCW	$I(\widetilde{T} > t)/G(t|X)$	Correct $G$
BJ	$$[S(t|X) - \Delta I(\widetilde{T} \le t) S(t|X)]/S(t\wedge\widetilde{T}|X)$$	Correct $S$
AIPCW (Doubly Robust)	$I(\widetilde{T} > t)/G(t|X)$ plus augmentation term	Correct $S$ or $G$

3. Statistical Properties and Oracle Guarantees

CUTs are constructed to yield unbiased estimation under specified conditions:

• Unbiasedness: If the relevant nuisance function (e.g., $G$ or $S$) is consistently estimated, the expectation of the CUT equals the counterfactual target (Xu et al., 2024, Sandqvist, 2024).
• Double Robustness: Doubly robust transformations are unbiased if either the event process or the censoring mechanism is correctly modeled (Sandqvist, 2024).
• Oracle Efficiency: With sample splitting, cross-fitting, and appropriately convergent nuisance estimators, plug-in regression on the pseudo-outcomes matches the oracle efficiency of uncensored procedures, even under arbitrary machine learning fits in the first stage (Sandqvist, 2024).
• Finite-sample bounds: The excess risk for regression or treatment effect estimation is tightly bounded by the oracle ensemble’s risk plus terms that vanish as nuisance estimation improves, detailed through learner-specific oracle inequalities (Xu et al., 2024).

4. Algorithmic Implementation and Application Workflows

Adaptations to practical data analysis settings follow a structured pipeline:

1. Nuisance estimation: Fit (possibly via ML) flexible models for the event process (survival, hazard) and/or the censoring distribution. For AIPCW, estimate both.
2. Transformation computation: Compute the CUT for each observation, using data-adaptive estimates for $S$ and/or $G$. For time-varying covariates, appropriately impute or restrict $S$ and $G$ to observed histories (Bartels et al., 2021, Sandqvist, 2024).
3. Regression or learning: Use CUTs as outcomes in standard regression, forest, or boosting algorithms. Double cross-fitting is recommended to avoid dependence between training and nuisance estimation (Xu et al., 2024).
4. Aggregation/oracle selection: Cross-validated selection, stacking, or convex-ensemble approaches yield model-averaged estimates with upper-bounded risk (Xu et al., 2024, Sandqvist, 2024).
5. Diagnostics and inference: The asymptotic normality of regression on CUTs justifies conventional confidence intervals and model selection criteria (Sandqvist, 2024, Hothorn et al., 2015).

In latent variable or endogeneity settings (e.g., instrumental variable survival analysis), strictly monotonic parametric transformations (e.g., Yeo-Johnson family) are used to induce sub-Gaussian error terms and homoscedasticity for consistent two-stage estimation (Willems et al., 2024).

5. Corrections in Fairness, Simulation, and Predictive Validity

CUTs play a central role in fairness-aware learning and simulation-based model checking under censorship:

• Fairness with Right-Censoring: Score transformations (random forest or re-ranking) "equalize" risk scores across protected groups even in the presence of censored labels, via empirical CDF estimation over comparable pairs (Zhang et al., 2022). These are unbiased under right-censoring and support constrained optimization for group fairness metrics such as Concordance Imparity (CI) and Fair Calibration (FC).
• Simulation and VPC Correction: In simulation-based visual predictive checks (VPCs), forced truncation at observed censoring times generically induces bias. Inverse probability of censoring (IPoC) weighted Kaplan-Meier estimators restore unbiased survival curves in each replicate. Marginal model resimulation further recovers variance properties (Bartels et al., 2021).
• Fairness by Orthogonalization: In the context of fair machine learning, orthogonal-to-bias (OB) transformations "censor out" the linear component of feature vectors correlated with sensitive attributes, achieving counterfactual fairness when (A, B) are jointly normal and Cov(A, B) = 0 (Chen et al., 2024).

6. Extensions: Likelihood-based and Transform Model Frameworks

Likelihood-based transformation models and optimal transformation approaches provide an efficient and general toolbox:

• Most Likely Transformation (MLT) Models: A monotonic parametric function $h(y|x)$ maps outcomes to a known error distribution (e.g., normal or logistic). Score equations for $h$ remain unbiased when integrating over censoring intervals; hence MLT estimation under arbitrary censoring/truncation is consistent and asymptotically normal (Hothorn et al., 2015).
• Optimal Transformation for Distributional Regression: In cases of discrete, interval-censored, or truncated response, a global monotonic transform $g(y)$ is fit to align marginal CDFs, preserving unbiasedness in finite samples. An alternating minimization over $g, f, s$ (location, scale) achieves internal consistency even with heavy data imperfections (Friedman, 2020).

7. Limitations, Assumptions, and Future Directions

The utility and validity of censoring unbiased transformations rely on several key assumptions:

• Coarsening at Random (CAR)/Independent Censoring: The censoring process must be conditionally independent of the outcome given observed histories and covariates (Sandqvist, 2024, Bartels et al., 2021).
• Correct Model Specification: All classes of CUTs (especially doubly robust forms) require consistent estimation of at least one source (event or censoring process), or else residual bias remains (Sandqvist, 2024).
• Numerical Stability and Finite Sample Practicalities: In IPCW-type weights, instability arises when estimated survival probabilities approach zero; regularization or truncation of extreme weights is often required (Bartels et al., 2021).
• Complex Data Types: Extensions to settings with dependent censoring, non-random truncation, dynamic covariates, and hidden confounding necessitate further development of transformation strategy, including joint modeling and control-function extensions (Willems et al., 2024).
• Fairness Orthogonalization Assumptions: Linear OB transformations guarantee counterfactual fairness only under joint normality and additive SCMs; strong nonlinearity or non-Gaussianity may result in only approximate fairness (Chen et al., 2024).

Ongoing research aims to generalize these transformations to more diverse data structures (competing risks, recurrent events, high-dimensional and deep learning settings), provide robust diagnostics for model misspecification, and operationalize these tools within automated machine learning pipelines. The theoretical framework unifies and extends classical missing data techniques, yielding scalable, interpretable, and provably unbiased methodologies for censored data across modern statistical science and machine learning (Hothorn et al., 2015, Xu et al., 2024, Sandqvist, 2024).