Nonlinear-in-Ranks Score Functions

• Nonlinear-in-ranks score functions are rank-based estimators that use nonlinear, bounded, and monotonic transformations to extend traditional Wilcoxon-type methods.
• They are pivotal in robust regression, survival analysis under censoring, and transformation models by providing flexible weighting of empirical ranks.
• Recent advances in their representation, optimality, and computational strategies offer a unified approach to handling censoring, partial observation, and pairwise-rank likelihood.

Nonlinear-in-ranks score functions constitute a class of rank-based estimators and associated score mappings that extend traditional linear (Wilcoxon-type) rank scores to general nonlinear, bounded, and monotone-increasing transformations of the empirical or imputed ranks. These constructions play a critical role in robust regression, survival analysis under censoring, transformation models, and semiparametric inference—permitting both flexible weighting of ranks and compatibility with the algebraic and asymptotic frameworks of classical R-estimation. Foundationally, these scores rely on user-specified functions over the residual distribution function, or on optimal transformations of ranks via reference distributions, as in D-rank methods. Recent advances have articulated the representation, optimality, and computational properties of nonlinear-in-ranks scores, providing unified perspectives across censoring, partial observation, and pairwise-rank likelihood formulations (Satten et al., 10 Jan 2026, &&&1&&&, Yu et al., 2021).

1. Definition and Key Properties of Nonlinear-in-Ranks Score Functions

Let $\psi : [0,1] \rightarrow \mathbb{R}$ denote a “score” or weight function defined on the CDF scale of residuals or covariates. Essential properties enforced in recent literature (Satten et al., 10 Jan 2026) are:

• $\psi$ is continuous on $[0,1]$ and continuously differentiable on $(0,1)$,
• $\psi$ is monotone increasing,
• $\psi$ is bounded: $\sup_{u\in[0,1]}|\psi(u)|<\infty$.

For each score function $\psi$, define its primitive $A(u) = \int_0^u \psi(s)\,ds$, on $0 \le u \le 1$. In practical construction, special cases are:

Score Family	$\psi(u)$	$A(u)$ (Primitive)
Wilcoxon (Linear-in-ranks)	$u$	$u^2/2$
Logrank (Extreme-value)	$-1-\log(1-u)$	$-(1-u)\log(1-u)$
Generalized F	$m_1m_2\frac{F^{-1}(u)-1}{m_2+m_1F^{-1}(u)}$	Bounded as $[-m_1, m_2]$

D-rank (distribution-guided) scores are defined as moments of order statistics of a reference law $Z$, that is, $S_n(r) = \alpha_{(r:n)} = \mathbb{E}[Z_{(r:n)}]$, where $Z_{(r:n)}$ is the $r$th order statistic of i.i.d. reference variables (Kim et al., 2017). For large $n$, these scores track quantiles $Q_Z(p_r)$ of $Z$.

2. Estimating Equations and Censored Data R-Estimation

For linear models under right-censored outcomes, observed data $(Y_i, \Delta_i, X_i)$, residuals $e_{i,\beta} = Y_i - X_i^T\beta$ are used with the Kaplan–Meier “self-consistent” estimator of survivor function $\hat{S}_\beta$, yielding a distribution function $\hat{F}_\beta = 1 - \hat{S}_\beta$. Rank-scores for each observation are computed by the following mechanisms (Satten et al., 10 Jan 2026):

• Generalized nonlinear-in-ranks score function for censored data is:

$$R_i^{(\psi)}(\beta) = \Delta_i\;\gamma_\psi(e_{i,\beta};\hat{F}_\beta) + (1 - \Delta_i)\;\Gamma_\psi(e_{i,\beta};\hat{F}_\beta),$$

where

$$\gamma_\psi(t; H) = \frac{A(H(t)) - A(H(t^-))}{H(t) - H(t^-)} \text{ or } \psi(H(t))$$

$$\Gamma_\psi(t; H) = \frac{A(1) - A(H(t))}{\hat{S}_\beta(t)}$$

The R-estimator $\hat{\beta}$ solves the vector equation $$\Psi_n(\beta; \psi) = \sum_{i=1}^n R_i^{(\psi)}(\beta) (X_i - \bar{X}) = 0$$.

In the uncensored, linear-in-ranks case, these formulas collapse to imputed mid-rank score assignments. The nonlinear construction generalizes the mid-CDF imputation [see Section 5].

3. Equivalence to Classical and Weighted Log-Rank Classes

Nonlinear-in-ranks score functions provide algebraic equivalence to established classes in robust regression with censored data:

• Ritov’s G-class: The estimating equation $S_{n,R}(\beta; \gamma)$ matches the nonlinear rank-score estimator with $\gamma(u) = \gamma_\psi(u; \hat{F}_\beta)$ (Satten et al., 10 Jan 2026).
• Tsiatis’ weighted log-rank: The weighted counting-process representation $S_{n,T}(\beta; w)$ coincides with the nonlinear-in-ranks approach when $$w(u; \beta) = \gamma_\psi(u; \hat{F}_\beta) - \Gamma_\psi(u; \hat{F}_\beta)$$.

This preserves rank-sum invariance and exact representation theorems for a wide class of bounded monotone nonlinear score functions, restoring core properties of linear-in-ranks R-estimation under general nonlinear scoring.

4. Theoretical Optimality and Asymptotics

Nonlinear-in-ranks scores deliver robust statistical efficiency, root-n consistency, and tractable limit distributions. For D-rank score functions (Kim et al., 2017):

• Optimality: Within the location–scale family, D-rank score assignments $\alpha_{(r:n)}$ maximize sample correlation between the transformed scores and responses, as formalized in Theorem 1.
• Asymptotic Normality: Under regularity, for estimator $\widehat{\rho}(s)$ computed over the top $[ns]$ ranks,

$$\sqrt{n}\left(\widehat{\rho}(s) - \rho\right) \xrightarrow{d} N\left(0, \frac{\Psi_\infty^{\mathrm{I}}(s)/\sigma_Y^2 + \rho^2 \Psi_\infty^{\mathrm{II}}(s)}{\Phi_\infty(s)^2}\right)$$

• Martingale Influence Function and Plug-in Variance: In censored R-estimation (Satten et al., 10 Jan 2026), the influence function and variance estimators are constructed via martingale representations and quasi-score plug-in formulas, directly leveraging the weight function $w_{β,\psi}(u)$.

5. Computational Algorithms and Nonlinearity in the Rank Mapping

Practical estimation with nonlinear-in-ranks scores (especially in transformation models) involves the following steps (Yu et al., 2021):

• Evaluate all pairwise differences $v_{ij} = (X_i - X_j)^T\beta$ and associated rank indicators $I(Y_i > Y_j)$.
• Fit $F_\beta$ by isotonic regression (PAVA), treating the projected pairwise differences as the x-axis and rank indicators as y-axis.
• Compute pseudo-score $$\psi_n(\beta) = \frac{1}{n^2}\sum_{i \ne j} (X_i - X_j)\left[I(Y_i > Y_j) - F_\beta(v_{ij})\right]$$.
• Solve $\psi_n(\beta) = 0$ under necessary constraints (e.g., $\|\beta\|_2 = 1$).

The score mapping $\psi_n(\beta)$ is highly nonlinear and piecewise constant, with "thresholds" at hyperplanes in projection space. This combinatorial structure results in discontinuities in score functionals as orderings change.

6. Diagnostics, Specification, and Empirical Implications

Diagnostic analysis for the specification of nonlinear-in-ranks score functions (notably in D-rank models (Kim et al., 2017)) involves:

• Residual analysis: After fitting, examine normalized residuals $e^*_{[r:n]}$ for independence and zero mean.
• Score-vs-residual plots: Lack of pattern suggests correct specification of reference distribution $Z$.
• Comparison of residual sum of squares and fitted-intercept tests across candidate score families enables empirical selection of the optimal nonlinear scoring function.

Empirical illustrations (e.g., top-30 Internet stock ranks versus next-day returns) demonstrate reduced MSE and improved goodness-of-fit when nonlinear score functions appropriately match the latent covariate distribution (Kim et al., 2017).

7. Comparison with Linear-in-Ranks and Related Methods

Nonlinear-in-ranks scores generalize Wilcoxon-type linear-in-rank scores ($\psi(u)=u$). The construction restores all algebraic identities in the linear case, while accommodating bounded, monotone, and flexible nonlinearity. Unlike pure rank-correlation methods or kernel approaches:

• Full pairwise information and covariate differences are exploited (Yu et al., 2021).
• There is no need for smoothing parameters or bandwidth selection.
• Optimal weights and asymptotic efficiency can be tuned or derived from reference families or empirical likelihood.
• Semi-parametric or nonparametric model structures can be incorporated with robust inference.

In summary, nonlinear-in-ranks score functions provide a mathematically rigorous and computationally tractable extension of rank-based estimation, supporting robust, optimal inference across regression, survival, and transformation modeling contexts (Satten et al., 10 Jan 2026, Kim et al., 2017, Yu et al., 2021).