---
title: Smoothed Wilcoxon Rank Scores
url: https://www.emergentmind.com/topics/smoothed-wilcoxon-rank-scores
type: topic
---

# Smoothed Wilcoxon Rank Scores

The smoothed Wilcoxon rank scores refer to a family of nonparametric statistics and estimators in which the classical discrete rank indicators in Wilcoxon-type tests are replaced with smooth (kernel-based) functions of the data. This approach yields statistics that are continuous with respect to the data, inherit the fundamental distribution-free properties of Wilcoxon procedures, and offer practical benefits in terms of handling ties and improving efficiency under monotone but non-Gaussian associations. The method has been developed in several directions, including robust correlation estimation, one-sample and two-sample location inference, and hypothesis testing, providing a high-accuracy approximation to orthodox signed-rank and rank-sum procedures [2511.08979], [1610.02145], [1704.07977].

## 1. Smoothed Empirical Cumulative Distribution Functions and Kernelifying Ranks

The core step is the replacement of the empirical cumulative distribution function (ecdf) with a smoothed or “kernelized” ecdf. The classical ecdf for a sample $\{X_j\}_{j=1}^n$ is
\[
F_n(x)=\frac{1}{n}\sum_{j=1}^n\mathbf{1}\{X_j\leq x\}.
\]
The smoothed version substitutes the indicator with a continuous cumulative distribution function (CDF) $H$, typically a kernel CDF such as the standard normal:
\[
\widetilde F_n(x)=\frac{1}{n}\sum_{j=1}^n H\left(\frac{x - X_j}{h}\right),
\]
with bandwidth $h=h_n>0$ satisfying $h_n\to 0$, $n h_n\to\infty$, $n h_n^4\to 0$ as $n\to\infty$. For each sample point $X_i$, the smoothed rank is then
\[
\widetilde R_i = n\,\widetilde F_n(X_i) = \sum_{j=1}^n H\left(\frac{X_i - X_j}{h}\right).
\]
Setting $h\to 0$ recovers the integer-valued ranks. Thus, the smoothing operation produces real-valued, tie-robust ranks that approach classical ranks in the limit [2511.08979].

## 2. Construction of Smoothed Wilcoxon Rank Scores and Correlation Estimators

The Wilcoxon linear score function for rank $r\in\{1,\dots,n\}$ is
\[
a(r) = \sqrt{12}\left(\frac{r}{n+1} - \frac{1}{2}\right),
\]
with analogous extension to the smoothed case:
\[
a(\widetilde R_i) = \sqrt{12}\left(\frac{\widetilde R_i}{n+1} - \frac{1}{2}\right).
\]
These scores are used to build generalized inner-product statistics. For estimating rank correlations, the smoothed Wilcoxon correlation estimator is
\[
\widehat\rho_{sa} = \frac{1}{s_a} \sum_{i=1}^n a(\widetilde R_i^X)a(\widetilde R_i^Y), \qquad s_a = \frac{n(n-1)}{n+1},
\]
which, after algebraic manipulation, is equivalent to the classical Spearman correlation but evaluated on smoothed ranks:
\[
\widehat\rho_{sa}
= 
\frac{\sum_{i=1}^n (\widetilde R_i^X - \frac{n+1}{2})(\widetilde R_i^Y - \frac{n+1}{2})}{n(n^2-1)/12}.
\]
This approach can be interpreted as a "smoothed Spearman-type estimator" or a continuous extension of Wilcoxon’s statistic, handling ties and preserving the nonparametric spirit [2511.08979].

## 3. Smoothed Wilcoxon-Type Tests for One-Sample and Two-Sample Problems

In the one-sample signed-rank scenario, the smoothed Wilcoxon statistic for a sample $\{X_i\}$ symmetric about $0$ is
\[
W_n^* = \sum_{1\le i<j\le n} K\left(\frac{X_i+X_j}{2h_n}\right) + \frac{1}{2}\sum_{i=1}^n K\left(\frac{2X_i}{2h_n}\right),
\]
where $K(u)$ is a kernel CDF. Under the null, the mean and variance match the classical statistic up to $O(n h_n^2)$ and the leading order does not depend on the parent distribution [1610.02145].

For two-sample inference, the discrete sum in the Wilcoxon rank-sum statistic
\[
W_2 = \sum_{i=1}^m\sum_{j=1}^n \mathbf{1}\{Y_j > X_i\}
\]
is replaced with its smoothed analogue:
\[
\widetilde W_2 = \sum_{i=1}^m\sum_{j=1}^n K\left(\frac{Y_j - X_i}{h}\right).
\]
The key effect is that the statistic becomes real-valued, its distribution under the null is close to normality (enabling accurate normal approximation), and it avoids the lattice-related discreteness artifacts that distort $p$-values in small samples [1704.07977].

## 4. Asymptotic Properties and Efficiency

Across all smoothed Wilcoxon variants, asymptotic expectations and variances under the null hypothesis are free of the underlying distribution to first order. For the smoothed Spearman-type rank correlation estimator $\widehat\rho_{sa}$:
- Under independence, $E[\widehat\rho_{sa}] = 0$ and $\operatorname{Var}[\widehat\rho_{sa}]\sim 1/(n-1)$.
- More generally, for a fixed value of the true association parameter $\rho$, a CLT holds:
\[
\sqrt{n}\left(\widehat\rho_{sa} - \rho\right) \rightsquigarrow N\left(0,\,\sigma_{sa}^2(\rho)\right).
\]
- The asymptotic variance for Wilcoxon linear scores is strictly smaller than for classical Spearman’s $\rho$ under many monotonic but non-Gaussian settings; simulated MSE reduction up to $10$–$50\%$ is observed [2511.08979].

For the smoothed Wilcoxon signed-rank and rank-sum tests, the Pitman asymptotic relative efficiency (ARE) with respect to their classical analogues is $1$; the two statistics are asymptotically equivalent:
\[
Z_n^* - Z_n \stackrel{p}{\longrightarrow} 0.
\]
Refined Edgeworth expansions with remainder $o(n^{-1})$ are available, leading to highly accurate $p$-value approximations even for moderate sample sizes [1610.02145], [1704.07977].

## 5. Handling of Ties and Robustness to Data Discreteness

Smoothed Wilcoxon rank scores automatically handle ties via the kernel function. If $X_i = X_j$, then $H(0)=1/2$, ensuring both observations receive the same, non-integer smoothed rank without resorting to ad-hoc average or random tie-breaking. This feature eliminates small bias present in classical rank-based methods [2511.08979]. In the two-sample context, smoothing removes gaps in attainable $p$-values caused by the discreteness of the rank-sum statistic, yielding continuous $p$-values with accurate calibration [1704.07977].

## 6. Implementation: Kernel and Bandwidth Choices

The choice of kernel and bandwidth is central for the practical performance of smoothed Wilcoxon procedures:
- The kernel $k$ should be symmetric, typically of higher order (e.g., 4th-order) to eliminate $O(n^{-1/2})$ bias in the Edgeworth expansion for $p$-values.
- Bandwidth $h$ must satisfy $h_n\to 0$, $n h_n\to\infty$ for asymptotic normality; typical choices include $h_n = n^{-1/4}$, $h_n = n^{-1/3}$, or $h_n = n^{-1/3}(\log n)^{-1}$ for refined Edgeworth expansions [1610.02145].

A practical computation path:
1. Construct the smoothed ecdf $\widetilde F_n(x)$ with kernel $H$ and bandwidth $h$.
2. Compute smoothed ranks $\widetilde R_i$ for all data points.
3. For correlation: apply Wilcoxon linear scores and form the inner-product estimator.
4. For tests: compute the smoothed sum and studentize according to the limiting variance.
5. Approximate $p$-values using the normal (or Edgeworth-corrected) distribution [2511.08979], [1610.02145], [1704.07977].

## 7. Applications and Empirical Efficiency Gains

Simulation studies document that under data with strong monotone but non-Gaussian association, the smoothed Wilcoxon correlation estimator outperforms both classical Spearman $\rho$ and Kendall $\tau$, reducing MSE by $10$–$50\%$ while matching performance under Gaussian data [2511.08979]. In testing scenarios, smoothed procedures exhibit empirical size close to nominal significance levels and avoid biases seen with classical Wilcoxon tests. Under heavy-tailed alternatives, smoothed medians can outperform the smoothed rank-sum, while under light-tailed alternatives, the smoothed Wilcoxon rank scores exhibit optimal power [1610.02145], [1704.07977].

In summary, the smoothed Wilcoxon rank score constructions offer a principled nonparametric approach yielding continuous, tie-robust, asymptotically normal statistics. They preserve efficiency and distribution-free properties while resolving issues associated with data discreteness and bias in classical rank-based methods [2511.08979], [1610.02145], [1704.07977].

Source: https://www.emergentmind.com/topics/smoothed-wilcoxon-rank-scores