Wilcoxon Signed-Rank Test

• Wilcoxon Signed-Rank Test is a nonparametric method that uses ranks and signs of paired differences to determine if the median shifts from zero.
• It offers high efficiency compared to the paired t-test, especially under mixture or non-Gaussian conditions, making it well-suited for heterogeneous data.
• Modern extensions such as smoothed, multivariate, differentially private, and Bayesian variants enhance its applicability in high-dimensional and privacy-sensitive research.

The Wilcoxon Signed-Rank Test is a classical, distribution-free statistical hypothesis test for assessing whether the center of a paired or one-sample distribution shifts away from zero, most commonly applied to before-and-after experiments or paired measurements. It combines sign information with ranks of magnitude, yielding superior efficiency compared to the sign test while remaining robust under broad non-Gaussian conditions. The test has spawned numerous modern extensions, including smoothed, differentially private, Bayesian, and multivariate optimal-transport-based variants, all grounded in its foundational framework.

1. Classical Formulation and Statistical Properties

Given $n$ paired data points $(u_i, v_i)$, the Wilcoxon signed-rank test evaluates the null hypothesis $H_0$: the distribution of differences $d_i = v_i - u_i$ is symmetric about zero. The alternative hypothesis is that the differences are systematically positive (one-sided) or lack such symmetry (two-sided).

Test Statistic:

1. Compute $d_i = v_i - u_i$ and $s_i = \mathrm{sign}(d_i) \in \{-1,0,1\}$.
2. Discard pairs with $d_i = 0$, or in the Pratt variant, retain them with $s_i = 0$.
3. For the nonzero $d_i$, assign absolute ranks $r_i$ (averaging ties).
4. Form the statistic:

$(u_i, v_i)$0

where $(u_i, v_i)$1 is the number of nonzero $(u_i, v_i)$2.

Distribution Under $(u_i, v_i)$3:

Exactly, enumerate all $(u_i, v_i)$4 sign patterns for $(u_i, v_i)$5; asymptotically, $(u_i, v_i)$6 with variance $(u_i, v_i)$7. The continuity-corrected normal approximation is routinely used for $(u_i, v_i)$8.

p-Value Calculation:

• Exact: $(u_i, v_i)$9
• Normal: $H_0$0

This nonparametric test is invariant under strictly monotone transformations and robust to outliers. Under Gaussian shift alternatives, its asymptotic relative efficiency (ARE) to the paired $H_0$1-test is $H_0$2, i.e., with only ~5% power loss under ideal normality—for heavier-tailed distributions, the ARE can exceed 1 (Rosenblatt et al., 2013).

2. Power under Shift and Mixture Alternatives

The standard claim is that, under pure location-shift in a Gaussian population, the $H_0$3-test is optimal, with the Wilcoxon signed-rank test slightly less efficient. However, when multi-modal or mixture alternatives are present—that is, only a fraction $H_0$4 of subjects experience a real effect while the rest remain unaffected—the Wilcoxon signed-rank test can surpass the $H_0$5-test in power.

Consider the mixture model:

$H_0$6

where $H_0$7 is the normal density.

For concentrated subpopulations (small $H_0$8 relative to $H_0$9), the Pitman ARE of Wilcoxon to $d_i = v_i - u_i$0-test can be $d_i = v_i - u_i$1, and practical scenarios (such as clinical trials with heterogeneous response or fMRI group analysis) confirm this efficiency advantage (Rosenblatt et al., 2013). As a rule of thumb, if $d_i = v_i - u_i$2, Wilcoxon is superior.

3. Extensions: Smoothed and Multivariate Signed-Rank Tests

Smoothed Wilcoxon Test

To remedy the discrete “lattice effect” of classical rank tests and to achieve finer distributional approximations, smoothing via kernel estimators is employed (Maesono et al., 2016). Let $d_i = v_i - u_i$3 and $d_i = v_i - u_i$4 a symmetric kernel. Define:

$d_i = v_i - u_i$5

where $d_i = v_i - u_i$6 is the integrated kernel and $d_i = v_i - u_i$7 controls smoothing.

Properties:

• $d_i = v_i - u_i$8 is asymptotically normal after centering and scaling.
• Pitman ARE versus the classical $d_i = v_i - u_i$9 is unity.
• Edgeworth expansions to $d_i = v_i - u_i$0 are available when a 4th-order kernel is used.
• Simulations indicate smoother p-value behavior and preserved power, with $d_i = v_i - u_i$1 recommended.

Multivariate Extension

The univariate Wilcoxon signed-rank test is generalized to arbitrary dimensions via optimal transport (Huang et al., 2023). Given $d_i = v_i - u_i$2 vectors $d_i = v_i - u_i$3, assign signs and ranks by mapping each sample via optimal assignment (with respect to a compact group $d_i = v_i - u_i$4 and a reference measure $d_i = v_i - u_i$5), inducing multivariate signed ranks.

The Generalized Wilcoxon Signed-Rank (GWSR) statistic

$d_i = v_i - u_i$6

retains exact distribution-freeness under the null and achieves ARE $d_i = v_i - u_i$7 relative to Hotelling’s $d_i = v_i - u_i$8 under broad shift alternatives, with local maximin optimality available via appropriate score transformation. Implementation requires solving an assignment problem (Hungarian/Auction method, $d_i = v_i - u_i$9) and is suitable for general notions of symmetry (central, sign, spherical).

4. Privacy-Preserving and Bayesian Variants

Differential Privacy

In contexts where dataset summaries can pose disclosure risks, a differentially private Wilcoxon signed-rank test is constructed (Couch et al., 2018). The Pratt variant is used, incorporating zeros ($s_i = \mathrm{sign}(d_i) \in \{-1,0,1\}$0) and ranking among all entries. The $s_i = \mathrm{sign}(d_i) \in \{-1,0,1\}$1 sensitivity is $s_i = \mathrm{sign}(d_i) \in \{-1,0,1\}$2, so Laplace noise of scale $s_i = \mathrm{sign}(d_i) \in \{-1,0,1\}$3 is added to $s_i = \mathrm{sign}(d_i) \in \{-1,0,1\}$4. Since the null distribution of the privatized statistic is a convolution of normal and Laplace, p-values are estimated via Monte Carlo, fully accounting for sampling and added noise.

Empirically, for $s_i = \mathrm{sign}(d_i) \in \{-1,0,1\}$5 and effect size $s_i = \mathrm{sign}(d_i) \in \{-1,0,1\}$6, 80% power is achieved at $s_i = \mathrm{sign}(d_i) \in \{-1,0,1\}$7 for the private test versus $s_i = \mathrm{sign}(d_i) \in \{-1,0,1\}$8 for the public version and $s_i = \mathrm{sign}(d_i) \in \{-1,0,1\}$9--$d_i = 0$0 for prior private approaches. Even at moderate privacy budgets and with up to 30% ties, the loss in power is mild, and practical recommendations include exclusive budget allocation to computation of the privatized statistic and Monte Carlo p-value estimation.

Bayesian Signed-Rank Testing

Bayesian inference is enabled via a latent normal model (Doorn et al., 2017), positing that observed $d_i = 0$1 are rank-revealing projections of $d_i = 0$2, with $d_i = 0$3 the location shift parameter. The observed signs and ranks are enforced as deterministic constraints on the latent variables. The hypotheses are $d_i = 0$4 versus $d_i = 0$5. The Savage-Dickey density ratio yields the Bayes factor:

$d_i = 0$6

A data-augmentation Gibbs sampler (truncated normal steps for $d_i = 0$7, Gaussian updates for $d_i = 0$8, inverse-gamma for $d_i = 0$9) is employed. Bayesian and classical tests are both rank-based and robust; however, the Bayesian approach yields direct evidence quantification and posterior intervals for effect sizes.

5. Practical Guidance and Application Scenarios

• For independent, paired data with a null hypothesis of symmetry about zero, the Wilcoxon signed-rank test is suitable when Gaussianity is dubious or outliers may be present (Rosenblatt et al., 2013).
• The Pratt variant avoids the need to privately estimate the number of nonzero differences and should be used in privacy-constrained analyses (Couch et al., 2018).
• In two-component or mixture alternatives common in heterogeneous responses (e.g., partial clinical trials, fMRI voxels), the signed-rank test can outperform the $s_i = 0$0-test, particularly for small shift/variance ratios (Rosenblatt et al., 2013).
• Smoothing the statistic supports higher-order distributional approximations and reduces discreteness in p-values for moderate samples (Maesono et al., 2016).
• Bayesian and differentially private versions are well-developed, with efficient computational schemes and robust error/power behavior (Doorn et al., 2017, Couch et al., 2018).
• The test extends to the multivariate setting with provable distribution-freeness and asymptotic optimality using optimal transport (Huang et al., 2023).

6. Table of Key Wilcoxon Signed-Rank Variants

Variant	Principle	ARE (vs. parametric)
Classical	Ranks + signs (1D), exact finite-sample	$s_i = 0$1
Smoothed	Kernel smoothing of indicators	$s_i = 0$2
Multivariate (GWSR)	Optimal transport signs/ranks (multi-D)	$s_i = 0$3
Bayesian	Latent normal, posterior/Bayes factor	≈ classical
Differentially Private	Laplace mechanism (Pratt), Monte Carlo p	Close to classical

7. Summary and Theoretical Significance

The Wilcoxon signed-rank framework unifies several branches of nonparametric inference via its central use of symmetric rank-and-sign statistics. Its large-sample efficiency remains remarkably close to classical parametric tests, and under mixture alternatives, it can exceed that efficiency. Modern theoretical advancements—including optimal-transport multivariate generalizations, smoothing for high-order accuracy, Bayesian evidence quantification, and formal differential privacy—demonstrate the flexibility of the signed-rank principle for contemporary methodological demands. Each extension maintains the foundational distribution-freeness and robustness, while targeting specific needs such as high-dimensional inference, privacy, or interpretability in scientific research (Rosenblatt et al., 2013, Maesono et al., 2016, Couch et al., 2018, Huang et al., 2023, Doorn et al., 2017).