Wilcoxon Rank Sum Test

The test evaluates the stochastic equality between two groups. The null hypothesis (H₀) states that a randomly selected value from population A has an equal probability of being greater or less than a randomly selected value from population B. Formally: P(X > Y) = P(Y > X). This is distinct from testing equality of medians unless the distributions are identical in shape. Rejecting H₀ indicates one group tends to produce larger values—a concept known as stochastic dominance.

Unlike parametric tests, the Wilcoxon test does not assume a specific data distribution (e.g., normality). The core assumptions are:

• Independence: Observations within and between groups are independent.
• Ordinal Scale: The variable is continuous or at least ordinal, allowing meaningful ranking.
• Shape: For testing medians specifically, the two distributions must have the same shape and variance. Without this, the test remains valid for stochastic dominance but not for median shift. This robustness makes it a default choice in single-cell RNA-seq where zero-inflated negative binomial data violates normality assumptions.

The test statistic U quantifies the degree of separation between groups. The procedure:

1. Pool and Rank: Combine all N = n₁ + n₂ observations and assign ranks from 1 to N, handling ties with mid-ranks.
2. Sum Ranks: Calculate R₁, the sum of ranks for group 1.
3. Compute U₁: U₁ = R₁ - n₁(n₁ + 1) / 2. The smaller of U₁ and U₂ is compared against a critical value. For large samples, a z-approximation using the mean and variance of U under H₀ provides the p-value. The effect size is often reported as the rank-biserial correlation.

High-throughput expression data frequently contains ties—identical values across conditions, especially for lowly expressed or undetected genes. The standard Wilcoxon test adjusts the variance of the U statistic downward to account for ties, preventing an inflated Type I error rate. The correction factor involves:

• tⱼ: The number of observations tied at rank j.
• Variance Adjustment: Var(U) is reduced by a term proportional to Σ(tⱼ³ - tⱼ). In extreme cases with massive zero-inflation, specialized methods like the two-part hurdle model or the Wilcoxon test on the positive counts only may be more appropriate.

In single-cell analysis, the Wilcoxon Rank Sum Test is a popular default method in tools like Seurat (FindMarkers function). Its advantages in this context:

• Robustness to Outliers: Ranks limit the influence of extreme expression values.
• Zero-Inflation Tolerance: Does not assume a parametric count distribution.
• Computational Speed: Rank-based calculations are fast for large cell-by-gene matrices. However, it treats each cell as an independent replicate, potentially inflating significance in pseudoreplication scenarios. The pseudobulk approach aggregates cells by sample first, then applies the test, to address this.

Statistical significance alone is insufficient for biomarker selection; effect size quantifies the magnitude of difference. The rank-biserial correlation (r) is the recommended effect size for the Wilcoxon test:

• Formula: r = 1 - (2U) / (n₁n₂)
• Interpretation: r ranges from -1 to +1, where 0 indicates complete overlap and ±1 indicates perfect separation. Benchmarks: small (|r| ≈ 0.1), medium (|r| ≈ 0.3), large (|r| ≈ 0.5).
• Relationship to AUC: r is mathematically equivalent to 2 × AUC - 1, where AUC is the area under the receiver operating characteristic curve, directly linking it to classifier performance.

The fundamental distinction driving test selection in genomics:

• Parametric tests (e.g., t-test, ANOVA) assume data are drawn from a specific distribution, typically a normal distribution, and estimate population parameters like the mean and variance.
• Non-parametric tests (e.g., Wilcoxon Rank Sum) operate on ranked data and test hypotheses about the distribution's shape or location without distributional assumptions.
• Key advantage: The Wilcoxon test is robust to outliers and heavy-tailed distributions common in single-cell RNA-seq data, where a few cells may exhibit extreme expression values.
• Trade-off: When data are truly normal, parametric tests have slightly higher statistical power.