Wilcoxon Signed-Rank Test vs Paired t-Test: When to Use Each

Use the paired t-test when you compare two related measurements (before/after, left/right, matched pairs) and the DIFFERENCES are approximately normal — it tests whether the mean difference is zero. Use the Wilcoxon signed-rank test, its non-parametric counterpart, when the differences are skewed, have outliers, are ordinal, or your sample is small enough that normality cannot be trusted — it tests whether the distribution of differences is symmetric about zero (in practice, whether the median difference is zero). Both require that pairs are independent of one another. The trade-off is simple: when normality holds, the t-test has slightly more power; when it does not, the t-test can be badly misled by a single outlier while the signed-rank test barely flinches. The signed-rank test keeps about 95% of the t-test’s power under normality, so the cost of choosing it defensively is small.

Collapse each pair to a single difference d = (after − before). You now have one sample of differences. Compute the mean difference d̄ and the standard deviation of the differences s_d. The test statistic is t = d̄ / (s_d / √n) with df = n − 1. Compare to the t-distribution. The only normality assumption is on the differences — not on the original two measurements, which is a point students routinely get wrong. Check it with a histogram or Q-Q plot of the d values. With n ≥ ~30 the Central Limit Theorem makes the t-test fairly forgiving; with n = 8 a single skewed difference can dominate d̄ and s_d.

Step 1: compute each difference d. Step 2: discard any pair with d = 0 (a zero carries no directional information) and reduce n accordingly. Step 3: rank the ABSOLUTE differences from smallest to largest, assigning average ranks to ties. Step 4: reattach the sign of each difference to its rank. Step 5: sum the positive ranks (W+) and the negative ranks (W−). The test statistic W is typically the smaller of W+ and W−. For small n compare W to a signed-rank critical-value table; for n larger than about 20 use the normal approximation: mean = n(n+1)/4, standard deviation = √[n(n+1)(2n+1)/24], so z = (W+ − n(n+1)/4) / sd, with a continuity correction of 0.5 improving accuracy.

Ten patients have systolic BP measured before and after a 6-week program. The ten differences (after − before, mmHg) are: −8, −12, −5, −15, −3, −10, −7, −9, −6, −11. The differences look roughly symmetric with no wild outlier, so the paired t-test is reasonable. d̄ = −8.6 mmHg. s_d ≈ 3.57. SE = 3.57/√10 = 1.13. t = −8.6 / 1.13 = −7.61 with df = 9. The two-tailed p-value is far below 0.001, so the program produced a statistically significant mean reduction of about 8.6 mmHg. A 95% CI for the mean change is −8.6 ± 2.262 × 1.13 = (−11.2, −6.0) mmHg — clinically meaningful, not just statistically significant.

Now eight patients rate pain on a 0–10 ordinal scale before and after treatment. Differences (after − before): −2, −1, −4, 0, −3, +1, −2, −5. The scale is ordinal and one pair shows a 0, so the signed-rank test fits better than a t-test. Drop the zero (n becomes 7). Absolute differences: 2, 1, 4, 3, 1, 2, 5. Ranks of absolute values (averaging ties): the two 1’s get rank 1.5 each; the two 2’s get rank 3.5 each; 3 gets 5; 4 gets 6; 5 gets 7. Reattach signs: negatives at ranks 3.5, 1.5, 6, 5, 3.5, 7 and the single positive at rank 1.5. W+ = 1.5 (the lone positive). W− = 26.5. W = 1.5. For n = 7 the critical value at α = 0.05 (two-tailed) is 2, and W = 1.5 ≤ 2, so reject the null: pain dropped significantly. Note the t-test would have been questionable here — the data are ordinal, not interval.

Zeros: the standard Wilcoxon procedure discards them, though some implementations (Pratt’s method) keep and rank them — check which your software uses, because the p-value can shift. Ties in absolute differences: assign average ranks and apply a tie correction to the variance in the normal approximation. For effect size, report the matched-pairs rank-biserial correlation r = (W+ − W−) / (W+ + W−), or the standardized r = Z/√N from the normal approximation, where N is the number of non-zero pairs; r around 0.1, 0.3, and 0.5 maps loosely to small, medium, and large. A p-value alone tells you an effect exists; the effect size tells you whether anyone should care.

Snap a photo of a paired-data problem and StatsIQ checks the differences for normality, recommends the paired t-test or the Wilcoxon signed-rank test accordingly, then runs the chosen procedure step by step — including the ranking table, sign attachment, and the normal-approximation z with continuity correction. It also reports the matching effect size so your conclusion is about magnitude, not just significance. This content is for educational purposes only.