What Are Parametric and Nonparametric Tests?
The word "parametric" refers to a population parameter — a fixed numerical characteristic of a distribution, like the mean. A parametric test assumes the data follow a distribution that can be fully described by those parameters. The normal distribution, for example, is completely described by its mean and standard deviation. When those assumptions hold, parametric tests are the most statistically efficient choice.
"Nonparametric" does not mean assumption-free. It means the test avoids assumptions about the distributional form of the data. These methods convert raw values to ranks before analysis, which makes them insensitive to the exact shape of the distribution and more resistant to the influence of outliers.
Parametric tests ask: "Is the population mean different?" Nonparametric tests typically ask: "Is one group's distribution shifted relative to another?" or "Is one group's median different?" The hypotheses are similar but not identical.
The Four Assumptions of Parametric Tests
Every parametric test rests on a set of assumptions about the data. Violating those assumptions — particularly with small samples — can produce p-values that are either too small or too large, leading to incorrect conclusions. There are four core assumptions to check before running any parametric test.
Parametric Test Assumptions
- Normality. The dependent variable is approximately normally distributed in the population (or within each group). For large samples (n > 30), the Central Limit Theorem makes this assumption less critical.
- Homogeneity of variance. For two-sample and multi-group tests, the variance should be roughly equal across groups. Test with Levene's test before running ANOVA or an independent samples t-test.
- Independence. Each observation must be independent of every other. This is violated in repeated-measures designs unless you explicitly account for it.
- Continuous measurement. The outcome is measured on an interval or ratio scale — not ordinal or nominal categories.
Nonparametric Assumptions
- Independence. Observations must still be independent within and between groups (or appropriately paired in the paired case). This assumption is shared with parametric tests.
- Ordinal or continuous data. Most nonparametric tests require the variable to be at least ordinal — ranks need to be meaningful. They do not require interval or ratio measurement.
- Continuous distribution (for some tests). The Mann-Whitney U and Wilcoxon tests technically assume continuous underlying distributions to rule out ties, though they handle ties reasonably well in practice.
- No distributional form required. You do not need to assume normality, homogeneity of variance, or any other distributional shape.
How to Test the Normality Assumption
Before choosing a parametric test, it makes sense to check whether your data are approximately normal. The two most widely used formal tests are the Shapiro-Wilk test (best for n < 50) and the Kolmogorov-Smirnov test (better suited to larger samples). Visually, a Q-Q (quantile-quantile) plot or a histogram can reveal departures from normality at a glance.
| Normality Test | Best Sample Size | How to Interpret |
|---|---|---|
| Shapiro-Wilk test | n < 50 (most reliable) | p < 0.05 → evidence against normality; consider nonparametric |
| Kolmogorov-Smirnov test | n ≥ 50 | p < 0.05 → distribution differs from normal; use with caution on large n |
| Anderson-Darling test | Any size | p < 0.05 → non-normal; sensitive to tail behavior |
| Q-Q plot | Any size | Points fall near the diagonal line → approximately normal |
| Histogram | n ≥ 20 for visual reliability | Bell-shaped curve → approximately normal; skewed or bimodal → non-normal |
A non-significant Shapiro-Wilk test (p ≥ 0.05) does not prove normality — it means you lack evidence against it. With very small samples (n < 10), the test has low power and will almost always return p ≥ 0.05 regardless of the true distribution. Base your decision on visual inspection too, and consider your sample size alongside the test result. See our full guide to normality tests for worked examples.
Parametric vs Nonparametric Tests: Complete Comparison Table
For every common parametric test, there is a nonparametric equivalent that handles the same research question without the normality requirement. The table below maps each parametric test to its nonparametric counterpart and tells you when each is the better choice.
| Parametric Test | Nonparametric Alternative | Design | Use Parametric When | Use Nonparametric When |
|---|---|---|---|---|
| Independent samples t-test | Mann-Whitney U test | 2 independent groups | Data are continuous and approximately normal | Ordinal data, non-normal distribution, or substantial outliers |
| Paired samples t-test | Wilcoxon signed-rank test | 2 related groups (repeated measures) | Differences between pairs are normally distributed | Differences are non-normal or ordinal |
| One-sample t-test | Wilcoxon signed-rank test | 1 group vs hypothesized median | Data are normal, testing mean vs a specific value | Data are skewed, testing median vs a specific value |
| One-way ANOVA | Kruskal-Wallis test | 3+ independent groups | Data are normal with equal variances | Data are non-normal, ordinal, or have unequal variances |
| Repeated measures ANOVA | Friedman test | 3+ repeated measures | Differences across conditions are normally distributed | Ordinal outcome or non-normal differences |
| Pearson correlation | Spearman rank correlation | Linear association between two variables | Both variables are continuous and normally distributed | Ordinal variables, monotonic but non-linear relationship, outliers present |
| McNemar's test (for related proportions) | Chi-square test of independence | Categorical data | Adequate expected cell counts (≥ 5) | Small expected cell counts; Fisher's exact test is the alternative |
Statistical Test Selection: Decision Tree
Work through these questions in order to identify whether a parametric or nonparametric test is appropriate for your data, and which specific test to use.
Step-by-Step Test Selection Guide
When n > 30 per group, the Central Limit Theorem makes the sample mean approximately normal regardless of the underlying distribution. In this case, parametric tests are generally robust to mild non-normality. With n < 15, check the normality assumption carefully — and with n < 10, most statisticians recommend nonparametric methods by default unless normality is well established from prior research on the same variable.
Statistical Power: Which Test Detects Effects Better?
Statistical power is the probability that a test correctly rejects a false null hypothesis — the probability of detecting a real effect when one exists. A test with low power misses real differences; a test with high power finds them.
Under ideal conditions (data are continuous and normally distributed), parametric tests have higher power than their nonparametric equivalents. The table below shows the asymptotic relative efficiency (ARE) — the ratio of sample sizes needed by the nonparametric test compared to the parametric test to achieve the same power.
The key takeaway: you lose very little power by using nonparametric tests even when the normality assumption holds. You lose about 5% efficiency with the Mann-Whitney U compared to the t-test. But when the normality assumption is violated — particularly with heavy-tailed distributions or outliers — nonparametric tests can be more powerful than their parametric equivalents. This was formally demonstrated by Frank Wilcoxon (1945) and later by Henry Mann and Donald Whitney (1947), who showed that rank-based tests preserve power even in difficult conditions.
Original Dataset: Parametric and Nonparametric Results Side by Side
The best way to understand the difference between these approaches is to run both on the same data. The dataset below contains exam scores from 16 students — eight in a traditional teaching group and eight in a flipped-classroom group. The data include one outlier (student T8) that creates a mild non-normal distribution.
| Student | Group | Exam Score (/100) | Rank (all 16) |
|---|---|---|---|
| T1 | Traditional | 62 | 3 |
| T2 | Traditional | 68 | 5 |
| T3 | Traditional | 71 | 7 |
| T4 | Traditional | 65 | 4 |
| T5 | Traditional | 74 | 9 |
| T6 | Traditional | 69 | 6 |
| T7 | Traditional | 58 | 2 |
| T8 | Traditional | 41 | 1 |
| F1 | Flipped | 78 | 11 |
| F2 | Flipped | 83 | 13 |
| F3 | Flipped | 76 | 10 |
| F4 | Flipped | 88 | 15 |
| F5 | Flipped | 79 | 12 |
| F6 | Flipped | 91 | 16 |
| F7 | Flipped | 84 | 14 |
| F8 | Flipped | 73 | 8 |
Parametric Approach: Independent Samples t-Test
Is there a significant difference in mean exam scores between teaching methods?
Hypotheses: H₀: μ_Traditional = μ_Flipped | H₁: μ_Traditional ≠ μ_Flipped
Descriptive stats: Traditional: mean = 63.5, SD = 10.2, n = 8 | Flipped: mean = 81.5, SD = 5.9, n = 8
Shapiro-Wilk on Traditional group: W = 0.832, p = 0.061. The outlier (T8 = 41) creates left skew but p ≥ 0.05 — borderline normality.
Test statistic: t = (63.5 − 81.5) / √[(10.2²/8) + (5.9²/8)] = −18.0 / 4.17 ≈ −4.32 | df ≈ 11 (Welch)
p-value: p = 0.0012 | α = 0.05
✓ Decision: Reject H₀. The flipped classroom group scored significantly higher (mean difference = 18 points, t(11) = −4.32, p = .001).
Nonparametric Approach: Mann-Whitney U Test
Do the two teaching method groups have different score distributions?
Hypotheses: H₀: The distributions are equal | H₁: The Flipped group scores tend to be higher
Rank sum for Traditional group (ranks 1–16 assigned across all 16 scores): Traditional ranks = 1+2+3+4+5+6+7+9 = 37. Expected R₁ = n₁(N+1)/2 = 8(17)/2 = 68.
U statistic: U₁ = R₁ − n₁(n₁+1)/2 = 37 − 36 = 1 | U₂ = n₁·n₂ − U₁ = 64 − 1 = 63 | Use U = min(1, 63) = 1
Critical value at α = 0.05 two-tailed, n₁ = n₂ = 8 from the Mann-Whitney U table: critical U = 13. Since U = 1 < 13 → reject H₀.
✓ Decision: Reject H₀. The flipped classroom group scores are significantly higher (U = 1, p < .001). Both methods reach the same conclusion here, though the nonparametric result is more defensible given the outlier in the traditional group.
The t-test and Mann-Whitney U test can reach different conclusions when outliers heavily distort the group mean. In this dataset, removing T8 (score = 41) would raise the traditional group mean from 63.5 to 66.7 — a 3.2-point shift. The Mann-Whitney U test is unaffected because it works on ranks. If your conclusions depend on a single outlier, the nonparametric result is more trustworthy. See our guide on handling outliers for more detail.
Real-World Case Studies: Choosing the Right Test
Case Study 1 — Healthcare
Clinical Trial: Pain Scores After Two Analgesic Treatments
Researchers measured pain intensity on a 0–10 numeric rating scale (NRS) for 20 patients — 10 receiving Drug A and 10 receiving Drug B. The data were measured once per patient (independent samples). A Shapiro-Wilk test on Drug A's pain scores returned W = 0.801, p = 0.012, indicating non-normality.
Decision: Because the pain scale is ordinal in nature (NRS does not have equal intervals between points) and the normality assumption is violated, the Mann-Whitney U test is the appropriate choice. The Mann-Whitney U test ranks the pain scores across both groups and tests whether one treatment tends to produce lower scores. Researchers in pain research routinely use this method because subjective pain scales are inherently ordinal.
Case Study 2 — Education
Pre-Test and Post-Test Scores: Measuring Learning Gain
A statistics instructor measured student exam scores before and after a six-week intervention. With 25 students and a clear pre/post design, this calls for a paired analysis. The differences between post- and pre-test scores were approximately normal (Shapiro-Wilk p = 0.31).
Decision: The normality assumption is satisfied for the differences (which is what the paired t-test checks). The paired samples t-test is the right choice. If the differences had been skewed — perhaps because most students improved by a small amount but a few made dramatic gains — the Wilcoxon signed-rank test would be preferred.
Case Study 3 — Marketing
Customer Satisfaction Across Three Store Formats
A retail chain asked customers to rate their satisfaction on a 5-point Likert scale (1 = Very Dissatisfied, 5 = Very Satisfied) across three store formats: in-store, curbside pickup, and online. The outcome is clearly ordinal — you cannot assume equal differences between rating categories. Sample size per format: n = 30–45.
Decision: Despite the adequate sample sizes, the ordinal measurement scale rules out ANOVA. The Kruskal-Wallis test is the correct choice. It tests whether the distributions of satisfaction ratings differ across the three store formats. If the test is significant, post-hoc pairwise Mann-Whitney U comparisons (with Bonferroni correction) identify which format pairs differ. Learn more about multiple comparisons in our Bonferroni correction guide.
Case Study 4 — Manufacturing / Quality Control
Product Weight Consistency Across Production Shifts
A food manufacturer weighs 40 randomly selected products per shift across four shifts. The outcome (weight in grams) is continuous and approximately normally distributed within each shift (all Shapiro-Wilk tests: p > 0.18). Levene's test for homogeneity of variance returns p = 0.43.
Decision: Both normality and equal variances hold. One-way ANOVA is the appropriate test. The large sample size (40 per group) and the continuous nature of the outcome make this a clear case for the parametric approach, which will have higher statistical power than the Kruskal-Wallis alternative.
Running the Tests: R, Python, and SPSS
All six major parametric-nonparametric test pairs covered in this guide can be run in a few lines of code. The examples below use the exam score dataset from Section 6 — substitute your own variable names as needed.
# Normality check shapiro.test(traditional) # Shapiro-Wilk # Parametric: independent samples t-test t.test(flipped, traditional, var.equal = FALSE) # Nonparametric: Mann-Whitney U (Wilcoxon rank-sum) wilcox.test(flipped, traditional, alternative = "two.sided") # Paired t-test t.test(post_scores, pre_scores, paired = TRUE) # Wilcoxon signed-rank (nonparametric paired) wilcox.test(post_scores, pre_scores, paired = TRUE) # One-way ANOVA result <- aov(score ~ group, data = df) summary(result) # Kruskal-Wallis (nonparametric ANOVA) kruskal.test(score ~ group, data = df) # Spearman correlation (nonparametric) cor.test(x, y, method = "spearman")
from scipy import stats import numpy as np # Normality check stat, p = stats.shapiro(traditional) print(f"Shapiro-Wilk: W={stat:.4f}, p={p:.4f}") # Parametric: Welch's t-test (unequal variances) t_stat, p_val = stats.ttest_ind(flipped, traditional, equal_var=False) # Nonparametric: Mann-Whitney U u_stat, p_val = stats.mannwhitneyu(flipped, traditional, alternative='two-sided') # Paired t-test t_stat, p_val = stats.ttest_rel(post_scores, pre_scores) # Wilcoxon signed-rank w_stat, p_val = stats.wilcoxon(post_scores, pre_scores) # One-way ANOVA f_stat, p_val = stats.f_oneway(group1, group2, group3) # Kruskal-Wallis h_stat, p_val = stats.kruskal(group1, group2, group3) # Spearman correlation rho, p_val = stats.spearmanr(x, y)
/* Shapiro-Wilk normality test */ EXAMINE VARIABLES=score BY group /PLOT NPPLOT /STATISTICS DESCRIPTIVES. /* Independent samples t-test */ T-TEST GROUPS=group('Traditional' 'Flipped') /VARIABLES=score /CRITERIA=CI(.95). /* Mann-Whitney U (nonparametric) */ NPAR TESTS /M-W=score BY group('Traditional' 'Flipped'). /* Wilcoxon signed-rank (paired) */ NPAR TESTS /WILCOXON=pre_score WITH post_score. /* One-way ANOVA */ ONEWAY score BY group /STATISTICS DESCRIPTIVES HOMOGENEITY. /* Kruskal-Wallis */ NPAR TESTS /K-W=score BY group(1 3). /* Spearman correlation */ NONPAR CORR /VARIABLES=var1 var2 /PRINT=SPEARMAN.
The CHOOSE Framework: A Six-Step Decision Guide
The CHOOSE framework consolidates the entire decision process into six sequential steps. Apply them in order and you will arrive at the right test for your data every time.
C — Check Your Data Type
Is the outcome variable nominal/categorical, ordinal, or continuous (interval/ratio)? Categorical outcomes → chi-square or Fisher's exact. Ordinal data → lean nonparametric. Continuous data → proceed to step H.
H — Handle the Assumptions
Run a Shapiro-Wilk test on each group. If p < 0.05 with a small sample, normality is suspect. Also check Levene's test for equal variances when comparing two or more groups. Document both results before deciding.
O — Observe Your Distribution
Plot a histogram and Q-Q plot for each group. Look for skewness, heavy tails, bimodality, or substantial outliers. Statistical tests and visual inspection together give a more reliable picture than either alone.
O — Outline Your Sample Size
With n > 30 per group, the Central Limit Theorem provides robustness and parametric tests are usually safe even with mild non-normality. With n < 15, assumptions matter more. With n < 10, default to nonparametric unless the variable is known to be normal from prior research.
S — Select the Appropriate Test
Based on steps C through O: how many groups? Independent or paired? Use the decision tree in Section 4 to identify the specific test. Record your reasoning — reviewers and instructors will ask.
E — Explain and Report Results
For parametric tests, report the test statistic (t, F), degrees of freedom, p-value, and effect size (Cohen's d, η²). For nonparametric tests, report the test statistic (U, W, H), sample sizes, p-value, and a rank-based effect size (r = Z/√N). Always write a plain-English conclusion.
Common Misconceptions About Parametric and Nonparametric Tests
| The Misconception | The Reality |
|---|---|
| ❌ "Nonparametric tests have no assumptions." | ✅ They assume independence and usually require ordinal or continuous data. Some also assume a continuous underlying distribution. |
| ❌ "A large sample means I must use a parametric test." | ✅ Large samples allow parametric tests to be robust to non-normality, but nonparametric tests remain valid and are sometimes preferred (e.g., for ordinal data). |
| ❌ "Nonparametric tests are always less powerful." | ✅ They are less powerful when normality holds, but more powerful than parametric equivalents when the distribution has heavy tails or many outliers. |
| ❌ "The p-value from a t-test and Mann-Whitney U test measure the same thing." | ✅ The t-test tests the difference in means. The Mann-Whitney U tests whether one group's values tend to be larger — the hypotheses are subtly different. |
| ❌ "If my data fail the Shapiro-Wilk test, I must use nonparametric." | ✅ Context matters. With n > 30, parametric tests are robust to mild non-normality. Significant Shapiro-Wilk results with very large samples often reflect trivially small departures from normality. |
| ❌ "Likert scale data should always be analyzed with ANOVA." | ✅ Likert scales are ordinal. Whether you use parametric or nonparametric methods is debated; most methodologists prefer the Mann-Whitney U or Kruskal-Wallis for single-item Likert measures. |
Entity and Term Glossary
| Term | Definition | Type |
|---|---|---|
| Parametric test | A test that assumes data follow a specific distribution (usually normal) and infers about population parameters | Method |
| Nonparametric test | A distribution-free test that uses ranks or signs instead of raw values | Method |
| Normal distribution | A symmetric bell-shaped distribution defined by mean (μ) and standard deviation (σ) | Distribution |
| Homogeneity of variance | The assumption that two or more groups share the same population variance | Assumption |
| Shapiro-Wilk test | A test of the null hypothesis that a sample comes from a normally distributed population | Diagnostic test |
| Levene's test | A test of the null hypothesis that population variances are equal across groups | Diagnostic test |
| Mann-Whitney U test | Nonparametric test for comparing two independent groups; uses rank sums | Nonparametric test |
| Wilcoxon signed-rank test | Nonparametric test for two related samples or one sample vs a hypothesized median | Nonparametric test |
| Kruskal-Wallis test | Nonparametric equivalent of one-way ANOVA; compares 3+ independent groups using ranks | Nonparametric test |
| Friedman test | Nonparametric equivalent of repeated measures ANOVA; compares 3+ related conditions | Nonparametric test |
| Spearman correlation | Rank-based measure of monotonic association; nonparametric alternative to Pearson r | Nonparametric measure |
| Statistical power | The probability of correctly rejecting a false null hypothesis (1 − β) | Concept |
| Effect size | A quantitative measure of the magnitude of an effect; independent of sample size | Concept |
| Central Limit Theorem | The principle that sample means approach normality as sample size increases, regardless of the population distribution | Theorem |
| ARE (Asymptotic Relative Efficiency) | The ratio of sample sizes needed by two tests to achieve equal power at large n | Concept |
Frequently Asked Questions
Where These Tests Are Used: Field Applications
Healthcare & Clinical Research
Clinical trials often use the Mann-Whitney U test for pain scores (ordinal) and the Wilcoxon signed-rank test for before/after measurements on the same patients. Survival data typically require specialized nonparametric methods like the log-rank test.
Education Research
Paired t-tests are standard for pre/post test comparisons. When class sizes are small (< 20 students) or test scores are heavily skewed, the Wilcoxon signed-rank test is preferred. Likert-based course evaluations use Kruskal-Wallis.
Business & Marketing
A/B testing with continuous metrics (revenue, session duration) and large samples suits parametric tests. Customer satisfaction surveys on ordinal scales use Mann-Whitney U or Kruskal-Wallis. NPS comparisons between segments often use nonparametric methods.
Psychology & Social Science
Most psychological scales are ordinal, making nonparametric tests the default. Spearman correlation is widely used for self-report questionnaire data. Friedman tests compare multiple conditions within the same participants in repeated-measures designs.
Manufacturing & Quality
Product dimensions and weights measured with calibrated instruments tend to be normally distributed — parametric t-tests and ANOVA are the natural choice. For sensory evaluations (appearance, taste ratings), nonparametric methods apply.
Finance & Economics
Financial returns often have heavy tails and are non-normally distributed. Nonparametric tests like the Wilcoxon signed-rank test are used to compare returns across periods. Spearman correlation measures the rank-based relationship between economic variables.
Quick Reference Summary
- Choose parametric when data are continuous, approximately normal, and sample sizes are adequate (n > 20–30 per group). Parametric tests have higher power when assumptions are met.
- Choose nonparametric when data are ordinal, clearly non-normal (especially with small samples), or contain influential outliers. Use them by default for Likert-scale single items.
- Independent groups (2): t-test (parametric) vs. Mann-Whitney U (nonparametric)
- Paired samples (2): Paired t-test (parametric) vs. Wilcoxon signed-rank (nonparametric)
- Three or more independent groups: One-way ANOVA (parametric) vs. Kruskal-Wallis (nonparametric)
- Three or more repeated conditions: Repeated measures ANOVA (parametric) vs. Friedman test (nonparametric)
- Linear association: Pearson r (parametric) vs. Spearman ρ (nonparametric)
- Test normality first: Shapiro-Wilk (n < 50) or Kolmogorov-Smirnov (n ≥ 50), plus a Q-Q plot
- Power trade-off: Nonparametric tests lose only ~5% power vs parametric tests under normality; under violations, they can be more powerful
Related Topics on Statistics Fundamentals
This guide focuses on the parametric vs nonparametric decision. The topics below go deeper into the individual tests and the statistical concepts that underpin them.
Mann-Whitney U Test
Step-by-step guide to the most widely used nonparametric test for two independent groups, with worked examples and a table of critical values.
Wilcoxon Signed-Rank Test
The nonparametric alternative to the paired t-test — how it works, when to use it, and how to interpret the W statistic.
One-Way ANOVA
Comparing means across three or more groups — assumptions, F-statistic calculation, post-hoc tests, and effect size reporting.
Spearman Rank Correlation
How to measure monotonic association without assuming linearity or normality — formula, interpretation, and comparison with Pearson r.
Normality Tests Guide
Shapiro-Wilk, Kolmogorov-Smirnov, Anderson-Darling, Q-Q plots — which to use and how to interpret results.
Statistical Test Selector
Interactive tool that guides you through the full test selection process based on your data type, design, and assumptions.