One p-value cannot check every assumption.
How to Test Data for Normality: p-Values, Q-Q Plots and Decisions
A Shapiro-Wilk p-value above 0.05 does not prove normality. Record what you tested, what the plot shows and whether the test conditions fit your data before deciding what to report. [1] [2]
AcademicFix Editorial Team · Sources reviewed:

Three distinctions that change the interpretation
Use the output to document a decision, rather than treating it as a certificate.
Is that enough to call the data normal?
Not crossing a threshold and confirming a hypothesis are different conclusions.
Check what p can tell you02 / Name the convention3 or 0?Which kurtosis did the software report?
Pearson and excess kurtosis use different normal-distribution reference values.
Compare the definitions03 / Trace one observation35 → 80What changes when one value changes?
Two invented datasets have the same n, but different final observations and test results.
Inspect the computed exampleFor researchers reporting normality checks in a thesis or article. This workflow concerns continuous measurements and relevant model diagnostics. It does not select an analysis for an unseen dataset, assess multivariate normality or certify every model assumption.
On this page
1. State exactly what you are examining
Replace “Are the data normal?” with a named variable or diagnostic target. Record its units, missing observations and any fitted model. NIST describes a common use of normal probability plots for residuals from fitted models; R also provides a residual Q-Q plot among its linear-model diagnostics. [3] [4]
Our practical recommendation: when checking a linear model, do not stop after inspecting pooled raw scores. Match the residual plot and other diagnostics to the model documentation. Preserve group information in the record when comparing groups. This is a documentation recommendation, not a claim that every analysis has the same normality assumption.
2. Do not turn p into a probability of normality
The Shapiro-Wilk null hypothesis concerns sampling from a normal distribution. SciPy explains that a small p-value can provide evidence against that hypothesis, while the reverse inference does not hold. ASA separately warns that p does not give the probability that a hypothesis is true and that scientific decisions require more than a threshold. [2] [1]
For p < 0.05, report evidence of incompatibility with the normal model. For p ≥ 0.05, report that normality was not rejected at this level. Do not rewrite the latter as “95% likely to be normal”. Choose the decision threshold before examining the results. [2] [1]
Neither result alone determines whether a t-test is appropriate or every parametric method is ruled out. Return to the study design and the other assumptions of the intended analysis.
3. Describe the Q-Q pattern, not just a verdict
An approximately straight normal probability plot is consistent with an approximate normal model. Systematic curvature and tail departures indicate differences in shape. NIST uses these plots to investigate skewness and shorter or longer tails. In model diagnostics, identify the residuals shown; R's residual Q-Q plot serves a different purpose from its residuals-versus-fitted plot. [3] [4]
A useful three-line plot note: describe the middle, the direction of departure in each tail and the record corresponding to any striking observation. This is more traceable than “the plot was normal”. Our data-audit recommendation is to check the original record and research context before removing a value, rather than deleting it to obtain a preferred p-value.
4. Check software conditions before choosing a test
Shapiro-Wilk limits depend on the implementation. SciPy requires at least three observations and warns that above 5000 observations its W statistic remains accurate but the p-value may not. R's shapiro.test accepts 3–5000 non-missing values. These are software-specific conditions, not a universal rule that Shapiro-Wilk becomes invalid above n = 50. [9] [10]
For the standard one-sample Kolmogorov-Smirnov test, parameters matter. The comparison distribution must be specified in advance. Estimating its mean and standard deviation from the same sample and using the unadjusted standard K-S p-value does not meet that condition. NIST explains the limitation; R notes that estimated-parameter theory is not implemented in ks.test. [5] [6]
If the output specifies a Lilliefors correction, retain that name in the report rather than merging it with unadjusted K-S. This guide does not calculate a Lilliefors p-value or assert equivalence between software packages. Also inspect the documentation for ties caused by rounding. [6]
5. Identify the definition before comparing kurtosis values
For a normal distribution, Pearson kurtosis has reference value 3. Excess kurtosis subtracts 3, giving reference value 0. NIST distinguishes both conventions; SciPy uses the 0 reference with fisher=True and the 3 reference with fisher=False. Record bias-correction settings separately. [7] [8]
Our decision rule for this guide: do not treat a skewness or kurtosis interval as automatic permission for every t-test, regression, thesis or journal tier. Name the convention and analysis context, and report graphical and model checks together. We are not attributing a universal acceptance threshold to a textbook. [1] [3]
Same n, a different tail: a 12-observation example
All values are invented for teaching; these are not research findings. Dataset A is 12, 15, 16, 17, 18, 19, 20, 21, 22, 24, 25, 35. Dataset B replaces only the final value, 35, with 80. These are separate illustrative cases, not an account of editing a real research record.
| Dataset | Change | W | p | Bounded interpretation |
|---|---|---|---|---|
| A: starting values | Final observation: 35 | 0.9186 | 0.2744 | At α = 0.05, normality was not rejected. This does not prove a normal distribution. |
| B: one value changed | 80 replaces 35 | 0.5315 | < 0.001 | At α = 0.05, the result indicates incompatibility with the normal model. Investigate the value 80 separately. |
Recalculated with shapiro.test in R 4.6.1. W and the p-value for A are rounded to four decimal places; B is reported as p < 0.001. Preserve the approximate p-value with the implementation and version. [10] [2]
What the example changes: identical sample sizes do not imply identical distributional shapes or test results. In a real dataset, investigate the origin of a value such as 80. For A, do not claim that normality was proved. Neither example establishes a rule for other sample sizes or analyses.
Reproduce the calculation in R
a <- c(12, 15, 16, 17, 18, 19, 20, 21, 22, 24, 25, 35)
b <- c(a[-length(a)], 80)
shapiro.test(a)
shapiro.test(b)
R.version.string6. Report the result, the plot and the remaining decision
- Name the variable or model residual, units, group and n after missing-data handling.
- Give the full test name, any correction, W or D and the p-value.
- Include the software, version and material calculation options.
- Describe the Q-Q pattern you actually inspected and any observations investigated.
- Separate the pre-specified decision level from model assumptions still unchecked.
“For the 12 illustrative observations, the Shapiro-Wilk test gave W = 0.9186, p = 0.2744. Normality was not rejected at the pre-specified α = 0.05 level. This was not interpreted as proof of normality or confirmation of every assumption of a subsequent analysis.”
In an actual report, add the software/version and your observed graphical findings. Do not copy a plot assessment that you have not performed.
This checklist is an editorial reporting recommendation. Check current requirements for your field and target journal separately.
Two shortcuts to avoid
Does p > 0.05 prove that my data are normal?
No. Failure to reject at a chosen level does not establish that the null hypothesis is true. Consider the plot, data context and assumptions of the intended analysis.
Should I switch to K-S automatically when n exceeds 50?
This guide recommends no automatic switch. Check test and software conditions, especially whether the K-S distribution parameters were estimated from the same sample.
Sources and review boundaries
The relevant official source sections were read on 19 October 2026. This guide distinguishes methodological claims, specific software conditions and editorial recommendations. No real thesis dataset or complete set of model assumptions was assessed.
- ASA: principles for interpreting p-values
Principles 1–3 and 6: model incompatibility, hypothesis probabilities and the limits of threshold decisions.
- SciPy: interpreting Shapiro-Wilk
Null hypothesis, W and p-values; failure to reject does not establish the null hypothesis.
- NIST: normal probability plots
Approximately straight patterns, tail departures and residuals from fitted models.
- R: diagnostic plots for linear models
Description and Details: residual Q-Q plots and other model diagnostics.
- NIST: K-S test limitations
Characteristics and Limitations: continuous distributions and fully specified parameters.
- R: ks.test documentation
Details: pre-specified parameters, ties and rounding; estimated-parameter theory is not implemented.
- NIST: skewness and kurtosis
Definition of Kurtosis and Alternative Definition: Pearson versus excess kurtosis.
- SciPy: Fisher and Pearson kurtosis
Definition and fisher/bias parameters: normal-distribution references of 0 and 3.
- SciPy: shapiro implementation limits
Parameters and Notes: minimum three observations; p-value warning above 5000.
- R: shapiro.test implementation limits
Arguments, Value and Source: 3–5000 non-missing observations and approximate p-values.
From the diagnostic check to the analysis plan
- Read p-values alongside effect sizes and confidence intervals
- Check the analysis plan against your question and design
- Compare two independent means with Welch’s t-test
- Check before-and-after differences with a paired t-test
- Place verified results in an APA 7 table