What are the SAMPL guidelines?

SAMPL (Statistical Analyses and Methods in the Published Literature) is a set of guidelines by Thomas Lang and Douglas Altman for reporting basic statistics in biomedical articles. It covers how to describe the statistical methods and how to report descriptive statistics, hypothesis tests, associations, regression, analysis of variance, survival and Bayesian analyses. It complements study-design checklists such as CONSORT and STROBE rather than replacing them.

Does SAMPL have a numbered checklist?

SAMPL is written as guidance in points under headings, not as a numbered checklist like CONSORT. Tools and journals that use it as a checklist number the points for reference. In AI for Research the points are numbered M1 to M10 for methods and R1 to R12 for results, so each can be checked and cited.

Should I report exact P values?

Yes. SAMPL asks for exact P values, such as P = 0.03 or P = 0.41, rather than only P < 0.05 or 'NS'. Very small values can be given as P < 0.001. The P value should come with the effect size and its confidence interval, because the P value alone does not tell the reader how large or how precise the effect is.

Should I use the standard deviation or the standard error?

Use the standard deviation to describe how spread out the data are, and a confidence interval to describe how precise an estimate is. SAMPL warns against using the standard error of the mean to describe variability, because it makes data look less variable than they are.

Does SAMPL apply outside medicine?

SAMPL was written for biomedical journals, but its rules on describing methods, reporting effect sizes with confidence intervals and giving exact P values apply to any field that uses statistics. Psychology, education and the social sciences have similar guidance, such as APA's journal article reporting standards.

SAMPL guidelines: what to report for every statistical analysis

Q: Should I use the standard deviation or the standard error?

Use the standard deviation to describe how spread out the data are, and a confidence interval to describe how precise an estimate is. SAMPL warns against using the standard error of the mean to describe variability, because it makes data look less variable than they are.

Q: Does SAMPL apply outside medicine?

SAMPL was written for biomedical journals, but its rules on describing methods, reporting effect sizes with confidence intervals and giving exact P values apply to any field that uses statistics. Psychology, education and the social sciences have similar guidance, such as APA's journal article reporting standards.

By Mubarak MutesasiraUpdated November 13, 20268 min read

The SAMPL guidelines (Statistical Analyses and Methods in the Published Literature) say how to report basic statistics in a research article. In the methods, state the purpose of each analysis, which test answered which question, how you checked its assumptions, the alpha level, whether tests were one- or two-sided, and the software. In the results, give the numbers analyzed, suitable descriptive statistics, the effect size with a confidence interval, and exact P values. They were written by Thomas Lang and Douglas Altman and published in 2013.

This guide is for anyone writing the methods and results of a quantitative study, and for supervisors and reviewers checking one. It belongs to our series on clinical studies and reporting.

What SAMPL is and where it fits

Design checklists such as CONSORT 2010 (trials) and STROBE 2007 (observational studies) tell you what to report about how the study was done. SAMPL covers the layer underneath: how to report the statistics in any of them. Use it alongside the design checklist, not instead of it.

SAMPL is written as guidance, in points under headings, not as a numbered list. The table below is our condensed version, numbered M (methods) and R (results) for reference, as in the Reviewer guideline check. Our reporting checklists template has a shorter one-page summary to fill.

#	Area	Report
M1	Data preparation	Any change to the raw data before analysis: transformations, categorization, imputation of missing values
M2	Purpose	The purpose of each analysis, tied to an objective
M3	Variables	The variables in each analysis, each summarized with suitable descriptive statistics
M4	Important difference	Where possible, the smallest difference considered clinically or practically important
M5	Methods	The main methods, clear about which method was used for which analysis
M6	Assumptions	That the data met each test's assumptions, and how that was checked
M7	Alpha and sidedness	The alpha level and whether tests were one- or two-sided
M8	Software	The software and its version
M9	Supplementary	Methods for sensitivity analyses and missing data
M10	Post hoc	Unplanned analyses, including subgroups, identified as exploratory
R1	Sample	The number analyzed in each group, with losses and exclusions and their reasons
R2	Descriptives	Mean and SD for roughly normal data; median and interquartile range (or range) for skewed data
R3	Percentages	Numerators and denominators with every percentage
R4	Effect	For main comparisons, the size of the difference with a confidence interval
R5	P values	Exact P values, with the test used
R6	Association	The measure of association with its confidence interval, and the direction
R7 to R9	Regression	How variables were chosen and entered; checks of assumptions, collinearity and fit; coefficients with confidence intervals
R10	ANOVA	The model, the factors, the between-group tests and effect sizes
R11	Survival	The event, how censoring was handled, survival estimates with confidence intervals
R12	Bayesian	Priors and why, the model, posterior estimates with credible intervals, sensitivity to the priors

What to report for each common test

SAMPL gives principles; this table applies them to the tests most papers use. For choosing the test in the first place, see which statistical test to use.

Test	Methods should say	Results should give	Example (invented)
Independent t-test	Two-sided, alpha, how normality and equal variances were checked, Welch or Student	Mean (SD) per group, mean difference with 95% CI, t, df, exact P, effect size	"Mean difference 4.2 points (95% CI 1.1 to 7.3); t(86) = 2.69, P = 0.009; d = 0.57"
Paired t-test	That observations were paired and how	Mean of the differences with 95% CI, t, df, P	"Mean change −3.1 kg (95% CI −4.4 to −1.8); t(29) = 4.87, P < 0.001"
One-way ANOVA	The factor, levels, assumption checks, post hoc method and correction	Group means (SD), F, both df, P, effect size (eta squared), post hoc differences with CIs	"F(2, 117) = 5.41, P = 0.006, η² = 0.08"
Chi-square	Expected counts checked; Fisher's exact test if small	Counts and percentages with denominators, chi-square, df, P, and a risk difference or ratio with CI	"18/60 (30%) vs 9/62 (15%); risk ratio 2.07 (95% CI 1.01 to 4.24)"
Mann-Whitney U	Why non-parametric	Medians (IQR) per group, U, P, and an effect estimate (such as the Hodges-Lehmann difference) with CI	"Median 12 (IQR 8 to 17) vs 9 (6 to 13); U = 1,412, P = 0.02"
Correlation	Pearson or Spearman, and why	r with 95% CI, n, P, and a scatterplot	"r = 0.42 (95% CI 0.24 to 0.57), n = 104"
Linear regression	Outcome, predictors, how chosen, checks of residuals and collinearity	Coefficients with 95% CIs, R², n	"Each additional year of age: 0.8 mmHg (95% CI 0.5 to 1.1)"
Logistic regression	Outcome coding, predictors, events per variable, fit	Odds ratios with 95% CIs, the number of events	"OR 1.9 (95% CI 1.2 to 3.0); 74 events"
Survival	Event, censoring, Kaplan-Meier, log-rank or Cox, proportional hazards check	Median survival with CI, hazard ratio with CI, numbers at risk	"HR 0.71 (95% CI 0.55 to 0.92)"

All numbers in the example column are invented to show the format.

Effect sizes and confidence intervals

The most important SAMPL point for most papers: report the size of the effect and its precision, not only whether it was "significant".

Use the natural units first. "4.2 points on a 0 to 100 scale" or "a 3.1 kg loss" means more to readers than a standardized effect. Add a standardized effect (d, eta squared, r) for comparisons across studies.
Give a confidence interval for every main estimate. The 95% CI shows the range of effects the data are compatible with. "Mean difference 2.0 (95% CI −0.5 to 4.5)" tells the reader the study cannot rule out no effect or a fairly large one.
Compare with the important difference (M4). If you set a minimal important difference, say whether the CI includes it.
For ratios, give the absolute numbers too. A risk ratio of 2 can mean 1 in 1,000 versus 2 in 1,000, or 30% versus 60%.

P values

Give exact values: P = 0.03, P = 0.41. Use P < 0.001 for very small ones. Not "P < 0.05", not "NS".
Name the test that produced each P value.
Do not read P > 0.05 as "no effect". Absence of evidence is not evidence of absence; look at the CI. If you need evidence for the null, a Bayesian analysis can provide it; see Bayes factor explained.
Correct for multiple comparisons when you make many, and say how.
Mark exploratory analyses (M10) so readers do not treat them as confirmatory.

Common mistakes

Mean ± SE to describe data. The SE shrinks with sample size and makes data look tidier than they are. Use the SD for spread and the CI for precision.
Mean (SD) for skewed data. Length of stay, costs and counts are usually skewed; give the median and interquartile range.
Percentages without denominators. "45% improved" of how many?
Only P values. No estimate, no interval.
"Trend toward significance" for P = 0.07. Report the estimate and CI and let them speak.
Unnamed tests. "Data were analyzed with appropriate tests" tells the reader nothing.
Assumptions never checked. Normality, equal variances, proportional hazards. See how to test for normality.
Subgroup claims from separate P values, rather than a test of interaction.
Software without a version.

A worked example (invented)

Example (invented for illustration). A draft results paragraph reads: "The intervention group improved significantly more than controls (p<0.05). Anxiety scores were 8.2 ± 0.6 and 10.4 ± 0.7."

Revised to SAMPL:

"At 8 weeks, 44 of 48 participants in the intervention group and 45 of 50 in the control group were analyzed (Figure 1). Mean GAD-7 scores were 8.2 (SD 4.1) and 10.4 (SD 4.9). The mean difference was −2.2 points (95% CI −4.1 to −0.3; Student's t(87) = 2.29, P = 0.02), smaller than the 4-point difference we prespecified as important. Analyses used R version 4.4."

The revision gives the numbers analyzed (R1), the SD instead of the SE (R2), the effect and CI (R4), an exact P with its test (R5), the comparison with the important difference (M4) and the software (M8). It also reads more honestly: the effect is real but smaller than the team hoped.

Checking your statistics in AI for Research

Two parts of AI for Research help here.

The Reviewer guideline check now includes SAMPL (2013) and SAGER (2016), with CONSORT 2010, STROBE 2007, PRISMA 2020, CARE 2013, STARD 2015, COREQ 2007, SRQR 2014 and SQUIRE 2.0. Choose SAMPL, and it lists items M1 to R12 with the passage that appears to report each, or "check by hand" when it finds none. It costs 5 credits per 10,000 words. The check drafts the checklist; you confirm it, and it does not guarantee a journal's acceptance. More: reporting guideline checker.
Data Lab runs the tests themselves. The statistics are calculated in your browser, not generated by the AI, and free. Each result comes with an effect size (Cohen's d, η², r or Cramér's V) and an APA 7 sentence, and the "Which test?" chooser includes a reporting template. Not every result shows a confidence interval, so add the CIs SAMPL asks for where they are missing. An AI write-up of the results section costs 2 credits; check it against the table above.

The Free plan has 45 credits a month; see pricing.

Learn more

Which statistical test to use
T-test vs ANOVA vs chi-square
SAGER guidelines for sex and gender reporting
Site initiation visit checklist, the pillar of this series
Reporting checklists template