Why reporting needs common rules
A statistical result is only useful if a reader can understand what was done, judge how much to trust it and, ideally, reproduce it. When methods and results are reported incompletely, others cannot evaluate or build on the work, and progress in both basic and clinical research slows (Prager et al., 2019). Reporting guidelines exist to close that gap. They do not tell you which analysis to run; they tell you what readers need in order to evaluate the analysis you ran.
This guide is the overview of our Statistical reporting series. It sets out principles that apply to every quantitative analysis, from a t test to a Bayesian model. The nine guides that follow it apply these principles to particular families of analyses and work through an example of each, so detailed points about, for example, hazard ratios or fit indices are left to them.
Start from the right guideline
Two general documents form a sensible baseline for any thesis or article. The SAMPL guidelines set out basic statistical reporting for biomedical journals, covering how to describe data, tests, models and results (Lang & Altman, 2015). The APA Journal Article Reporting Standards for quantitative research (JARS-Quant) do the same for psychology and the wider social sciences, with modules for different designs (Appelbaum et al., 2018).
On top of these, use the design-specific guideline your field and target journal expect. Randomised trials should follow CONSORT 2025, which updated the 2010 statement, drew in items from several of its extensions and added a section on open science (Hopewell et al., 2025). Cohort, case-control and cross-sectional studies should follow STROBE (von Elm et al., 2007). The EQUATOR Network maintains a searchable library of these and many other guidelines, including extensions for particular designs. Reading the relevant checklist before you write, rather than ticking it off afterwards, is the single most effective way to avoid omissions.
| Guideline | Use it for | What it adds |
|---|---|---|
| SAMPL (Lang & Altman, 2015) | Any biomedical or health analysis | Basic rules for describing data, tests, models and results |
| APA JARS-Quant (Appelbaum et al., 2018) | Psychology, education, social sciences | Design modules; sample, measures, analysis and results |
| CONSORT 2025 (Hopewell et al., 2025) | Randomised trials | Participant flow, analysis populations, absolute and relative effects, harms |
| STROBE (von Elm et al., 2007) | Observational studies | Confounding, unadjusted and adjusted estimates, missing data |
Describe the data before you test it
Readers need to know who or what was studied before they can interpret a comparison. Report the number of participants or units at each stage and in each group, with reasons for exclusions and losses; a flow diagram does this compactly for trials and many cohort studies (Hopewell et al., 2025). Give counts alongside percentages (for example, 42/120, 35%) so that the denominator is never in doubt, and state the denominator whenever it changes because of missing data.
Choose summary statistics that suit the distribution. For roughly symmetric continuous data, the mean and standard deviation (SD) are appropriate; for skewed data, the median with interquartile range (IQR) usually describes the typical value and spread better (Lang & Altman, 2015). Do not use the standard error of the mean (SEM) to describe variability: the SEM measures the precision of the estimated mean and shrinks as the sample grows, whereas the SD describes how much individual observations vary (Altman & Bland, 2005).
Match precision to what the data support. Cole (2015) suggests, as a general rule, two or three significant digits for effect sizes and one or two for measures of variability, with percentages as whole numbers (or one decimal place below 10%). A mean age of 34.6 years is informative; 34.5833 years is noise.
Tables are best for exact values that readers may want to look up or reuse. Figures are best for patterns, distributions and comparisons. Bar charts of means hide the distribution and can mislead with small samples; Weissgerber et al. (2015) recommend plots that show individual data points or the full distribution, such as dot plots, box plots or violin plots. Label axes with units, avoid truncated axes that exaggerate differences, and state what error bars show.
Report effects, uncertainty and tests together
The core of any result is an estimate of the effect or association you set out to measure, with a measure of its uncertainty. Report the effect size in units readers understand (a mean difference in points, a risk difference, a regression slope per unit) and, where helpful, a standardised effect size, each with a 95% confidence interval (Lang & Altman, 2015; Appelbaum et al., 2018). For binary outcomes in trials, CONSORT 2025 asks for both absolute and relative effects (Hopewell et al., 2025).
When you report a hypothesis test, give the test statistic, its degrees of freedom and the exact p-value (for example, p = .016 rather than p < .05; write p < .001 only for smaller values). Name the test precisely, such as Welch's t test or a two-sided Fisher's exact test, and say whether it was one- or two-sided. The degrees of freedom let readers check that the test matches the design and sample.
Interpret p-values with care. The American Statistical Association's statement stresses that a p-value does not measure the probability that a hypothesis is true or the size of an effect, and that conclusions should not rest on whether it crosses a threshold (Wasserstein & Lazar, 2016). Greenland et al. (2016) catalogue common misreadings, including treating a non-significant result as proof of no effect. A later task force appointed by the ASA president affirmed that p-values and significance tests remain valid when properly applied and interpreted, alongside other measures of uncertainty (Benjamini et al., 2021). Proposals to lower the conventional threshold for claims of new discoveries to .005 have also been debated (Ioannidis, 2018). The practical lesson is the same across these positions: report the estimate, its interval and the exact p-value, and let the size and precision of the effect drive the interpretation.
Explain the analysis so it can be checked
The methods section should let a competent reader repeat the analysis. That means stating, for each research question, the outcome, the comparison, the test or model and why it fits the design, and any adjustment variables (Lang & Altman, 2015). Say which assumptions mattered, how you checked them (for example, residual plots) and what you did when they were not met. Avoid choosing between tests on the basis of formal pre-tests alone; robust methods such as Welch's test are often a better default.
- Sample size. Explain how it was determined. For a power calculation, give the target effect and its justification, the significance level, the power and any allowance for attrition (Appelbaum et al., 2018).
- Missing data. Report how much was missing for each key variable, what assumptions you made about why, and how you handled it. Complete-case analysis can be biased, and multiple imputation needs its model and number of imputations described (Sterne et al., 2009).
- Multiplicity. If several outcomes, groups or time points are tested, say how the risk of false positives was managed: a single pre-specified primary outcome, a family-wise method such as Holm's, or false discovery rate control, which limits a different error rate. Whatever you chose, state it and why.
- Software. Name the software, version and key packages, and share code where you can (Prager et al., 2019).
Pre-specified, exploratory and deviations
Readers interpret a result differently depending on whether it tested a hypothesis fixed in advance or emerged from exploring the data. Preregistration, or a registered protocol and statistical analysis plan, makes that distinction verifiable and reduces the scope for selective reporting (Nosek et al., 2018). Trials should report their registration and where the protocol can be found (Hopewell et al., 2025).
Label exploratory analyses as such, and report any change from the planned analysis with the reason for it. Subgroup analyses deserve particular care: state whether they were pre-specified, and compare subgroups with a test of interaction rather than by noting that the effect was significant in one subgroup and not in another. Describe the characteristics of the sample, such as age and sex, so that readers can judge to whom the results apply.
A worked example
The study below is hypothetical and purely illustrative. Suppose 108 students were randomised to a feedback intervention or usual teaching, and 100 (50 per group) completed the end-of-term test. A first draft reads: "The intervention group scored significantly higher (p < 0.05)." That sentence gives no effect size, no interval, no test and no sample size.
| Element | Weak version | Improved version |
|---|---|---|
| Sample | Not stated | Of 108 randomised, 100 completed the test (4 missing per group) |
| Descriptives | Not stated | Intervention 72.4 (SD 10.1); usual teaching 67.3 (SD 10.7) |
| Effect | "significantly higher" | Mean difference 5.1 points, 95% CI 1.0 to 9.2 |
| Standardised effect | Not stated | Hedges' g = 0.49, 95% CI 0.09 to 0.88 |
| Test | "p < 0.05" | Welch's t test, t(97.7) = 2.45, p = .016, two-sided |
| Missing data | Not stated | Complete-case analysis; sensitivity analysis with multiple imputation |
The improved version lets a reader see that the effect is moderate but imprecise: the interval runs from a difference of about one point, which might matter little, to about nine points, which would matter a good deal. That is a more honest summary than "significant", and it tells the next researcher how large a study is needed to narrow the estimate. The fractional degrees of freedom show that Welch's correction was used.
Where to go next in the series
Each guide below applies these principles to one family of analyses, with a worked example and a what-to-report table.
| If you ran | Read |
|---|---|
| t tests, chi-square and other significance tests | Reporting hypothesis tests |
| Pearson or Spearman correlations | Reporting correlation analyses |
| Linear, logistic or other regression models | Reporting regression analyses |
| Comparisons of several group means | Reporting ANOVA and ANCOVA |
| Associations between exposures and outcomes | Reporting association analyses |
| Risks, rates, odds or hazards | Reporting risks, rates and ratios |
| Time-to-event analyses | Reporting survival analyses |
| A questionnaire measurement model | Confirmatory factor analysis |
| Priors and posterior distributions | Reporting Bayesian analyses |
Common mistakes
- Reporting p-values without effect sizes and confidence intervals.
- Describing a non-significant result as showing "no difference" or "no effect".
- Using the SEM to describe variability, or error bars without saying what they show.
- Percentages without counts, or denominators that silently change because of missing data.
- Testing for baseline differences between randomised groups, which by design differ only by chance.
- Claiming a subgroup effect because one subgroup was significant and another was not.
- Reporting many decimal places, or "p = 0.000" copied from software output.
- Omitting the software version, analysis code or deviations from the planned analysis.
A useful final test is to give your results section to a colleague with the relevant checklist and ask whether they could reconstruct each headline number and its uncertainty. Where they cannot, the reporting needs more work.
Key takeaways
- Use SAMPL or APA JARS-Quant as a baseline and the design-specific guideline, such as CONSORT 2025 or STROBE, on top (Lang & Altman, 2015; Appelbaum et al., 2018).
- Describe data with the SD or IQR, not the SEM, and give counts with every percentage (Altman & Bland, 2005).
- Lead with effect sizes and confidence intervals; give test statistics, degrees of freedom and exact p-values (Wasserstein & Lazar, 2016).
- Explain sample size, assumptions, missing data and multiplicity so that the analysis can be checked (Sterne et al., 2009).
- Separate pre-specified from exploratory analyses and report deviations from the plan (Nosek et al., 2018).
