Blazon Scientific logo

Statistical reporting guide

How to Report Statistical Results in a Scientific Paper

Written by Thomas Solomon, PhD

A statistical analysis can be perfectly reasonable and still be reported badly. A sentence announcing a statistically significant result says something about a hypothesis test, but remarkably little about effect magnitude, uncertainty or whether the finding matters.

A P value is not a scientific importance detector. Convenient though that would be.

Good statistical reporting identifies the outcome and comparison, explains how the estimate was obtained, communicates uncertainty and keeps the interpretation proportionate to the design. This guide focuses on reporting quantitative results in clinical, biomedical and health research, using a hypothetical exercise-and-hypertension trial to connect the examples.

Table of contents
The research workflow

Reporting is connected to decisions made throughout a study. The final stage translates planned analyses and observed findings into a transparent research paper.

  1. Develop the research question
  2. Plan the study
  3. Prepare and conduct the study
  4. Collect and manage data
  5. Analyse and interpret
  6. Write, submit and publish

The reported estimate should address the research question and reflect the actual data, analysis and assumptions—not just the output that looks most favourable.

1. Start with the research question and analysis plan

Before reporting a statistical result, identify the scientific question it answers. In our continuing hypothetical trial, researchers compare 12 weeks of supervised exercise with usual care in adults with hypertension. The primary objective is to estimate the between-group effect on clinic systolic blood pressure at 12 weeks. The prespecified hypothesis predicts lower blood pressure in the exercise group.

The primary result should therefore report an appropriate between-group comparison. Separate tests of change from baseline within each group would answer different questions, even though they use some of the same measurements. A small P value from the wrong comparison doesn’t rescue the analysis.

Define what the estimate represents

A study may aim to estimate a mean difference, an association, a risk ratio, a survival probability or predictive performance. In some clinical trials, the target is described as an estimandA precisely defined quantity the study aims to estimate, including the population, outcome, comparison and relevant handling of post-randomisation events.. Report the comparison and effect measure that correspond to the study’s objective, not just a statistic that the software happened to produce.

Distinguish prespecified and exploratory findings

Where available, the protocol, registration and statistical analysis plan help establish the primary outcome, analysis population, comparison, covariates and handling of missing data. Important departures should be explained. Exploratory subgroup or sensitivity analyses can be informative, but they should not be presented as though they had always been the confirmatory plan.

For more on establishing these decisions in advance, see How to Write a Statistical Analysis Plan.

2. Report sample sizes, denominators and descriptive statistics

Statistical estimates are difficult to understand if readers cannot see what contributed to them. Report the relevant number of participants, samples, clusters or experimental units for each important outcome. The numbers recruited, randomised, measured and analysed may differ, particularly when follow-up observations are missing.

Be explicit about denominators and independence

“Eighteen participants experienced an adverse event” is less useful than “18 of 90 participants (20%)”, assuming 90 is the relevant denominator. For longitudinal research, distinguish participants from their repeated observations. Likewise, three technical measurements from one specimen do not necessarily constitute three independent biological replicates. A large number of rows in a spreadsheet is not a substitute for an appropriately large independent sample.

Mean and standard deviation

The mean describes the arithmetic average; the standard deviation (SD) describes dispersion among observations. A mean blood pressure of 132 mmHg (SD 12 mmHg) identifies a group average and the observed spread, not the precision of the mean as an estimate of a population quantity. For markedly skewed data, show informative distributional summaries or plots rather than assuming the mean alone tells the story.

Median and interquartile range

The median is the middle value in an ordered distribution, and the interquartile range (IQR) runs from the 25th to the 75th percentile. These can be useful for skewed outcomes, but they are not an automatic response to a failed normality test. Choose summaries that describe the data; choose statistical analyses that address the research objective and design.

SD, SE and CI describe different things

SD, SE and CI are commonly confused statistical quantities. The table below helps define them:

MeasureWhat it communicates
Standard deviation (SD)Dispersion among observations
Standard error (SE)Estimated sampling variability of a statistic
Confidence interval (CI)Uncertainty about a parameter using a specified procedure and model

For a conventional independent-sample mean, the estimated SE is the sample SD divided by the square root of the sample size. That formula doesn’t apply unchanged to every model or dependent-data design. If the intention is to show variation between participants, an SD or distribution plot is generally more useful than an SE; if the intention is to show uncertainty around an estimate, an appropriate interval may be preferable.

SD, SE and CI are related, but they’re not interchangeable guests at the statistical family reunion.

3. Report effect estimates, not only statistical significance

An effect estimate expresses the magnitude and direction of a difference, association or other scientific quantity. Examples include mean differences, risk differences, risk ratios, odds ratios, rate ratios and hazard ratios. Report a measure that answers the intended question, with an uncertainty measure suited to the analysis.

Keep meaningful units where possible

For our fictional trial, an adjusted difference of −4 mmHg, defined as exercise minus usual care, indicates an estimated 4 mmHg lower mean systolic blood pressure in the exercise group after the specified adjustment. That is much more interpretable than “exercise significantly improved blood pressure”. The latter neither identifies the size of the difference nor states precisely what was compared.

Absolute and relative effects

In a separate fictional example, an event occurs in 10 of 100 people in an intervention group and 20 of 100 in the comparator. The observed risks are 10% and 20%; the risk difference (intervention minus comparator) is −10 percentage points, and the risk ratio is 0.50. These measures convey complementary information. The unadjusted odds ratio, however, is approximately 0.44. An odds ratio is not generally interchangeable with a risk ratio, particularly for common outcomes.

For randomised trials with binary outcomes, CONSORT 2025 recommends reporting absolute and relative effect estimates with measures of precision, where applicable.2 The simple percentages above provide arithmetic examples, not inferential intervals or an analysis of a real trial.

When do standardised effects help?

A standardised mean difference can facilitate comparisons across studies using different measurements of a common construct. Its meaning still depends on the standardising SD, underlying scale and design. Where original units are clinically interpretable, they often deserve pride of place; standardisation isn’t an automatic upgrade in scientific importance.

4. Report confidence intervals and uncertainty correctly

A point estimate gives one value, not a complete account of its uncertainty. Under a specified model and procedure, a confidence interval helps communicate the precision of the estimated effect.

For the illustrative trial, suppose the baseline-adjusted difference is −4.0 mmHg (95% CI −7.0 to −1.0 mmHg). A suitable sentence is: “At 12 weeks, adjusted mean systolic blood pressure was lower in the exercise group than in the usual-care group (between-group difference −4.0 mmHg, 95% CI −7.0 to −1.0 mmHg).” State the subtraction direction and the model in the text, table footnote or Methods.

What does a 95% confidence interval mean?

In conventional frequentist inference, 95% refers to the long-run coverage of the interval-generating procedure under its assumptions. It does not assign a 95% probability to the fixed true parameter lying within this particular observed interval. The interval helps assess the effect values comparatively compatible with the data under the chosen model, but cannot account automatically for selection bias, measurement error or every other uncertainty.

What if the interval includes the null?

For a mean difference the conventional null value is zero, whereas for a ratio it is one. If the fictional blood-pressure difference were −4 mmHg (95% CI −10 to +2), the interval would span effects of different sizes and directions. Reporting “there was no effect” solely because the interval includes zero would conceal important uncertainty.

Consider meaningful-effect thresholds

Where possible, define clinically or scientifically relevant effect magnitudes independently of the observed results. An interval excluding large benefits and harms has different implications from a wide interval including both. The precision of an estimate and the scientific importance of an effect are separate considerations.

5. Report and interpret P values

A P value is calculated for a specified hypothesis test under a set of model assumptions. Broadly, it describes how unusual the observed test statistic, or one more extreme, would be under the null model. It does not tell us the probability that the null hypothesis is true, nor the size or clinical importance of an effect.5

State what was tested

A P value for an adjusted between-group contrast is not interchangeable with a P value testing within-group change from baseline. The Methods should identify the procedure; the Results should identify the corresponding outcome and comparison. A P value is not a scientific importance detector. Convenient though that would be.

Use appropriate precision

Where relevant, report the numerical value, such as P = 0.032, rather than only P < 0.05, subject to journal conventions. For very small values, a threshold such as P < 0.001 or scientific notation may be suitable. Avoid P = 0: display rounding and numerical underflow do not establish an exactly zero probability. Nor does copying twelve decimal places make an analysis twelve times more trustworthy.

Statistical significance is not scientific importance

A small P value can accompany a negligible estimated difference; a larger P value can accompany an imprecisely estimated effect that might matter. Context, effect magnitude, uncertainty and study quality are necessary for interpretation. If a decision threshold was prespecified, explain its role without treating it as a universal separator between scientifically meaningful and meaningless results.6

Do not equate non-significance with no effect

When an effect is estimated imprecisely, report the estimate and interval rather than simply announcing “no difference”. Evidence of equivalence or non-inferiority requires a suitable question, margin, design and analysis. Observed or post hoc power calculated from the observed effect generally adds little to a completed study compared with reporting the estimate and its uncertainty.

6. Match reporting to the outcome type

No reporting template is appropriate for every outcome. Identify what is measured, the relevant denominator or follow-up period, the effect scale, and the model or summary used.

Continuous outcomes

For a continuous outcome such as blood pressure, present suitable group summaries and the scientifically relevant comparison in meaningful units. A difference in change scores and a baseline-adjusted follow-up difference are not the same estimator; label them accurately.

Binary outcomes

Report event counts and denominators in each group alongside relevant risk differences, risk ratios or odds ratios and their uncertainty. Define the assessment period, outcome and comparison direction. Absolute and relative estimates may convey different aspects of the practical importance of the same result.

Counts and event rates

When participants contribute different follow-up times, an event rate per person-year or other justified exposure unit may be informative. A rate ratio compares rates, not simply the proportion of participants who experienced at least one event. Specify follow-up, event definition and relevant model assumptions.

Time-to-event outcomes

State the number of events and relevant follow-up information. A hazard ratio is not automatically a risk ratio or a difference in cumulative event probability. Report its model basis and assumptions, and consider absolute survival or risk estimates at clinically meaningful timepoints where appropriate.

Ordinal outcomes

For ordered categories, define which direction represents improvement and identify the model and reference categories. If an ordinal model yields an odds ratio, its interpretation depends on outcome coding and model assumptions. Without that context, the direction of a reported effect can become a rather elaborate guessing game.

7. Report regression and other model-based results

Model output needs context. Identify the outcome, model family, comparison or predictor increment, effect scale, adjustment variables and uncertainty. Readers should not have to infer the scientific quantity from a coefficient name in the software’s output.

Adjusted and unadjusted estimates

For the exercise trial, a linear regression models systolic blood pressure at 12 weeks using treatment group and baseline blood pressure as predictors. Its treatment coefficient is an adjusted between-group comparison. Observed group means are still useful descriptively, but they are not automatically adjusted means. In observational research, covariate adjustment may address confounding under assumptions; in a randomised trial, baseline adjustment can improve precision.

Interpret coefficients on the correct scale

A linear-model coefficient, an odds ratio from logistic regression and a hazard ratio from a survival model mean different things. Define the predictor change or reference category and whether a coefficient was transformed. With interaction terms, a coefficient may be conditional on other predictors; meaningful contrasts or predicted values may communicate the intended comparison better.

Report material model diagnostics

Depending on the analysis, diagnostics may concern residuals, influential observations, non-linear relationships, overdispersion, proportional hazards or prediction-model calibration. Report problems and important analytical changes when they affect interpretation; a manuscript need not reproduce every diagnostic plot.

Bayesian results use different terminology

Bayesian credible intervals and posterior probabilities require explanation of their model and prior assumptions. A credible interval is not simply another name for a frequentist confidence interval. Report the inferential quantity actually estimated rather than translating it into P value terminology by habit.

8. Report paired, repeated and clustered observations

Dependence between observations affects both uncertainty and interpretation. A row in a dataset is not always an independent research unit.

Paired measurements

For paired observations, report the number of relevant pairs, the paired difference or other targeted estimate, and appropriate uncertainty. Separate pre- and post-measurement SDs generally do not reveal the uncertainty of paired differences without information about within-person association.

Longitudinal comparisons

Repeated measures can address several questions: changes over time, differences at a specific follow-up, or differences in trajectories. A significant time effect does not establish that an intervention outperformed its comparator. Likewise, a significant change in one group and a non-significant change in another does not itself establish a between-group difference. Report the direct contrast that addresses the research question.

Clustered observations

When participants are nested within clinics, schools or other clusters, report the clustering structure and how it entered the analysis. Relevant details may include the number and sizes of clusters, allocation unit and method used for within-cluster dependence. Twenty clinics containing 1,000 participants are not necessarily equivalent to 1,000 independent experimental units for every inferential question.

9. Report missing data, multiplicity and sensitivity analyses

Missing observations and analysis populations

Report how much information was unavailable, which participants contributed to each analysis and how missing values were handled. Complete-case analysis, imputation and likelihood-based approaches use information differently and depend on different assumptions. Don’t use “intention-to-treat” as a substitute for explaining actual inclusion and treatment of missing outcomes or relevant post-randomisation events.

Multiple tests and outcomes

Many outcomes, timepoints, subgroups and contrasts create opportunities for chance findings to receive disproportionate attention. Describe prespecified multiplicity procedures and which tests they address. Exploratory work does not always require the same correction as confirmatory testing, but the number and nature of the analyses and the strength of claims should remain transparent.

Sensitivity analyses

Explain the purpose and findings of important analyses testing alternative defensible assumptions, such as missing-data mechanisms, model specifications or exclusion rules. If they materially alter the primary result, say so. Highlighting only the most favourable version is not a sensitivity analysis; it is selective presentation.

10. Present statistical results in tables and figures

Statistical software may happily produce 47 columns of output. Your readers probably don’t need all 47. A good table keeps the scientifically relevant result visible, distinguishes observed summaries from model-based estimates, and explains uncertainty without forcing readers to interrogate the software manually.

The worked example: a complete reporting table

The values in Table 1 describe a deliberately constructed teaching dataset, not a real clinical trial. There are 200 fictional participants, with 100 in each group. The 12-week outcome is analysed by a linear regression with treatment group and baseline systolic blood pressure as predictors. Group columns show observed means and SDs; the primary comparison is model-adjusted.

Table 1. Systolic blood pressure at 12 weeks in the fictional exercise trial

OutcomeExercise (n = 100)Usual care (n = 100)Adjusted difference (95% CI)Two-sided P
Clinic systolic blood pressure, mmHg132.0 (12.1)136.0 (11.1)−4.0 (−7.0 to −1.0)0.0096

Table note: Group summaries are observed mean (SD). The adjusted difference is exercise minus usual care, estimated by linear regression including randomised group and baseline blood pressure. The 95% CI and P value are based on the corresponding treatment coefficient and a two-sided t reference distribution. Numbers are rounded; the synthetic design is constructed so the adjusted estimate and descriptive difference coincide, which need not occur in real data.

Compare weak and informative reporting

Table 2 describes how the reporting changes when the effect size and its uncertainty are included:

Table 2. Reporting is more informative when an effect and uncertainty are included

Less informativeMore informative
“Exercise significantly reduced blood pressure.”“The adjusted difference at 12 weeks was −4.0 mmHg (95% CI −7.0 to −1.0 mmHg), exercise minus usual care.”
“The groups did not differ (P > 0.05).”Report the estimated contrast and interval; explain which meaningful effects remain compatible with the analysis.
“Mean ± SEM.”Identify whether the aim is to show individual variation (for example, SD) or uncertainty about an estimate (for example, CI).
“There was a significant time effect.”Identify the time contrast and, if relevant, the direct treatment-by-time or between-group comparison.

How the Results paragraph could read

At 12 weeks, mean systolic blood pressure was 132.0 mmHg (SD 12.1) in the exercise group and 136.0 mmHg (SD 11.1) in the usual-care group. The baseline-adjusted between-group difference was −4.0 mmHg (95% CI −7.0 to −1.0 mmHg; P = 0.0096).

This is a reporting illustration, not evidence of treatment effectiveness. The Methods would explain the analysis, and the Discussion would interpret clinical importance and study limitations. The Results prose need not reproduce every value in a table; it should make the primary finding unmistakable.

Error bars, legends and independent units

State whether plotted bars or intervals represent SD, SE, confidence intervals, credible intervals or another statistic. An unadjusted mean plot is not the same as an adjusted treatment-effect plot. Legends should identify the outcome, units, comparison and relevant numbers, and should not imply that multiple observations from one experimental unit are independent.

11. Interpret the results in relation to the hypothesis

The Results communicate the numerical finding; the Discussion evaluates what it means. Return explicitly to the original aims and, where applicable, the prespecified hypothesis. In our example, an adjusted difference below zero is in the predicted direction, but that does not prove the hypothesis or establish the persistence of benefit beyond 12 weeks.

Distinguish statistical and clinical importance

The magnitude and precision of an estimated effect should be considered alongside intervention burden, potential harms, the patient population and external evidence. A small P value cannot make these judgements. Neither does a large point estimate become convincing when it is based on sparse or biased data.

Keep causal claims proportionate

An appropriately designed and analysed randomised trial may support causal conclusions about an intervention under its target treatment comparison. In observational research, regression adjustment alone does not automatically remove confounding or establish causation. For exploratory or descriptive studies without a formal hypothesis, return to the defined objectives and research questions.

Distinguish absence of evidence from evidence of little effect

Very wide uncertainty may leave important benefit and harm compatible with the findings, whereas a narrower interval may rule out some effects of practical relevance. Ordinary non-significance is not equivalent to equivalence; that conclusion requires a suitably framed question, margin and analysis.

12. Check statistical consistency throughout the manuscript

A sound statistical analysis can become difficult to evaluate when the same outcome is reported differently in the abstract, Results and figures. Before submission, trace each principal result through objective → outcome → analysis population → model → estimate → display → abstract → Discussion.

Reconcile the numbers

Check group labels, outcome units, timepoints, sample numbers, effect scales, model adjustments, P values and uncertainty intervals. If −4.0 mmHg appears in one location and −4.6 mmHg elsewhere, determine whether the figures refer to different models, rounding or a transcription error. Never silently pick whichever seems most plausible.

Check basic calculations and plausibility

Where sufficient information is available, recheck percentages, denominators and straightforward conversions. Review whether implausibly precise or extreme values may reflect unit mistakes, rounding or model differences. Statistical-method problems, reporting errors and genuine integrity concerns are distinct categories. An unusual number warrants investigation, not an automatic accusation.

Compare the report with the analysis plan

Where a protocol or statistical analysis plan exists, confirm that primary outcomes, analysis populations and methods agree with what was planned, or that deviations are transparently reported. Use relevant guidance such as CONSORT 2025, STROBE and the SAMPL Guidelines together with the target journal’s current instructions.1,2,4

For specialised help with analysis and reporting, see Blazon Scientific’s research writing and statistical analysis services.

13. Final statistical reporting checklist

Before submission, check whether a reader unfamiliar with the dataset could identify the scientific target, understand the reported estimates and evaluate the corresponding uncertainty. The objective isn’t to display every calculated number; it’s to supply enough information for meaningful interpretation.

  • The primary result corresponds to the objective and hypothesis where applicable.
  • Outcome, timepoint, comparison direction and reference group are clear.
  • Prespecified and exploratory analyses are distinguishable.
  • Relevant denominators, sample sizes, analysis populations and experimental units are reported.
  • Descriptive summaries suit the data; SD, SE and CI are labelled correctly.
  • Important effects or associations include interpretable estimates and appropriate uncertainty.
  • Absolute and relative effects are included where useful or required.
  • P values, when used, identify the tested hypothesis and are not overinterpreted.
  • Model types, predictors, adjustment variables, effect scales and material assumptions are clear.
  • Pairing, repeated observations and clustering are handled and reported appropriately.
  • Missing data, multiplicity and relevant sensitivity analyses are described.
  • Tables and figures identify units, statistics, denominators and uncertainty.
  • Numbers agree across the abstract, Methods, Results, tables, figures and supplements.
  • The Discussion returns to the aims and hypotheses without overstating causality or importance.
  • Applicable reporting guidance and journal requirements have been checked.

Even a careful checklist cannot certify statistical validity; it can reveal omissions that prevent readers from judging it.

Frequently asked questions

Should I report confidence intervals as well as P values?

Effect estimates with appropriate uncertainty generally provide information P values alone cannot communicate. Whether a P value is also appropriate depends on the objective, method and journal requirements. Explain what was estimated and what any reported P value tests.

Should I report mean ± SD or mean ± SE?

Use the measure suited to the purpose: SD describes dispersion of observations and SE describes estimated sampling variability of a statistic. An appropriate confidence interval is often more useful for showing uncertainty about a treatment difference or model effect.

How should I report a non-significant finding?

Report the estimate and uncertainty, then explain which scientifically meaningful effects remain compatible with the data under the model. Do not conclude that an effect is absent or that two treatments are equivalent merely because a conventional significance threshold was not crossed.

How should I report regression or mixed-effects results?

Identify the model, outcome, relevant contrast or predictor change, effect scale, uncertainty and adjustments. For clustered or repeated observations, explain how the dependence was handled and report the estimate that addresses the study’s question.

How many decimal places should I report?

Choose precision appropriate to the measurement, estimate and journal instructions. Excessive digits imply unjustified precision. Reporting blood pressure to six decimal places might make a sphygmomanometer rather nervous.

References and further reading

  1. Lang TA, Altman DG. Basic statistical reporting for articles published in biomedical journals: the “Statistical Analyses and Methods in the Published Literature” or the SAMPL Guidelines. International Journal of Nursing Studies. 2015;52(1):5–9. https://doi.org/10.1016/j.ijnurstu.2014.09.006.
  2. Hopewell S, Chan AW, Collins GS, et al. CONSORT 2025 statement: updated guideline for reporting randomised trials. BMJ. 2025;389:e081123. https://doi.org/10.1136/bmj-2024-081123.
  3. CONSORT 2025 explanation and elaboration: updated guideline for reporting randomised trials. BMJ. 2025;389:e081124. https://doi.org/10.1136/bmj-2024-081124.
  4. von Elm E, Altman DG, Egger M, et al. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement: guidelines for reporting observational studies. PLoS Medicine. 2007;4(10):e296. https://doi.org/10.1371/journal.pmed.0040296.
  5. Greenland S, Senn SJ, Rothman KJ, et al. Statistical tests, P values, confidence intervals, and power: a guide to misinterpretations. European Journal of Epidemiology. 2016;31:337–350. https://doi.org/10.1007/s10654-016-0149-3.
  6. Wasserstein RL, Lazar NA. The ASA’s statement on P-values: context, process, and purpose. The American Statistician. 2016;70(2):129–133. https://doi.org/10.1080/00031305.2016.1154108.
  7. Wasserstein RL, Schirm AL, Lazar NA. Moving to a world beyond “P < 0.05”. The American Statistician. 2019;73(sup1):1–19. https://doi.org/10.1080/00031305.2019.1583913.
  8. Lakens D. Improving Your Statistical Inferences. Open educational resource; consult the current online version. https://lakens.github.io/statistical_inferences/.
Home > Research Resources for Medical Researchers > How to Report Statistical Results