How to Write Statistical Results for Each Statistical Test: Complete Medical Research Guide

Normally Distributed Continuous Variables

Non-normal continuous variables should never be summarized with mean and SD alone — report median and interquartile range (IQR) instead: "Median length of stay was 4 days (IQR 3–7)." APA 7 permits either format but requires you to state explicitly which one you used and why (i.e., the result of your normality test). For a STROBE-compliant cohort or cross-sectional study, this Table 1 should also report the number and percentage of missing values for every variable, since undisclosed missingness is one of the most frequently cited STROBE deficiencies in peer review. Categorical baseline variables follow the same logic in reverse: report n (%) for every category, using Valid Percent once any missing data exists, exactly as covered in our SPSS output interpretation guide.

The p-value alone cannot tell a reader whether a 1.8 g/dL difference in hemoglobin is clinically meaningful — the mean difference, its CI, and Cohen's d together do that job, which is why all three appear in the template above rather than the p-value in isolation. Use the independent t-test calculator to generate this output directly. For an RCT comparing two treatment arms, CONSORT explicitly requires the between-group difference and its CI to be the headline figure for both the primary and secondary outcomes — not the p-value alone, and not merely "significant" or "not significant" as a verdict.

This is the single most common reporting error in before-after studies, and it matters because the two numbers tell entirely different stories: a correlation of 0.71 only says patients kept roughly the same relative ranking from pre- to post-treatment, while the paired t-test is what actually establishes whether the mean level changed. See our paired vs unpaired t-test guide for choosing the correct design, and the paired t-test calculator to compute this directly.

Report medians consistently throughout — mixing means with a non-parametric test signals the analysis and the write-up do not match, which is exactly the kind of inconsistency a statistical reviewer is trained to catch on a first read. Because the Mann-Whitney U test does not directly estimate a mean difference, the effect size r (calculated as Z divided by the square root of N) is the standard way to convey magnitude, and should be interpreted using the same small/medium/large bands used for Pearson's r. Use the Mann-Whitney U calculator, and see Mann-Whitney vs t-test for when this test applies.

The Wilcoxon signed-rank test is the non-parametric counterpart to the paired t-test, and the same principle applies: the p-value only tells you a change occurred, while the medians and effect size tell the reader how large that change was. Compute this directly with the Wilcoxon signed-rank calculator.

Eta-squared (η²) is not produced automatically by most software and must be calculated as the between-groups sum of squares divided by the total sum of squares — worth the extra step, since a large F with a tiny η² in a very large sample tells a different clinical story than the same F with a large η². Use the one-way ANOVA calculator. For a three-arm CONSORT trial, report each pairwise comparison's own effect size and CI, not only the omnibus F.

Reporting the sphericity check is not optional decoration — a violated-but-unreported sphericity assumption means the uncorrected F-test's p-value cannot be trusted, and an examiner who spots the omission will reasonably question every other result in the same analysis. Use the repeated measures ANOVA calculator for the full breakdown, including the automatically corrected degrees of freedom.

Epsilon-squared (ε²) is the recommended effect size for Kruskal-Wallis because, unlike eta-squared, it is derived from the rank-based H statistic itself rather than assuming normally distributed sums of squares. Run this with the Kruskal-Wallis calculator.

Kendall's W ranges from 0 (no agreement in ranking across time points) to 1 (perfect agreement), and functions as the Friedman test's effect size in the same way η² does for ANOVA — report it whenever the omnibus result is significant. Use the Friedman test calculator for repeated ordinal or non-normal measurements across 3+ time points.

Cramér's V is preferred over the raw phi coefficient for any table larger than 2×2, and gives readers a bounded 0-to-1 sense of association strength that the chi-square statistic alone (which scales with sample size) cannot provide. Use the chi-square calculator. For a STROBE-compliant cohort study, also report the raw counts and row percentages in an accompanying table, not only the summary statistic in the text.

Notice that the underlying counts and percentages are still reported in full even though the difference did not reach significance — a non-significant result is still a result, and CONSORT and STROBE both expect adverse-event and safety-outcome data to be reported completely regardless of statistical significance. Run this with the Fisher's Exact Test calculator, and see Chi-Square vs Fisher's Exact for the decision rule.

A correlation coefficient can never establish direction of causation on its own, no matter how small the p-value — only a study design built to isolate cause (an RCT, or a longitudinal design with appropriate confounder adjustment) can support causal language, and journal reviewers routinely reject manuscripts that overstate a correlational finding this way. Use the correlation calculator to compute r with its 95% CI.

Because Spearman's rho is calculated on ranks rather than raw values, it is also the appropriate choice whenever one or both variables are ordinal (Likert-type items, disease stage) rather than truly continuous, even if both happen to be normally distributed. See Pearson vs Spearman correlation for choosing between the two, computed via the same correlation calculator.

Adjusted R², not plain R², is the figure to headline once more than one predictor is in the model, because plain R² mechanically increases every time a predictor is added — even a clinically meaningless one — while the adjustment penalizes model complexity and gives a more honest sense of how much variance the model genuinely explains. Use the linear regression calculator, and see multivariate analysis for building multi-predictor models.

The Hosmer-Lemeshow test follows the same reversed logic as a normality test: a non-significant result (p > 0.05) is the desired outcome, meaning the model's predicted probabilities are well calibrated against observed outcomes — reporting it without explaining this direction is a common source of reviewer confusion. Use the logistic regression calculator and see the full logistic regression guide.

If the median survival time is not reached in one or both groups by the end of follow-up — common in trials with high cure rates or short follow-up — report the survival rate at a clinically meaningful fixed time point instead (e.g., "12-month survival was 78% vs 61%"), which CONSORT's survival-outcome extension treats as an acceptable and often more informative alternative. Use the Kaplan-Meier calculator and see the full Kaplan-Meier guide.

A hazard ratio describes the instantaneous risk of the event at any given moment, conditional on having survived to that moment — it is not the same quantity as a relative risk or an odds ratio, and using those terms interchangeably is a factual error a statistical reviewer will flag immediately. Use the Cox regression calculator to fit the model and check proportionality.

The Youden index (sensitivity + specificity − 1) is the standard, defensible method for choosing a single cutoff to report, rather than eyeballing the curve or choosing whichever cutoff produces the most favorable-looking numbers after the fact. Use the ROC curve calculator and see Sensitivity, Specificity, PPV & NPV for the surrounding vocabulary.

The same raw data can produce meaningfully different ICC values depending on whether a one-way or two-way model is chosen and whether consistency or absolute agreement is specified, so naming the exact model is not a formality — it is the only way a reader can judge whether your 0.91 is comparable to a 0.91 reported elsewhere. Use the ICC calculator and see the full ICC guide.

Two measurement methods can be very strongly correlated yet systematically biased relative to one another by a clinically important amount — correlation is blind to that kind of fixed offset, which is exactly the failure mode Bland-Altman analysis is designed to catch. Use the Bland-Altman calculator and see the full Bland-Altman guide.

Raw percent agreement systematically overstates true reliability whenever one category is much more common than another, because two raters can agree by chance alone a large fraction of the time on the dominant category — kappa corrects for exactly this, which is why journals in diagnostic and rater-reliability research require it over simple percent agreement. Use the Cohen's Kappa calculator and see the full Kappa guide.