Blog
Manifesto · Start here
EBM taught doctors to cite statistics, not read them. Five numbers that explain the problem, and why we built StatsPlease. Read this first.
June 27, 2026 · ~5 min read
Read the manifesto →Regression
R² is a share of variance, not a percentage correct and not a grade on the model. A slope can be highly significant at R² under 20%.
August 15, 2026 · ~7 min read
Read guide →One sample
Comparing a case series mean against a published figure is an inferential claim, and it quietly treats someone else’s estimate as exact.
August 15, 2026 · ~6 min read
Read guide →Reliability & agreement
Chance agreement climbs whenever one category dominates, and kappa falls with it even when the raters are reading cases identically.
August 15, 2026 · ~7 min read
Read guide →Browse by topic
Filters the topic sections below. Start here and Latest stay put.
Test selection
A decision framework organized by research question: name what you are comparing, check the one assumption that decides the routing, and follow the link to the worked example.
July 11, 2026 · ~6 min read
Read guide →Interpretation
Not the probability the hypothesis is true, and not the chance your result was a fluke. The ASA's definition, a computed example, and the correct sentence to write.
July 2, 2026 · ~11 min read
Read guide →Assumptions
Moderate non-normality is well tolerated. Unequal variance paired with unequal group sizes is not, and it doesn't announce itself. Two real comparisons, computed.
July 12, 2026 · ~5 min read
Read guide →Parametric vs not
A flowchart tells you which test to run. It doesn't tell you what sentence you're licensed to write once it finishes. The same real trial data, two different claims.
July 12, 2026 · ~7 min read
Read guide →Data handling
Every package decides what counts as missing and which columns are numbers before a test runs, and most decide it silently. What stops an import, what gets repaired, and where a tool should refuse to guess.
August 3, 2026 · ~9 min read
Read guide →Interpretation
One outcome score can carry four separate thresholds, and they do not collapse into each other. A 2024 shoulder arthroplasty study established all four in the same 100 patients.
July 28, 2026 · ~7 min read
Read guide →Interpretation
A change can exceed the clinically important difference and still sit inside the instrument's own measurement error. In 128 patients with lumbar spinal stenosis, one of three instruments failed exactly that test.
July 28, 2026 · ~8 min read
Read guide →Interpretation
A confidence interval that misses significance is not automatically uninformative. If the whole interval sits below the threshold, that is a positive finding. Worked on a real rotator cuff trial.
July 31, 2026 · ~7 min read
Read guide →Interpretation
Applying an individual-level threshold to a group average is more contested than it looks. It also has a documented failure mode, and a genuine regulatory defence.
July 31, 2026 · ~8 min read
Read guide →One sample
Comparing a case series mean against a published figure is an inferential claim, and it quietly treats someone else’s estimate as exact.
August 15, 2026 · ~6 min read
Read guide →Two groups
When an independent t-test is the right test, and how to report a non-significant result in full with its effect size and confidence interval.
July 11, 2026 · ~5 min read
Read guide →Before vs after
Paired t-test or Wilcoxon signed-rank? The choice depends on whether the difference scores are normal, and on real data it can flip significance.
July 11, 2026 · ~5 min read
Read guide →Before vs after
The standard sentence — “a significant difference in medians” — holds only when the paired differences are symmetric. What the rank sum actually licenses you to write.
August 14, 2026 · ~6 min read
Read guide →Two groups, skewed
How to choose between the two most common comparison tests, with normality testing and reporting guidance.
June 26, 2026 · ~8 min read
Read guide →Three or more groups
A significant ANOVA says something differs, not which groups. See why the post-hoc test that follows can tell a different story.
July 11, 2026 · ~5 min read
Read guide →Repeated measures
When the same patients are measured at 3+ timepoints and the data are skewed, the Friedman test replaces repeated-measures ANOVA. Computed on real trial data.
July 12, 2026 · ~5 min read
Read guide →Three or more groups
A significant H statistic confirms a difference exists somewhere and localises nothing. Dunn's test with a correction is the second step reviewers check for.
August 14, 2026 · ~7 min read
Read guide →Association
Pearson and Spearman answered the same dataset 0.22 points apart. How to check normality and pick the right correlation coefficient.
July 11, 2026 · ~5 min read
Read guide →Reporting
Writing r when you ran Spearman is a notation error. The correct symbol, AMA format, and a real result computed with scipy.
July 13, 2026 · ~6 min read
Read guide →Regression
R² is a share of variance, not a percentage correct and not a grade on the model. A slope can be highly significant at R² under 20%.
August 15, 2026 · ~7 min read
Read guide →Binary outcome
What an adjusted odds ratio actually means: a real three-predictor logistic regression shows what holding the other variables constant does.
July 11, 2026 · ~4 min read
Read guide →2x2 tables
Chi-square and Fisher's exact test the same question. Expected cell counts decide which one is valid. A worked example computed both ways on real data.
July 11, 2026 · ~6 min read
Read guide →R by C tables
An R×C table doesn't automatically qualify for chi-square just because a 2×2 table would. When it needs Fisher's exact instead, computed on real cohort data.
July 12, 2026 · ~7 min read
Read guide →Ordered categories
A standard chi-square discards the order in an ordered exposure entirely. The Cochran-Armitage trend test doesn't, computed on real data.
July 12, 2026 · ~7 min read
Read guide →2x2 tables
The same patients measured twice on a yes/no outcome break the independence a chi-square test assumes. Only the discordant pairs carry the signal.
August 14, 2026 · ~7 min read
Read guide →Survival
How to read a survival curve, what the log-rank test really compares (the whole curve, not one time point), the proportional-hazards trap, and how to report it in AMA format. Worked on the classic leukaemia remission trial.
July 18, 2026 · ~9 min read
Read guide →Reliability
Cohen's kappa needs categories. A screw length in millimeters doesn't have any. ICC has ten forms, and each one gives a different number from the same data.
July 12, 2026 · ~6 min read
Read guide →Reliability & agreement
Chance agreement climbs whenever one category dominates, and kappa falls with it even when the raters are reading cases identically.
August 15, 2026 · ~7 min read
Read guide →Meta-analysis
I² and Cochran's Q measure real disagreement between pooled studies, not just noise. A real 13-trial result where they disagree sharply with a naive read.
July 12, 2026 · ~7 min read
Read guide →Study planning
The four inputs of an a-priori power analysis, a worked example (64 per group), how to report it, and why post-hoc power is the reviewer trap to avoid.
July 18, 2026 · ~8 min read
Read guide →Trial design
The operating surgeon always knows which procedure they are performing. Who can actually be blinded, what rides on it, and the one sentence that reports it properly.
July 11, 2026 · ~6 min read
Read guide →Workflow
Export raw, not labels. Match REDCap field types to statistical variable types, then let StatsPlease pick the test. A real result, computed.
July 13, 2026 · ~9 min read
Read guide →Perspective
Same model, same data, four accuracy rates. Why the methods-section standard is verification, not trust in whoever (or whatever) ran the numbers.
July 2, 2026 · ~5 min read
Read essay →Tools
SPSS, R, Stata, Prism, JASP, jamovi, Julius, and StatsPlease compared honestly: price, learning curve, what journals accept, and which to choose for your situation.
2026-08-08 · ~12 min read
Read guide →Put it to work
Every result in these guides was computed with scipy, not generated by ChatGPT or any other language model. Do the same with your own data: upload the file, open the Analysis tab, set Outcome and Group, then press the preset that matches your design, Group Comparison, Before vs After, or Correlation. Pressing the preset is what runs the analysis; there is no separate run step. The Results tab returns the statistic, the P value, and the effect size, written in AMA or APA format.
Start freeFree during open access. No credit card.