Longer write-ups of my papers, with the methods, the R software and the numbers behind
the conclusions, and practical notes on statistics in biomedical research.
A confidence interval is read off a sampling distribution, and the bootstrap estimates that distribution by resampling the data. Three interactive figures show how, why the simplest bootstrap interval is too narrow in small samples, and how often the percentile, BCa, bootstrap-t and t intervals cover the truth. Tables cover what to resample in each design and how many resamples are enough, and a checklist covers what to report.
Every p-value is a tail area of a null distribution. Shuffling the group labels builds that distribution from the data. Three interactive figures show how, how the t distribution approximates it and where it fails, and how the p-value itself varies from study to study. Tables cover how many shuffles are enough, what a shuffle assumes and what to shuffle in each design, and a checklist covers what to report.
A p-value, an effect size and a confidence interval answer three different questions about one result. Three figures (two of them interactive) show how they behave, and tables show how to read an interval against the smallest effect that matters and what to report for common analyses.
The SD describes the data; the SE describes an estimate. Two interactive figures show why SE bars mislead, what overlapping error bars do and do not tell you, and a table maps common situations in a manuscript to what to report.
Leakage in cross-validation comes from study structure, from procedure, and from artefacts only the data reveal. This methods note shows how three R packages take one source each, how they compose into one pipeline with a feedback loop, and what each layer buys in a 30-replicate simulation.