Everything on this page is a preprint: a manuscript that has been submitted to a journal but has not yet completed peer review. These are author versions. They may contain errors, their findings may change substantially during review, and they should not be treated as established evidence or used to guide clinical or policy decisions.
Where a manuscript is later published, the peer-reviewed version of record supersedes the file here and is linked from the Publications page. Please cite the version of record whenever one exists.
I post preprints because I think work should be readable while it is being judged, not only after. It is the same reasoning behind FAIR Press Journals, which I co-founded: research is more useful when the reasoning, the code and the data are open to inspection.
Are normality tests applied to model residuals or to raw data in physiotherapy and rehabilitation research, and what are the consequences for statistical power? A full-text meta-research audit and simulation study
Normality tests such as Shapiro–Wilk or Kolmogorov–Smirnov are widely used to choose between parametric and non-parametric analyses — but the assumption they concern applies to the residuals of the fitted model, not to the raw outcome data. This full-text meta-research audit covers articles published in 16 physiotherapy and rehabilitation journals between 2015 and 2024 that reported a named normality test. Of 1,155 definitively codable papers, correcting for classifier labelling error gives a conservative prevalence of 89.0 % (95 % CI 81.4–95.2) applying the test to raw data rather than residuals, rising to 90.4 % among determinable papers; only 12 papers (1.0 %) tested residuals. The error exceeded 85 % in every journal with at least ten coded papers and showed no material change over the decade. A simulation across four distributions and sample sizes from 10 to 100 per group shows the power cost never exceeded about three percentage points — the test eventually chosen mattered far more than the residual-versus-raw distinction. Misapplication is the norm, the consequences are small, and correct practice is essentially free.
Does the choice of Sleep Regularity Index calculator or sleep/wake input change the conclusions? A re-analysis of the NHANES 2011–2014 sleep regularity–obesity association
An earlier study reported an inverse cross-sectional association between the Sleep Regularity Index (SRI) and body mass index in NHANES 2011–2014, computed from the released sleep/wake classification. I re-estimate it under different calculators, smoothed versions of the released classification and series derived anew from the raw signal (published sample, n = 7,085). The released classification switched between sleep and wake a median of 16 times per day. Smoothing raised the median SRI from 61 to 68–76, while on the same input the calculators agreed to 0.1 points. The fully adjusted BMI difference between the top and bottom SRI quintile was −4.5 % for the released series and −2.8 % to −3.7 % after smoothing. Series derived anew from the 80 Hz raw signal (random sample of 1,000) scored 10.1 to 17.7 points higher and gave −0.7 % to +0.4 % against −4.8 % for the released classification. The SRI level depended on the input and only marginally on the calculator, and the inverse association with BMI weakened with every smoothed and newly derived input.
Moving the goalposts on purpose: almost sure hypothesis testing and what it would do to physiotherapy trials
Clinical trials declare their results at a single fixed threshold, a p-value below 0.05, and that line does not move with the sample size. Under a true null the standardised statistic crosses 1.96 infinitely often, so the line can never be consistent. Almost sure hypothesis testing lets the level shrink with the sample size (αn = n−p, p > 1), so that only finitely many errors of either kind occur with probability one. I ask what this does to real physiotherapy trials, first with the operating characteristics at a fixed sample size and then with three empirical applications: three open-access trials (n = 140 to 2,249), the primary outcome of 21 trials with openly deposited individual participant data, and 256 musculoskeletal trials reanalysed from published summary statistics. At the sizes physiotherapy trials reach, the bar sits at |z| ≈ 2.5 to 3.9, and 80 % power for a moderate effect costs roughly twice the sample size. Of 256 comparisons, 127 clear 0.05 and 82 clear the almost sure bar. The rule is best used as a pre-specified sensitivity analysis beside the effect estimate and its confidence interval.