Comparable-strata designer: Continuous-outcome trials (the coefficient-of-variation screen)

In Continuous-outcome trials (the coefficient-of-variation screen), bands are comparable on both the coefficient-of-variation (dispersion) ratio and the outcome level, so a pooled-variance contrast on the log scale is valid. Click the axis to add a cut, drag a cut to move it, click a cut to remove it.

Disclaimer. This planner is a research and educational aid that illustrates a statistical design methodology; it is not a validated clinical, diagnostic, or regulatory instrument, and is provided "as is" without warranty of any kind. Its suggestions rest on assumptions and tolerances that you are responsible for judging appropriate to your own data, and they do not replace review by a qualified statistician or the design, ethical, institutional, and regulatory requirements that govern your study. It is not medical advice and must not be used to direct the care of any individual patient. The author accepts no responsibility or liability for any decision, analysis, or outcome arising from use of this tool or its results, which are relied upon solely at the user's own risk.
Reference dataset behind “Parameters — published defaults”: Cushing's syndrome urinary steroid data — continuous, right-skewed urinary-steroid concentrations (log-normal). Aitchison & Dunsmore (1975); R: MASS::Cushings. Reproduce the data ↗ ·
Second example in this tool: Stamey prostate / PSA cohort — serum PSA (right-skewed, log-normal). Stamey et al. (1989); R: faraway::prostate. Reproduce ↗
How to cite. This planner is free and open-source software (MIT License), © 2026 William J. Dwyer. If you use it or build upon it, please retain this notice and cite the accompanying methodology paper together with the archived software — Dwyer WJ, Comparable-strata designer (v2.1.0) [Software], Zenodo, https://doi.org/10.5281/zenodo.20709963. You may use, modify, and redistribute it under the MIT License terms.

Confidence-interval calculator

Accurate 95% intervals for positive, right-skewed data at small N — built where the model is honest (the log scale, or a gamma GLM's log link) and back-transformed to the natural domain, using the screen's identity σ² = ln(1 + CV²). The eq-7 check routes to one of three rescues — stay natural, the log-scale t (log-normal), or a gamma GLM (gamma) — and two groups get a ratio of means with its p-value and CI reported in natural units. Load Example G for a real-data gamma rescue.

Load an example:
Example G is real public data: serum bilirubin (mg/dL) for patients without vs with ascites in the Mayo Clinic primary-biliary-cirrhosis trial — bilirubin is this paper’s running example, and the small ascites arm reads gamma on the eq-7 check. Dickson et al. (1989); R survival::pbc. Reproduce ↗
Runs in your browser.
What this is doing, and the one thing to watch

Why the log domain. If the data are log-normal, ln(x) is exactly normal, so a t-interval on the logs is valid without the central limit theorem, even at n = 5–10; back-transforming gives an asymmetric natural-domain interval that respects the skew.

The catch. The naive back-transform exp(ŷ ± t·sy/√n) is a CI for the geometric mean = median, not the arithmetic mean. For the arithmetic mean this uses Cox's method on ŷ + sy²/2 (Land's method is the exact version).

Heteroscedasticity & the fallback. Constant CV means SD ∝ mean; the log is the variance-stabilising transform for that case. For two groups, ρ = ln(1+CV&sub2;²)/ln(1+CV&sub1;²) selects pooled vs Welch on the logs. When the sy² ≈ ln(1+CV²) check diverges, or the data contain zeros/negatives, abandon the parametric route for a BCa bootstrap or a wider distribution-free interval.