Statistics guide for bench scientists
A short map of what Quantify does after the N1 numerical blockers closed. Use the result panel for the number you publish.
Pick the comparison that matches the design
Unpaired two groups with possibly unequal variance: Welch t-test (runline default). Equal-variance Student t is an explicit choice, not a silent fallback. Paired observations: paired t or Wilcoxon signed-rank (exact for any n ≥ 1). Two independent groups that fail normality or are ranks: Mann–Whitney (exact when there are no ties and n1·n2 < 10,000; otherwise normal + continuity, named). Three or more groups: Welch ANOVA when variances differ; Kruskal–Wallis with Dunn-family post-tests when ranks are the right scale.
P values, tails, and failed fits
α and one- vs two-sided options change the stars and the P that Quantify reports. Upper tails use complementary CDFs so tiny P values do not collapse to 0. If an analysis cannot run (too few points, singular design, non-convergence), Quantify returns a typed “not computed” result with a reason — never F = 0 / P = 1 dressed up as a statistic.
Multiple comparisons are named, not implied
Post-tests reuse the omnibus MSE unless the dialog says otherwise. Dunn is the Kruskal–Wallis / Friedman family; Tukey and Dunnett are for ANOVA. Adjusted P uses stage-2 Benjamini–Hochberg (BKY) when that convention is selected — Methods text names it. Do not report “significant after Tukey” if the engine ran Dunn.
Dose–response: say whether X is already log
4PL / 3PL fits parameterise log10(EC50). The dialog asks whether X is concentration or log(concentration). A 1 nM EC50 is recoverable; older absolute-step floors that broke molar units are gone. Convergence is relative (step and SSE) plus a gradient check. n ≤ p is refused instead of inventing 1 residual df.
Quartiles: type 7 (R) vs type 6 (star-style)
Default quantiles are type 7 (R). Type 6 — the (n+1)p method — is available and labelled on the result. Small-n boxes will differ between the two; pick one and keep it in Methods.
What this guide is not
This is runline Quantify’s behaviour after the N1 numerical blockers closed (NIST StRD, R, and SciPy goldens in CI). It is not a competitor-package cross-check (withdrawn; public references only). Mixed-model Kenward–Roger / Satterthwaite details belong on the analysis Methods line for that run. Copy numbers from the result panel, not from this page.
Digit-by-digit NIST / R / SciPy scores: Quantify numerical report. Open Quantify
Guides: Technique hubs by category