testflow

Bug fixes (scientific validity)

• sample_size_bioequivalence(method = "iterative_tost") now computes the two-one-sided-test (TOST) power exactly via the noncentral t distribution (Phillips 1990), instead of a normal approximation that treated the within-/between-subject variance as known. The previous version under-sized bioequivalence studies by up to ~20% at small n or CV/GMR combinations away from GMR = 1 (e.g. 13 vs. the correct 16 subjects for a 2x2 crossover at GMR = 0.95, CV = 0.15, 90% power). Verified in package tests against PowerTOST::power.TOST(), the regulatory-standard, Owen's-Q-based implementation, matching exactly for balanced designs. **If you sized a bioequivalence study with testflow before this fix, re-run it** - the corrected n may be larger. method = "normal_approx" is unaffected.
• test_agreement()'s hypothesis test of kappa = 0 now uses the null-hypothesis standard error (Fleiss, 1981), instead of incorrectly reusing the confidence-interval standard error (Fleiss, Cohen & Everitt, 1969, evaluated at the observed kappa) as the test denominator. The CI itself was already correct and is unchanged. Because the CI-based SE understates the correct null SE whenever agreement is materially above chance (the typical case), the previous z-statistic was inflated and p-values were anti-conservative (too small); in one check this moved z from 17.2 to the correct 14.4, matching irr::kappa2() exactly. **If you relied on a test_agreement() p-value near your significance threshold, re-run it** - the corrected p-value is larger (less significant).

Sample-size precision planning

sample_size_precision(endpoint = "binary", design = "one_sample") gained

rare-prevalence support (e.g. newborn-screening / rare-event studies):

• New method = c("wald", "wilson", "exact"). "wald" is unchanged and remains the default, so existing calls return identical numbers. "wilson" and "exact" (Clopper-Pearson) instead run a deterministic doubling-bisection integer search for the minimum n whose confidence interval has maximum one-sided half-width no greater than width - more reliable than the normal approximation for small or extreme proportions. The achieved half-width is not perfectly monotone in n (integer event counts round differently as n changes), so the search includes a bounded backward repair scan to guard against converging on a non-minimal n; verified against an exhaustive brute-force scan in the package tests.
• New criterion = c("expected", "worst_case") controls whether the search targets precision at the single anticipated event count, or requires it to hold across a prevalence-local band of plausible counts.
• New min_expected_events and max_n control the rare-event diagnostic threshold and the search's upper bound, respectively.
• x$diagnostics now reports rare-event planning quantities that a CI-width target alone does not guarantee: expected_events, probability_zero_events, probability_at_least_one_event, complete_case_n vs. adjusted_n (dropout-inflated), and achieved precision at complete_case_n. An approximation_warning field (and a matching single warning()) fires when the Wald approximation is used or when expected event/non-event counts are small.
• conservative = TRUE (worst-case p = 0.5) remains method = "wald" only and now errors informatively if combined with "wilson"/"exact", since there is no well-defined worst-case p for the asymmetric search.
• print(), summary(), and as_tibble() on sample_size objects gained an optional "Diagnostics" section / columns, populated only when diagnostics are available; all previously existing fields are unchanged.

Binary design = "two_sample" precision planning (risk-difference

half-width) now implements unequal allocation correctly instead of silently

returning the equal-allocation result whenever allocation != 1. This is a

**behavior change** for existing calls with allocation != 1; `allocation

= 1` (the default) is numerically unaffected.

A developer-facing numerical validation script

(inst/validation/validate-sample-size-precision.R) independently

recomputes achieved confidence intervals and checks the minimum-n property

across a grid of prevalences, widths, and alpha levels.

Bug fixes

• test_repeated()/test_repeated_long() no longer error with "missing value where TRUE/FALSE needed" on R-devel. summary() on an aovlist fit walks every Error() stratum, including the degenerate intercept-only stratum (a single fitted value), and a recent change to stats:::summary.aov()'s perfect-fit check can produce NA there. We now summarize each stratum independently and skip any that fail, since the intercept-only stratum was never used for the reported test.

Assumption checking

Closed remaining gaps where an assumption was labeled "not checked"

because no diagnostic had been implemented (as opposed to being genuinely

uncheckable, e.g. too few observations, which is unchanged). Every

assumption status now reflects either a real check or an honest "not

applicable"/"not enough data" label - never an unimplemented one.

• car is now a hard dependency (Imports, not Suggests): regression homoscedasticity (Breusch-Pagan), multicollinearity (VIF), and ANOVA/factorial variance homogeneity (Levene's test) no longer fall back to "not checked" when car isn't installed, and test_factorial(type = 2/3) no longer errors for the same reason.
• Added a real sphericity check for test_repeated()/test_repeated_long() with 3+ conditions, via Mauchly's test (stats::mauchly.test() on the wide-format multivariate model, X = ~1) - verified to match mauchly.test() and rstatix::anova_test() exactly. Reports "not applicable" for exactly two conditions, where sphericity isn't defined.
• Added a real symmetry-of-deviations check for test_one_sample() and test_paired() (used to justify the Wilcoxon signed-rank test), via the closed-form Cabilio & Masaro (1996) test - verified to match lawstat::symmetry.test(option = "CM", boot = FALSE) exactly.
• Added a real linearity check for test_correlation(method = "pearson"), via a quadratic-term F-test (anova() comparing y ~ x to y ~ x + I(x^2)) - equivalent to Ramsey's RESET test with a single quadratic regressor term, verified to match lmtest::resettest(power = 2, type = "regressor") exactly.
• test_correlation_matrix() now applies a Benjamini-Hochberg correction to its pairwise p-values by default (p.adj column in the correlation table), instead of reporting pairwise p-values with no correction applied.

Sample size planning

• Removed sample_size_ordinal(method = "whitehead"). Unlike every other formula in the sample-size module, the proportional-odds formula had not been checked against a published worked example or an external reference implementation (only against itself), and the input convention for probs (which arm's category distribution to supply) is ambiguous across secondary sources describing Whitehead (1993). sample_size_ordinal() now implements only Noether's method, which was independently verified. It no longer takes method, probs, or odds_ratio arguments.