Roadmap¶

The plan to 1.0.0 and the work done towards it, from the package's ROADMAP.md.

This document tracks where the package is going. It is updated alongside meaningful code changes. Milestones are ordered, not dated. The history of what changed and why lives in NEWS.md; the statistical evidence lives in inst/validation/.

Where things stand (0.17.0)¶

On CRAN. 0.11.2 was accepted in September 2026, after one review round on 0.11.1; 0.12.0, 0.13.0, 0.14.0 and 0.15.0 are GitHub releases. 0.16.0 was accepted on 2026-10-29, after its auto-check. From 0.16.0 releases go to CRAN again, about one to two months apart (decided 2026-10-28), so not every GitHub release goes. CRAN's additional check of 0.16.0 with Intel MKL as BLAS failed on tests that compared two computations of the same numbers bit for bit. 0.17.0, published on CRAN on 2026-11-01, fixes that check: on CRAN the tests compare such numbers to a tolerance of 1e-10, off CRAN still bit for bit, and cran-blas.yml checks the package under MKL, OpenBLAS, BLIS, R without long doubles and a noisy BLAS on the pull requests that change it (#172, under 0.11.x).
The validation matrix is complete for every exported perform*Test(). The sweep found six broken procedures, all corrected or withdrawn, and twenty-five of the twenty-six heteroscedasticity tests hold their nominal level (inst/validation/README.md). The tests added since have studies of their own: inst/validation/harrison-mccabe-size-power.R and inst/validation/glejser-skewed-errors.R.
Weighted fits were outside that matrix, and wrong. Every design in it fitted an unweighted model. On a weighted fit the tests read the raw residuals, whose variance differs across observations by assumption, and rejected a correctly weighted model every time. 0.12.0 corrects that and adds weighted designs to the matrix (inst/validation/weighted-fits-size.R). In 0.13.0 the bootstrap tests accept a weighted fit too, with designs of their own in the same study, and since 0.15.0 they estimate the weights of a fitWLS() fit again in each replicate rather than take them as known.
Reference equivalence is asserted to 1e-8 against lmtest, car and plm wherever a reference implementation exists. lmtest::bptest() is not a reference for weighted fits: it tests the raw residuals.
The public surface has been reviewed. 0.14.0 gave each of the 129 exports of 0.13.0 a verdict (the export inventory under 0.14.0): 104 are core API, 22 are deprecated and warn, and 3 move to a companion package that has no name yet. With bootstrapTestStatistic(), simulate_size() and simulate_power(), and runAdvancedDiagnostics() in 0.15.0, the package exports 133 objects, and test-public-api.R asserts the 107 planned for 1.0.0. 0.16.0 adds no export: it registers eight S3 methods, print() and summary() of the suites of runHeteroTests() and print() and coef() of their summaries. 0.17.0 changes no export, argument or default: NAMESPACE holds the same 133 exports.
The interface matches its description. In 0.15.0 every test takes (model, data, ...), performBoxMTest(data, group) being the documented exception, studentize chooses the studentized statistic throughout, ?test_results documents what every test returns, and every test refuses the same bad inputs with one message. The old call forms warn in 0.15.x and stop working in 1.0.0. The defaults changed with it: 999 replicates for every resampling function, White and Koenker as the default battery, and the corrected Glejser statistic.
The tests read the model they are given. In 0.16.0 the tests of a model, but the resampling tests, read the response, the design, the residuals, the weights, the offset and the rank from one validated model context per call, built from the stored fit, and match data to its observations, named data in any order and numbered data in the order of the fit (steps 1 to 7 of docs/design/model-context.md). White's degrees of freedom, Box's correction factor, Szroeter's null moments, Pesaran's test on an unbalanced panel and the ties of the Fligner-Killeen test are corrected, and the caches follow a contract. In 0.17.0 the other four steps are done: the resampling tests refit their replicates on the stored design (step 8), a fit with na.action = na.exclude is read as the same fit with na.omit (step 9), the runners and the functions outside the tests match data as the tests do (step 10), and a MASS::rlm() fit is refused and a penalised mgcv::gam() fit tested or refused by what each test assumes (step 11). What stands between 0.17.0 and 1.0.0 is the internals and the defect backlog (0.18.0), the evidence at the release gate and the open decisions (0.19.0), and the freeze.

What 1.0.0 means¶

1.0.0 is a promise rather than a feature count. From 1.0.0:

Exported names, argument names and order, and defaults change only in a major release, and only after a minor release in which the old form still works and warns.
Every exported test returns an htest object with a documented, uniform set of fields.
A correction to a statistic, its degrees of freedom or its p-value is a bug fix. It is never deferred for the sake of compatibility. It bumps the minor version and gets its own NEWS.md section, as 0.11.0 did.

Version numbers follow from that: a patch release changes no reported value; a minor release adds functionality or corrects a statistic; a major release breaks the API.

Everything below exists to make that promise safe to give. The sequencing rule this file has always followed still holds: statistical correctness first, then the API, and never both in one release. 0.12.0 reopened correctness for weighted fits and closed it again, and 0.13.0 corrected what was found after it, so the API is next.

Guiding principles¶

Statistical correctness first. Every test must reproduce an established reference (or a defensible derivation) and be covered by a test that checks behaviour (size and power), not just object structure.
A reference comparison is not a substitute for a size check. Four of the six faults the sweep found were in procedures with no reference to compare against, and the two that had one agreed with it.
Guards are simulated, not stored. A stored-value regression test would have frozen each of those faults rather than caught it.
One consistent interface. All diagnostics return base-R htest objects and follow the perform*Test(model, data, ...) convention. (Since 0.15.0 every exported test does, except performBoxMTest(data, group), which tests the data rather than residuals; see 0.15.0.)
Scale and robustness as first-class concerns: streaming implementations for large data, resampling and robust variants for small or non-normal samples.

Milestones¶

0.11.x: CRAN maintenance (now)¶

Watch the CRAN check results across all flavours and fix anything flagged within the deadline CRAN sets.
README: install from CRAN first (install.packages("heteroTests")), with GitHub as the development route; add a CRAN badge and use the canonical URL https://CRAN.R-project.org/package=heteroTests.
The README feature list is out of date. It still advertises an HC0–HC4 covariance test and Cameron–Trivedi, both removed in 0.8.0. Bring it in line with the exports, and extend test-documentation-consistency.R so a README or vignette that names a removed function fails. The guard went into test-public-api.R, beside the list of removed exports, and also matches the names prose gives the removed tests.
Documentation drift: the performStudentizedBPTest() help page describes performKoenkerTest() as "the absolute-residual variant", which predates the Breusch–Pagan/Koenker correction. Koenker is the studentized n R^2 form.
Check that the Zenodo record picked up 0.11.2, and cite the CRAN release in inst/CITATION and CITATION.cff. Zenodo archives 0.11.2 and 0.12.0; only the 0.12.0 record links to CRAN.
Close tracking issue #38 as On CRAN. It was closed on 2026-09-27.
Reinstate a release gate. R CMD check --as-cran can stay weekly rather than per merge, but it must pass in full on the tagged commit (vignettes and --run-donttest included) before any submission, alongside win-builder (R-devel). release.yml now runs --as-cran --run-donttest on the tarball it is about to tag and creates no tag if the check warns or errors; win-builder stays a manual step (cran-comments.md).
The gate failed. The weekly --as-cran job stopped with an error on 2026-09-28 and on 2026-10-05: the \donttest{} example of performWhiteTestBootstrap(parallel = TRUE) starts detectCores() - 1 processes, three on the runner, and --as-cran allows two. The example is skipped on Windows, so a local check there does not see it. Fixed in 0.13.0 by changing the example, which now runs the call only where it starts at most two processes; the "CRAN submission checks" job passed on the branch on 2026-10-11 and on main on 2026-10-12. The function still uses detectCores() - 1 workers by default; since 0.15.0 n_cores sets them, and the result no longer depends on how the draws are split.
CRAN's check of 0.16.0 with Intel MKL as BLAS failed on 2026-10-30 (https://www.stats.ox.ac.uk/pub/bdr/Rblas/MKL/heteroTests.out, R-devel r90786): [ FAIL 179 | WARN 17 | SKIP 96 | PASS 15260 ], where the eight regular flavours that had checked it were OK. Every failure was an expect_identical() between two computations of the same numbers: a cached result and an uncached one, a call with data and one without, an lm and the Gaussian glm of the same model, an old call form and the new one, the streaming runner and the exact one, a test and its registered name. They agree to the last bit on R's reference BLAS; under MKL they differed by up to 1.6e-12 relative (performVarianceFormTest(), 0.008329376120413 against 0.008329376120400), and MKL gave different last bits for the same call made twice. In one the last bits changed a discrete outcome: on every row of iris twice, which absolute residuals of Fligner-Killeen's test tie, 8.133 for the lm against 8.149 for the glm, a fragility of performFlignerKilleenTest() itself (under 0.18.0). Fixed in #172. The tests compare such numbers with expect_identical_numbers() (tests/testthat/helper-numbers.R), bit for bit off CRAN, as in CI and on the author's machine, and on CRAN, as testthat's on_cran() decides it, with expect_equal() to a tolerance of 1e-10: waldo's, the mean absolute difference of the elements that differ relative to their mean expected size, or absolute where that size is at most 1e-10. That is 64 times the largest difference under MKL. It is used at 235 sites in 30 files, found by running the suite as CRAN runs it with every expect_identical() of two non-integral numbers recorded, and by the noisy BLAS below. Six blocks whose ties, merged residuals or cache hits turn on the last bits run off CRAN only, the Fligner block among them: 27 of the expectations CRAN ran on main. On CRAN, the F of the variance form test on the cells of the matrix with missing rows and White's test on a design 1e6 from zero are compared to 1e-8 (blas_tolerance()), Harrison-McCabe's two-sided p-value with twice the one-sided one on the scale of the integral both come from, and the p-values of Box's M, about 1.8e-10, by their logs. cran-blas.yml runs R CMD check --as-cran with NOT_CRAN unset in the R-hub containers for MKL (CRAN's version, 2026.2), OpenBLAS, BLIS and R without long doubles, and with a BLAS that moves its results by up to one unit in the last place (tools/noisy-blas.c, on three seeds), on pushes to main, weekly, and on the pull requests that change the package or its tests. On main before the fix MKL passed, 2025.3 and 2026.2 alike: on the runners' AMD processors MKL runs its generic code, whose last bits do not depend on where the data sit in memory, where CRAN's machine runs its code for Intel processors. The first version of the noisy BLAS failed 1,923 expectations, at the sites of all 179 of CRAN's; OpenBLAS, BLIS and noLD passed. Since the review of #172 the noisy BLAS leaves alone the integral results no IEEE BLAS changes and what the reference BLAS leaves exact, and moves equal results of one call alike. A real BLAS can break such a tie, when a row falls to a peel or tail loop whose kernel rounds otherwise; the tests whose ties turn on it were reviewed by hand, and those of Szroeter's test run off CRAN only. With the fix every job passes, the noisy BLAS on each of its three seeds.

0.12.0: weighted fits and the shape of the variance (done)¶

This release was not on the roadmap. It took the 0.12.0 number because a corrected statistic bumps the minor version, and the milestones below moved up by one.

Weighted fits. On an lm fitted with weights, every residual-based test is computed from the Pearson residuals sqrt(w) * e. Goldfeld–Quandt and RESET refit with the weights. Unweighted fits return bit-identical values.
performVarianceFormTest(), a test whose null hypothesis is a variance function (exponential or power) rather than constant variance.
fitWLS() takes var_formula and form, and records the variance function it fitted.
Evidence. inst/validation/weighted-fits-size.R and inst/validation/variance-form-size-power.R, with simulated guards in test-weighted-fits.R and test-variance-form.R.

Found on the way and left for their own changes:

Resampling on weighted fits. performWildBootstrapTest(), performWhiteTestBootstrap(), rbootstrap_test_statistic() and performQuantileRegressionTest() refused a weighted fit, because they refitted without the weights. Resolved in 0.13.0 for the first three: they resample the Pearson residuals, return each draw to the scale of its observation and refit with the weights. That change leaves an unweighted fit as it was; the corrections 0.13.0 also makes to unweighted refits are listed under 0.13.0 below. In inst/validation/weighted-fits-size.R, at 2000 replications with B = 499, they reject a correctly weighted fit 0.045 to 0.054 of the time in two designs. performQuantileRegressionTest() still refuses one: on the weighted scale its slopes differ across quantiles under the null hypothesis, and a version for weighted fits needs its own construction. A lead: the quantile regressions of sqrt(w) y on sqrt(w) X with a column of ones added have equal slopes under the null hypothesis, which is the location-shift form the test assumes. It needs a derivation of the reference distribution and a size study before it can ship.
runSurveyHeteroTests() did not use the survey weights. It called stats::lm(formula, data = data, weights = weights), where weights was a local variable. lm() looks the name up in data and then in the environment of the formula, found stats::weights, failed, and the tryCatch() refitted without weights. Resolved in 0.13.0 by making the function say what it does, because using the weights would have been worse: sampling weights are not inverse error variances, and a test on a fit that uses them rejected a homoscedastic model 97% to 100% of the time when the weights varied with the regressor. runHeteroTests(formula, design) did exactly that through a svyglm() fit, and now takes the unweighted route too. Evidence: inst/validation/survey-designs-size.R. A design-based test is under Candidate tests.
.ht_fit_from_formula() had the same scoping fault and always failed with invalid type (closure) for variable '(weights)', so the per-group refits built from an lm, glm or parsnip fit could not run. Fixed in 0.13.0: each group is refitted by evaluating the model's own call on it, which gives it its own rows of the weights and keeps the family of a glm.
tidy() returned one row per degree of freedom. A result with two parameters, such as the F test of performVarianceFormTest() or the wild bootstrap with its df and B, was tidied into two rows, and the grouped and suite methods stacked them. Fixed in 0.13.0: one row per test, with the second parameter in a column of its own, parameter2.
autoplot() of a test suite did not show significant results. The bars were p-values on a linear axis from 0 to 1, so a test with p = 0.002 had no visible bar, and the chart was empty exactly when the tests rejected. The highlighting of p < 0.05 never applied either: the condition was written isTRUE(df$p.value < 0.05), which is FALSE for a suite of more than one test. The grouped method had the same condition and added a facet for the suggestions column. Fixed in 0.13.0: the bars are -log10(p) on an axis labelled with p-values, with a line at alpha and the p-value printed beside each bar.
An additive variance function, sigma^2 = a + z'b, was implemented for fitWLS() and the form test and withdrawn before release. Under its own null, at 5000 replications, the test rejected 6.1% of the time at the 5% level with 150 and with 400 observations, outside the release gate, and the fitted variances were not all positive in 17% of samples at 50 observations and 6% at 150. It needs an estimator that keeps the fitted variances positive before it can come back.
Tests of constant variance on a fitWLS() fit take the estimated weights as known. With the variance function right and Gaussian errors, Koenker rejects 8% of the time at the 5% level with 150 observations and 11% with 600, White 6% and 8%, and Harvey never. That is documented, and performVarianceFormTest() is the calibrated alternative, but a correction for the estimation step would let compareModelDiagnostics() report calibrated p-values for a weighted remedy.

The bootstrap tests hold the weights fixed and share the error: the wild bootstrap rejects 0.086 with 150 observations and 0.105 with 600, where it rejects 0.054 and 0.048 with the true weights in the same design. A bootstrap that runs fitWLS() again on every replicate is such a correction. In the same designs it rejects 0.042 to 0.060 with 150 observations for the wild bootstrap, the White bootstrap and rbootstrap_test_statistic() with Koenker's statistic, and 0.051 with 600 for the wild bootstrap (inst/validation/weighted-fits-size.R). Resolved in 0.15.0 for the bootstrap tests, as proposed: the argument reestimate_weights of the three functions refits the variance function fitWLS() recorded on the model, with the same var_formula and form, in each replicate, by default on a fit that carries that record; weights the user supplied are known and stay fixed (0.15.0, below). The tests of constant variance themselves still take the weights as known.

0.13.0: what was found after 0.12.0 (done)¶

This release was not on the roadmap either. It took the 0.13.0 number because it corrects statistics and adds functionality, and the milestones below moved up by one again.

Values that changed. The bootstrap tests keep an offset given as an argument when they refit an unweighted model, rbootstrap_test_statistic() matches the rows of the data to the fit by name and leaves out rows the fit dropped, and performWhiteTestBootstrap() regenerates a transformed response, where its p-value was 1. fitWLS() keeps the offset of the model it refits, and performVarianceFormTest() changes with it on a model with an offset. The Szroeter test that runHeteroTests() runs by name orders the observations by the fitted values instead of by the second column of the data. runHeteroTests() and runDiagnostics() no longer test a survey design through a fit weighted by its sampling weights. tidy() returns one row per test.
What was found on the way to 0.12.0: the resampling tests accept a weighted fit, and the survey designs, the grouped refits and autoplot() are fixed (above).
Two tests from the candidates: performHarrisonMcCabeTest() and performGlejserTest(robust = TRUE). Two more were measured and not adopted, the modified Wald test and stage 1 of the Dette, Neumeyer and Van Keilegom test (Candidate tests).
listDiagnostics(), and the release gate in release.yml (0.11.x, above).

0.14.0: review the public surface, deprecate and remove nothing (done)¶

Goal: every one of the 129 exports gets a verdict, which is one of core API, renamed, internal or moved out. Nothing is removed in this release. Everything that will go warns.

The author accepted the four proposals for this release on 2026-10-14: the dashboard and the benchmark suite move to a companion package and the recommendation engine stays, with its return format frozen; performCookWeisbergTest() stays, with an optional var_formula, and performNCVTest() is deprecated; test() is folded into summary.HeteroDiagnostic(); and simulate_*() and sigma_*() stay as a documented snake_case family while the one-offs are renamed. The companion package has no name yet, so the dashboard and benchmark functions are not deprecated until it has one.

Duplicate exports. Each duplicate below is deprecated in favour of the name it duplicates. performStudentizedBPTest() agrees with performKoenkerTest() to 1e-8 except on a model with an offset, which it drops (listed under 0.16.0). It keeps its implementation, so no value changed in 0.14.0; on a fitWLS() fit it read stale weights until 0.15.0 fixed that (listed under 0.16.0, done in 0.15.0).
- performBreuschPaganTest() is performBPTest(), assigned at R/performBPTest.R:232.
- performStudentizedBPTest() is a separate implementation that matches performKoenkerTest() in every column of the full sweep. Confirm the statistics agree to 1e-8, then keep one.
- performCookWeisbergTest() delegates to performNCVTest() with the fitted values as the variance regressor. performCookWeisbergTest() becomes the one entry point for the Cook–Weisberg score test, with var_formula as an optional argument, and performNCVTest() is deprecated (decided above).
- *Enhanced pairs: checkModel()/checkModelEnhanced(), plotResidualsFitted()/plotResidualsFittedEnhanced() and plotDiagnosticSuite()/plotDiagnosticSuiteEnhanced(). Merge each improvement into the base name.
Internal machinery that is exported: std_error(), std_warning(), validateTestInputs(), checkData(), TestFactory/test_factory(), the cache functions (cachedTest(), clearTestCache(), clearAnalysisCache()) and the ht_* logging functions. Unexport them unless a user-facing reason exists; for logging, keep at most ht_set_log_level(). Deprecated in 0.14.0 and unexported in 1.0.0, except the cache functions, which stay for now (decided), and ht_set_log_level().
The five r-prefixed exports (rbootstrap_test_statistic(), rcalculateEffectSize(), restimate_test_power(), rrunAdvancedDiagnostics() and rvalidate_against_reference()) follow no naming scheme in the package. Rename each to the house convention or unexport it. Four are done in 0.14.0: rbootstrap_test_statistic() is renamed bootstrapTestStatistic() and the other three are deprecated as internal helpers. rrunAdvancedDiagnostics() stays core API (decided), and was renamed runAdvancedDiagnostics() in 0.15.0.
test() is a generic with a very common name. Attached after devtools, it masks devtools::test(). Fold it into summary.HeteroDiagnostic() (decided above). Deprecated in 0.14.0, naming runDiagnostics(x$model, x$data, tests); summary.HeteroDiagnostic() calls runDiagnostics() and returns what it returned.
Naming convention. camelCase dominates (performXTest(), fitWLS(), plotX()). The simulation family (simulate_hetero(), sigma_*()) is declared a deliberate snake_case family and documented as one (decided above). The one-offs (run_benchmark_suite(), generate_benchmark_report(), simulate_type_I_errors(), simulate_power_analysis()) are renamed. The last two are simulate_size() and simulate_power() in 0.14.0; the first two move out with the benchmark suite instead.
Scope. The dashboard (R/dashboard.R, 698 lines) and the benchmark suite (R/benchmarking_system.R, 798 lines) move to a companion package (decided above), once it has a name. Between them they account for most of the optional Suggests (shiny, DT, plotly, htmlwidgets, bench). Open: the package has no name, so they are not deprecated in 0.14.0 (see Decisions needed).
Related model checks (performRESETTest(), performVIFDiagnostic(), performInfluenceDiagnostics(), performScatterDiagnostic()) are not heteroscedasticity tests. Keep them, grouped in the reference index as model checks, so they are not mistaken for tests of the variance.
Mechanism. Use base .Deprecated(), so no new dependency is needed. Each call warns and names its replacement. NEWS.md carries an old → new mapping table.
Exit criterion. An export inventory with a verdict for every one of the 129 exports, and test-public-api.R extended to assert the planned 1.0 export list, so that an accidental new export fails CI.

Export inventory¶

Every export in NAMESPACE on 2026-10-14, with a verdict. Of the 129, 104 are core API, 10 are renamed, 12 become internal and 3 move out. The S3 methods registered in NAMESPACE are not exports and are not listed; S3method(test, HeteroDiagnostic) goes with test(). The verdicts apply the decisions of 2026-10-14 as written, and the author's decisions on this inventory, taken the same day, are marked "decided".

For each renamed or internal export, the last column lists every caller found by a search of R/ (code, not comments or the definition), vignettes/ and inst/validation/, adds README.md where it names the export, and counts the lines in tests/ that call it. The notebooks in inst/tutorials/ call none of these exports. The article was not searched: it moves to the 1.0 API after 1.0.0. An export that becomes internal keeps working for its callers in R/, and for the tests, which run inside the namespace; a renamed export keeps working under its old name and warns. What has to change in each pull request are the callers outside R/, and the package's own calls of a name that will warn.

Export	Area	Verdict	Evidence that the change is safe
`performWhiteTest()`	test	core API
`performWhiteTestStreaming()`	test	core API
`performWhiteTestRobust()`	test	core API
`performWhiteTestBootstrap()`	test	core API
`performBPTest()`	test	core API
`performBPTestStreaming()`	test	core API
`performBPTestRobust()`	test	core API
`performBreuschPaganTest()`	test	renamed: `performBPTest()`, the same object	R/: none (an alias, assigned in `performBPTest.R`). Vignettes: `using_heteroTests.Rmd:128`. inst/validation/: `full-sweep-size-power.R:169`; 2 lines of `README.md`. Tests: 3 lines.
`performKoenkerTest()`	test	core API
`performKoenkerTestStreaming()`	test	core API
`performStudentizedBPTest()`	test	renamed: `performKoenkerTest()`, once item 2 shows they agree to `1e-8` (provisional, decided)	R/: registry name `"student_bp"` (`diagnosticsRegistry.R:70`, `test_factory.R:105`); `performBPTestRobust(studentized = TRUE)` (`robust_implementations.R:249`); `rvalidate_against_reference()` (`:455`). Vignettes: none. inst/validation/: `full-sweep-size-power.R:171`, `weighted-fits-size.R:138`; 4 lines of `README.md`. Tests: 8 lines.
`performGQTest()`	test	core API
`performParkTest()`	test	core API
`performSzroeterTest()`	test	core API
`performHarrisonMcCabeTest()`	test	core API
`performWildBootstrapTest()`	test	core API
`performQuantileRegressionTest()`	test	core API
`performRankPermutationTest()`	test	core API
`performHighDimensionalTest()`	test	core API
`performSpatialHeteroTest()`	test	core API
`performSpearmanTest()`	test	core API
`performLeveneTest()`	test	core API
`performGlejserTest()`	test	core API
`performBrownForsytheTest()`	test	core API
`performFlignerKilleenTest()`	test	core API
`performBartlettTest()`	test	core API
`performHarveyTest()`	test	core API
`performArchLMTest()`	test	core API
`performMcLeodLiTest()`	test	core API
`performHartleyFmaxTest()`	test	core API
`performOBrienTest()`	test	core API
`performCookWeisbergTest()`	test	core API
`performNCVTest()`	test	renamed: `performCookWeisbergTest()`, which gains an optional `var_formula` (decided)	R/: registry name `"ncv"` (`diagnosticsRegistry.R:67`); `performCookWeisbergTest()` delegates to it (`performCookWeisbergTest.R:69`); the benchmark suite (`benchmarking_system.R:368`). Vignettes: none. inst/validation/: `full-sweep-size-power.R:165`, `pass-a-size-power.R:87`, `weighted-fits-size.R:133,196`; 6 lines of `README.md`. Also `README.md:34`. Tests: 20 lines.
`performSpreadLevelTest()`	test	core API
`performPesaranTest()`	test	core API
`performBPRandomEffectsTest()`	test	core API
`performDavidianCarrollTest()`	test	core API
`performVarianceFormTest()`	test	core API
`performBoxMTest()`	test	core API
`rbootstrap_test_statistic()`	test	renamed: `bootstrapTestStatistic()` (decided)	R/: `performWhiteTestRobust()` and `performBPTestRobust()` (`robust_implementations.R:135,270`). Vignettes: none. inst/validation/: `offset-refits-size.R`, `weighted-fits-size.R`, labels in `make-table.R`; 22 lines of `README.md`. Also the Decisions needed row on `fitWLS()` fits. Tests: 60 lines.
`performScatterDiagnostic()`	model check	core API
`performVIFDiagnostic()`	model check	core API
`performRESETTest()`	model check	core API
`performInfluenceDiagnostics()`	model check	core API
`checkModel()`	model check	core API
`checkModelEnhanced()`	model check	renamed: merged into `checkModel()`	R/: `performWhiteTestStreaming()` (`performance.R:267`). Vignettes: none. inst/validation/: none. Tests: 8 lines.
`runHeteroTests()`	runner	core API
`runHeteroTestsParallel()`	runner	core API (decided)
`runDiagnostics()`	runner	core API
`runDiagnosticPlots()`	runner	core API
`registerDiagnostic()`	runner	core API
`listDiagnostics()`	runner	core API
`registerPlot()`	runner	core API
`runMultivariateTests()`	runner	core API
`runTimeSeriesTests()`	runner	core API
`runPanelTests()`	runner	core API
`runSurveyHeteroTests()`	runner	core API
`compareModelDiagnostics()`	runner	core API
`compareTestResults()`	runner	core API (decided)
`analyzeMLResiduals()`	runner	core API
`HeteroDiagnostic()`	runner	core API
`test()`	runner	renamed: its warning names `runDiagnostics(x$model, x$data, tests)`; `summary()` is unchanged (decided)	R/: `summary.HeteroDiagnostic()` calls it (`HeteroDiagnostic.R:59`). Vignettes: `real_world_case_studies.Rmd:110`, `tutorial.Rmd:32`, `using_heteroTests.Rmd:100,105`. inst/validation/: none. Also `README.md:143`. Tests: 19 lines. `summary.HeteroDiagnostic()` will call `runDiagnostics()` directly.
`fitWLS()`	remediation	core API
`fitRobust()`	remediation	core API
`autoTransform()`	remediation	core API
`suggestRemediation()`	remediation	core API
`autoCompareRemediations()`	remediation	core API
`analyseDatasetCharacteristics()`	recommendation	core API
`suggestDiagnosticsForProfile()`	recommendation	core API
`interpretHeteroTestResults()`	recommendation	core API
`recommendRemediationStrategies()`	recommendation	core API
`warnDiagnosticAssumptions()`	recommendation	core API
`buildDiagnosticDecisionTree()`	recommendation	core API
`composeDiagnosticNarrative()`	recommendation	core API
`generateHeteroRecommendations()`	recommendation	core API
`simulate_hetero()`	simulation	core API
`simulate_arch1()`	simulation	core API
`sigma_linear()`	simulation	core API
`sigma_exponential()`	simulation	core API
`sigma_group()`	simulation	core API
`sigma_piecewise()`	simulation	core API
`sigma_poly()`	simulation	core API
`sigma_sin()`	simulation	core API
`sigma_multiplicative()`	simulation	core API
`sigma_spatial()`	simulation	core API
`sigma_logistic()`	simulation	core API
`sigma_inverse()`	simulation	core API
`sigma_power()`	simulation	core API
`sigma_step()`	simulation	core API
`sigma_u_shape()`	simulation	core API
`sigma_exp_decay()`	simulation	core API
`sigma_gaussian_peak()`	simulation	core API
`sigma_piecewise_linear()`	simulation	core API
`simulate_type_I_errors()`	simulation	renamed: `simulate_size()` (decided)	R/: none. Vignettes: none. inst/validation/: none. Tests: 3 lines.
`simulate_power_analysis()`	simulation	renamed: `simulate_power()` (decided)	R/: none. Vignettes: named in the prose of `troubleshooting.Rmd:44`. inst/validation/: none. Tests: 3 lines.
`plotResidualsFitted()`	plotting	core API
`plotSpreadLevel()`	plotting	core API
`plotResidualDensity()`	plotting	core API
`plotResidualQQ()`	plotting	core API
`plotBubbleVariance()`	plotting	core API
`plotDiagnosticSuite()`	plotting	core API
`plotBeforeAfter()`	plotting	core API
`plotResidualsFittedEnhanced()`	plotting	renamed: merged into `plotResidualsFitted()`	R/: `plotDiagnosticSuiteEnhanced()` (`plotDiagnostics.R:279`). Vignettes: none. inst/validation/: none. Tests: 5 lines.
`plotDiagnosticSuiteEnhanced()`	plotting	renamed: merged into `plotDiagnosticSuite()`	R/: `generateDiagnosticReport()` (`reporting.R:53`); the dashboard (`dashboard.R:536`). Vignettes: none. inst/validation/: none. Tests: 4 lines.
`theme_hetero()`	plotting	core API
`scale_colour_hetero_diagnostic()`	plotting	core API
`scale_fill_hetero_diagnostic()`	plotting	core API
`generateDiagnosticReport()`	reporting	core API
`downloadTeachingData()`	reporting	core API
`ht_set_log_level()`	infrastructure	core API
`std_error()`	infrastructure	internal	R/: 92 lines in 24 files, all internal callers, which keep working. Vignettes: none. inst/validation/: none. Tests: 9 lines.
`std_warning()`	infrastructure	internal	R/: `performance.R:181`, `performWhiteTest.R:232`, `validation.R:675`. Vignettes: none. inst/validation/: none. Tests: none.
`validateTestInputs()`	infrastructure	internal	R/: `generateHeteroRecommendations()` (`intelligent_recommendations.R:686`), the test factory (`test_factory.R:80`). Vignettes: none. inst/validation/: none. Tests: 21 lines.
`checkData()`	infrastructure	internal	R/: 16 lines in 7 files (`advancedDiagnostics.R`, `advancedRemediation.R`, `fitWLS.R`, `HeteroDiagnostic.R`, `intelligent_recommendations.R`, `reporting.R`, `runPanelTests.R`). Vignettes: none. inst/validation/: none. Tests: 9 lines.
`TestFactory`	infrastructure	internal	R/: only `test_factory.R:98`, which builds `test_factory`. Vignettes: none. inst/validation/: none. Tests: 2 lines.
`test_factory`	infrastructure	internal	R/: through its alias `.test_factory`, in `.run_registered_diagnostic()` (`performance.R:610-612`) and the dashboard (`dashboard.R:53`). Vignettes: none. inst/validation/: none. Tests: 10 lines.
`cachedTest()`	infrastructure	core API for now (decided); revisited at 1.0.0, once 0.16.0 bounds the caches
`clearTestCache()`	infrastructure	core API for now (decided); revisited at 1.0.0, once 0.16.0 bounds the caches
`clearAnalysisCache()`	infrastructure	core API for now (decided); revisited at 1.0.0, once 0.16.0 bounds the caches
`ht_enable_log_capture()`	infrastructure	internal	R/: none. Vignettes: none. inst/validation/: none. Tests: none.
`ht_log_history()`	infrastructure	internal	R/: none. Vignettes: none. inst/validation/: none. Tests: none.
`ht_clear_log_history()`	infrastructure	internal	R/: none. Vignettes: none. inst/validation/: none. Tests: none.
`rcalculateEffectSize()`	infrastructure	internal	R/: `performWhiteTestRobust()`, `performBPTestRobust()` (`robust_implementations.R:158,292`), `restimate_test_power()` (`bootstrap_methods.R:599`). Vignettes: none. inst/validation/: none. Tests: 4 lines.
`restimate_test_power()`	infrastructure	internal	R/: `performWhiteTestRobust()`, `performBPTestRobust()` (`robust_implementations.R:159,293`). Vignettes: none. inst/validation/: none. Tests: 3 lines.
`rrunAdvancedDiagnostics()`	runner	core API (decided); its name is still open, see the item on the `r`-prefixed exports
`rvalidate_against_reference()`	infrastructure	internal	R/: none. Vignettes: none. inst/validation/: none. Tests: 2 lines. The reference comparisons live in `tests/testthat/`.
`launchDiagnosticDashboard()`	dashboard and benchmarks	moved out (companion package not named yet; not deprecated in 0.14.0)	R/: none outside `dashboard.R`. Vignettes: none. inst/validation/: none.
`run_benchmark_suite()`	dashboard and benchmarks	moved out (companion package not named yet; not deprecated in 0.14.0)	R/: none outside `benchmarking_system.R`. Vignettes: none. inst/validation/: none.
`generate_benchmark_report()`	dashboard and benchmarks	moved out (companion package not named yet; not deprecated in 0.14.0)	R/: none outside `benchmarking_system.R`. Vignettes: none. inst/validation/: none.

Notes on the verdicts:

performStudentizedBPTest() is provisional. Item 2 first asserts that it agrees with performKoenkerTest() to 1e-8 over the designs of the full sweep, weighted, rank-deficient and fitWLS() fits included. Disagreement anywhere is a finding, not a rename, and is reported before anything else.
test() warns and names runDiagnostics(x$model, x$data, tests), which returns what test() returns. summary.HeteroDiagnostic() keeps its return value, so no value changes in 0.14.0.
Registry names. "student_bp" and "ncv" stay registered and keep calling what they call now, so their output does not change. Each warns once per call of runHeteroTests() or runDiagnostics(), not once per group or bootstrap replicate, and goes in 1.0.0 with the function behind it.
The cache functions (cachedTest(), clearTestCache(), clearAnalysisCache()) stay exported and are not deprecated in 0.14.0. They are revisited at 1.0.0, once 0.16.0 bounds the caches.
rrunAdvancedDiagnostics(), runHeteroTestsParallel() and compareTestResults() stay core API. The name of rrunAdvancedDiagnostics() is still open under the item on the r-prefixed exports.
The simulation renames. simulate_size() and simulate_power() join the documented snake_case family, since both simulate under the sigma_*() functions.
bootstrapTestStatistic() takes a test function as its first argument, so it is not a perform*Test(model, data, ...); the new name says what it does in the house camelCase.
The dashboard and the benchmark suite are not deprecated in 0.14.0, because the companion package has no name yet. They call plotDiagnosticSuiteEnhanced(), the test factory and performNCVTest(), which change before the code moves.
analyzeMLResiduals() and analyseDatasetCharacteristics() spell the verb two ways. Both are core API; the roadmap names no change for them.

0.15.0: make the interface match its description (done)¶

These changes break existing calls. They have to happen before 1.0.0, not after it.

The (model, data, ...) convention. Nine exported tests did not follow it:
- no data argument: performArchLMTest(), performMcLeodLiTest(), performCookWeisbergTest(), performNCVTest(), performSpearmanTest(), performSpreadLevelTest(), performDavidianCarrollTest() and performHarveyTest();
- no model argument: performBoxMTest(data, group).
Decided on 2026-10-15 as proposed: the eight take data = NULL in second position, recovered from the model frame when omitted, and performBoxMTest() is the documented exception, since it tests the data rather than residuals. A call that passes the old second argument by position (performArchLMTest(m, 3), performCookWeisbergTest(m, ~x), performHarveyTest(m, "fitted")) is recognised by its type, works in 0.15.x, warns with the new call and stops working in 1.0.0; NEWS.md has the table. No value changes: every old and new form returns what the old form returned in 0.14.0, checked with identical() against 0.14.0. Extended on 2026-10-15 to every other test: performRESETTest(model, power) as the eight, and the 29 tests that required data now default it to NULL, the model frame. A call that worked returns what it returned, except performRESETTest(m, NULL) and performRESETTest(m, <data frame>), which read the second argument as power and returned NaN or a meaningless F. - [x] Argument names. Harvey used studentize and performBPTestRobust() studentized. Standardised on studentize, which matches lmtest::bptest(); studentized is accepted with a deprecation warning in 0.15.x and stops working in 1.0.0, and the robust_details$studentized field of the result keeps its name. No value changes. - [x] The name of rrunAdvancedDiagnostics(). Renamed runAdvancedDiagnostics() (decided 2026-10-15), the house convention for a runner. The old name warns with .Deprecated() in 0.15.x and goes in 1.0.0 with the names deprecated in 0.14.0. No value changes. - [x] Resampling defaults. B was 499 for the wild bootstrap, 999 for rank permutation, and 1000 for the White bootstrap and bootstrapTestStatistic(). Under the (1 + #) / (B + 1) convention, a test at level alpha is exact only when alpha * (B + 1) is an integer, and 1000 fails that at 5%. Done (decided 2026-10-15: B = 999): every resampling function draws 999 by default, the robust tests with bootstrap = TRUE and runAdvancedDiagnostics() (500 before) and performSpatialHeteroTest() (499 permutations before) included. Default p-values change; NEWS.md has an example for each. Of the recorded evidence only the spatial row of the full sweep used a default, and it was regenerated: size 0.045 to 0.050, power 0.362 to 0.350. - [x] Wild bootstrap multiplier (Rademacher or Mammen). Settle it by a size study at small n with skewed errors, and record the result in inst/validation/.

Measured on 2026-10-15 in inst/validation/wild-multiplier-size.R: the cross-sectional design of Pass A at 30, 50 and 100 observations, with Gaussian, t5 and centred exponential errors, 5000 replications and B = 999, both multipliers on the same samples. The decision rule was written into the script before the full run started; the script was first committed with its results. Neither multiplier holds the level in more cells, and no paired difference is outside its 99% interval (within Monte Carlo error). Both are inside the release gate in the same six of the nine size cells, those with Gaussian and t5 errors, and outside it in the same three, those with exponential errors, where Rademacher's multiplier rejects 0.082, 0.073 and 0.061 of the time and Mammen's 0.078, 0.070 and 0.059: Mammen's rate is lower in all three cells, each difference inside its 99% interval. The interval around 0.049, the rate of an exact test with B = 999, gives the same split. Against sigma^2 = exp(0.4 x1) the power of the two differs by less than one point at each n. Done (decided 2026-10-16: keep "rademacher", as the study recommended); no value changes. The help page said the bootstrap controls the size for heavy-tailed errors; that holds for the t5 errors of the study and not for the skewed ones, with either multiplier, and #125 corrected the page. A wild bootstrap without a multiplier, which may hold the level better under skewed errors, is to be measured in 0.16.0 (below). - [x] Reproducibility under parallel = TRUE. A search of R/ finds no clusterSetRNGStream() and no L'Ecuyer set-up on the mclapply() and parLapply() paths, so set.seed() does not make parallel resampling results reproducible. Done: replicate i draws from stream i of a L'Ecuyer-CMRG generator seeded by one draw from the session's, on both paths, in bootstrapTestStatistic(), performWhiteTestBootstrap() and runHeteroTestsParallel(), so the value does not depend on the number of workers. Serial calls are unchanged; parallel and serial values differ and are each valid (test-parallel-reproducibility.R). runHeteroTestsParallel() with one worker runs in the session on the same streams, so its value does not depend on n_cores either, and changes. The White bootstrap runs on a socket cluster on Windows, where it stopped, and gains n_cores. The forked path is exercised by CI on Linux and macOS only. - [x] Bootstrap tests on a fitWLS() fit. Accepted for 0.15.0 as proposed under Decisions needed, and moved from 0.14.0 because it changes values. performWildBootstrapTest(), performWhiteTestBootstrap(), bootstrapTestStatistic() and its deprecated name take reestimate_weights. NULL, the default, runs fitWLS() again on each replicate of a fit that carries its record of the variance function and holds the weights of any other model fixed; FALSE holds them fixed, as 0.13.0 and 0.14.0 did; TRUE on a model without the record is an error. Each replicate is fitted as the model fitWLS() was given, the recorded variance function is fitted again on the recorded regressors, and the replicate is refitted with the weights that gives; one whose variance function cannot be fitted counts as a failed replicate. The default p-values on a fitWLS() fit change, announced in bold in NEWS.md. An lm, a weighted lm and a Gaussian glm return identical() values, and so does a fitWLS() fit with FALSE. In inst/validation/weighted-fits-size.R, at 2000 replications with B = 499, the wild bootstrap rejects 0.0605 and 0.0515 of the time at 150 and 600 observations, the White bootstrap 0.0415 and bootstrapTestStatistic() with Koenker's statistic 0.045, against 0.0865, 0.105, 0.064 and 0.0825 with the weights held fixed. Those four rows were measured by the prototype of 0.12.0 and not run again: the package reproduces it on 200 samples of each, p-values, replicates and random draws (inst/validation/weighted-fits-refit-reproduction.csv). The wild bootstrap at 150 observations is outside the release gate's interval [0.042, 0.058] and the White bootstrap just below it; a study at 5000 replications is listed under 0.19.0. - [x] The default battery. runHeteroTests(), test()/summary() on HeteroDiagnostic, the advanced diagnostics and the dashboard all defaulted to c("white", "breusch_pagan"), and "breusch_pagan" is the classical statistic. The package's own sweep shows it rejecting 28.75% of the time under a t5 null, against 4.75% for Koenker. Done (decided 2026-10-15: White and Koenker): the default is c("white", "koenker") in those functions and in runDiagnostics(), compareModelDiagnostics(), compareTestResults() and runSurveyHeteroTests(), and the classical test runs by name. suggestRemediation() reads "koenker" as well as "breusch_pagan", and suggestDiagnosticsForProfile() recommends Koenker's test as the baseline. runAdvancedDiagnostics() already computed the studentized statistic by default and is unchanged. No recorded number moves. - [x] The default of robust in performGlejserTest(). It was FALSE, Glejser's (1969) statistic, which in inst/validation/glejser-skewed-errors.R rejects a true null hypothesis 8.8% to 17.7% of the time at the 5% level when the errors are skewed, at every sample size from 50 to 1000. Done (decided for 0.15.0 as proposed: TRUE): the default is the corrected statistic of Im (2000) and Machado and Santos Silva (2000), added in 0.13.0. From 150 observations it holds its level under every error distribution measured, inside the release gate in all 27 of its null cells; at 50 observations it rejects 5.7% to 6.6% of the time under skewed errors, four of those six cells above the gate, and gives up 1.1 and 1.9 points of power under Gaussian errors. Default values change; NEWS.md has an example. robust = FALSE returns what the default returned in 0.14.0, checked with identical() against 0.14.0. Of the recorded evidence the full sweep and the weighted fits call the default, and their two rows were regenerated from the same seeds: size 0.0600 to 0.0575 and 0.0475 to 0.0450. Pass A and the Glejser study name Glejser's statistic and now give robust = FALSE; their numbers do not move. - [x] Return contract. Document the fields every perform*Test() returns, both the htest core and the package's additions, on one help page. Check every exported test against it in a single parametrised test. This also retires the hand-maintained printer and metadata drift listed under the old technical debt. Done: ?test_results and test-return-contract.R. Five tests gained the name df of their parameter and performBoxMTest() its data.name; no number changed. - [x] Uniform input validation across all exported tests. The README lists its absence as a known limitation. test-input-validation.R gives every test a glm that is not a least-squares fit, a model with an NA coefficient, data that lack rows of the fit and data that lack a variable, and every test stops on each with the same message. Refusing the inputs some tests returned a value for was decided on 2026-10-15: 36 test names refuse a glm of a family other than gaussian or a link other than the identity, five refuse an NA coefficient, and the variance-form, Pesaran and random-effects tests refuse data that cannot be matched to the fit (NEWS.md lists the calls). A Gaussian glm with the identity link is accepted and gives the result of the lm. With prior weights the GQ, RESET, White bootstrap and wild bootstrap tests, which read the weights of an lm, read them since then, decided on 2026-10-15, where they took the glm for an unweighted fit, and the quantile regression test refuses it as it refuses a weighted lm; those calls change value. No recorded number moves.

0.16.0: the model context, part 1 (increments 1–7), and statistics fixes (done)¶

Planned as one milestone for the statistics evidence, the dependencies and the internals, with no API change. The author split it on 2026-10-26: 0.16.0 is the first seven steps of the model context and the statistics fixes found on the way, and the rest of the plan moves to 0.17.0 (the model context, part 2), 0.18.0 (internals, dependencies and the defect backlog) and 0.19.0 (release-grade evidence and the open decisions). No export or argument changes; the default of chunk_threshold_mb in runHeteroTests() and runDiagnostics() becomes Inf, and the print and summary methods of the suites, approved on 2026-10-21, are added.

0.17.0: the model context, part 2 (increments 8–11) (done)¶

Moved from 0.16.0 on 2026-10-26: the last four steps of the model context, and the defects found with it that those steps, or a pull request of their own, settle. Among those are N3, White's streaming test on collinear terms (decision 22 of the design), and the threshold of the streaming tests on squared residuals that do not vary. Done in #164 (N3 and the threshold), in #165 and #167 (step 8), #168 (step 9), #169 (step 10) and #171 (step 11). The release also fixes CRAN's additional check of 0.16.0 with Intel MKL (#172, under 0.11.x). No export, argument or default changes; hardhat joins the suggested packages. What the steps found and did not settle moved to 0.18.0 at the release (Left from steps 8 to 10 of the model context).

0.18.0: internals, dependencies and the defect backlog¶

Moved from 0.16.0 on 2026-10-26, with no API change: the internals and dependencies planned for it, and the defects recorded before and during it that the model context does not settle. Each fix that changes a value is announced in bold in NEWS.md.

The reverse-dependency job of R-CMD-check.yml has never passed. On every weekly run since at least 2026-10-05, and on the manual runs on the 0.16.0 release branch, it stopped in setup-r-dependencies with "Could not solve package dependencies" (pkgdepends), before revdepcheck::revdep_check() ran. It is not part of the release gate (release.yml). Done in #166: the job asked for any::revdepcheck, and revdepcheck is not on CRAN; it is installed from GitHub, r-lib/revdepcheck. The package has had no reverse dependencies on CRAN so far, so the job checks none, but it runs: on the manual run of the 0.17.0 release branch it passed, with no reverse dependency to check.
The DOI of Szroeter (1978) on the help page of performSzroeterTest(). The page gave doi:10.2307/1913833, which Crossref resolves to Dagenais (1978), "The Computation of FIML Estimates as Iterative Generalized Least Squares Estimates…", in the same issue of Econometrica. Szroeter's paper is doi:10.2307/1913831 (Econometrica 46(6), 1311–1327). Fixed in the roxygen (R/additional_tests.R), the help page and its page on the site. In the same way, inst/CITATION named Fligner's co-author "Thomas J. Killeen"; Crossref, and the article in JASA 71, give Timothy J. Killeen. The help pages give only the initials. Found by the R Journal article's additions (heteroTests-paper #6) and their review.
man/ is not what roxygen writes from R/. On main after 0.17.0 (a3709bc), devtools::document() with roxygen2 7.3.3, the version of RoxygenNote, changes 36 files of man/:
- It deletes ht_logging.Rd and rbootstrap_test_statistic.Rd, and writes 15 pages that man/ lacks. Among them are bootstrapTestStatistic.Rd, for the @rdname bootstrapTestStatistic of R/bootstrap_methods.R, and heteroTests_ggplot.Rd, a topic with no page today.
- It rewrites 19 more. Some of the changes are in the text: ht_log.Rd still has the wording of September 2025, which R/utils.R has since replaced. Others are in the form: the usage is wrapped one argument per line, \link{} replaces \code{\link{}}, and \code{} replaces the \file{} of paths.
So some pages were edited by hand, or written by another roxygen setup, and a help page can lag its source. That includes the pages of the CRAN tarball, which R CMD build takes from man/. Find which, make man/ the output of devtools::document(), review the help pages that change, and add a CI step that fails when document() changes man/. Found while fixing the DOI above, which was fixed in the roxygen and in the help page alone. Done in #180: the pages had been edited by hand, or written by hand under a roxygen header, as ht_logging.Rd was in #7, and document() had not been run on them since; neither the roxygen2 version nor the collation locale makes a difference. The edits the roxygen lacked were carried into R/ (the .recover argument of safe_lm(), the arguments of print.htest_enhanced(), the model of performQuantileRegressionTest()), with the faults of the roxygen markdown they hid: `r ~ z` evaluated as inline R code, \% written as \\%, which drops the rest of the line, and a link to parallel::mclapply() that roxygen resolves to mcdummies on Windows. One sentence was dropped as no longer true ("Neither strategy accepts a glm", on the page of bootstrapTestStatistic()), and the eight helpers documented on generateHeteroRecommendations.Rd name it as their page. The block that opens R/ggplot_extensions.R documents no object and is @noRd, so heteroTests_ggplot gets no page. The 86 pages roxygen writes are now what document() writes, and a step of run-checks.yml fails when roxygen2::roxygenise(), with the roxygen2 of RoxygenNote, changes man/ or NAMESPACE; the 52 it does not write are the next item. - [ ] 52 help pages and NAMESPACE are written by hand. roxygen2 does not overwrite a file without its "Generated by roxygen2" line, and 52 pages of man/ have none, among them those of runHeteroTests(), runDiagnostics(), performWhiteTest(), performBPTest(), performKoenkerTest() and most of the tests, and generateHeteroRecommendations.Rd. 45 of them are pages of a roxygen block in R/, which users do not read and the CI step above cannot see; the other 7 (test_results, diagnostic_data, hetero_test_suite, simulate_hetero and the methods of HeteroDiagnostic) have no block. The page and the block have been kept up side by side. Given the roxygen line, the 45 pages are rewritten by +3,332 / -1,451 lines (git diff --numstat). Split into sentences, their rendered text holds 550 to 715 sentences of the page that the block lacks (the Supported models sections, the DOIs, aliases such as generate_benchmark_report and the glance() and augment() methods) and 570 to 670 of the block that the page lacks (references, see-also lists, examples), by the two ways #180 and its review matched them. Merge each page into its block, delete the page and let roxygen write it; write the 7 as blocks. The blocks of performBPTest(), performKoenkerTest() and performWhiteTest() have 28 lines with \%, which markdown writes as \\%; two blocks have no @name (R/error_handling.R, R/intelligent_recommendations.R). generateHeteroRecommendations.Rd documents nine functions, eight of whose blocks name it with @rdname: written by roxygen, it would take its \name and title from the first of them, analyseDatasetCharacteristics(), and one data argument, that of generateHeteroRecommendations(), while the page's data, updated for 0.17.0, is in no block. NAMESPACE has no roxygen line either: written by roxygen it would lose 59 exports whose blocks have no @export (runHeteroTests(), performBPTest(), the sigma_*() functions among them), the importFrom() of stats, MASS, R6 and utils and importFrom(ggplot2, scale_fill_manual), and gain a, Register and heteroscedasticity and the registration of print.htest_enhanced(), which NAMESPACE lacks today (a method for a class no function of the package returns). A first run with no NAMESPACE also exports the two autoplot() methods as functions, since the package then loads without importFrom(ggplot2, autoplot); a second run registers them again. Found by #180. - [ ] The site's reference pages do not render \tabular. tools/build-docs.R has no case for it, so render() renders its children: the column format becomes text, \tab is dropped and \cr becomes a line break. The tables of four pages come out as lines of cells separated by two spaces, after the format as a stray token: docs/reference/performBPTest.md, performKoenkerTest.md and performWhiteTest.md (rrrrrl, lines 42, 41 and 67), the rejection rates with fixed effects for small groups, from man/performBPTest.Rd, performKoenkerTest.Rd and performWhiteTest.Rd, written by hand; and performWildBootstrapTest.md (line 112, "0.058.llrrrr", the format glued to the sentence before the table), the size of the wild bootstrap by errors and multiplier, from the markdown table of the roxygen of performWildBootstrapTest() (R/modern_diagnostics.R). R's help renders them. build-docs.R would need a "\\tabular" case that writes a pipe table: the alignment of each column from the format (l, r and c as :---, ---: and :---:), the rows split at \cr and the cells at \tab, each cell rendered inline on one line, with | escaped, the first row as the header, which each of the four tables has, and a blank line before and after, since the table of performWildBootstrapTest.Rd starts on the line of the sentence before it. Found by #182; recorded, not fixed. No value changes. - [ ] Dependencies. - Move curl to Suggests: only downloadTeachingData() uses it. - R6 goes if TestFactory leaves the API, since it is the only user. - Suggests shrinks with the companion split. - Settle whether digest belongs in Imports or Suggests, depending on whether caching stays exported. - [ ] Pass R CMD check with _R_CHECK_DEPENDS_ONLY_=true. With 39 packages in Suggests, every test, example and vignette that uses one must degrade gracefully. Audit the skip_if_not_installed() coverage; 20 of the 69 test files currently use it. - [ ] The R floor. DESCRIPTION claims R >= 4.1, but CI tests only devel, release and oldrel-1. Either add R 4.1 to the matrix or raise the floor to what is tested. - [ ] Replace non-ASCII characters in R/ with escapes (89 lines across 27 files). - [ ] The memory of a cache key. rcache_hash() (R/cache.R) serializes the model frame, the design, the vectors of the fit and the data into one raw vector on every call of cachedTest() and runHeteroTests(), about twice their size, which a wide design makes large, and frees it at once. Hash them part by part, or stream the serialization into the hash, so that the temporary allocation stays small (found by the review of #144). No value changes. - [ ] A runner call changes how the workflow it is given serializes. The names of the residuals and fitted.values of a workflow's engine fit are ALTREP deferred string conversions of the row numbers, which expand when they are read, and the runners read them: after runHeteroTests() of a workflow fitted to 120 observations they take 2,236 bytes where they took 1,209. No value of the workflow changes, but a cache that hashes objects, such as that of targets, sees a workflow changed by the call. Found by the review of #169, which saw it on main as well; recorded, not fixed. Read the names without expanding them, or work on a copy of the vectors. - [ ] Replicates lost without a warning. A replicate of bootstrapTestStatistic() whose refit fails, or, under pairs resampling, whose design cannot be built on the rows it drew (rpairs_design(), since the review of step 8 of the model context), is left out silently: only effective_samples shows it, as for a failed refit before. The replicates of a fitWLS() fit whose variance function cannot be fitted are left out with a warning that gives their number. Consider warning whenever replicates are lost (found by the review of #167). No value changes. - [ ] Split R/validation.R (1,441 lines) into cache, result-type, assumption-checking and requirement-dispatch units, and factor out the repeated boilerplate: scalar validators, model and data preparation, and intercept stripping. - [ ] Retire TODO.md, SUGGESTIONS.md and cran.md. The open items in the first two all concern setup.sh and bootstrapping renv, and cran.md is a generic guide. With CRAN now the install route, decide whether setup.sh, renv.lock and the Dockerfile remain supported. If they do, complete renv.lock against Imports/Suggests so the Docker build uses it instead of install_local(). If they do not, remove them. - [ ] Keep the CRAN test run short. Long simulations run under skip_on_cran(), and the full versions run on CI. - [x] runHeteroTestsParallel() names its results "model". It returns what the registered functions return, so White's test and the four others of the registry report data.name = "model" there, where runHeteroTests() reports the formula of the model since 0.16.0 (NEWS.md). Give it the rule of the runner, or say on its help page that it does not. Done in #182: it has the rule, and each of its results is the one runHeteroTests() gives for the test on the same call, extras$streaming = FALSE included, its numbers computed the same way: identical() to it where the BLAS gives the same last bits on every call, as R's reference BLAS does, and equal to the last bits of the BLAS where it does not, as under Intel MKL or the noisy BLAS of cran-blas.yml, which the test allows for (from the review of

182). On socket workers the model a result carries is the copy that¶

travelled to the worker. Four names report "model": "white", "white_bootstrap", "student_bp" and "szroeter". A resampling test still draws from its own stream, and the list is still not a hetero_test_suite. cachedTest(), which returned "model" too, has the rule as well; it gives no extras$streaming, which says how the runner ran a test, and its help page says so. - [x] A grouped run of a multivariate lm. runHeteroTests() refuses lm(cbind(y, x2) ~ x1) with the class of the model context ("needs a model with one response"), but on grouped data it refits the model in each group and records each refusal as a failure of the group: it returns a grouped suite whose every test is NA, with status "error" and issue "model_context", where the refusal of the context should be final (decision 6). A grouped MASS::rlm() fit behaved the same in the first version of step 11 of the model context, and with MASS attached main returned the values of an rlm in each group; step 11 refuses it before any group is refitted, beside a fitWLS() fit. Found by the review of step 11; the same on main at the merge of step 10. Done in step 11 (#171), from its second review, for every refusal of the context: one in a group stops the grouped run with its class and message, a multivariate lm and a penalised gam given a test that needs least squares among them; other failures are recorded in their group, as before. A suite of NA becomes a refusal. - [ ] performInfluenceDiagnostics() and performVIFDiagnostic() on a multivariate lm. Every test, the runners, fitWLS() and bootstrapTestStatistic() refuse lm(cbind(y, x2) ~ x1) with "needs a model with one response"; these two, which are not built on the model context, return a value. On main, with d from contract_data() (tests/testthat/helper-test-recipes.R), performInfluenceDiagnostics(lm(cbind(y, x2) ~ x1, data = d)) returns the Cook's distances as a 120 by 2 matrix, a column for each response; influential as 12 positions among its 240 cells (31, 32, 46, 59, 91, 122, 152, 155, 164, 166, 185 and 201, the seven above 120 rows of the column of x2); and a cutoff 4 / (n - p) that counts the 4 coefficients of both responses, 0.0345, where lm(y ~ x1) gives 4 / 118. performVIFDiagnostic() of the same fit returns the VIF of the design the responses share, 1 for x1, the value of lm(y ~ x1) (and with a second regressor, d$x3 <- rnorm(120) drawn straight after contract_data(), which sets the seed 20261015, 1.005058 for each, as for lm(y ~ x1 + x3)). The decision is whether they refuse it as the tests do. Found in the 0.17.0 release audit; recorded, not fixed. A value becomes a refusal, or none. - [ ] recommendRemediationStrategies() without data. recommendRemediationStrategies(results, model, NULL) stops with "argument is of length zero" for every model, an lm included: identifyVariancePatterns() indexes the NULL it is given. Found by increment 11 of the model context, recorded, not fixed; the same on main at the merge of increment 10. Read the model frame without data, or say that data is needed. A refusal becomes a value, or a clearer refusal. - [x] The deprecation messages of 0.14.0. The names deprecated in 0.14.0 that have a replacement (performBreuschPaganTest(), performNCVTest(), performStudentizedBPTest(), the three *Enhanced functions, rbootstrap_test_statistic(), test(), simulate_type_I_errors(), simulate_power_analysis(), and "student_bp" and "ncv" in runHeteroTests()) warn with the default text of .Deprecated(), which does not say that they go in 1.0.0, as the internal helpers, the 0.15.0 forms and rrunAdvancedDiagnostics() do. Found in the 0.15.0 release audit; no value changes. Done in #182: the ten functions give the message of rrunAdvancedDiagnostics(), now written by one helper for the eleven, as "'performBreuschPaganTest' is deprecated and leaves the exports in heteroTests 1.0.0. Use 'performBPTest' instead.", and the two test names that of the studentized argument of performBPTestRobust(), "Test name "ncv" is deprecated and stops working in heteroTests 1.0.0. Use "cook_weisberg" instead.". The tests of each pin the 1.0.0, and test-public-api.R checks that every deprecated function gives a message of its own. - [x] ht_analyse_failure() labels a failure by its message, ignoring case. It calls a failure a rank deficiency when the message matches "rank|singular" and a data-quality issue when it matches "NA|NaN|missing|Inf", both with ignore.case = TRUE (R/error_handling.R:269-275), and ht_data_cleaning_suggestions() picks its advice by the same patterns. The refusals of a glm that is not a least-squares fit name the test, and two names match: on m <- glm(am ~ wt + hp, data = mtcars, family = binomial), runHeteroTests(m, mtcars, tests = "rank_permutation") stops with "Rank deficiency detected for Rank Permutation: performRankPermutationTest() needs a model fitted by stats::lm(): ..." and the advice to remove collinear predictors, and tests = "high_dimensional" with "Data quality issue detected", from the "na" of performHighDimensionalTest(), and the advice to remove missing values. The other tests refuse it as "Diagnostic execution failed". Match whole words, and NA only in capitals, or give the refusal of such a glm an issue of its own. Recorded, not fixed, in 0.15.0; only the message and the recorded issue change. Done in #182, both ways. The refusal has the class heteroTests_glm_family_error, which the failure analysis reads as the issue "glm_family", final as those of the model context and of the rows of data: on the glm above, runHeteroTests(m, mtcars, tests = "rank_permutation") stops at once with "Model is not a least-squares fit for Rank Permutation: performRankPermutationTest() needs a model fitted by stats::lm(): ...", that class and no advice, and so do the other tests, where White's test tried its four fallbacks first; a grouped run records the issue in each group. The other messages are read word by word, NA, NaN and Inf in capitals only, and the names of the rank tests ("Rank permutation", "rank-based") are not a rank deficiency, so that too few observations for the rank permutation test are no longer one. The notes of summary() give the new issue, and drop the prefix of White's exact fit, "Test cannot be computed", which they kept. From the review of #182: the refusal is no longer a simpleError, as those of the model context are not; catch it by its class or as an error. A diagnostic a user registers under the name of a fallback, as registerDiagnostic("breusch_pagan", f) for White's test, is no longer tried after this refusal, nor after one of data that lack rows of the fit (the item below), as it was not after the other final refusals. The plurals "NAs" and "NaNs" are read as NA and NaN, "singularity" and "singularities" as "singular", and "Spearman rank" is a name too. - [ ] performBoxMTest() with a missing value in group or data returns NA: a row whose group is NA becomes a row of NAs in every group (data[group == gr, ]), and a missing value in data makes its group's covariance matrix NA. Found by the review of #158 (2026-10-26), where the test now refuses the groups on which it is not defined. Either refuse such data with a message that names the rows, or leave the rows out with a warning, as the group tests do ("Removed N observations due to missing values"). Either changes what such calls return. - [ ] performBoxMTest() on a group with a constant variable, on macOS. The check of #158 decides that the covariance matrix of a group is singular by the rank of its data centred at colMeans() (rboxm_refuse_singular(), R/performBoxMTest.R:196). Where colMeans() sums in double precision, as on macOS on arm64, whose long double is a double, the mean of 50 values of 0.2 is 0.19999999999999993, the centred column is a constant of -8.3e-17 rather than zero, and qr() gives the four columns rank 4; the log-determinant is -Inf, and the test stops with the message for a determinant that under- or overflows ("its variables have full rank, and their scale makes the determinant under- or overflow. Rescale the variables.") instead of the one that names the constant variable. test-boxm.R:157 fails so on the macOS (R release) job of every run on main since #158 (runs 40326050314, 40361728314, 40382244911 and 40392607338, the last on 67bf41a), and passes on Linux and Windows. The call stops either way; only the message and its advice differ. Look for a variable that is constant in the group before the rank, or decide the rank on the scale of the uncentred data. Found in the 0.16.0 release audit; recorded, not fixed. - [ ] performFlignerKilleenTest() and ties of decimal arithmetic between different observations. Two different observations can have deviations that are equal in exact decimal arithmetic, as when two pairs of observations of a regressor recorded with few digits differ by the same amounts; the binary values of the data break such a tie by a few units in the last place of the response, which can exceed the tolerance of 8 * eps * s, scaled by the residuals. On iris with every row twice, lm(Sepal.Length ~ Petal.Length) by Species, one of the 106 sets of exact ties stays broken, by 10.4 times eps * s, and the statistic is 8.14863 where the residuals computed exactly, scaled to integers, give 8.14287 (8.13778 before 0.16.0); with + Species, 0.479152 against 0.479156. Two forms of one model can differ by several percent: on 70 observations that repeat 30 values of x recorded with one decimal, y ~ poly(x, 3) + g gives 1.21547, and y ~ x + I(x^2) + I(x^3) + g and y ~ poly(x, 3, raw = TRUE) + g give 1.30280, 7% more: repeated values of x in different groups are different observations, rightly not given one residual, and poly() breaks their decimal ties in its last bits (the second review of #156). On iris itself and on the one-way fits of sleep, PlantGrowth, warpbreaks, InsectSprays and iris[-c(50, 100, 150), ] the value is exact. Found in 0.16.0 with the two items on the test under 0.16.0; recorded, not fixed. A tolerance scaled by the response would cover it, at the cost of merging deviations of continuous data that differ (the review of #156). - [ ] performFlignerKilleenTest() depends on the last bits of the BLAS. On iris with every row twice, lm(Sepal.Length ~ Petal.Length) by Species gives 8.14863 on R's reference BLAS; with its coefficients moved by one unit in the last place, as another BLAS may compute them, it gives 8.13333, 8.14287, 8.14863 or 8.14867, and 8.13333 is what CRAN's check of 0.16.0 with Intel MKL gave for the lm (8.149 for the glm). With + Species the value moves in its sixth digit (0.479151 to 0.479156); the poly() form does not move. So the statistic and the p-value a user sees depend on their BLAS. The likely cause: rfligner_rounding_ties() ties deviations within 8 eps times the largest absolute residual of their group, while the rounding of y - X b scales with |y| + |X b|, the size that rfligner_residuals() already computes; scale the tolerance by it. Found by the review of #172 (whose test of these fits runs off CRAN only); recorded, not fixed. Changes a value on such data. - [x] Two comparisons in the tests failed under the noisy BLAS after #172. The noise a comparison draws depends on every BLAS call made before it in the process, so a test, or a change to code, that runs earlier moves it. On CRAN the F of the variance form test is compared to 1e-8 (blas_tolerance()) on every cell of the model-context matrix, not only on those with missing rows (#184). And the Gaussian glm against lm(..., tol = 1e-11) on a design within 1e-9 of collinear (test-model-eligibility.R) is compared to 1e-4 on CRAN, 1e-6 off it. That design amplifies the last bits: one unit in the last place of the data moves the p-values of the glm and of the lm apart by up to about 5e-7, in the Harrison-McCabe test and in Koenker's. On #185, which changes the streaming code that earlier tests run, the noisy BLAS moved them past 1e-6: Noisy BLAS 2 once (Koenker's), Noisy BLAS 3 twice. The statistics of the default lm, which aliases x3, are at least 4.6e-3 from those of the glm, so the test still tells the two ranks apart. Tests only; no value changes. Done in #186. - [ ] Squared residuals that are equal but for rounding. On lm(y ~ x) with y = 1 + 0.5 x + 0.3 s, where s is 1 or -1, the residuals are 0.3 and -0.3, and their squares differ only in their last bits. With x = rep(1:20, each = 2) and s alternating, Koenker's test gives 8.13 on 1 (p = 0.004); with x = 1:40 and s in the pattern (1, -1, -1, 1), 15.25 on 1 (p = 9e-5). White's test gives 8.2 and 27.6, from the rounding alone. The streaming forms refuse both, saying the test "requires variation in squared residuals", and the studentized statistic stops on non-finite values. The same on 0.17.0 and with #184, whose rule rightly does not fire, since the regressors do not fit the rounding exactly. Decide the variation of the squared residuals in memory as the streaming forms do, on the scale of the residuals. Found by the review of #184; recorded, not fixed. Makes values refusals on such data. - [ ] performVIFDiagnostic() gives a number for an infinite VIF. On a design with exactly collinear columns, such as x3 = x1 * 3 beside x1 (the wf_local_variable workflow of helper-model-context.R), the VIF of both is infinite, and the function returns the diagonal of solve(cor(X)): of the order of 1 / eps where solve() succeeds, its digits set by the last bits of the BLAS, and Inf for every column where it fails, that of x2, 1.0, among them. Under the noisy BLAS of cran-blas.yml the two computations that test compares gave 1.5e15, 1.8e15, 2.25e15 or Inf for x1 and x3, and 1.0 or Inf for x2, as the seed and the call decided. Take the rank of the design first and give Inf for the columns it aliases only. Found by the noisy BLAS in #172, whose test leaves such VIFs out on CRAN; recorded, not fixed. Changes a value on such designs. - [ ] Left from step 1 of the model context (under 0.16.0), on the rows of data matched by number: - Follow-up, from the review of step 1: try the rows in the order of the fit also when the rows picked by number fail the comparison and data has one row per observation. That accepts a fit to d[order(x), ] given the same data numbered again, and character row names that look like numbers. - Known limit, as in 0.15.0: the rule on ties covers data taken in the order of the fit only. If the fit used numbered data with ties, and the same data is given permuted within those ties, it is matched by number and accepted, both before and after step 1. A test that reads a column the model frame does not hold then sees the rows exchanged: the rank permutation test with order_by = "z" gives -0.02142 instead of -0.02619 in the order of the fit. Extend the rule on ties to rows matched by number when the call reads such a column. - [x] Data that lack rows of the fit are refused without the class of the other refusals of rows. rstop_data_rows() (R/interface.R:673, called at R/model_context.R:678) refuses them with a plain stop(). Every other refusal of the rows of data has the class heteroTests_data_rows_error (rstop_data_rows_unmatched()), which the runners treat as final (R/error_handling.R:306). On a fit to the 120 rows of d, runHeteroTests(m, d[1:100, ]) tries the four fallbacks of White's test, each of which refuses the same rows. It then stops with "Data quality issue detected for White" and the advice to "Remove or impute missing values", which the "missing" of "Missing rows" selects (see the item on ht_analyse_failure() above). A tidymodels workflow and a parsnip fit_xy() fit already give this refusal the class (R/ecosystem_integration.R:636-645). Give it the class and keep its message, which the suite pins (tests/testthat/test-model-context.R:287), so that the runners raise it at once. Only the class, and the message of the runners, change. Found by the review of the R Journal article's subsection on the model context. Done in #182: the refusal keeps its message and has the class, so runHeteroTests(m, d[1:100, ]) and runDiagnostics() stop at once with "Data do not match the model for White: White test requires data to contain the rows used to fit the model. Missing rows: 101, 102, 103", that class and no advice. A grouped run, whose groups are refitted on their own rows and so do not meet it, would record it in its group with the issue "data_rows" and no fallback, as it records the other refusals of rows. The workflows and parsnip fits, which told it by its message, now tell it by its class. The refusal is no longer a simpleError, as the other refusals of rows are not; catch it by its class or as an error (from the review of #182). - [x] The streaming tests could overflow on terms that are combinations of others. chunked_regression_r2() (R/performance.R) stacked each chunk under the factor of the chunks before it by qr.R(qr(rbind(factor, rows), tol = 0)), the decomposition of LINPACK (dqrdc2) without pivoting, which makes a Householder reflection from every column, however little is left of it. What is left of a column of the auxiliary design that is a combination of others, as the squares and products of dummies in White's design are, is rounding error; the next such column leaves about the square of that, and so on down. dqrdc2 divides a column by its norm, and when that norm is below the reciprocal of the largest double the result is not finite: the next chunk, or the qr() that reads the factor, stopped on "NA/NaN/Inf in foreign function call (arg 1)". The defect predates #184, which found it: it came with the stacking, in 0.17.0, and is on main at 01164f6. The noisy BLAS jobs of cran-blas.yml on #184 stopped there, on all three seeds, in White's streaming test on a one-way layout of nine groups of ten observations and one of two, a design that pull request tests; R's reference BLAS does not reach it there, but a multiple of a regressor of size 1e-300 does it on any BLAS, in chunks of any size (test-streaming.R). Done in #185: the stack is decomposed by qr() at a tolerance of 1e-13 (rchunk_factor()), which moves to the end a column of which less than 1e-13 of its norm is left after the columns before it, and the factor keeps the rows up to the rank, its columns put back in their order. The rows left out hold only what is left of the columns of the design moved, so the cross-products of the other columns do not change; where no column falls below the tolerance, the factor is that of 0.17.0 to the last bit. What is left of the response is never left out: when dqrdc2 moves it, the factor keeps it as one more row, the norm of its effects past the rank after the reflections of the columns kept (qr.qty(), with the reflections dqrdc2 goes on to make from the columns it moved set to 0, as they can be not finite). The first version of #185 left it out too, and the review of #185 found that this lost the exact-fit decision: dqrdc2 compares the response with its norm about zero over the rows so far, which squared residuals that vary little are almost all of, so what is left below the tolerance can be the whole of the residual sum of squares, and it adds up over the chunks. In 10,000 chunks of 7 of squared residuals 1 + 2.5e-8 x and such a part, whose total sum of squares about their mean is 6e-16 of their sum of squares, the residual sum of squares was 0, an exact fit for White's streaming test, where one chunk and 0.17.0 give 3.16e-18, five times sqrt(.Machine$double.eps) of the total; it is now 3.16e-18 in chunks of either size (test-streaming.R). The factor can now have fewer rows than columns: the rank, the effects, the residual sum of squares and rfactor_regressors_exact() read it through qr(), as before. On 21 fits (continuous regressors, factors, a one-way layout, near-collinear and distant regressors, poly(), weights, a Gaussian glm and three mgcv::gam() fits, the basis of s(x1) among them) in chunks of 7 and 10,000, 106 of the 126 statistics of the three streaming tests are identical() to those of 0.17.0 and the others, each on a stack where a column fell below the tolerance, within 8.4e-16 relative (their p-values within 4.4e-15), with the same degrees of freedom and refusals; the factor is identical() on the 55 stacks where no column fell below it. On the 17 fits of the review (100 streaming results), one of 100,000 observations among them, the largest change is 2.0e-14 relative, less than the 2.2e-14 to 4.5e-14 by which White's streaming test differs from performWhiteTest() there in 0.17.0 as since. A decomposition of LAPACK with column pivoting, tried first on the branch of #184, moved White's streaming statistic on the basis of an mgcv::gam() fit by 1.5e-8, and was not used. - [ ] A column of very small norm can still overflow the stacking of the streaming tests. rchunk_factor() (R/performance.R) keeps a column of which at least 1e-13 of its norm is left after the columns before it, so its guarantee is relative: a column whose norm is below about 5.6e-296, 1e13 times the reciprocal of the largest double, can leave less than that reciprocal, and dqrdc2 divides by it. On lm(y ~ x1 + x2) with x1 = 1e-150 * u and x2 = 1e-150 * (u + 1e-6 * w), 60 observations, whose squares and product in White's design are of size 1e-300 and nearly collinear, performWhiteTestStreaming() stops on "NA/NaN/Inf in foreign function call (arg 1)" in chunks of 7 and of 10,000, in 0.17.0 and since

185 alike, where performWhiteTest() gives 25.476; at 1e-140 it gives¶

25.476 too. Found by the review of #185. Scale the columns of the stack before decomposing it, or refuse such a design with a message that says why; changes a value only where the test stops today. - [ ] Left from steps 8 to 10 of the model context, recorded, not fixed. Moved from 0.17.0 at its release (2026-10-30), where the steps that found them are done; each changes what a call returns: - From step 8: bootstrapTestStatistic() refits each replicate at the tolerance of lm(), 1e-7, as its refits of the formula did, so a fit whose rank was decided at a smaller tolerance and differs, an lm fitted with a smaller tol or a Gaussian glm (1e-11) with a nearly collinear column, is refused ("data does not reproduce the fit"), as such an lm was before. Refitting at the tolerance of the fit would give them values. A refusal becomes a value. - From step 8, left by step 11: fitWLS() of an mgcv::gam() fit whose smooths are all fixed (s(x, fx = TRUE)), the least-squares fit of its design, refits that design but returns an lm whose terms rebuild the variables of the formula, not the basis of the smooths, so the tests refuse it as a fit whose design does not reproduce its residuals (M8), as since step 8. Refuse such a fit in fitWLS() itself, as a penalised gam is since step 11, or keep its basis there, so that it can be tested. A refusal moves, or becomes a value. - From step 9: performInfluenceDiagnostics(), and so runDiagnostics()$influence, returns the Cook's distances padded with NA, one per row, as step 9 pads an output; but influential holds their positions among the rows, where the same fit with na.omit gives their positions among the observations (the names are the same). Decide which the positions count. A value changes, or none. - From the review of step 10: augment() of the results of a workflow whose preprocessing drops rows of new data, a recipe step applied to them (step_filter() or step_naomit() with skip = FALSE), refuses its data, since the preprocessing does not say which rows of data it kept; the runners take those data, whose rows the alignment confirms. Find the rows the preprocessing kept, or keep the refusal. A refusal becomes a value, or none. Measured on the second review of step 10: forging each row on its own, which says which rows a step that maps each row on its own kept, costs 2.75 ms a row (0.33 s for 120 rows, about 14 s for 5,000, where forging the whole data takes 3 ms), and a step that makes a row from other rows gives other values row by row; not adopted, the refusal kept. - From the third review of step 10: augment() refuses the data of a workflow whose recipe makes a row from other rows (step_mutate(cs = cumsum(x)), dplyr::row_number(), a mean over dplyr::lag(), step_lag()) or sorts data already in its order (step_arrange()), where 0.16.0 bound each row of data in the order of the fit its own values; the runners take them. A way to accept them: data in the order of the fit whose columns the preprocessing leaves unchanged identify each row (no two rows alike in them) and match the mapped rows in that order, so that each mapped row is the row of data at its position. A refusal becomes a value, or none. - [ ] Found by the model-eligibility work (the Supported models page), recorded, not fixed, as they change what a test computes rather than which models it accepts. Those below are not settled by the model context; the others are under 0.16.0 and 0.17.0: - rvalidate_against_reference("white", ...) compares White's statistic with car::ncvTest(), the Cook-Weisberg score test, so it reports a mismatch on every model. Compare with lmtest::bptest() on the squares and cross products, as test-model-eligibility.R does. Deprecated; changes what it returns. - performInfluenceDiagnostics() counts the aliased coefficients of a rank-deficient fit in p for its default cutoff, 4 / (n - p), and so runDiagnostics() does; the rank is the number of coefficients. Changes a value. - generateHeteroRecommendations() refuses a rank-deficient fit (it checks the model with the checks of validateTestInputs()), which the tests and runHeteroTests() accept since 0.16.0. Accepting it makes a refusal a value. - [x] Koenker's and the Breusch-Pagan tests fit the squared residuals exactly on the same designs, since their auxiliary regression is on the regressors: on the paired design of White's exact fit (under 0.16.0) performKoenkerTest() gives 60 on 30 degrees of freedom (p = 0.0009), the number of observations, as lmtest::bptest() does, and performBPTest() 63.25 on 30 (p = 0.0004), a function of the spread of the squared residuals alone, whatever their pattern. Found by the review of the follow-up of N2. Refuse such a fit in both, at the threshold of White's test, in a pull request of their own. Refuses calls that return a value. Since 0.16.0 both warn on that design, whose 30 levels of two observations are fixed effects for small groups (the item on them, under 0.16.0); the refusal is still to do. Since #183 performKoenkerTest() no longer calls summary.lm(), so on such a fit it no longer gives that function's warning "essentially perfect fit: summary may be unreliable": on lm(y ~ x1 + f) with 10 levels of 2 observations 0.17.0 warned and #183 gives the same 20 on 10 without it. Done in #184: both tests, and their streaming forms with the same message, refuse an auxiliary regression whose residual sum of squares is at most sqrt(.Machine$double.eps) of the total, the rule of White's test, now one function that White's test calls too (rauxiliary_exact(), R/performWhiteTest.R). On the paired design they refuse where they gave 60 and 63.25 on 30, and on lm(y ~ x1 + f) where they gave 20 and 28.03 on 10, before the warning on fixed effects for small groups, as White's test does. The message is White's, built by the same code (rrefuse_exact_fit()): "The auxiliary regression of Koenker's test fits the squared residuals exactly: the squared residuals are a combination of its 31 independent terms with its intercept, for 60 observations. Its R-squared is 1 and its statistic the number of observations, whatever the data. The design forces the exact fit, as a factor with two observations per level does, however many levels it has; use more observations per level."; the Breusch-Pagan test's says its statistic is "a function of the spread of the squared residuals alone, whatever their pattern". Its class is a general one, heteroTests_exact_fit_error, not White's, whose name and field regressors_exact are about White's auxiliary design; White's refusal carries it too, after its own heteroTests_white_exact_fit, with its message unchanged, so that one class catches every exact fit. ht_final_exact_fit() reads the general class, and the new refusals say that the regressors fit the squared residuals exactly, so in runHeteroTests() they are final: "breusch_pagan" and "koenker", streamed or not, stop with the message where they returned the statistic, without the Cook-Weisberg test, their last fallback, under their name; the runner drops the class, as it does on White's refusal (keeping it, as #182 does for the refusal of rows, is left to decide). A fallback that refuses an exact fit ends the search (a refusal of class heteroTests_exact_fit_error whose regressors_exact is not FALSE; other refusals still go on to the next fallback, as White's refusal of a saturated design, with regressors_exact = FALSE, goes on to the Breusch-Pagan test, 3.63 on 6 on 22 observations). So "high_dimensional", which refuses 20 observations, no longer returns the Breusch-Pagan test's 28.03 on 10 under its name on lm(y ~ x1 + f), and "spatial_hetero" without listw stops with its own refusal on 12 subjects of two, where it returned the Breusch-Pagan test's 12.37 on 12. The audit of the other tests of an auxiliary regression of the squared residuals, on the paired design: the studentized statistic (performStudentizedBPTest(), performBPTestRobust(studentize = TRUE) with or without the bootstrap, "student_bp"), 60 on 30, performBPTestRobust(studentize = FALSE), 63.25, and bootstrapTestStatistic() of either test (p = 1 for Koenker's) refuse it with the refusal of the test they compute; the wild bootstrap, which gave 60 with p = 1, every replicate an exact fit too, refuses it before the replicates; the Cook-Weisberg test with var_formula = ~ treat + subject (63.25 on 30), Harvey's test of the regressors (63.07 on 30, and an infinite F studentized), the high-dimensional test with variance_threshold = 1, which keeps every component (60 on 30), and the variance form test with a variance regressor constant in each pair and against = ~ subject (F = 4.06e29 on 28 and 30) refuse it by the same rule, with regressors_exact unknown (NA) for the Cook-Weisberg, Harvey and variance form tests, whose regression need not be on the regressors. Not exact fits there, and unchanged: the Cook-Weisberg test of the fitted values (0.608 on 1), Harvey's of the fitted values and their square, the high-dimensional test with its default 26 components (56.84 on 26), Glejser's and Park's tests of one variable, and Bartlett's test of groups of two; Levene's, Brown-Forsythe's and Fligner-Killeen's refuse groups of two already. autoTransform(), which chooses by the p-value of the Breusch-Pagan test, stops with its refusal on such a design. One step away, with a subject observed a third time or 20 subjects of three, every test gives the value it gave. In inst/validation/fixed-effects-size.csv the T = 2 cells of the two tests, 10 to 200 subjects, are such fits (Koenker's mean statistic is n, and the Breusch-Pagan test rejects in 25.8% to 100%): the README says so, and the rows are kept as measured; no T = 3 or other cell is one. One test of the warning moved from two observations a cell to three (test-fixed-effects-warning.R, the crossed a * b). Refuses calls that returned a value. - Found by the audit of #184: on the paired design Glejser's test of treat, which takes the values 0 and 1 in each pair, gives 1.1e-15 on 58 degrees of freedom (p = 1), since the absolute residuals of a pair are equal and their regression on treat has a slope of 0 whatever the data: a test that cannot reject. Any test of a variable that varies only within pairs, balanced as treat is, gives 0 there. Not an exact fit, and not changed. Refuse such a variable, or say so on the help pages; changes a value, or none. - [x] To fix before 0.18.0 is released: the rule refuses a fit that the data make exact, not the design, and its message blames the design. It compares the residual sum of squares of the auxiliary regression with the total sum of squares of the squared residuals, and one gross error in a level of two observations can dominate that total. On lm(y ~ g) with 9 groups of 10 and one of 2 (y normal with mean 50 and sd 5, set.seed(1)), one value of the pair multiplied by 100 leaves a residual of 2336 against an sd of 4.33; the dummy of the pair fits its two equal squared residuals, and the ratio falls below sqrt(.Machine$double.eps). 0.17.0 gives Koenker's test 92 on 9 and the Breusch-Pagan test 2069.35 on 9, and autoTransform() chooses boxcox. #184 refuses both tests, their streaming forms and the runner, and autoTransform() stops, saying "The design forces the exact fit … use more observations per level", which is false here. White's test has refused this case since 0.16.0. With y ~ x + f a shift of 200 sd is still computed and one of 500 is refused; if x varies within the small level, nothing is refused. Refuse only a fit the design forces: push a fixed pseudo-response through the fit, and refuse only if the same regressors also fit its squared residuals exactly, so that White's test gives its value there too. Found by the review of #184; to be fixed in its own pull request. Makes refusals values. Done in #187: the tests refuse an exact fit only when the design forces it. Each test pushes a fixed pseudo-response, standard normal draws from a fixed seed by R's default generators, after which the state and the kinds of the session's generator are put back, through the QR decomposition of the fit, with its weights, and refuses only when the same auxiliary regressors fit the same function of its residuals exactly too (rforced_exact_fit()): their squares, their logs for Harvey's test, their centred squares for the studentized statistic. When the squares of those residuals do not vary, whether the design forces the fit is not known, and the test refuses it without saying so. The first version of #187 used the fractional parts of i times the golden ratio, whose pattern its review found the designs could follow, by the order of their rows: on 72 subjects of two visits sorted by visit a fit the design forces gave Koenker's test 144 on 72, and on 4 units of 13 periods sorted by unit a fit the data make exact was refused as forced, where with the rows shuffled it gave 52 on 15. The streaming forms carry those squares as one more response in the pass they make, and decide as the exact forms do; the stacking of #185 keeps what is left of both responses (rchunk_factor()). On the layout above the dummies fit the squared residuals to within 1.1e-9 of their total, and Koenker's test gives 92 on 9 and the Breusch-Pagan test 2069.35 on 9, streamed or not, through runHeteroTests() too, and autoTransform() chooses boxcox, the values of 0.17.0. White's test gives its value there too, 92 on 9, and n R^2 = 80 on the pairs of test-performWhiteTest.R whose residuals are +(1 + x) and -(1 + x): a refusal becomes a value. Whether its regressors alone fit the squared residuals, which decides the runner's fallback, is asked of the squared pseudo-residuals too. A fit without a QR decomposition, an mgcv::gam() tested on its design or an lm fitted with qr = FALSE, is refused when the regression fits exactly, with a message that says only that it does, not that the design forces it. autoTransform() leaves out a candidate the Breusch-Pagan test refuses, since a candidate drops the rows where its response is not finite and so can fit another design, and stops with the first refusal only when every candidate is refused. A design that forces the fit, the paired design or a factor of two observations a level, is refused as in #184. - [ ] Left from the review of #187, recorded, not fixed. An mgcv::gam() fit keeps no QR decomposition, so on an exact fit it is refused without the reason and the advice, on a paired design too, which forces it; for a gam whose smooths are fixed the QR decomposition of its weighted design would tell. And on a fit the data make exact the F statistics come back as values, as in 0.17.0: Harvey's studentized test gives Inf on 1 and 78 on pairs whose log squared residuals are x exactly, and the F of the variance form test can be a ratio to rounding error. Say so on their help pages, or refuse them. - [x] The degrees of freedom of performBPTest() and performKoenkerTest() are the number of columns of the design but its intercept, not the rank of their auxiliary regression less its intercept, as White's are since N2: on a model without an intercept whose columns span the constant, such as lm(y ~ 0 + group + x1) with a factor of four levels, the auxiliary regression adds an intercept that the dummies already span, and the tests count 5 where the rank says 4. In the design of contract_data() the Breusch-Pagan test gives 8.653 on 5 and Koenker's 10.52 on 5, where lm(y ~ x1 + group), the same column space, gives the same statistics on 4; the streaming forms refuse that model ("auxiliary regression became rank deficient"), as before the pull request of N3. Found by that pull request. Take the rank, as White's test does; changes a p-value, and makes the streaming refusals values. Done in #183: both tests take the rank of the auxiliary regression less its intercept, decided by the code of White's test on the regressors centred one by one (rauxiliary_columns(), R/performWhiteTest.R), and their streaming forms the rank they decide in chunked_regression_r2() (R/performance.R) by the same rule, where they refused a rank deficiency. On lm(y ~ 0 + group + x1) the Breusch-Pagan test gives 8.653 on 4 (p = 0.0704, was 0.1237 on 5) and Koenker's 10.52 on 4 (p = 0.0325, was 0.0617 on 5), the values of lm(y ~ x1 + group) and of lmtest::bptest(), which takes the rank of its auxiliary fit too, so the package now agrees with it there; the streaming forms give them where they refused, and performBPTestRobust(studentize = FALSE), whose interval follows the degrees of freedom, and the registered "breusch_pagan" and "koenker" follow: runHeteroTests() asked to stream them (chunk_threshold_mb) gives the streamed values, where it fell back on the Cook-Weisberg test, 3.79 on 1, under their names. Changes a p-value. The auxiliary regression is fitted from the matrix of the columns kept (rauxiliary_fit()), on the centred columns where those as they are lose one, not by lm() from a formula on the columns as they are, and that changes the statistic on three kinds of design of full rank, from a wrong value to that of the streaming forms, which had it already, and of lmtest::bptest() but on the glm below, whose design it refits by lm.fit(), dropping the regressor too (from the review of #183): a model without an intercept with a regressor far from zero, which that fit dropped as a combination of its intercept while the degrees of freedom counted it, lm(y ~ 0 + I(x1 + 1e8) + x2) in the design of contract_data() giving 2.850 on 2 (p = 0.241) for 16.72 (p = 0.0002) and Koenker's 2.372 (p = 0.305) for 13.92 (p = 0.001), the values of lmtest::bptest() with varformula = ~ x1 + x2; a Gaussian glm, which keeps at a tolerance of 1e-11 a regressor that lm() would alias, glm(y ~ I(1e6 + x1 / 10) + x2) giving 2.005 for 12.45 and 1.457 for 9.044, the values of lm(y ~ x1 + x2); and a regressor named scaled_sq or e_squared, which the formula of the Breusch-Pagan or the Koenker test took for its response, lm(y ~ scaled_sq + x2) giving 102.0 for 12.45 and lm(y ~ e_squared + x2) 0.014 for 9.044. Changes a value there; on the 16 other fits the review compared, one of them a rank-deficient lm tested on the estimable part of its design, every result was identical() to 0.17.0's. A model whose only regressor is constant, lm(y ~ 0 + one), has no regressor beyond the intercept and is refused by all four ("requires at least one predictor beyond the intercept"), where the exact tests gave 0 on 1 and the streaming ones refused it as rank deficient; the exact and streaming forms refuse it with the error of a model with an intercept alone, a simpleError with the same message, as the "rank deficient" refusal was a simpleError too, of the rassumption_violation template of .std_error(), whose "Results may be unreliable" does not fit a refusal. Koenker's streaming test named the Breusch-Pagan test in that refusal, and names its own. runHeteroTests() on that model, which returned those values, stops with the refusal, as on lm(y ~ 1), and advises "Fit a model with a regressor that varies", where it advised removing one, its only regressor, as it did on lm(y ~ 1) when data held a constant column: on a model whose design is constant the constant columns of data are no longer advised away (ht_constant_design(), R/error_handling.R). The audit of the other tests of an auxiliary regression of the residuals: performWildBootstrapTest() reported the number of regressors as its degrees of freedom and reports the rank, and computed its statistic, for the data and each replicate, by the formula of lm() too, and computes it on the auxiliary regression of Koenker's test, so that on the far-from-zero model it gives 13.92 where it gave 2.372, the statistic of x2 alone, and a p-value of that statistic; its values on other fits are those of 0.17.0 (four fits at a fixed seed, to the last bit), and it refuses the model whose regressors are all constant, and a model with an intercept alone, lm(y ~ 1), on which it gave 0 on 1, p = 1, a test that cannot reject. The studentized statistic (performStudentizedBPTest(), performBPTestRobust(studentize = TRUE), "student_bp"), the Cook-Weisberg test (var_formula too), Harvey's test and the variance form test took the rank already; Glejser's, Park's, the spread-level and the Davidian-Carroll tests regress on one variable; and the high-dimensional test counts the principal components it keeps. No study of inst/validation/ fits such a model. - The studentized statistic (performStudentizedBPTest(), performBPTestRobust(studentize = TRUE), "student_bp") regresses the centred squared residuals on the design of the fit without adding an intercept, as lmtest::bptest() does, where performKoenkerTest(), whose statistic its help page says it computes, adds one. On a model without an intercept whose columns do not span the constant, lm(y ~ 0 + x1 + x2) in the design of contract_data(), it gives 1.507 on 1 degree of freedom (p = 0.220), the value of lmtest::bptest(), and Koenker's test 5.636 on 2 (p = 0.060). With an intercept, or columns that span the constant, the two agree. Found by the audit of #183. Add the intercept, or say on its help page that it differs there; changes a value, or none. - White's auxiliary design is built from the regressors centred one by one, so its rank does not depend on where a regressor that enters the model linearly is located, but a product or a power of a variable located far from zero inside the model stays ill-conditioned, and can lose a column of the auxiliary design while lm() keeps every coefficient. On 200 observations of two standard normal regressors (set.seed(1); x1 <- rnorm(200); x2 <- rnorm(200), y <- 1 + x1 + x2 + rnorm(200, sd = exp((x1 + x2) / 4))), lm(y ~ I(x1 + 2e3) * I(x2 + 2e2)) gives 46.83 on 7 degrees of freedom where lm(y ~ x1 * x2), the same model, gives 51.90 on 8, and lm(y ~ I(x1 + 2e3) + I((x1 + 2e3)^2)) 8.05 on 3 for 11.29 on 4. What counts is the shift relative to the spread of the regressor. On ten data sets of 200 observations (five seeds of standard normal and of U(0, 1) regressors, the response as above), with x1 shifted by k standard deviations and x2 by k or k / 10 of its own, and k from 100 to 40,000 in steps of a factor of 1.12: up to k = 700 the degrees of freedom held in all three designs and the statistic moved by at most 3e-5 relative (1.1e-7 up to k = 200); White's test lost one column from k = 790 to 2,000, the square first and the uniform regressors before the normal ones; and lm() aliased the term itself from k = 3,200 (the square, and the interaction with both shifted by k) to 11,000 (x2 shifted by k / 10), before White's test sees it. With I(x1 + 1e4) * I(x2 + 1e3) on the data above the fit is additive, and the test gives 35.02 on 5, the value of lm(y ~ x1 + x2). So shifts of 1e3 and 1e2 keep the degrees of freedom on standard normal regressors, 1,000 standard deviations, and lose a column on U(0, 1) ones, 3,300 to 3,700: 7 for 8 in three seeds, the statistic 0.5% to 6.8% lower. A year from 1990 to 2020, about 220 standard deviations from zero, and its square, year * z and income * age give the value of the centred model to seven digits or more. Found by the review of the follow-up of N2; the examples and thresholds corrected by the review of #149. An optional improvement: build Z from the orthonormal basis Q of qr(X_c), whose squares and products span the same columns as those of the centred regressors when the cross products are in, and only then. Changes a value on such designs. - On lm(y ~ poly(x1 + 1e6, 2) + x2, data = d) the model context refuses d, the data of the fit: the basis of the polynomial computed again from d through the predvars of the terms differs from that of the model frame by more than the tolerance of the comparison, and the tests stop with "cannot match the rows of data to the observations of the model". Without data, or with poly(x1 + 1e3, 2), White's test gives the value of poly(x1, 2). Found by the review of the follow-up of N2, as on main before it. A fix makes such calls return a value. - [ ] Defects found during the 0.12.0 and 0.13.0 work, left for 0.16.0 so that 0.14.0 changes no value. Each fix that changes a value is announced in bold in NEWS.md. The one below is not settled by the model context; the others are under 0.16.0 and 0.17.0. - The unexported performWhiteTestEnhanced(bootstrap = TRUE), through bootstrap_white_test(), takes the log of a log response twice, drops an offset given as an argument and takes rows by position. Remove it or fix it. No exported value changes.

0.19.0: release-grade evidence and the open decisions¶

Moved from 0.16.0 on 2026-10-26: the evidence for 1.0.0 at the grade of the release gate in inst/validation/README.md, measured on the code that will be frozen, and the decisions still open.

Re-run the full sweep at N_MC = 5000, so that the release gate in inst/validation/README.md ([0.042, 0.058]) applies to every exported test, not only Passes A and B. At 400 replications the Monte Carlo standard error is 1.1%, so the Gaussian sizes between 0.068 and 0.072 (Cook–Weisberg and NCV, Spearman, Levene) cannot be told apart from noise. Pass A, at 5000 replications, put Cook–Weisberg at 0.044. Moved from 0.16.0 (decided by the author on 2026-10-26): 0.16.0 does not need it, and run on the code before the freeze it certifies that code.
The bootstraps that estimate the weights again, at the release gate. The 0.15.0 default of reestimate_weights rests on rates measured at 2000 replications with B = 499 (inst/validation/weighted-fits-size.csv), where the 99% interval around 0.05 is [0.037, 0.063] and all four cells are inside it. Against the release gate's interval at 5000 replications, [0.042, 0.058], the wild bootstrap with 150 observations, 0.0605, is outside and the White bootstrap, 0.0415, just below it; against [0.041, 0.057], around 0.049, the rate of an exact test with B = 999, the wild bootstrap is outside. Measure the four cells of block 4 at 5000 replications with B = 999, the default, beside the true weights in the same design. From the time per sample measured in 0.15.0 that costs about 36 to 50 CPU-hours for the four cells, and about 7.6 for the two cells of the wild bootstrap alone. If the wild bootstrap stays outside the gate at 150 observations, say so on its help page.
The validation rows the R Journal article reads, on the code to be frozen. The article and its supplement read rates at render time from ten studies in inst/validation/:
- seven since the article's update to 0.17.0: weighted-fits-size.csv, glejser-skewed-errors.csv, harrison-mccabe-size-power.csv, modified-wald-size-power.csv, offset-refits-size.csv, variance-form-dnvk-size-power.csv and szroeter-design-null-size.csv;
- three since its additions of 2026-11-02 (heteroTests-paper #6): pass-a-size-power.csv, wild-multiplier-size.csv and fixed-effects-size.csv.
pass-a-size-power.csv was last written in 0.7.1 (2026-08-31). Five of the first seven were last written in 0.13.0 or 0.15.0, and so was wild-multiplier-size.csv (2026-10-15). offset-refits-size.csv and fixed-effects-size.csv were written in the work on 0.16.0 (2026-10-18 and 2026-10-23), and szroeter-design-null-size.csv the day before 0.16.0 was released (2026-10-26). None was measured on 0.17.0. The full sweep at N_MC = 5000 (above) may replace the Pass A rates the article reads.

The README says which rows predate a later change to their procedure, not the version each block was measured on. One such row is the Szroeter row of the known-weights block of weighted-fits-size.csv, measured before 0.16.0 took the test's null moments from the design, and not run again.

Re-measure every row that predates a change to its procedure on the code to be frozen, and record in the README the version on which each block was measured, so that the article can say it. Found by the review of the article's update to 0.17.0; the last three studies were added after the article's additions.

The article's range for fixed effects for small groups, 22% to 98%, comes from fixed-effects-size.csv, whose T = 2 cells of the Breusch-Pagan and Koenker tests are refusals since 0.18.0 (#184; the README says so, and the rows are kept as measured). The article reads only the cells of 30 subjects of 3 and of 4, 100 of 8 and 200 of 4, none of them an exact fit. - [ ] The performance study the R Journal article reads, on the code to be frozen. The article and its supplement read their timings and memory allocations from three extracts in the article's repository, paper/generated/performance_*.csv. paper/scripts/performance_study.R wrote them under heteroTests 0.11.2, and they were not written again, so the article says they do not measure the current code. Since then the exact Breusch-Pagan and Koenker tests have come to read the fit through the model context, the cache key has changed, and the streaming code has been rebuilt twice: on the stored design in 0.16.0 and on a stacked QR in 0.17.0. No milestone ran the study again. The script calls run_benchmark_suite() and generate_benchmark_report(), which move to the companion package (see the row of Decisions needed on its name, due in 0.19.0). The suite's ncv entry also times the internal .ncv_test(model) (R/benchmarking_system.R:363-368), where in 0.11.2 it timed the exported performNCVTest(model). Rewrite the script so that it times the exported tests and calls no function that leaves the package. Run it on the code to be frozen, and record the run in the article's PROVENANCE.md. Found by the review of the article's restructure to the journal's 20 pages. - [ ] Provenance of a study run in chunks. Commit the script before a chunked run is launched, and record its md5 in the settings of each part, so that MERGE refuses parts written by another version of the script. wild-multiplier-size.R predates the convention: its parts carry no md5, and adding one now would make MERGE refuse them. - [ ] The open decisions. This release settles the rows of Decisions needed (below) that are still open and due before 1.0.0, so that none is open at the freeze.

1.0.0: freeze¶

After 1.0.0¶

First-class panel and spatial heteroscedasticity workflows.
Helpers that bridge detection to conditional-variance modelling (ARCH/GARCH) for the time-series path.
An experiment/report artefact (model-card style) summarising a diagnostic run.
Streaming for more tests where it is meaningful, under a single chunk_threshold_mb policy.
Stronger feasible-WLS and auto-transform helpers: iterated feasible GLS, an estimator of the additive variance function that keeps its fitted variances positive, and a Box–Cox search.
A form of performVarianceFormTest() that does not need the standardized errors to have a constant fourth moment, following the robust statistic of Wooldridge (1991).
A variance-regressor argument for performBPTest() and performKoenkerTest(), the varformula of lmtest::bptest(), so that a fit with fixed effects for small groups, on which the tests warn since 0.16.0, can be tested against its regressors without the dummies. On the balanced designs of inst/validation/fixed-effects-dummies.R, where the tests with the dummies reject up to 100% of the time at the 5% level, the squared residuals regressed on x alone kept the level: 4.6% to 5.5% for Koenker's statistic and 4.3% to 5.6% for the Breusch-Pagan statistic. Unbalanced designs are untested. A new argument.
The test of Juhl and Sosa-Escudero (2014, Journal of Econometrics 178(3), 484-494) of heteroskedasticity after fixed effects estimation, for panels of many units observed a fixed number of times, the setting where White's, Koenker's and the Breusch-Pagan tests warn. It needs the user to name the factor that is the unit, which no test of the package takes today. A new export; its reference implementation, and whether it requires a balanced panel, are to be checked first (only the abstract could be read in 0.16.0).
A single parametrised accuracy-validation test, with reference coverage broadened to skedastic and sandwich.
The companion package for the dashboard and the benchmark suite (decided on 2026-10-14; not yet named).

Candidate tests¶

None of these is scheduled. Each would be a new export, so each has to clear the release gate in inst/validation/README.md before it ships: agreement to 1e-8 with an established implementation or with a reconstruction of the primary reference, simulated size inside the gate, and power against the alternative it is built for.

The list is arranged by where the test can be found today, because that decides both what there is to validate against and what the package would add. Availability was checked on 2026-10-09 by searching CRAN, the R documentation sites and the web for each test by name. "No R implementation found" reports that search. It is not proof that none exists.

No R implementation found¶

A first implementation in R. Where another system has one, it is the reference; where none does, the guard is a reconstruction of the primary reference and simulated size.

Modified Wald test for groupwise heteroscedasticity in a fixed-effects panel (Greene 2000; Baum 2001). The panel tests cover an individual effect and cross-sectional dependence, not unequal variances across units. Reference implementation: Stata's xttest3. The Python package panelbox (1.0.2) has a function of that name that computes a different statistic, a likelihood-ratio form. A second search on 2026-10-10 found no R implementation either. Measured for 0.13.0 and not adopted: it does not hold its level. In the designs of inst/validation/modified-wald-size-power.R, at 5000 replications, the test as xttest3 computes it rejects a true null hypothesis at the 5% level this often:

Errors	Units	T = 5	T = 10	T = 30	T = 100
Gaussian	10	0.927	0.681	0.282	0.100
Gaussian	30	1.000	0.966	0.584	0.183
Gaussian	100	1.000	1.000	0.928	0.361
t5	10	0.966	0.876	0.621	0.367
t5	30	1.000	0.999	0.944	0.717
t5	100	1.000	1.000	1.000	0.982

Two unbalanced panels are at 0.809 and 1.000 under Gaussian errors. Each unit's term is divided by a variance estimated from the same few residuals, and the excess grows with the number of units. In long panels the error reverses, because the statistic tends to a chi-squared with one degree of freedom fewer than its reference: with 2 units and 1000 periods it rejects 0.014 of the time. Three of the study's 40 null cells fall inside the release gate, all at 300 or 1000 periods, where the two errors happen to cancel, and under t5 errors it still rejects 0.129 of the time with 10 units and 0.235 with 30 at 1000 periods. Before a test of this hypothesis can ship it needs a different variance estimate or a small-sample correction, which makes it a different test. Two leads, neither part of the study: Bartlett's test on the within residuals held its level under Gaussian errors in an exploratory run but, being a normal-theory test, not under t5 errors; and Juhl and Sosa-Escudero (2014, Journal of Econometrics 178, 484-494) derive studentized tests for heteroscedasticity after fixed-effects estimation with the number of periods fixed, whose fit to this hypothesis has not been checked. Greene's page 598, which Baum follows, could not be obtained, so whether Greene's own form differs from Baum's was not checked. - A joint LM test of homoscedasticity in a one-way error-component model (Baltagi, Bresson and Pirotte 2006), against heteroscedasticity in the individual effect and in the remainder error together. - Pagan–Hall (1983), the test of constant variance after instrumental-variables estimation, where Breusch–Pagan and White are valid only under further conditions. Reference implementation: Stata's ivhettest. The package accepts no instrumental-variables fit today, so this one starts with the input layer. - Kelejian–Robinson (1998), a joint test for spatial autocorrelation and heteroscedasticity. sphet::kpjtest() is a different test by Kelejian and Piras. - Newey–Powell (1987), a test built on asymmetric least squares: under constant variance the expectile regressions have equal slopes. expectreg fits expectiles and has no such test. - Tests in nonparametric regression, which have power against any smooth departure from constant variance and need no variance regressors to be named: Eubank and Thomas (1993), Dette and Munk (1998), Zheng (2009), and Chown and Müller (2018). The cost is a bandwidth or a difference sequence to choose. - A consistent test of a variance function: Dette, Neumeyer and Van Keilegom (2007), Wang and Zhou (2007), and Samarakoon and Song (2011). performVarianceFormTest() has power against the terms it is given; these have it against any departure, at the cost of smoothing. Stage 1 of the first measured for 0.13.0, and not exported: as defined, it does not hold its level. The 2007 article could not be obtained; a prototype in plain R follows the authors' submitted manuscript of 2005, whose Tables 4.1 and 4.5 it reproduces in level and to within a few points in power. In the designs of inst/validation/variance-form-size-power.R, with the variance function in one variable and the mean a linear model, the Cramér-von Mises statistic is inside the release gate in 4 of its 12 Gaussian null cells at 5000 replications, all four with constant variance at 150 or 400 observations; it rejects a true exponential or power variance function 0.019 to 0.036 of the time at 50 to 400 observations. The bootstrap draws its data from the least-squares fit of the variance function, on which the statistic's null distribution depends: drawn from the true shape, on the same samples, three of the four such cells come to 0.047 to 0.051. Against a departure the added terms of performVarianceFormTest() do not span it has the power that test lacks, 0.750 against 0.089 at 400 observations; against the other alternatives it has less. Design note, prototype and evidence: inst/validation/variance-form-dnvk-design.md and the variance-form-dnvk-* files beside it. What to do next is under Decisions needed. Wang and Zhou (2007) assume a parametric mean, which is this package's setting; Samarakoon and Song (2011) could not be obtained. Stage 1 of Wang and Zhou (2007) is parked on the branch feat/wang-zhou-stage1, to be resumed after 1.0.0: pass 1 of its size and power study is done, and the Gaussian null cells at 5000 replications that decide it have not been run. The branch's wip/wang-zhou-stage1/README.md says how to resume. - Bai, Pan and Yin (2018), a test of homoscedasticity when the number of regressors grows with the sample. Li–Yao, below, is a different test for the same setting. - A design-based test of constant variance for survey data. The tests assume an independent sample with equal weights, and runSurveyHeteroTests() rejects a homoscedastic model 37% to 100% of the time when selection depends on the response or the sample is clustered. The natural candidates regress the squared residuals of a survey::svyglm() fit on the variance regressors using the design, as a Wald test (survey::regTermTest()) or as a Rao score test. Both were measured in inst/validation/survey-designs-size.R and neither clears the gate: 6% to 25% under the null at 300 and 1000 observations when the weights vary with the regressor. The variance estimate needs a small-sample correction first.

In another R package, not in this one¶

These add no test to R. They add it to the common interface, and each comes with a reference implementation to agree with.

Sign and size bias tests (Engle and Ng 1993): whether the squared residuals respond differently to negative and to positive lagged residuals. The time-series tests detect ARCH effects and say nothing about asymmetry. rugarch::signbias() computes the tests for a fitted GARCH model.
A break in the variance at an unknown date: the CUSUM of squares test (Brown, Durbin and Evans 1975), the statistic of Inclán and Tiao (1994), or the kappa tests of Sansó, Aragó and Carrion (2004), which stay valid under heavy tails and conditional heteroscedasticity. Goldfeld–Quandt and Szroeter need the ordering and the split to be chosen in advance. Reference implementation for the last: micss::kappa_test().
Simonoff–Tsai (1994) and Verbyla (1993). Versions of the Cook–Weisberg score test built on modified profile likelihood and on residual maximum likelihood, for small samples and high-leverage designs. Reference implementations: skedastic::simonoff_tsai() and skedastic::verbyla().
Bickel (1978) and Anscombe (1961). Tests of the variance against the fitted values. Bickel's is the robust one. Reference implementations: skedastic::bickel() and skedastic::anscombe().
Li–Yao (2019) for regressions with many regressors. performHighDimensionalTest() regresses the squared residuals on principal components and cites no heteroscedasticity test as its source. Reference implementation: skedastic::li_yao().
A spatial Breusch–Pagan test (Anselin 1988) for fitted spatial models. performSpatialHeteroTest() applies Moran's I to the squared residuals, which detects variance that clusters in space and does not test it against regressors. Reference implementation: spatialreg::bptest.Sarlm().
The rest of skedastic. Of its 25 tests, the package has counterparts for eight. Besides those named above, skedastic provides bamset(), carapeto_holt(), diblasi_bowman(), dufour_etal(), evans_king(), godfrey_orme(), honda(), horn(), rackauskas_zuokas(), wilcox_keselman(), yuce() and zhou_etal(). None is a priority by itself; each is a candidate when a user asks for it.

Order of work¶

The first group is where the package can contribute something R does not have, which is also what a reader of the article asks of it. Within that group:

The modified Wald test. The panel functions exist, and it is the test Stata users look for under the name xttest3. Measured for 0.13.0 and not adopted, because it does not hold its level (above).
The Dette, Neumeyer and Van Keilegom test. It removes the main limitation of performVarianceFormTest(), and is the most work of the two. Stage 1 done for 0.13.0: as the manuscript defines it, it does not hold its level when the variance is not constant (above).

Done¶

A Glejser test that is valid under skewed errors (Im 2000; Machado and Santos Silva 2000), in 0.13.0, as performGlejserTest(robust = TRUE). In the designs of inst/validation/glejser-skewed-errors.R the uncorrected statistic rejects a true null hypothesis 9% to 18% of the time at the 5% level when the errors are skewed, at every sample size, and skedastic::glejser() uses the uncorrected absolute residuals as well. The corrected statistic is the default since 0.15.0 (the 0.15.0 milestone, above).
Harrison–McCabe (1979), in 0.13.0, as performHarrisonMcCabeTest(), beside Goldfeld–Quandt and Szroeter. The 1979 article was not read; the definition rests on Harrison (1981), by its first author, whose published exact critical values the function reproduces, and on lmtest::hmctest(), with which the statistic agrees to 1e-8. The p-value is exact under normal errors given the regressors. Neither reference computes that p-value: lmtest::hmctest() simulates it from a regression on a constant alone, and skedastic::harrison_mccabe() 2.0.3 drops repeated eigenvalues and, with deflator, leaves the design in the order of the rows. In the designs of inst/validation/harrison-mccabe-size-power.R it holds its level under Gaussian errors, rejects 12% to 16% of the time under t5 errors as the other two ordered tests do, and has less power than either.

Decisions needed¶

Question	Proposal	Needed by
Ties in the ordering of `performSzroeterTest()`	Average the ranks within ties and adjust the null variance for them. Today ties keep the order of the rows the model was fitted to, which holds the level but makes the statistic depend on that order when `order_by` is discrete, and for the registered test on a model whose design rows repeat, as with factors, binary or other discrete regressors. It changes the values the function returns. Since 0.16.0 the null mean and variance come from the design, and they hold for any weights: with averaged ranks `a_i` in place of `i`, `h = sum a_i e_(i)^2 / sum e_(i)^2` has `E h = tr(M A) / nu` and `Var h = 2 (nu tr(M A M A) - tr(M A)^2) / (nu^2 (nu + 2))` with `A = diag(a_i)`, so the averaged ranks carry into the moments with no separate adjustment; `rszroeter_moments()`, which takes `1, ..., n` today, would take the `a_i`	0.19.0
Supported R floor	Test the floor in CI, or raise it to oldrel-1	0.18.0
The name of the companion package for the dashboard and the benchmark suite	None recorded. Once it has one, `launchDiagnosticDashboard()`, `run_benchmark_suite()` and `generate_benchmark_report()` warn here for a minor release, naming it, and the code moves	0.19.0
A consistent test of the variance function after stage 1	Do not ship the Dette, Neumeyer and Van Keilegom test as defined: its prototype rejects a true non-constant variance function 0.019 to 0.036 of the time. Choose between a new stage 1 for a modified procedure (a better estimate of the variance function for the bootstrap data, or a studentized statistic) and one for Wang and Zhou (2007), and compare the published article with the manuscript first. If a test ships, a function of its own rather than `performVarianceFormTest(against = "any")`	Before any stage 2
`setup.sh` / `renv` / Docker	Keep Docker for reproducibility; drop `setup.sh` now CRAN is the install route	0.18.0
Register the tests now callable as `(model, data)`	None recorded. Since 0.15.0 `performArchLMTest()`, `performMcLeodLiTest()`, `performSpearmanTest()`, `performDavidianCarrollTest()` and `performHarveyTest()` can be called as `(model, data)` with no further argument, which is when a test goes in `R/diagnosticsRegistry.R`; registering one adds a name `runHeteroTests()` accepts and changes the article	1.0.0