Technical reference: testing

Summary

Counts below are from a fresh full run on the installed package: all files green, 0 failures. “Tests” = test_that() blocks; “Assertions” = passing expect_* checks.

All 37 testthat files run standalone under Rscript (each carries library(growthcleanr) at the top). Most also carry skip_on_cran(). The exceptions are test-cran-smoke.R, which runs on CRAN by design, and test-adult-clean.R, test-child-sde-charid.R, test-consolidated-input-warning.R, and test-gc-runtrial.R.

Running Tests

# Single file
NOT_CRAN=true Rscript -e 'testthat::test_file("tests/testthat/test-child-algorithms.R")'

# All files
NOT_CRAN=true Rscript -e 'testthat::test_dir("tests/testthat/")'

IMPORTANT: Tests use library(growthcleanr), which loads the installed package. After any code changes, reinstall before testing — otherwise tests run against stale code:

devtools::install_local(".", force = TRUE, upgrade = "never")

Development testing not distributed. Some testing used in developing the algorithms is not part of the distributed test suite, because it depends on large or non-public fixtures and is being rebuilt. It includes a stress test on a 400-subject dataset with induced errors, code locks on preterm catch-up and preterm same-day cases, tests of the recumbent/standing offset for gestational-age-corrected children, a comparison of ext_bmiz() against the CDC SAS program, and a regression harness for the adult algorithm that checks every row of a 1,508-row fixture at all four permissiveness levels.

Per-file detail

The sections below give a per-test breakdown for the files whose individual tests are worth enumerating. Files without a section here (for example the input-format, hostile-input, flexible-coding, CLI, and recentering suites) are fully described by their row in the Summary table above — the Scope column carries their coverage rather than repeating it as a per-test table.

test-cleangrowth.R

The cleangrowth() entry point — dispatches to child and adult algorithms and assembles combined output.

test-cran-smoke.R

The only test file that intentionally runs on CRAN (no skip_on_cran()), because almost every other algorithm file is skipped there for runtime. A sub-second end-to-end cleangrowth() sanity check on tiny inline fixtures, asserting only definitional/unambiguous cases so it cannot fail on a platform-specific trajectory judgment — it guards build/load failure, output-assembly regressions, and gross numeric divergence.

test-data-frame-api.R

The data-frame style entry point (the primary calling style) and the id auto-generation/validation behavior.

test-child-regression.R

Frozen-count and stability tests on syngrowth data. These break on intentional behavior changes — update expected values (and document the re-baseline) when that happens.

test-partial-run.R

Tests the gc_partial_run() incremental-run wrapper. The core guarantee is that a partial run equals a full run of the same data under the same options (gc cleans each subject independently after reference recentering).

test-child-algorithms.R

Constructed-scenario tests organized by algorithm step.

• CF rescue (Step 6): standard rescues more than none; cf_status populated for rescued; none mode rescues none; all mode rescues all; cf_detail adds cf_status/cf_deltaz/cf_nextz; CF-E rows (CF sharing an Include day) labeled and cf_nextz skips temp SDEs; cf_detail columns absent by default.
• CF rescue rule (Step 6, deltaz/nextz): a carried value followed by a continued decline is rescued (CF-R); one followed by a jump up (true CF) is excluded (CF-NR); all CFs in one identical-value string share a single status (whole-string decision on the last CF); an end-of-series CF (no cf_nextz) falls back to the age rule.
• Evil Twins / OTL (Step 9): single/two/three consecutive unit errors (×2.54) excluded as Exclude-C-Evil-Twins; collateral damage on clean neighbors bounded.
• Error Load (Step 21): threshold and mincount parameters control behavior; a many-error subject triggers Unevaluable-Trajectory on remaining Includes.
• Age boundary: HEADCM cleaned ≤5y, Exclude-HC-Out-of-Range >5y.
• Parallel execution: parallel = TRUE matches parallel = FALSE.
• GA correction (Step 2b): very low birth weight gets correction (ctbc.sd ≠ tbc.sd); birth z ≥ −2 does not.
• Birth EWMA2 (Steps 15/16): extreme birth weight excluded; normal included; extreme birth height excluded.
• LENGTHCM: results identical to HEIGHTCM for the same measurements.

test-child-parameters.R

Each configurable parameter is accepted and behaves as documented.

test-child-edge-cases.R

Minimal, degenerate, and boundary-condition data.

test-child-et.R

Child Evil Twins on a deterministic fixture (the real ET membership coverage — test-child-algorithms.R Tests 5–7 only assert “excluded somehow”).

test-child-birth-ewma2.R

Birth HT/HC EWMA2 changes (Steps 15/16).

test-child-corroboration.R

test-child-unit-error-range.R

The default-ON informational unit_error_range output column and its internal helpers.

test-exclusion-code-crosswalk.R

Locks the Detailed→Summary crosswalk directly (guards the silent-NA-factor class: a new internal code with no crosswalk entry fails here). Runs a syngrowth subset + constructed rows guaranteeing shared/child/adult folds.

test-adult-clean.R

64 unit tests calling cleanadult() directly (via growthcleanr:::), organized by step. 224 assertions. Covers PIV (1H/1W incl. BMI-PIV), minimal/edge cases, Weight Cap (4W), same-day (9H/10W), Evil Twins (9Wa anchor rule), Extreme EWMA (9Wb), Height Distinct (10H 2D band + 3+D window + ht_band), Moderate EWMA (11Wb), 1D (13, incl. BMI limits + custom limits, the no-BMI height minimum being equal at loosest and looser, and a single-value height kept when any same-day BMI is in range), Error Load (14), parameter validation (incl. invalid and non-scalar permissiveness → clean error, not a length-mismatch crash; and hostile setting overrides — NA / non-finite / non-scalar / wrong-type overall_ht_min / allow_ht_gain / repval_handling etc. → clean type errors, not scalar-if() crashes), output structure, repeated values (2W), 2D ordered/non-ordered weight (11Wa/11Wa2), cap influence, the full-fixture regression (against inst/testdata/adult-gc-test-ALL-PHASES.csv, 1508 rows; the WT and HT rows carrying expected codes are compared), imperial units, mean height (computed from finally-included heights only, populated on excluded height rows, NA when no height is included), multi-subject isolation, the permissiveness framework (4 levels + overrides + allow_ht_gain/mod_ewma_f/allofus15 effects), and permissiveness_presets() completeness.

test-adult-cutpoint.R

Routing at adult_cutpoint, the adult-cutpoint advisory, and adult-specific input handling, all end-to-end through cleangrowth(). 24 assertions.

test-adult-et.R

Adult Evil Twins anchor rule (Adult Step 9Wa), the adult-side companion to test-child-et.R.

test-checkpoint.R

Checkpointing via cleangrowth(checkpoint_dir = ...) (the folded interface) plus the deprecated cleangrowth_checkpoint() shim.

test-utils.R

Post-cleaning utilities (outside the algorithm). 524 assertions (loop-heavy).

test-visit-date-agedays.R

Fractional agedays is cleaned as whole days but returned as submitted, with the whole-day value in gc_internal_agedays (internal_detail). Computing agedays from a date of birth plus a visit date (data-frame style). Covers route resolution (agedays column wins when present; age_source = "dates" forces the date route; supplying both an agedays column and dates uses agedays silently, no warning), the both-required rule for computing ages from dates (a lone date column with no agedays column errors), the parse-only path (a lone visit_date or dob alongside an agedays column is parsed and surfaced as gc_internal_visit_date/gc_internal_dob under internal_detail without driving ages, and is gated behind that flag), and the expanded set of accepted text-date formats: order-free year-first numeric (2013-03-04, 2013/03/04, 20130304) and English spelled-month (03mar2013, March 15, 2024), plus ambiguous numeric day/month (3/4/2013) only when date_order is set. Also covers the anti-misparse guards — one format class per run (jointly across both columns), four-digit 19xx/20xx years only (two-digit and out-of-range rejected), English-locale-only month names, and pattern-based classification so a content-unambiguous 25/12/2013 still needs date_order — rejection of unrecognized/numeric (epoch-ambiguous) columns, the gc_internal_dob/gc_internal_visit_date audit columns under full_detail, blank/NA dates yielding a missing age, and a visit date before the date of birth.

test-carry-through.R

carry_through for long input: columns growthcleanr does not consume as role columns are parked and returned appended after the growthcleanr columns, in their original order with values and types intact. Covers reserved-name collisions (the user’s column returned as userinput_<name>, with a warning; a user column named line carried unchanged), carry_through = FALSE restoring the drop-all behavior, dob/visit_date echoing back this way, the no-id path (reattachment keys on input row position, not the user’s id), carried columns named like a remapped role column (e.g. a column subjid when subjid = "patid") returned with the user’s values in single and checkpointed runs alike, a role column named line returned intact alongside carried columns, and the guarantee that the input table is not modified.

test-consolidated-input-warning.R

The consolidated input warning (Option B) and the three Class-2 defect warnings it introduced. Covers: clean data with ids is silent; measurement == 0, negative measurement, and negative agedays each warn and are set aside (per-parameter-group N/% for the measurement value problems, all-rows N/% for age); several issues collapse into a single warning() whose lead line counts the union of distinct set-aside rows (not the sum); a missing id alone yields only the Generated ids section (no Unusable data lead line) with the tie-breaking explanation; and the unrecognized-sex bullet leads with the sex_coding fix.

test-semiwide-input.R

input_format = "semi_wide": one row per visit, one column per parameter, reshaped to canonical long before cleaning. The keystone is equivalence – the same measurements submitted semi-wide and long must produce identical exclusion codes, checked on a hand-built fixture and a syngrowth slice. Also covers semiwide_cols remapping, structural column remapping, empty cells versus a literal 0, same-day duplicates resolving as same-day extraneous values, the missing-weight/height warnings (the head-circumference-absent warning was removed 2026-07-24), the date route, run-to-run determinism, the error catalog, and long output returning the user’s id under its own name with growthcleanr’s number in gc_internal_id, the long-only length rule (length.adjust is rejected in semi-wide input; a LENGTHCM/LENGTHIN value at any age gets default handling, matching a length column mapped to HEIGHTCM, including at 19 years and at days 729 and 730).

test-wide-input.R

input_format = "wide": one row per subject, repeated observation blocks marked by a _N column suffix, reshaped to canonical long before cleaning. The keystone is again equivalence – the same measurements submitted wide and long must produce identical exclusion codes – checked on a 40-subject syngrowth slice and on a hand-built ragged fixture; a companion test asserts wide and semi-wide input agree on the same measurements, and the hand-built case also pins the minted-id emission order (subject, block, then weight -> height -> head circumference). Also covers block discovery from the data (occupied cells only, so ragged subjects need no padding), default parameter-named prefixes, wide_cols and structural/age-prefix remapping, imperial and length prefixes (pounds/inches clean identically to the metric equivalents while param/measurement come back verbatim), two blocks on one day resolving as same-day extraneous values, the date route (dob + visit_date_N), a value in a block with no age flagged Exclude-Missing-Info, the four strictness warnings (block-numbering gap, max_blocks exceeded, non-monotonic block ages, a subject on more than one row), the missing-parameter warnings, run-to-run and row-order determinism, the input table left unmutated, the error catalog (no block columns, prefix with no _N columns, unrecognized parameter, two parameters sharing a prefix, missing subjid, no age source, supplied id, leading-zero block collision), the round-trip spec the re-widened output layer consumes, and the long-only length rule (length.adjust rejected; a LENGTHCM/LENGTHIN value at 2 years or older gets default handling, matching a prefix mapped to HEIGHTCM).

test-semiwide-output.R

output_format = "semi_wide": the submitted table is returned with the requested growthcleanr columns slotted in beside each value column. The keystone closes an equivalence chain in both directions – long -> long == semi -> long == melt-back of semi -> semi – on a hand-built fixture and on a syngrowth slice with same-day duplicates preserved as separate visit rows; the melt-back is written in the test rather than reusing package code, so an adapter bug cannot cancel itself out. Also covers echo fidelity (identical() per submitted column, input not mutated, verdict placed immediately right of its value column), widened_emit control and its switch-gated errors, imperial conversion reported through gc_internal_meas/gc_internal_param, the two-directional length offset on long-format fixtures (both directions and the 729/730-day boundary checked through sd.orig, conversion-before-offset, gc_internal_meas left at the entered metric value, adult LENGTHCM relabeled HEIGHTCM but not offset, missing ages left alone; length.adjust in semi-wide input is no longer supported – see the long-only length rule under the input files), carry_through in mirror mode, the three scope errors, and determinism.

test-wide-output.R

output_format = "wide": the submitted subject rows are returned with the requested growthcleanr columns slotted in beside each value block column (exclude_wt_1 right of wt_1) – direct long -> wide, echoing the input table rather than rebuilding it. The keystone closes the same both-directions equivalence chain as semi-wide output – long -> long == wide -> long == melt-back of wide -> wide – on a hand-built ragged fixture and on a syngrowth slice (blocks built from distinct visit days); the melt-back is written in the test. A separate weight-only test exercises same-day extraneous values at scale – injected duplicate same-day weights become consecutive same-agedays blocks (the case the WT+HT keystone must dedupe away) and are asserted to resolve (>= 8 Exclude-Extraneous) and round-trip through wide output unchanged. Also covers echo fidelity (identical() per submitted column, input not mutated, verdict immediately right of its value column, dropping the generated columns reproduces the input), ragged blocks getting NA verdicts, widened_emit control including the wide-specific per-block gc_internal_agedays_N level (one column per block, not per measurement, NA for absent blocks) and switch-gated errors, imperial values echoed verbatim while reported in metric through gc_internal_meas/gc_internal_param, carry_through in mirror mode, same-day blocks resolved as extraneous values with verdicts landing on their own columns, the scope errors (long/semi-wide -> wide refused) and the pre-cleaning name collision, and determinism.

test-growthcleanr-app.R

The bundled Shiny app: that it ships and loads, plus the app’s mapping-stage guards that reject bad column mappings before cleaning runs. The mapping-guard tests exercise the app’s helper logic directly (not a live browser session).

Coverage Gaps

Summary of the standing gaps: