RSearch Statistical Validation
How RSearch qualifies statistical analysis
This page describes the validation system behind released RSearch statistical capabilities. It intentionally reports validation at the system and analysis-family level rather than publishing a repetitive page for every individual method.
What qualification means
Qualification covers execution routes, input contracts, statistical evidence, reporting, privacy, and failure behavior. Registry counts reflect recorded qualification status, not a guarantee that every configuration or end-to-end workflow is free of defects. Reported issues are investigated and retested.
Numerical evidence
Reference fixtures and method-specific checks compare expected statistics, uncertainty, diagnostics, or model evidence against independent or native-R references where appropriate.
Eligibility and input contracts
Variable roles, minimum data requirements, unsupported combinations, missing-data policies, and negative inputs are tested so unsupported work is blocked rather than guessed.
Evidence-bound reporting
Released methods must produce normalized evidence that can support tables, interpretation, assumptions, provenance, and reproducible code without inventing unavailable results.
Privacy boundary
Qualified local methods execute in the browser. The capability contract explicitly requires raw research rows not to leave the browser through statistical execution.
Reproducibility
RSearch preserves the method identity, configuration, evidence, and generated R code needed to inspect or reproduce the analysis outside the guided interface.
Release gates
Qualification is enforced in automated release checks. Specialized families, including SEM and intensive methods, have additional method-specific gates.
Validation by analysis family
FamilyPrimary validation focus
Descriptive and diagnosticsNumeric reference, missing-data policy, diagnostic boundaries, negative inputs, reporting.
Group comparisons and associationReference statistics, assumptions, effect evidence, confidence intervals, role validation, claim boundaries.
Regression and predictive modelsCoefficient evidence, diagnostics, uncertainty, held-out validation where relevant, tuning boundaries, reproducibility.
Measurement and SEMMeasurement quality, structural evidence, model diagnostics, uncertainty, qualified package/runtime behavior, model-specific safeguards.
Causal, longitudinal, survival and BayesianDesign-specific assumptions, uncertainty, estimator-specific boundaries, convergence or censoring policies where relevant, reproducible execution.
Machine learning and clusteringInput contracts, out-of-sample or stability evidence where applicable, importance boundaries, resource controls, reproducibility.
Inspectable test evidence and reference sources
These are concrete reference inputs and test procedures used or available for comparison. A source fixture, a passing build, and a certified numerical match are different kinds of evidence. We do not present an unverified comparison as a pass.
ExampleEvidence and current disclosure
NIST Longley, linear regressionThe repository includes
nist_longley.csv and regression case R1. The automated analysis-regression suite exercises Longley through browser R. NIST's certified reference R-squared is 0.995479004577296. The suite includes explicit checks of the native-R baseline's certified intercept (absolute tolerance 0.1) and R-squared (absolute tolerance 1e-8). Current measured output and pass/fail ledger: not published here. This is not a certified comparison of every RSearch method or a NIST endorsement. NIST certified results.R-native regression and statisticsThe regression qualification code checks examples against native R statistical procedures with defined numerical tolerances, including 1e-7 and 1e-6 for selected checks. This is internal reference testing, not NIST certification. R reference manuals.
Structural equation modelsSeparate SEM and CFA qualification checks exist in the codebase. Reference model specifications, estimators, and package versions must be matched before any numerical agreement is claimed. lavaan documentation.
End-to-end and release checksBrowser, offline/PWA, recovery and release checks run in GitHub Actions. An individual green workflow demonstrates only the checks it actually executed; skipped tests and known failures must not be counted as passes. Frontend CI history ยท Combined release CI history (repository access may be required).
Known failures and gaps in public validation evidence
Public results ledger not yet available: This page does not currently publish a method-by-method dated report of observed values, certified expected values, tolerances, pass/fail status and environment versions. No complete passing numerical count should be inferred from the registry summary above.
Documented release qualification failures: Prior staging checks have failed on browser navigation timeouts, transient R package downloads, and release-version or rollback identity mismatches. Those failures required investigation and were not evidence of numerical validity. Release checks can be inspected in the linked workflow histories.
Unresolved or untested configurations: A passing reference fixture cannot establish correctness for every dataset, estimator, sample size, missingness pattern, or model. Unsupported configurations must remain labelled or blocked. See Known Limitations.
RSearch is not certified or endorsed by NIST. References to NIST denote independent benchmark data, not approval of RSearch.
What this page does not claim
Validation does not mean every possible dataset, model specification, or researcher decision is automatically appropriate. RSearch can qualify software behavior and statistical procedures, but researchers remain responsible for study design, construct meaning, measurement quality, data provenance, and substantive interpretation.
When a capability has not completed its required qualification, RSearch should block or label that boundary rather than silently switch to an unqualified substitute.