Numerical accuracy and limitations¶
rparity implements statistical methods independently in Python and compares their outputs against recorded observations from R packages. Comparisons cover estimates, covariance, predictions and inference, with declared tolerances for each quantity. A passing case establishes agreement for that case and configuration.
Recorded evidence¶
The synthetic corpus records 6,155 passes out of 6,256 cases. Its 101 unresolved differences remain recorded as failed comparisons. Sixteen examples from public R-package documentation provide additional checks. Use the option-level coverage to see which model configurations are covered.
| Model group | Passing comparisons | Recorded differences |
|---|---|---|
| Mixed models, GLS and inference | 3,799 / 3,856 | 57 |
| GAMs | 1,184 / 1,200 | 16 |
| Distributional models | 1,172 / 1,200 | 28 |
Detailed evidence is kept in the repository's development directory: mixed-model report, GAM and distributional-model report, and case-level discrepancies. This material documents reproducibility and numerical exceptions. It is not required to fit a model or use the package.
Before interpreting your fit¶
- Review convergence warnings and the model summary. An optimizer success flag or finite likelihood alone does not establish reliable uncertainty.
- Inspect singular random-effect covariance and weak zero-inflation or dispersion components. Their individual coefficients and Wald tests can become poorly identified.
- For GAMs, check basis size and residual patterns. Coefficients depend on the spline coordinates; fitted functions can agree while native coefficient coordinates differ.
- Choose the prediction and inference scale explicitly. Setting random effects to zero differs from integrating over their distribution.
- Use sum contrasts for Type III main-effect tests in interaction models.
The migration guides state the relevant options and limits for each model family. GAM standard errors condition on the fitted smoothing parameters. Distributional-model prediction standard errors are unavailable. Dense numerical calculations target models with modest basis dimensions and small or medium grouping structures.
Runtime and platform scope¶
The fitting and inference runtime uses no R installation or rpy2.
The published wheel has been executed and checked on macOS ARM64. Other
platforms still need execution validation. Cubic shrinkage smooths use a
bundled native OpenBLAS backend, with dependency license notices included in
the wheel. The Python implementation is MIT licensed.
R is used only to regenerate development references. See the clean-room record for the implementation and reference-generation boundaries.