Common questions

Does analysis require R?

No. NumPy and SciPy perform analysis. R is used only for Gaussian reference checks.

Should I supply effect estimates or raw outcomes?

Use set_data and fit for estimates and standard errors from an earlier analysis. Use fit_student_t for a specified native t error model. Use fit_regression or its family wrappers for numeric predictors and outcomes. These routes use different likelihoods and inference; see methods.

Which regression families are available?

Gaussian, binomial/logistic, Poisson, negative-binomial (NB2), multinomial, ordinal, Poisson/NB2 zero-inflated and hurdle, Cox, and Weibull proportional hazards. Gaussian random intercepts or slopes can be added with one grouping factor. Residual SD, NB2 dispersion, ordinal thresholds and Weibull shape are supplied and held fixed.

Are credible intervals frequentist confidence intervals?

They are posterior credible intervals under the stated model. Summary-statistic intervals condition on fitted prior parameters. Regression intervals also depend on finite posterior sampling. Repeated-dataset coverage needs a separate study. Cox survival intervals omit additional plug-in baseline-estimation uncertainty.

Can values be missing?

Summary effect estimates may use NaN. Each row needs an observed condition; all scale inputs must remain finite and positive. Inference uses observed Gaussian or t blocks. Pattern learning requires complete selected rows. Direct regression rejects missing outcomes and predictors.

Why did a fit stop without convergence?

Covariance learning can reach an iteration limit or reject a constrained proposal. Mixture fitting reports its objective and Karush–Kuhn–Tucker (KKT) residual. Student-t inference also reports quadrature refinement. Regression reports chain traces, split R-hat, effective sample sizes and component switches. Successful execution alone does not establish convergence.

What does prediction do for a new group?

Known group labels use their posterior effects. Unknown labels require allow_new_groups=True, which draws group effects using the retained random scale. Omitting group labels sets their effect to zero. For a nonlinear link, this differs from averaging responses over a population of groups.

Why is a sharing entry NaN?

No rows met the selection threshold in either condition, so its denominator is zero. Sharing describes posterior means; it is not the posterior probability that two effects are equal.

How can memory use be reduced?

For Gaussian summaries, use fit_prior and prior.iter_apply(data, batch_size=...). Reduce batch size for many conditions or components. Student-t inference stores quadrature components; regression stores retained draws. Their allocation guards are not total process-memory limits.