pymargins.GLMAdapter

class pymargins.GLMAdapter

Base adapter for any model with prediction μ = g⁻¹(Xβ + offset).

Covers all GLM families uniformly. Concrete subclasses for statsmodels’ GLM and similar frameworks fill in framework-specific extraction.

Implementations of predict typically use either:

• jax-native: write the link inverse in JAX (logit, log, identity, probit are all available in jax.scipy.special)

• custom JVP with link.inverse_deriv: use _gradients.make_glm_jvp_wrapper

Both produce identical numerical results; choose based on the discussion in the design docs.

__init__()

Methods

__init__()
attach(session)	Attach this adapter to a Margins session.
bootstrap_state()	Replay state for a refitted adapter.
coefficients()	Return β̂ as a 1D JAX array.
column_index_of_variable(name)	Return the design-matrix column index corresponding to a variable.
covariance([vcov_spec])	Return Σ̂ as a 2D JAX array.
design_matrix_from_df(df)	Build a design matrix from a concrete DataFrame of evaluation rows.
predict(beta, X[, offset])	Prediction on the response scale, JAX-compatible.
refit(resampled_data, *[, index])	Refit the model on resampled data, returning a new adapter.
variable_metadata()	Return per-variable metadata used by averaging and validation.

Attributes

gradient_backend_recommendation	Recommend pure autodiff for GLM adapters.
n_outcomes	Number of outcome classes for multi-outcome models, default 1.
outcome_labels	Outcome class labels for multi-outcome models, or None.
supported_inference_methods	GLM adapters support delta, simulation, and bootstrap inference.
supports_jax_autodiff	GLM adapters typically use JAX-native link inverses or analytical JVP.
training_data	The training data used to fit the model.