pymargins

Expert-mode marginal effects for Python. Session-level analytical pre-commitment, JAX-native autodiff, and a κ-driven simulation fallback when the delta method is unsafe.

pymargins wraps a fitted statistical model in a Margins session, then computes adjusted predictions, slopes, contrasts, and arbitrary differentiable estimands — with uncertainty from the delta method, Krinsky–Robb simulation, or bootstrap — across statsmodels, linearmodels, lifelines, and scikit-learn model classes.

• Source code: https://github.com/huntermills707/pymargins

• Issue tracker: https://github.com/huntermills707/pymargins/issues

• Changelog: https://github.com/huntermills707/pymargins/blob/main/CHANGELOG.md

• PyPI: https://pypi.org/project/pymargins/

Getting started

• Introduction

Tutorials — learning by doing

• Getting started
• GLM — Binomial logit
• GLM — Poisson count
• OLS — linear regression
• Multinomial logit
• Cox proportional hazards
• Accelerated failure time models
• Instrumental variables (2SLS / LIML / GMM)
• Panel data — fixed effects
• Generalised Estimating Equations (GEE)
• Linear Mixed Effects
• Contrasts and difference-in-differences
• Inference — delta, simulation, bootstrap
• Inference scales and the κ diagnostic
• scikit-learn models
• Survey designs — weights, strata, and clusters
• Multiple imputation via reimpute
• Pooling precomputed imputations (pool_imputations)

How-to guides — task-focused recipes

• Robust and clustered standard errors
• Bootstrap inference
• Cluster and block bootstrap
• Simultaneous confidence intervals and multiple-testing correction
• Scenario helpers
• Grid predictions
• Effects by subgroup with over=
• Difference-in-differences on the response scale
• Elasticities and semi-elasticities
• Discrete changes for binary / categorical regressors
• Per-observation influence
• Reading and controlling the κ fallback
• Writing a custom adapter
• Matching support
• Exporting and persisting results
• Plotting predictions and effects
• Nonlinear estimands with evaluate
• Linear contrasts with contrasts
• Contrasts vs evaluate — choosing the right tool
• Formula interface for array-fit models

Demos — end-to-end analyses

• Mroz (1987) — Female labor force participation
• Fair (1978) — Extramarital affairs, two ways
• Spector & Mazzeo (1980) — Did PSI work?
• Rossi recidivism — Cox proportional hazards end-to-end
• Wage panel — union premium with entity fixed effects
• Union premium heterogeneity — effects by subgroup
• ANES96 — Multinomial logit on party identification
• Travel mode choice — the value of travel time (WTP)
• California API — stratified survey design
• Bayesian marginal effects — posterior draws via from_posterior
• Transform pipelines — honest CIs for data-dependent preprocessing

Reference

• API reference
  • Quick reference
  • Public API
  • Session
  • Results
  • Scenario helpers
  • Adapter interface
  • Concrete adapters
  • Gradient helpers (for adapter authors)
  • Transform pipeline (bootstrap)
  • Multiple imputation (pooling)
  • Matching

Explanations — theory and design

• Mathematical motivation
• The session pre-commitment
• Inference scale (phi / phi_inv)
• The κ curvature diagnostic
• Gradient backend: autodiff vs wrapped-FD vs FD
• Delta vs simulation vs bootstrap
• The scenarios model — at vs atexog
• The adapter pattern
• How pymargins differs from marginaleffects and margins

Indices

• Index

• Module Index

• Search Page