PyXtabond2: Dynamic Panel Data Estimation in Python

pyxtabond2 is a Python package for estimating dynamic panel data models using the Generalized Method of Moments (GMM). It replicates the matrix algebra, options, and diagnostics of Stata's xtabond2 (David Roodman), and adds an original extension for panels with unobserved interactive fixed effects: IFE-GMM.

Ideal for applied econometrics and macroeconomic research, this package bridges the gap between Python's data science ecosystem and dynamic panel methodologies.

📖 Full documentation: docs/latex/user_manual.pdf (every option, Stata mapping, diagnostics) and docs/latex/methodological_note.pdf (full econometric derivations and validation record).

🌟 Key Features

• Difference GMM (Arellano-Bond 1991) and System GMM (Blundell-Bond 1998), one-step and two-step, with all four robust/non-robust combinations Stata supports.
• Windmeijer (2005) correction: exact finite-sample correction for two-step robust standard errors.
• Forward Orthogonal Deviations (FOD): Arellano-Bover (1995) transformation, maximizing sample size in unbalanced panels with gaps.
• Instrument collapsing and per-variable-group instrument suboptions (lag(), collapse, eq() via GMMStyle/IVStyle) to control instrument proliferation.
• Generalized h(), artests(), arlevels, and cluster() options, matching xtabond2.mata.
• Comprehensive diagnostics: Arellano-Bond AR(ℓ) tests, Sargan/Hansen J-tests, and Difference-in-Sargan/Hansen tests for instrument exogeneity.
• IFE-GMM: an original iterative PCA-defactoring GMM estimator for panels with interactive fixed effects (Bai 2009), with automatic factor-count selection (Bai & Ng 2002; Ahn & Horenstein 2013) and a split-panel jackknife bias correction (Dhaene & Jochmans 2015).
• Direct export: publication-ready tables to LaTeX or Microsoft Word, including multi-model comparison tables (GMMStargazer).

📦 Installation

pip install pyxtabond2

For Word export:

pip install "pyxtabond2[export]"

Development install from source:

git clone https://github.com/Kahindo048/pyxtabond2.git
cd pyxtabond2
pip install -e ".[dev,export]"

🚀 Quick Start

pyxtabond2 comes with integrated example datasets so you can start experimenting immediately.

from pyxtabond2 import PanelData, PyXtabond2, load_dataset

# 1. Loading the data
df = load_dataset('df_panel.csv')

# 2. Data preparation
panel = PanelData(df, id_col='Country', time_col='Year')
panel.data['L1_Growth'] = panel.get_lag('Growth', 1)
df_ready = panel.data.reset_index()

modele = PyXtabond2(df_ready,
                    id_col='Country',       # Group identifier (country, firm)
                    time_col='Year',        # Time identifier
                    dep_var='Growth',       # Dependent variable
                    x_vars=['L1_Growth', 'Capital', 'Labor', 'Wage', 'Investment'],
                    gmm_vars=['Growth', 'Capital'],  # Arellano-Bond instruments
                    iv_vars=['Ide'],                 # Standard IV instruments
                    model_type='system', twostep=True, robust=True, small=True)

# 3. Estimation
result = modele.fit()
result.summary()

# 4. Export results for publication
result.to_latex("gmm_results.tex", full_output=False)
result.to_word("gmm_results.docx", full_output=False)

See examples/examples.py for a full workflow with four specifications and comparative export (and examples/ife_exemples.py for the IFE-GMM companion workflow), and the User Manual for the complete option reference (including IFE-GMM).

Performance

pyxtabond2 vectorizes the panel transforms, instrument construction, and GMM engine with NumPy/Numba, and caches panel structure across the IFE-GMM defactoring loop instead of re-running pandas on every iteration. On a realistic Monte Carlo design (System GMM, N=200, T=46, collapse=True, comparing standard GMM / IFE-GMM / IFE-GMM with jackknife bias correction per replication), the current release fits that three-estimator trio roughly 2.4× faster than before the most recent optimization pass alone (gating diagnostics during IFE-GMM iterations, cheaper rank/inverse computations), and substantially faster still relative to earlier releases that re-ran the full data pipeline every IFE-GMM iteration. Always prefer collapse=True for simulation-heavy designs (see the User Manual's performance notes for parallelization guidance).

Known limitations

Weights (pweight/aweight/fweight) and Stata's multi-way cluster(var1 var2) are not implemented; a few narrow numerical edge cases (documented in the Methodological Note) can diverge from Stata. See the User Manual's "Known Limitations" section for the complete, current list.

📖 References & Methodology

This package implements the algorithms and corrections outlined in the following seminal papers (see docs/latex/methodological_note.pdf for the complete bibliography and full derivations):

Arellano, M., & Bond, S. (1991). Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations. The Review of Economic Studies.

Arellano, M., & Bover, O. (1995). Another look at the instrumental variable estimation of error-components models. Journal of Econometrics.

Blundell, R., & Bond, S. (1998). Initial conditions and moment restrictions in dynamic panel data models. Journal of Econometrics.

Windmeijer, F. (2005). A finite sample correction for the variance of linear efficient two-step GMM estimators. Journal of Econometrics.

Roodman, D. (2009). How to do xtabond2: An introduction to difference and system GMM in Stata. The Stata Journal.

Bai, J. (2009). Panel data models with interactive fixed effects. Econometrica.

Bai, J., & Ng, S. (2002). Determining the number of factors in approximate factor models. Econometrica.

Ahn, S. C., & Horenstein, A. R. (2013). Eigenvalue ratio test for the number of factors. Econometrica.

Dhaene, G., & Jochmans, K. (2015). Split-panel jackknife estimation of fixed-effect models. Review of Economic Studies.

🤝 Contributing

Contributions, issues, and feature requests are welcome! Feel free to check the issues page on the GitHub repository.

License

MIT License. See LICENSE.