dynpanelai

Machine learning and modern inference for dynamic panel data models.

🔗 Links

A unified Python implementation of eight methodologies for dynamic panels with high-dimensional controls — classical GMM, bias corrections, LASSO-based moment selection, double machine learning, orthogonal/debiased Lasso, optimal shrinkage, and neural lag discovery — behind one data container, one results object, and one publication-quality reporting layer.

pip install dynpanelai

Why this package

Dynamic panel models face a specific tension. Fixed effects are essential for persistent heterogeneity, but they interact with the lagged dependent variable to produce the Nickell bias. The classical fix — instrument with lagged levels — generates a number of moment conditions that grows like T², which reintroduces bias through overfitting. And modern applications add hundreds of controls, where naive post-LASSO inference is simply invalid.

Each module here solves one part of that problem, and they share enough infrastructure that you can run all of them on the same panel and compare.

Quick start

import dynpanelai as dp

df = dp.datasets.load_abond_employment()          # 140 UK firms x 9 years
panel = dp.PanelData(df, unit="id", time="year")

res = dp.diff_gmm(panel, y="n", lags=2,
                  predetermined=["w"], exogenous=["k"],
                  gmm_lags=(2, 4), steps=2)
print(res.summary())

==============================================================================
Difference GMM (2-step, FD)
Dependent variable: n
Observations = 611   Units = 140   Periods = 9
------------------------------------------------------------------------------
                    coef    std.err.         z     P>|z|
------------------------------------------------------------------------------
L1.n              0.3118      0.6357     0.490     0.624
L2.n             -0.0445      0.1825    -0.244     0.807
w                -1.2939      0.8251    -1.568     0.117
k                 0.3698      0.1580     2.340     0.019  **
------------------------------------------------------------------------------
instruments: 7
Hansen J: chi2(3) = 3.256, p = 0.354
AR(1): z = -3.079, p = 0.002
AR(2): z = -0.733, p = 0.464
==============================================================================

Known limitations

• system_gmm() raises NotImplementedError. The level-equation instrument block does not validate: on a mean-stationary simulated panel where the extra moments hold by construction, Hansen rejects at p<0.001 and the AR coefficient comes back 0.34 against a true 0.75. Use diff_gmm(..., collapse=True), which recovers 0.778 on the same design.
• AR(1)/AR(2) are labelled [approximate]. They are a simplified m-test that does not net out parameter-estimation error. Directionally reliable, not a substitute for the exact Arellano-Bond statistic.

Which method should I use?

Start from the shape of your panel, not from the method you have heard of.

Is T short (under ~15)?
├── Yes → gmm.diff_gmm (collapse=True), or biascorr.*
│         The ML estimators need sqrt(N)/T → 0 and will mislead you here.
└── No  → Do you have many controls, or nonlinear ones?
          ├── No  → Is m²/(NT) large?  (many moment conditions)
          │         ├── Yes → ablasso.ABLasso
          │         └── No  → gmm.diff_gmm
          └── Yes → What do you want to learn?
                    ├── One treatment effect       → dml.DMLDynamicPanel
                    ├── Effects across many groups → ortho.OrthogonalLasso
                    ├── All coefficients, uniformly → hdpanel.PanelLasso
                    ├── A forecast                 → shrink.PenalizedFE
                    └── Who responds over what horizon → neural.ACGate

Every estimator warns you when you are outside its assumptions — for example DMLDynamicPanel reports sqrt(N)/T and warns when it exceeds 1, and PanelLasso warns when the weak-sparsity assumption looks violated.

Comparing estimators

The comparison table is the point of a unified package:

res = {
    "FE":      dp.fixed_effects(panel, "n", lags=2, x=["w", "k"]),
    "DFE-A":   dp.debiased_fe(panel, "n", lags=2, x=["w", "k"]),
    "AH":      dp.anderson_hsiao(panel, "n", lags=1, exogenous=["w", "k"]),
    "Diff GMM": dp.diff_gmm(panel, "n", lags=2,
                            predetermined=["w"], exogenous=["k"]),
}
print(dp.comparison_table(res, params=["L1.n", "L2.n", "w", "k"]))
print(dp.comparison_to_latex(res, caption="Employment dynamics"))

Long-run effects with delta-method standard errors come for free:

lr, se = res["Diff GMM"].long_run("w", ["L1.n", "L2.n"])

Bundled real data

Both load as tidy long-format frames, ready for PanelData.

Documentation

• docs/user_guide.md — step-by-step, from a CSV to a finished table
• docs/syntax_reference.md — every estimator, every argument
• docs/methods.md — the econometrics behind each module
• examples/ — runnable scripts reproducing each paper's headline exercise

Installation

pip install dynpanelai              # core
pip install dynpanelai[plots]       # + matplotlib figures
pip install dynpanelai[neural]      # + PyTorch, for AC-GATE
pip install dynpanelai[all]         # everything

From source:

git clone https://github.com/merwanroudane/dynpanelai.git
cd dynpanelai
pip install -e ".[dev]"
pytest

Citation

@software{roudane_dynpanelai_2026,
  author  = {Roudane, Merwan},
  title   = {dynpanelai: Machine Learning and Modern Inference for
             Dynamic Panel Data Models},
  year    = {2026},
  url     = {https://github.com/merwanroudane/dynpanelai},
  version = {0.1.0}
}

Please also cite the paper behind whichever estimator you use; each module docstring gives the full reference.

License

MIT — see LICENSE.