Skip to content

Dynamic Models (GMM)

Dynamic panel models include lagged dependent variables as regressors, capturing persistence and adjustment dynamics. Standard fixed effects estimation of dynamic panels is inconsistent due to Nickell bias -- the correlation between the lagged dependent variable and the transformed error term. Generalized Method of Moments (GMM) estimators solve this problem by using lagged levels or differences as instruments.

PanelBox provides four GMM estimators covering the main approaches in the literature, with built-in diagnostic tests for instrument validity (Hansen J, Sargan) and serial correlation (AR(1)/AR(2)).

Why GMM?

In a dynamic panel model \(y_{it} = \rho y_{i,t-1} + X_{it}\beta + \alpha_i + \epsilon_{it}\):

  • OLS is biased upward (\(\hat{\rho}\) too high) because \(y_{i,t-1}\) is correlated with \(\alpha_i\)
  • Fixed Effects is biased downward (\(\hat{\rho}\) too low) due to the Nickell bias, especially when \(T\) is small
  • GMM produces consistent estimates by using lagged values as instruments

Rule of thumb

GMM is designed for panels with large N, small T (many entities, few time periods). With large T, the instrument count can explode; use collapse=True to control proliferation.

Available Models

Model Class Reference Key Feature
Difference GMM DifferenceGMM Arellano-Bond (1991) First-differenced equations with lagged level instruments
System GMM SystemGMM Blundell-Bond (1998) Adds level equations with lagged difference instruments
CUE-GMM ContinuousUpdatedGMM Hansen et al. (1996) Continuously updated weight matrix; more efficient
Bias-Corrected BiasCorrectedGMM Hahn-Kuersteiner (2002) Analytical correction for finite-sample bias

Quick Example

from panelbox.gmm import SystemGMM
from panelbox.datasets import load_abdata

data = load_abdata()

model = SystemGMM(
    "n ~ L.n + w + k",
    data, "id", "year",
    gmm_instruments=["L.n"],
    iv_instruments=["w", "k"],
    collapse=True,
    time_dummies=False
)
results = model.fit(two_step=True)
print(results.summary())

Key Concepts

One-Step vs. Two-Step

Step Weight Matrix Properties
One-step Identity or robust Consistent; less efficient
Two-step Optimal (from step 1 residuals) More efficient; SEs may be downward biased

Practical advice

Two-step estimates are more efficient, but report Windmeijer-corrected standard errors (applied automatically by PanelBox) to account for the downward bias of two-step SEs.

Instrument Proliferation

The number of GMM instruments grows quadratically with \(T\). Too many instruments:

  • Overfit endogenous variables
  • Weaken the Hansen J test (test has no power)
  • Bias coefficient estimates toward OLS

Solutions: Use collapse=True or limit the lag depth with max_lags.

Essential Diagnostics

Test What It Tests Good Result
Hansen J Instrument validity (overidentification) p > 0.10 (do not reject)
AR(1) First-order serial correlation in differences Reject (expected)
AR(2) Second-order serial correlation in differences Do not reject 
# Diagnostics are included in the summary
print(results.summary())

# Or access individually
print(f"Hansen J: p = {results.hansen_test.pvalue:.4f}")
print(f"AR(1):    p = {results.ar_tests[1].pvalue:.4f}")
print(f"AR(2):    p = {results.ar_tests[2].pvalue:.4f}")

Detailed Guides

  • Difference GMM -- Arellano-Bond estimator (detailed guide coming soon)
  • System GMM -- Blundell-Bond estimator (detailed guide coming soon)
  • CUE-GMM -- Continuously updated GMM (detailed guide coming soon)
  • Bias-Corrected GMM -- Finite-sample correction (detailed guide coming soon)
  • Instruments -- Choosing and validating instruments (detailed guide coming soon)

Tutorials

See GMM Tutorial for interactive notebooks with Google Colab.

API Reference

See GMM API for complete technical reference.

References

  • Arellano, M., & Bond, S. (1991). Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations. Review of Economic Studies, 58(2), 277-297.
  • Blundell, R., & Bond, S. (1998). Initial conditions and moment restrictions in dynamic panel data models. Journal of Econometrics, 87(1), 115-143.
  • Roodman, D. (2009). How to do xtabond2: An introduction to difference and system GMM in Stata. Stata Journal, 9(1), 86-136.
  • Windmeijer, F. (2005). A finite sample correction for the variance of linear efficient two-step GMM estimators. Journal of Econometrics, 126(1), 25-51.