CUE-GMM (Continuous Updating Estimator)¶

Quick Reference

Class: panelbox.gmm.ContinuousUpdatedGMM Import: from panelbox.gmm import ContinuousUpdatedGMM Stata equivalent: gmm (eq), instruments(...) winitial(unadjusted) wmatrix(hac) R equivalent: gmm(..., type = "cue")

Overview¶

The Continuous Updating Estimator (CUE), developed by Hansen, Heaton, and Yaron (1996), is an alternative to two-step GMM that jointly optimizes the coefficient vector \(\beta\) and the weighting matrix \(W(\beta)\). Unlike two-step GMM, which fixes \(W\) at first-step estimates, CUE continuously recomputes \(W\) as a function of the parameters being estimated.

CUE-GMM minimizes:

\[\hat{\beta}^{CUE} = \arg\min_\beta \; g(\beta)' \, W(\beta)^{-1} \, g(\beta)\]

where \(g(\beta) = \frac{1}{N} \sum_i Z_i' \varepsilon_i(\beta)\) are the sample moment conditions and \(W(\beta) = \frac{1}{N} \sum_i g_i(\beta) g_i(\beta)'\).

This approach offers lower finite-sample bias and invariance to linear transformations of the moment conditions, making it a valuable robustness check for standard GMM results.

Quick Example¶

from panelbox.gmm import ContinuousUpdatedGMM

model = ContinuousUpdatedGMM(
    data=panel_data,
    dep_var="y",
    exog_vars=["x1", "x2"],
    instruments=["z1", "z2", "z3"],
    weighting="hac",
    bandwidth="auto",
)
results = model.fit()
print(results.summary())

When to Use¶

Robustness check for two-step GMM results (do estimates change substantially?)
Small to moderate samples (N = 100-1000) where finite-sample bias matters
Sensitivity concerns about moment normalization
Cross-sectional or pooled GMM settings (not dynamic panel-specific)

Key Considerations

CUE-GMM is computationally expensive -- the weighting matrix is recomputed at each optimization iteration
Good starting values are critical -- defaults to two-step GMM estimates
Convergence is not guaranteed with weak instruments
Unlike DifferenceGMM/SystemGMM, CUE-GMM works with explicit instrument variables (not GMM-style dynamic instruments)

Detailed Guide¶

Motivation: Why Not Two-Step?¶

Two-step GMM estimates \(\beta\) in two stages:

Estimate \(\hat{\beta}_1\) with \(W = I\) (identity)
Construct \(\hat{W}\) from step-1 residuals, re-estimate \(\hat{\beta}_2\)

This creates a dependence on the first-step estimates through \(\hat{W}\), which can cause:

Finite-sample bias in moderate samples
Sensitivity to normalization of moment conditions
Suboptimal efficiency when \(\hat{W}\) is imprecise

CUE eliminates this by treating \(W\) as a function of \(\beta\) in a single optimization.

Theoretical Properties¶

Under standard regularity conditions (Hansen, Heaton, and Yaron 1996):

Consistent: \(\hat{\beta}^{CUE} \xrightarrow{p} \beta_0\)
Asymptotically normal: same limiting distribution as efficient two-step GMM
Invariant: results do not depend on how moment conditions are scaled
Lower finite-sample bias: better coverage of confidence intervals

Weighting Options¶

CUE-GMM supports three weighting matrix types:

HAC (Default)ClusterHomoskedastic

model = ContinuousUpdatedGMM(
    data=data, dep_var="y",
    exog_vars=["x1", "x2"],
    instruments=["z1", "z2", "z3"],
    weighting="hac",
    bandwidth="auto",  # Newey-West automatic bandwidth
)

Newey-West HAC kernel accounts for heteroskedasticity and autocorrelation. Bandwidth is set automatically using \(L = \lfloor 4(T/100)^{2/9} \rfloor\).

model = ContinuousUpdatedGMM(
    data=data,  # Must have MultiIndex (entity, time)
    dep_var="y",
    exog_vars=["x1", "x2"],
    instruments=["z1", "z2", "z3"],
    weighting="cluster",
)

Cluster-robust weighting by the first level of a MultiIndex (entity).

model = ContinuousUpdatedGMM(
    data=data, dep_var="y",
    exog_vars=["x1", "x2"],
    instruments=["z1", "z2", "z3"],
    weighting="homoskedastic",
)

Simple moment variance. Use only when errors are homoskedastic.

Standard Error Options¶

Analytical (Default)Bootstrap

model = ContinuousUpdatedGMM(
    ..., se_type="analytical"
)

Uses the sandwich formula \(\hat{V} = (\bar{G}' \hat{W}^{-1} \bar{G})^{-1}\).

model = ContinuousUpdatedGMM(
    ...,
    se_type="bootstrap",
    n_bootstrap=999,
    bootstrap_method="residual",  # or "pairs"
)
results = model.fit()
ci = model.conf_int(alpha=0.05, method="percentile")

Residual or pairs bootstrap for inference robust to distributional assumptions.

Comparing CUE with Two-Step GMM¶

from panelbox.gmm import ContinuousUpdatedGMM

# Estimate CUE-GMM
cue_model = ContinuousUpdatedGMM(
    data=data, dep_var="y",
    exog_vars=["x1", "x2"],
    instruments=["z1", "z2", "z3"],
)
cue_results = cue_model.fit()

# Compare with two-step (using two_step results as baseline)
# CUE automatically uses two-step as starting values
j_test = cue_model.j_statistic()
print(f"J-statistic: {j_test['statistic']:.4f}")
print(f"J p-value: {j_test['pvalue']:.4f}")
print(f"Overidentification df: {j_test['df']}")

Configuration Options¶

Parameter	Type	Default	Description
`data`	`pd.DataFrame`	required	Data (MultiIndex for cluster)
`dep_var`	`str`	required	Dependent variable name
`exog_vars`	`list[str]`	required	Exogenous regressors
`instruments`	`list[str]`	required	Instrumental variables
`weighting`	`str`	`"hac"`	`"hac"`, `"cluster"`, or `"homoskedastic"`
`bandwidth`	`str` or `int`	`"auto"`	HAC bandwidth (`"auto"` or integer)
`se_type`	`str`	`"analytical"`	`"analytical"` or `"bootstrap"`
`n_bootstrap`	`int`	`999`	Bootstrap replications
`bootstrap_method`	`str`	`"residual"`	`"residual"` or `"pairs"`
`max_iter`	`int`	`100`	Maximum optimization iterations
`tol`	`float`	`1e-6`	Convergence tolerance
`regularize`	`bool`	`True`	Ridge regularization for near-singular W

Diagnostics¶

Hansen J-Test¶

j_test = cue_model.j_statistic()
print(f"J-stat: {j_test['statistic']:.4f}, p-value: {j_test['pvalue']:.4f}")
print(f"Degrees of freedom: {j_test['df']}")
print(f"Interpretation: {j_test['interpretation']}")

Convergence Check¶

results = cue_model.fit()
print(f"Converged: {cue_model.converged_}")
print(f"Iterations: {cue_model.niter_}")
print(f"Criterion value: {cue_model.criterion_value_:.6f}")

If CUE Does Not Converge

Try providing better starting values, increasing max_iter, or relaxing tol. Convergence failures often indicate weak instruments or near-singular weighting matrices.

Tutorials¶

Tutorial	Description	Link
Complete GMM Guide	Overview of all GMM estimators	Complete Guide
GMM Diagnostics	Interpreting J-test and other diagnostics	Diagnostics

References¶

Hansen, L. P., Heaton, J., & Yaron, A. (1996). "Finite-Sample Properties of Some Alternative GMM Estimators." Journal of Business & Economic Statistics, 14(3), 262-280.
Newey, W. K., & West, K. D. (1987). "A Simple, Positive Semi-Definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix." Econometrica, 55(3), 703-708.
Hansen, L. P. (1982). "Large Sample Properties of Generalized Method of Moments Estimators." Econometrica, 50(4), 1029-1054.
Hall, A. R. (2005). Generalized Method of Moments. Oxford University Press.