Driscoll-Kraay Standard Errors¶

Quick Reference

Class: panelbox.standard_errors.DriscollKraayStandardErrors Convenience: panelbox.standard_errors.driscoll_kraay() Model integration: model.fit(cov_type="driscoll_kraay") Stata equivalent: xtscc (Hoechle 2007) R equivalent: plm::vcovSCC()

Overview¶

In many panel datasets, entities are subject to common shocks --- macroeconomic fluctuations, regulatory changes, or weather events that affect all entities simultaneously. This creates cross-sectional dependence: errors for different entities in the same time period are correlated.

Standard clustered SE (by entity) do not account for this. Driscoll & Kraay (1998) proposed a nonparametric covariance estimator that is simultaneously robust to:

Heteroskedasticity
Autocorrelation (serial correlation)
Cross-sectional dependence (spatial correlation)

This makes it ideal for macro panels and industry panels where common factors drive residual correlation.

When to Use¶

Panel data with common shocks (macroeconomic, policy, weather)
Large \(N\), moderately large \(T\) (need \(T\) large enough for HAC)
When the Pesaran CD test rejects the null of cross-sectional independence
Regional or industry panels where cross-sectional dependence is expected

When NOT to use

Single time series: Use Newey-West instead
Very short panels (\(T < 15\)): DK may not perform well
Spatial correlation with distance decay: Use Spatial HAC for explicit geographic weighting
Macro panels with \(T > N\): Consider PCSE as an alternative

Quick Example¶

from panelbox.standard_errors import DriscollKraayStandardErrors, driscoll_kraay

# Convenience function
result = driscoll_kraay(X, resid, time_ids, max_lags=3)
print(f"SE: {result.std_errors}")
print(f"Kernel: {result.kernel}")
print(f"Max lags: {result.max_lags}")
print(f"Time periods: {result.n_periods}")

# Class-based (more control)
dk = DriscollKraayStandardErrors(
    X=X, resid=resid, time_ids=time_ids,
    max_lags=5,
    kernel="bartlett",
)
result = dk.compute()

# Via model.fit()
from panelbox.models import FixedEffects
model = FixedEffects("y ~ x1 + x2", data, entity="country", time="year")
results = model.fit(cov_type="driscoll_kraay")
print(results.summary())

Mathematical Details¶

The Driscoll-Kraay Estimator¶

The covariance matrix is:

\[ V_{DK} = (X'X)^{-1} S_{DK} (X'X)^{-1} \]

where the meat \(S_{DK}\) is a HAC estimator applied to time-averaged cross-sectional moments:

\[ S_{DK} = \hat{\Gamma}_0 + \sum_{j=1}^{m} w(j) \left( \hat{\Gamma}_j + \hat{\Gamma}_j' \right) \]

The autocovariance matrix at lag \(j\) is:

\[ \hat{\Gamma}_j = \sum_{t=j+1}^{T} \left( \sum_i X_{it} \hat{e}_{it} \right) \left( \sum_i X_{i,t-j} \hat{e}_{i,t-j} \right)' \]

The key insight is that the cross-sectional sums \(\sum_i X_{it} \hat{e}_{it}\) aggregate information across entities at each time period, capturing both within-entity and cross-entity correlation.

Kernel Functions¶

The kernel weights \(w(j)\) ensure positive semi-definiteness of \(S_{DK}\):

Bartlett (default)ParzenQuadratic Spectral

\[w(j) = 1 - \frac{j}{m+1}\]

Linear decay. Most commonly used. Equivalent to the Newey-West kernel.

\[ w(j) = \begin{cases} 1 - 6z^2 + 6z^3 & \text{if } z \leq 0.5 \\ 2(1-z)^3 & \text{if } 0.5 < z \leq 1 \end{cases} \]

where \(z = j/(m+1)\). Smoother than Bartlett, with higher-order bias reduction.

\[w(j) = \frac{3}{x^2}\left(\frac{\sin x}{x} - \cos x\right)\]

where \(x = 6\pi j / (5(m+1))\). Optimal in a mean-squared error sense (Andrews, 1991).

Bandwidth Selection¶

The default bandwidth (maximum lags) follows the Newey-West rule:

\[ m = \left\lfloor 4 \left( \frac{T}{100} \right)^{2/9} \right\rfloor \]

\(T\)	Default \(m\)
10	1
20	2
30	2
50	3
100	4
200	5

Configuration Options¶

DriscollKraayStandardErrors Class¶

Parameter	Type	Default	Description
`X`	`np.ndarray`	---	Design matrix \((n \times k)\)
`resid`	`np.ndarray`	---	Residuals \((n,)\)
`time_ids`	`np.ndarray`	---	Time period identifiers \((n,)\)
`max_lags`	`int` or `None`	`None`	Maximum lags; if `None`, uses \(\lfloor 4(T/100)^{2/9} \rfloor\)
`kernel`	`str`	`"bartlett"`	Kernel: `"bartlett"`, `"parzen"`, `"quadratic_spectral"`

DriscollKraayResult¶

Attribute	Type	Description
`cov_matrix`	`np.ndarray`	DK covariance matrix \((k \times k)\)
`std_errors`	`np.ndarray`	DK standard errors \((k,)\)
`max_lags`	`int`	Lags used
`kernel`	`str`	Kernel function used
`n_obs`	`int`	Number of observations
`n_params`	`int`	Number of parameters
`n_periods`	`int`	Number of time periods
`bandwidth`	`float` or `None`	Bandwidth parameter

Kernel Comparison Example¶

from panelbox.standard_errors import DriscollKraayStandardErrors

kernels = ["bartlett", "parzen", "quadratic_spectral"]

for kernel in kernels:
    dk = DriscollKraayStandardErrors(
        X=X, resid=resid, time_ids=time_ids,
        max_lags=4, kernel=kernel,
    )
    result = dk.compute()
    print(f"{kernel:25s}: SE = {result.std_errors}")

Diagnostics¶

Diagnostic Summary¶

dk = DriscollKraayStandardErrors(X, resid, time_ids)
print(dk.diagnostic_summary())

Reports the number of observations, time periods, maximum lags, kernel function, and warnings about potential issues (e.g., few time periods, large max_lags relative to \(T\)).

Testing for Cross-Sectional Dependence¶

Before using DK, test whether cross-sectional dependence is actually present:

from panelbox.validation.cross_sectional_dependence.pesaran_cd import PesaranCDTest

cd_result = PesaranCDTest(results).run(alpha=0.05)
print(f"Pesaran CD: statistic={cd_result.statistic:.3f}, p={cd_result.pvalue:.4f}")
print(cd_result.conclusion)

Driscoll-Kraay vs Other SE Types¶

Feature	Clustered (entity)	Driscoll-Kraay	Newey-West
Heteroskedasticity	Yes	Yes	Yes
Within-entity autocorrelation	Yes	Yes	Yes
Cross-sectional dependence	No	Yes	No
Requires cluster structure	Yes	No	No
Asymptotics	\(G \to \infty\)	\(T \to \infty\)	\(n \to \infty\)
Best for	Micro panels	Macro panels	Single time series

Common Pitfalls¶

Pitfall 1: Too few time periods

DK requires \(T\) to be moderately large for the HAC estimator to work well. With \(T < 15\), DK standard errors may be unreliable. The diagnostic summary warns when \(T < 20\).

Pitfall 2: Bandwidth too large

Setting max_lags too large relative to \(T\) reduces the effective sample size for each autocovariance estimate. A rule of thumb is \(m < T/4\).

Pitfall 3: Highly unbalanced panels

DK works best with balanced or near-balanced panels. Severely unbalanced panels can distort the time-averaged moments.

References¶

Driscoll, J. C., & Kraay, A. C. (1998). Consistent covariance matrix estimation with spatially dependent panel data. Review of Economics and Statistics, 80(4), 549-560.
Hoechle, D. (2007). Robust standard errors for panel regressions with cross-sectional dependence. The Stata Journal, 7(3), 281-312.
Andrews, D. W. K. (1991). Heteroskedasticity and autocorrelation consistent covariance matrix estimation. Econometrica, 59(3), 817-858.
Vogelsang, T. J. (2012). Heteroskedasticity, autocorrelation, and spatial correlation robust inference in linear panel models with fixed-effects. Journal of Econometrics, 166(2), 303-319.