Driscoll-Kraay Inference with Short Time Series

Why the Number of Time Periods Matters

Driscoll–Kraay inference is useful when panel errors may be dependent across units as well as over time. Its covariance estimator first aggregates score contributions across units in each period, then estimates the time-series long-run covariance of those aggregates. That second step needs enough time periods to reveal how the aggregate scores vary and co-vary across lags.

This creates an important distinction: having many units does not by itself provide a long time series. A panel with thousands of units observed for five periods can estimate each period’s aggregate score precisely under suitable conditions, but it still contains only five time points for estimating the temporal covariance structure. Driscoll–Kraay standard errors in such a panel may be sensitive to individual periods and bandwidth choices, and their usual large-sample justification may be weak.

The issue concerns inference, not necessarily the fixed-effects coefficient itself. The coefficient may be well-defined and estimable, while the covariance estimate used to quantify its uncertainty remains fragile. A short time dimension is therefore a reason to examine the assumptions and sensitivity of inference, not a mechanical rule that one particular standard error must be used.

What the Driscoll–Kraay Estimate Uses

For clarity, consider a balanced panel with \(N\) units and \(T\) periods, and one coefficient of interest after removing fixed effects and other included regressors. Let \(\tilde{x}_{it}\) be the residualized regressor and \(\hat{u}_{it}\) the fitted residual. The estimated score contribution for unit \(i\) in period \(t\) is \(q_{it}=\tilde{x}_{it}\hat{u}_{it}\). Define the average per-unit bread and the normalized time-aggregated score as follows:

A_N=\frac{1}{N}\sum_{i=1}^{N}\sum_{t=1}^{T}\tilde{x}_{it}^{\,2}, \qquad G_t=\frac{1}{\sqrt{N}}\sum_{i=1}^{N}q_{it}.

For a maximum lag \(L\), the Bartlett-weighted Driscoll–Kraay variance estimate for this coefficient is

\widehat{\operatorname{Var}}_{\mathrm{DK}}(\hat{\beta}) = \frac{1}{N A_N^2} \left[ \sum_{t=1}^{T}G_t^2 + 2\sum_{\ell=1}^{L}w_{\ell} \sum_{t=\ell+1}^{T}G_tG_{t-\ell} \right], \qquad w_{\ell}=1-\frac{\ell}{L+1}.

The bracketed expression estimates the long-run variance of the time series of aggregate scores. Products \(G_tG_{t-\ell}\) retain dependence across periods; because each \(G_t\) sums over all units, the products also include cross-unit score products. This is the feature that allows the estimator to accommodate broad cross-sectional dependence under its assumptions. The standard error is the square root of the variance estimate.

The formula describes a common Bartlett-kernel version; finite-sample multipliers and details of the fitted model can differ across implementations. A reported result should identify the implementation sufficiently for a reader to reproduce it. The unit-clustered alternative instead allows arbitrary dependence over time within each unit and forms its variance from each unit’s time-summed score:

\widehat{\operatorname{Var}}_{\mathrm{unit}}(\hat{\beta}) = \frac{1}{N^2 A_N^2} \sum_{i=1}^{N} \left(\sum_{t=1}^{T}q_{it}\right)^2.

That estimator generally relies on cross-unit independence, or a condition making cross-unit dependence negligible for its limit. Neither covariance estimator dominates in every panel: they address different dependence structures and rely on different asymptotic arguments.

Definition: A short-\(T\) Driscoll–Kraay problem occurs when the time dimension is too limited for the time-series long-run covariance estimate to be dependable under the intended large-\(T\) approximation. In particular, with \(T\) fixed, increasing \(N\) alone does not supply an increasing number of time observations for estimating that long-run covariance.

Why Short \(T\) Weakens the Approximation

Driscoll–Kraay inference is commonly motivated by asymptotics in which the time dimension grows, together with conditions controlling temporal dependence and the behavior of the aggregate scores. The key point here is the need for \(T\) to grow: the lagged products in the formula must provide information about the aggregate score’s temporal covariance. When \(T\) is fixed, the estimator has only finitely many such products, even if \(N\) becomes very large.

The problem is not simply that the formula includes too many lags. Setting \(L\) to zero would remove the lagged products, but it would not establish that aggregate scores are uncorrelated over time. Conversely, choosing a smaller bandwidth can reduce the noisiness of distant-lag estimates, yet it cannot manufacture time-series information that the panel does not contain. Bandwidth choice is therefore part of the sensitivity analysis, not a remedy for a fundamentally short time dimension.

Identify the source of the approximation.
Driscoll–Kraay estimates a long-run covariance over the time-indexed aggregate scores \(G_t\), so its time-series justification requires enough periods.

Check what the panel actually supplies.
Record \(N\), \(T\), and the number of lags used. A large \(N\) does not change the number of observed aggregate-score time points.

Compare plausible alternatives and sensitivities.
Consider whether the dependence assumptions of unit clustering are credible, and examine whether the Driscoll–Kraay estimate changes substantially across defensible bandwidth choices.

Qualify the inferential claim.
With few periods, report the limitation and avoid presenting a conventional asymptotic confidence interval as though its nominal coverage were assured.

Worked Score-Based Comparisons

Worked Example: A Panel With Four Periods

Suppose a hypothetical balanced panel has \(N=4\), \(T=4\), and \(A_N=2\). Its estimated score contributions \(q_{it}\), arranged by unit across the four periods, are:

\begin{array}{c|rrrr|r} & t=1 & t=2 & t=3 & t=4 & \sum_t q_{it}\\ \hline i=1 & 1 & 0 & 1 & 2 & 4\\ i=2 & -1 & 0 & -1 & -2 & -4\\ i=3 & 1 & -1 & 1 & -1 & 0\\ i=4 & 1 & -1 & 1 & -1 & 0 \end{array}

The column sums across units are \(2,-2,2,-2\), so the normalized aggregate scores are \(G_t=(1,-1,1,-1)\). Use bandwidth \(L=1\), for which \(w_1=1-1/2=0.5\). The sum of squared scores is \(4\), and the lag-one product sum is \((-1)+( -1)+(-1)=-3\). Thus the bracketed Driscoll–Kraay quantity is

$$ 4+2(0.5)(-3)=1. $$

Therefore,

\widehat{\operatorname{Var}}_{\mathrm{DK}}(\hat{\beta}) =\frac{1}{4(2^2)}(1)=0.0625, \qquad SE_{\mathrm{DK}}=\sqrt{0.0625}=0.25.

For unit clustering, the squared unit-total scores sum to \(4^2+(-4)^2+0^2+0^2=32\). Hence \(\widehat{\operatorname{Var}}_{\mathrm{unit}}(\hat{\beta})=32/[4^2(2^2)]=0.5\), giving \(SE_{\mathrm{unit}}=\sqrt{0.5}\approx0.7071\). These numbers are computed from the same score contributions, but the estimates aggregate dependence differently. With only four periods, neither numerical result should be treated as automatically reliable: the Driscoll–Kraay estimate has very little time-series information, while the unit-clustered estimate has only four clusters and relies on its own cross-unit assumptions.

Worked Example: How Bandwidth Changes a Short-\(T\) Estimate

Now suppose a hypothetical panel has \(T=5\), \(N=100\), \(A_N=2\), and aggregate scores \(G_t=(1,0,1,0,1)\). The sum of squares is \(3\). For bandwidth \(L=1\), the lag-one product sum is zero, so the HAC bracket is \(3+2(1/2)(0)=3\).

For \(L=2\), the lag-one sum remains zero. The lag-two products are \(1\cdot1+0\cdot0+1\cdot1=2\), and \(w_2=1-2/3=1/3\). The bracket is therefore \(3+2(1/3)(2)=13/3\). The two variance estimates are

\widehat{\operatorname{Var}}_{\mathrm{DK},L=1} =\frac{3}{100(2^2)}=0.0075, \qquad \widehat{\operatorname{Var}}_{\mathrm{DK},L=2} =\frac{13/3}{100(2^2)}\approx0.01083.

The corresponding standard errors are approximately \(0.08660\) and \(0.1041\). They differ because the second bandwidth includes a lag-two covariance estimate. With only five periods, each available lag is based on few products; this calculation demonstrates sensitivity, not which bandwidth is correct. It also does not establish that either standard error has accurate finite-sample coverage.

Worked Example: Interpreting an Uncertain Coefficient Comparison

Suppose a hypothetical fixed-effects regression estimates a coefficient of \(0.60\) outcome units per unit increase in a regressor. The panel has \(T=5\). A chosen Driscoll–Kraay specification reports \(SE=0.20\), while unit clustering reports \(SE=0.35\). For illustration, the normal-reference statistics for testing \(H_0:\beta=0\) are \(0.60/0.20=3.00\) and \(0.60/0.35\approx1.714\).

Under a standard normal reference distribution, the corresponding two-sided p-values are approximately \(0.0027\) and \(0.0865\). The numerical difference changes the conventional 5% decision under those reference calculations. But the calculations do not show that one covariance estimator is valid and the other invalid. Driscoll–Kraay may be poorly approximated at \(T=5\); unit clustering may be unsuitable if cross-unit dependence is material, and its small number of clusters also matters. A responsible conclusion is that this panel’s inferential result is sensitive to covariance assumptions and short-\(T\) limitations—not that the coefficient is definitively different from zero based on whichever standard error is smaller.

Assumptions, Practical Meaning, and a Common Pitfall

Before using a Driscoll–Kraay standard error, state the dependence structure it is meant to accommodate and the conditions that support the time-series approximation. The usual motivation allows cross-sectional dependence through the aggregate scores, while requiring temporal dependence to be sufficiently controlled and the time dimension to grow for the long-run covariance estimate to stabilize. The precise regularity conditions depend on the model and implementation; a short time span is a direct warning that the large-\(T\) argument may be a poor guide to finite-sample behavior.

Unit-clustered inference is not a universal fallback. Its covariance estimate permits arbitrary within-unit serial dependence, but the conventional justification treats units as independent clusters (or requires cross-unit dependence to be asymptotically negligible). If units share shocks, that condition may fail. Conversely, selecting Driscoll–Kraay solely because cross-unit dependence is plausible does not remove its short-\(T\) limitation.

Common mistake: “There are many units, so the Driscoll–Kraay approximation must be accurate.”
Why it is wrong: The estimator’s long-run covariance is built from \(T\) time-indexed aggregate scores. Increasing \(N\) does not increase the number of observed time periods.
Correction: Assess the adequacy of \(T\), the temporal dependence conditions, and bandwidth sensitivity separately from the size of \(N\). Describe the limitations and the assumptions behind any comparison.

In applications, report \(N\), \(T\), the covariance estimator and bandwidth, and the standard errors used for the principal inference. When the time dimension is short, useful practice is to show how conclusions respond to defensible alternatives and to state clearly that such comparisons are sensitivity checks, not a guarantee of valid coverage. If the available panel cannot support the desired dependence-robust inference, that is a limitation of the information in the data, not a defect that can be repaired by reporting more decimal places.

Key takeaway: Driscoll–Kraay inference relies on a time-series approximation for aggregate scores. A large cross section cannot compensate for a very short time dimension, and agreement or disagreement with unit-clustered results must be interpreted in light of each estimator’s assumptions.

Check Your Understanding

Use the formulas and distinctions above to answer the following questions.

Why does increasing \(N\) without increasing \(T\) fail to supply more time observations for estimating the Driscoll–Kraay long-run covariance?
In the first worked example, what feature of the score aggregation allows Driscoll–Kraay products to include cross-unit score dependence?
For unit-clustered inference, what cross-unit condition is typically needed in addition to allowing within-unit serial dependence?
In the second worked example, why does the \(L=2\) variance differ from the \(L=1\) variance, and why does that difference not identify the correct bandwidth?
A researcher reports that a Driscoll–Kraay standard error is smaller than the unit-clustered standard error in a panel with six periods. What additional information and qualifications are needed before drawing an inferential conclusion?