Diagnosing Common Shocks and Spatial Dependence

Why the Shape of Dependence Matters

A finding of cross-sectional dependence says that errors for different panel units move together; it does not, by itself, reveal how that dependence is organized. A macroeconomic surprise may affect many regions at once, producing broad co-movement. A local disruption may affect nearby regions more than distant ones, producing dependence that fades with distance. These structures have different substantive explanations and can motivate different error specifications and inference strategies.

This tutorial treats residual dependence as a diagnostic problem. The goal is to compare patterns across unit pairs, across distance, and across the eigenvalues of the residual correlation matrix. These summaries can make a broad common component more plausible, or instead point toward localized spatial dependence. They cannot identify a unique mechanism from residuals alone: common shocks can have heterogeneous loadings, spatial spillovers can extend far, and both structures can occur together.

Start With Pairwise Residual Patterns

For a balanced panel with \(T\) observations per unit, calculate a sample correlation for each pair:

A heat map or matrix of these correlations is often more informative than a single average. Broadly positive or negative correlations across many pairs are compatible with a common component. A pattern in which correlations are strongest for neighboring units and weaken with separation is compatible with localized spatial dependence. Blocks can also matter: a common shock may affect one industry or region, while units in different blocks respond weakly or with opposite signs.

Residual correlations are estimated from a finite time series, and the residuals themselves depend on the fitted model. With short \(T\), individual correlations can be noisy; estimating many pairs does not make each estimate precise. Serial dependence also affects the uncertainty of these estimates. Treat correlation maps as diagnostic summaries, not as a collection of independent formal tests.

Summarize Dependence by Distance

When units have geographic coordinates or a meaningful network, assign each pair a distance \(d_{ij}\). Then group pairs into distance bands and calculate the mean residual correlation within each band. For a band \(B_k\) containing \(m_k\) pairs, define

A marked decline in \(\bar{\rho}(B_k)\) as distance grows is evidence of distance-related dependence. A relatively flat profile, with correlations of similar sign across near and far pairs, is more consistent with broad co-movement. Neither pattern is decisive: a common shock with geographically varying loadings can produce a distance gradient, and spatial dependence with a long range can appear broad.

Distance is not always physical. In a production network, units may be close if they share suppliers; for financial institutions, closeness might be based on exposures. The weights or distance measure should be chosen for a substantive reason, rather than selected after inspecting correlations. If the network is represented by weights \(w_{ij}\), the analyst should state whether weights are binary, normalized, or otherwise scaled.

Why a Common Shock Creates Broad Correlation

A simple covariance argument explains one common-shock signature. Suppose, at a fixed time \(t\), the error for unit \(i\) has the representation \(u_{it}=\lambda_i f_t+v_{it}\), where \(f_t\) is a scalar common shock, \(\lambda_i\) is the unit-specific loading, and \(v_{it}\) is the remaining disturbance. Assume the variables have finite second moments and zero means. For two distinct units \(i\) and \(j\), assume all of the following:

• \(\operatorname{Cov}(f_t,v_{it})=0\) and \(\operatorname{Cov}(f_t,v_{jt})=0\);
• \(\operatorname{Cov}(v_{it},v_{jt})=0\);
• \(\operatorname{Var}(f_t)=\sigma_f^2\).

Under these conditions, expand the covariance rather than assuming that cross-unit uncorrelatedness of the \(v\)'s is sufficient:

The factor-orthogonality conditions are essential to the last equality. If either remaining disturbance is correlated with \(f_t\), the corresponding cross term need not be zero, even if \(v_{it}\) and \(v_{jt}\) are uncorrelated with each other. If loadings have similar signs, the common component tends to create positive covariances across many pairs; loadings with opposite signs can create negative covariances for some pairs. Thus, a common shock need not produce uniformly positive correlations.

Use the Residual Correlation Spectrum Carefully

A further summary is the eigenvalue pattern of the \(N\times N\) residual correlation matrix. A broad component can make one or a few eigenvalues large because many units move together in a common direction. By contrast, under a network with bounded local neighborhoods and bounded dependence strength, dependence confined to neighbors often spreads variation over more eigenvalues rather than concentrating it in one direction. This is a diagnostic contrast, not a universal classification rule: dense spatial networks and multiple common shocks can produce more complicated spectra.

For an idealized illustration, suppose the population correlation matrix has \(N\) units, unit diagonal, and the same off-diagonal correlation \(\rho\). It can be written as \(R=(1-\rho)I_N+\rho\mathbf{1}\mathbf{1}'\), where \(I_N\) is the identity matrix and \(\mathbf{1}\) is the vector of ones. The vector \(\mathbf{1}\) has eigenvalue \(1+(N-1)\rho\); every vector orthogonal to \(\mathbf{1}\) has eigenvalue \(1-\rho\). When \(\rho>0\) is fixed, the leading eigenvalue grows with \(N\), which is the signature of a broad, pervasive direction of co-movement in this idealized case.

A Practical Diagnostic Sequence

Common Mistakes and Limits

• Equating dependence with a common shock. A rejection of cross-sectional independence establishes that some cross-unit association is present; it does not say whether it is broad, local, or caused by an omitted common factor.
• Inferring a factor from one high correlation. One strongly correlated pair can result from a local link or chance variation. A common-shock interpretation requires a broad pattern or other supporting evidence.
• Treating distance decay as proof of spatial spillovers. Nearby units may share policies, industries, or exposure to common events. The pattern is compatible with spatial dependence but does not identify a causal mechanism.
• Dropping covariance conditions in the factor argument. The equality \(\operatorname{Cov}(u_{it},u_{jt})=\lambda_i\lambda_j\sigma_f^2\) requires both factor-residual covariances to be zero, as well as zero covariance between the two remaining disturbances. Cross-unit uncorrelatedness alone is insufficient.
• Using a single summary as a verdict. Correlation averages can hide opposite-signed loadings, distance bins can conceal clusters, and a leading eigenvalue can be affected by a few influential units. Read the diagnostics together and in context.

These diagnostics matter because the geometry of dependence helps determine whether a broad common component or a local interaction is a plausible description of the residuals. It also cautions against choosing a covariance estimator solely from a generic label such as “cross-sectional dependence.” The pattern is evidence about structure, not a substitute for a model, substantive reasoning, or appropriate inference.