Common Errors in Interpreting Spatial Spillovers

Why Spatial Spillovers Are Easy to Misread

A spatial or network model can include coefficients for neighboring outcomes, neighboring covariates, or both. Those coefficients describe parts of the model, but they do not automatically equal the full response of outcomes to a change in a covariate. Feedback through the spatial multiplier can transmit a change across units and back again. Nor does a nonzero spatial coefficient, by itself, establish that one unit causally affected another: the model’s assumptions and identification matter.

The aim here is to keep three claims separate: what a coefficient means within a specified model, what the model implies for marginal effects across units, and whether those effects have a causal interpretation. The calculations use small hypothetical systems so that the distinctions are visible.

Definition: In a spatial-lag model, \(y=\rho Wy+X\beta+u\), \(y\) is the \(n\)-vector of outcomes, \(X\) is the matrix of covariates, \(\beta\) is the vector of covariate coefficients, \(W\) is a specified spatial-weights matrix, and \(\rho\) is the spatial-lag coefficient. If \(I-\rho W\) is invertible, the model can be written \(y=(I-\rho W)^{-1}X\beta+(I-\rho W)^{-1}u\). The coefficient \(\rho\) describes the model’s dependence on the weighted neighboring outcomes; it is not, in general, the marginal effect of a covariate.

The weights matrix represents the chosen definition of which units are connected and how their values are combined. It is part of the model, not a neutral description that can be ignored during interpretation. For the examples below, \(W\) is treated as fixed, and its entries are specified explicitly. In applied work, interpretation also depends on the units, scaling, and construction of \(W\).

Coefficient Versus Marginal Effect

For the spatial-lag model above, suppose \(X\) includes a covariate \(x\) with coefficient \(\beta_x\), and hold the other regressors fixed. The matrix of derivatives of outcomes with respect to the entries of \(x\) is

\frac{\partial y}{\partial x'}=\beta_x(I-\rho W)^{-1}.

Entry \((i,j)\) gives the model-implied change in outcome \(y_i\) for a one-unit increase in covariate \(x_j\), holding other covariates fixed. In general, a change in \(x_j\) can affect \(y_j\), outcomes at connected units, and—through further feedback—outcomes elsewhere. The coefficient \(\beta_x\) is the direct covariate term in the structural equation; it is not usually any one of those full derivatives.

The inverse matrix is often called a spatial multiplier. When the series converges, it has the expansion \((I-\rho W)^{-1}=I+\rho W+\rho^2W^2+\cdots\). The successive terms represent the initial response and additional rounds of propagation. The stability or invertibility conditions for a particular model must be checked; this expansion is not justified just by writing down the series.

Definition: For a covariate with impact matrix \(S=\partial y/\partial x'\), the direct impact is commonly summarized as \(\operatorname{tr}(S)/n\), the mean of the diagonal elements. The total impact is commonly summarized as the mean row sum of \(S\). The indirect impact is total impact minus direct impact. These are summaries of model-implied derivatives; report the convention used, and retain the individual matrix elements when unit-specific effects matter.

Worked Example: A Spatial-Lag Coefficient Is Not the Own-Unit Effect

Consider two hypothetical regions with \(W=\begin{pmatrix}0&1\\1&0\end{pmatrix}\), \(\rho=0.2\), and a covariate coefficient \(\beta_x=1\). Assume \(I-\rho W\) is invertible. For this matrix, the eigenvalues of \(W\) are \(1\) and \(-1\), so those of \(I-\rho W\) are \(0.8\) and \(1.2\), both nonzero. The impact matrix is

S=(I-0.2W)^{-1} =\frac{1}{1-0.2^2} \begin{pmatrix}1&0.2\\0.2&1\end{pmatrix} =\begin{pmatrix}1.041667&0.208333\\0.208333&1.041667\end{pmatrix}.

The own-unit derivative is \(1.041667\), not the structural coefficient \(1\). The cross-unit derivative is \(0.208333\). Both are per one-unit increase in the covariate, in the outcome’s units. Each row sums to \(1.25\), so the average total impact is \(1.25\); the average direct impact is \(1.041667\); and the average indirect impact is \(1.25-1.041667=0.208333\). Here the symmetry makes each unit’s row summary the same. These values describe the specified model, not evidence by themselves that changing the covariate causes the outcomes to change.

Spatial Durbin Models: More Than One Covariate Term

In a spatial Durbin model, neighboring covariates enter as well as neighboring outcomes. For one covariate, write \(y=\rho Wy+\beta_x x+\theta_x Wx+u\). Here \(\theta_x\) is the coefficient on the weighted covariate, while \(\beta_x\) is the coefficient on the unit’s own covariate. The derivative matrix is

S=(I-\rho W)^{-1}(\beta_x I+\theta_x W).

Consequently, neither \(\beta_x\) nor \(\theta_x\) alone is generally the direct, indirect, or total impact. The covariate terms and outcome feedback interact in the derivative matrix. The signs of the coefficients also need not determine the signs of every element or summary impact.

Worked Example: A Covariate-Lag Coefficient Is Not the Indirect Impact

Use the two-region weights matrix from the preceding example, and consider hypothetical parameters \(\rho=0.2\), \(\beta_x=1\), and \(\theta_x=0.5\). The invertibility check is unchanged: the eigenvalues of \(I-0.2W\) are \(0.8\) and \(1.2\). The covariate part of the derivative is \(\beta_x I+\theta_xW=\begin{pmatrix}1&0.5\\0.5&1\end{pmatrix}\). Multiplying gives

S=\frac{1}{0.96} \begin{pmatrix}1&0.2\\0.2&1\end{pmatrix} \begin{pmatrix}1&0.5\\0.5&1\end{pmatrix} =\frac{1}{0.96} \begin{pmatrix}1.1&0.7\\0.7&1.1\end{pmatrix} =\begin{pmatrix}1.145833&0.729167\\0.729167&1.145833\end{pmatrix}.

The average direct impact is \(1.145833\), and the average total impact is the row sum \(1.145833+0.729167=1.875\). Thus the average indirect impact is \(1.875-1.145833=0.729167\). The coefficient \(\theta_x=0.5\) is not the indirect impact: the latter includes feedback through \(\rho Wy\). As a check, increasing the covariate in both regions by one unit changes each outcome by \(1.875\) in this model. The common change has \(Wx=x\), so its response is \((\beta_x+\theta_x)/(1-\rho)=(1+0.5)/0.8=1.875\), matching the row sum.

When Is a Spillover Causal?

An impact matrix is a derivative implied by a model. Calling that derivative a causal effect requires more than calculating it. At minimum, the interpretation must define a meaningful intervention, specify the relevant connections and their timing, and justify why the covariate change is not confounded with unobserved determinants of outcomes. The disturbance structure, model specification, and treatment of the weights matrix must also support the proposed interpretation. When outcomes of connected units enter contemporaneously, their joint determination creates an identification problem: variation in a neighbor’s outcome is not automatically an exogenous treatment.

Peer-effect models make this distinction especially important. In a linear-in-means specification, an individual’s outcome may depend on peers’ outcomes as well as peers’ covariates and shared or correlated factors. A peer-outcome coefficient can reflect simultaneity or common influences unless the model has a defensible source of identifying variation and the needed assumptions. Evidence of association among connected individuals is not, on its own, evidence that one individual’s behavior caused another’s to change.

Worked Example: A Model-Implied Peer Response Needs Causal Assumptions

Suppose two hypothetical people are connected to one another, so \(G=\begin{pmatrix}0&1\\1&0\end{pmatrix}\), and their outcomes satisfy the structural equation \(y=0.25Gy+2x+u\). For the algebraic response calculation, assume \(I-0.25G\) is invertible; its eigenvalues are \(0.75\) and \(1.25\), so it is. If \(x_1\) increases by one unit and \(x_2\) stays fixed, the model-implied derivative is

\frac{\partial y}{\partial x_1} =(I-0.25G)^{-1} \begin{pmatrix}2\\0\end{pmatrix} =\frac{1}{1-0.25^2} \begin{pmatrix}1&0.25\\0.25&1\end{pmatrix} \begin{pmatrix}2\\0\end{pmatrix} =\begin{pmatrix}2.133333\\0.533333\end{pmatrix}.

Under this model, the first person’s outcome changes by about \(2.133\) outcome units per one-unit increase in \(x_1\), and the connected person’s outcome changes by about \(0.5333\) outcome units. The second response is a model-implied spillover. To interpret it causally, one would also need assumptions supporting the intervention on \(x_1\), the fixed and correctly specified connection structure, and identification of the peer-outcome coefficient apart from shared shocks and other confounding. Without those supports, the calculation is an implication of the fitted equation, not proof of a causal peer effect.

Common Interpretive Errors

Calling a structural coefficient a marginal effect. A coefficient is one component of a spatial model. Compute or otherwise characterize the derivative matrix before describing own-unit or cross-unit responses.
Calling the neighbor-outcome coefficient the spillover. The spillover is generally a derivative or a defined summary of derivatives, not simply \(\rho\). In models with spatially lagged covariates, the impact matrix also depends on those terms.
Equating association with causation. Spatial dependence, correlated shocks, reflection, and endogenous network formation can complicate identification. A plausible causal claim needs assumptions and an identification argument, not only a spatial coefficient.
Reporting an impact without its scale or summary convention. State the covariate change, outcome units, and whether the number is a direct, indirect, total, average, or unit-specific effect.
Treating statistical significance as a causal verdict. A small p-value for a coefficient does not establish that a causal spillover exists. Nor is inference for a nonlinear impact summary automatically identical to inference for a single coefficient; uncertainty for the requested impact must be assessed using an appropriate procedure.
Assuming one weights matrix is definitive. The weights matrix encodes a substantive choice about exposure or connection. Robustness across plausible alternatives is informative, but does not establish that any one matrix captures the true causal network.

Key takeaway: Keep the model coefficient, the derivative-based impact, and the causal interpretation distinct. State which impact is being summarized, calculate it from the full model, and justify the assumptions needed for a causal claim.

Check Your Understanding

For each question, distinguish what follows from a model calculation from what requires additional assumptions.

In a spatial-lag model, why does a covariate coefficient generally differ from the own-unit marginal effect?
For the two-region spatial Durbin example, identify the average direct, indirect, and total impacts. Which coefficient is not itself the indirect impact?
What does the \((i,j)\) entry of an impact matrix represent, and what are its units?
Name two assumptions or identification concerns that must be addressed before interpreting a peer-outcome coefficient as causal.
Why does finding a statistically significant spatial coefficient not by itself establish a causal spillover?