Spatial Durbin Models and Common-Factor Restrictions

Why Add Spatially Lagged Covariates?

A spatial-lag model allows outcomes to respond to neighboring outcomes. But in many settings, an outcome may also respond directly to neighboring units’ covariates: a region’s employment might depend on employment elsewhere and on nearby regions’ education or infrastructure. A spatial Durbin model includes both channels. Its coefficients therefore do not, by themselves, describe the full response to changing a covariate at one location; simultaneous feedback must also be accounted for.

The model also has a useful special case. Under a particular parameter restriction, the spatial lag and the covariate lags combine so that the model can be written as a spatial-error model. The restriction is an algebraic relationship among parameters, not a general claim that spatial Durbin and spatial-error models are interchangeable.

Specification and Assumptions

Definition: A spatial Durbin model is $$ y=\rho Wy+X\beta+WX\theta+\varepsilon. $$ Here $y$ is an $n\times1$ outcome vector, $X$ is an $n\times K$ matrix of covariates, $W$ is a known $n\times n$ spatial-weights matrix, $\rho$ is the spatial-lag parameter, and $\beta$ and $\theta$ are $K\times1$ coefficient vectors. The vector $\varepsilon$ is the disturbance. The term $Wy$ is the spatial lag of the outcome, and $WX$ contains spatial lags of the covariates.

For the interpretations and algebra below, treat $W$ as fixed and assume $I-\rho W$ is invertible. This condition permits a unique reduced form. Assume also that $E(\varepsilon\mid X)=0$ when interpreting the conditional mean. These assumptions do not require independent disturbances for the algebraic results, though additional assumptions would be needed for particular estimation and inference procedures. The weights convention—such as row normalization—matters for the intercept and should be stated in an application.

Solving the model for $y$ gives

y=(I-\rho W)^{-1}(X\beta+WX\theta+\varepsilon).

Thus the outcome reflects both the covariates’ direct contribution and the propagation of their contribution through spatial feedback. The inverse matrix is the spatial multiplier. Stability and invertibility are discussed in the earlier material on spatial multipliers.

Marginal Effects as a Matrix

For covariate $k$, let $\beta_k$ and $\theta_k$ denote the corresponding entries of $\beta$ and $\theta$, and let $x_k$ be the $k$th column of $X$. Holding the other covariates fixed, the conditional mean’s response to changes in $x_k$ is described by an $n\times n$ matrix:

S_k=\frac{\partial E(y\mid X)}{\partial x_k^\mathsf{T}} =(I-\rho W)^{-1}(\beta_k I+\theta_k W).

The entry $S_{k,ij}$ is the change in the conditional mean outcome at unit $i$ for a one-unit increase in covariate $k$ at unit $j$, holding the other covariates fixed. Diagonal entries are own-unit responses; off-diagonal entries are cross-unit responses. In an asymmetric weights structure, effects need not be symmetric: a change at $j$ can affect $i$ differently from the effect of a change at $i$ on $j$.

A commonly reported summary is the average direct effect, the mean of the diagonal entries of $S_k$. The average total effect is the mean of its row sums, equivalently the sum of all entries divided by $n$. The average indirect effect is average total minus average direct. These are summaries of the model’s marginal responses; interpreting them causally requires an appropriate structural and identification argument, not merely the presence of $W$.

Worked Example: Calculating Spatial Durbin Impacts

Consider two hypothetical regions, with region 1’s neighbor being region 2 and vice versa. Let $W=\begin{pmatrix}0&1\\1&0\end{pmatrix}$, let $\rho=0.2$, and suppose the coefficient for a covariate such as training expenditure is $\beta_k=2.5$ outcome units per expenditure unit, with spatial-lag coefficient $\theta_k=0.5$. The hypothetical covariate values are $x_k=(10,14)^\mathsf{T}$. Since $W^2=I$, the stability condition holds here because $I-0.2W$ is invertible (its determinant is $0.96$).

The impact matrix is

S_k=(I-0.2W)^{-1}(2.5I+0.5W) =\frac{(I+0.2W)(2.5I+0.5W)}{0.96} =\frac{2.6I+1.0W}{0.96}.

In the numerator, the coefficient on $I$ is $2.5+0.2(0.5)=2.6$, and the coefficient on $W$ is $0.5+0.2(2.5)=1.0$. Therefore,

S_k= \begin{pmatrix} 2.708&1.042\\ 1.042&2.708 \end{pmatrix}.

A one-unit increase in training expenditure in region 1 changes the conditional mean outcome by about $2.708$ units in region 1 and $1.042$ units in region 2. By symmetry, the average direct effect is $2.708$ outcome units per expenditure unit and the average indirect effect is $1.042$. The average total effect is the row sum, $2.708+1.042=3.750$ outcome units per expenditure unit. These values include feedback through $Wy$, not just the coefficients $2.5$ and $0.5$.

The Common-Factor Restriction

A spatial-error model has the form $y=X\beta+u$, where the disturbance follows $u=\lambda Wu+\varepsilon$. Equivalently, $(I-\lambda W)u=\varepsilon$. The common-factor restriction describes when the spatial Durbin equation can be rearranged into this form. For each covariate included in both $X\beta$ and $WX\theta$, the restriction is $\theta_k=-\rho\beta_k$; in vector form, $\theta=-\rho\beta$.

Theorem: If $\theta=-\rho\beta$, then the covariate terms satisfy $X\beta+WX\theta=(I-\rho W)X\beta$. Consequently, the spatial Durbin equation becomes $$ y=X\beta+(I-\rho W)^{-1}\varepsilon. $$ The composite disturbance $u=(I-\rho W)^{-1}\varepsilon$ satisfies $u=\rho Wu+\varepsilon$. Thus the model has a spatial-error representation with $\lambda=\rho$, subject to the intercept convention described below.

The restriction is called a common-factor restriction because the spatial-lag operator factors out of the covariate terms. To verify the result, substitute $\theta=-\rho\beta$ into the original equation:

\begin{aligned} y&=\rho Wy+X\beta-\rho WX\beta+\varepsilon\\ &=\rho Wy+(I-\rho W)X\beta+\varepsilon. \end{aligned}

Subtract $\rho Wy$ from both sides and multiply by $(I-\rho W)^{-1}$. This yields $y=X\beta+(I-\rho W)^{-1}\varepsilon$. If $u=(I-\rho W)^{-1}\varepsilon$, then $(I-\rho W)u=\varepsilon$, or $u=\rho Wu+\varepsilon$, which is precisely the spatial-error form.

This equivalence concerns the model’s form and parameters; it does not say that every spatial Durbin model is a spatial-error model. If the restriction fails, the covariate-lag terms do not factor into the error representation by this argument. Nor does the equivalence itself establish which specification is appropriate for a particular application.

Worked Example: Applying the Restriction With an Intercept

Suppose two hypothetical regions have $W=\begin{pmatrix}0&1\\1&0\end{pmatrix}$, so $W\mathbf{1}=\mathbf{1}$, and the model includes an unlagged intercept $\alpha\mathbf{1}$. For one covariate, set $\rho=0.3$, $\beta=1.8$, and $\theta=-0.54$. The restriction holds because $-\rho\beta=-0.3(1.8)=-0.54$. Let the hypothetical covariate vector be $x=(10,14)^\mathsf{T}$, and let $\alpha=0.6$.

The covariate contribution on the right-hand side is

1.8x-0.54Wx = \begin{pmatrix}18\\25.2\end{pmatrix} - 0.54\begin{pmatrix}14\\10\end{pmatrix} = \begin{pmatrix}10.44\\19.8\end{pmatrix}.

The factored expression gives the same result:

(I-0.3W)1.8x = \begin{pmatrix}18\\25.2\end{pmatrix} - 0.3\begin{pmatrix}25.2\\18\end{pmatrix} = \begin{pmatrix}10.44\\19.8\end{pmatrix}.

The intercept needs separate attention. Since $W\mathbf{1}=\mathbf{1}$, its transformed contribution is $\alpha(I-\rho W)^{-1}\mathbf{1}=\alpha(1-\rho)^{-1}\mathbf{1}$. Here that is $0.6/0.7=0.8571$ per region, not $0.6$. So, with a row-standardized $W$ and an unlagged intercept in the Durbin model, the equivalent error representation has the same slope $\beta$ and a transformed intercept $\alpha/(1-\rho)$. The slope restriction alone does not imply that the intercept coefficient is numerically unchanged.

When the Restriction Holds for Some Covariates

The common-factor relationship is covariate-specific. It may hold for some regressors but not others. A failed restriction for even one included covariate means that the entire covariate part does not reduce to the stated spatial-error representation with the same coefficient vector. The impact matrix remains useful regardless: it describes the effects under the Durbin model without imposing the restriction.

Worked Example: Checking Two Covariates

In a hypothetical model, suppose $\rho=0.25$, with coefficient vectors $\beta=(1.2,-0.8)^\mathsf{T}$ and $\theta=(-0.30,0.15)^\mathsf{T}$ for two covariates. The required spatial-lag coefficients would be

-\rho\beta =-0.25 \begin{pmatrix}1.2\\-0.8\end{pmatrix} = \begin{pmatrix}-0.30\\0.20\end{pmatrix}.

The first covariate satisfies the restriction, but the second does not: its required coefficient is $0.20$, not $0.15$. To see the implication, use the two-region swap matrix from the earlier examples. For the first covariate, $\theta_1=-\rho\beta_1$, so

S_1=(I-\rho W)^{-1}\beta_1(I-\rho W)=1.2I.

Its own-unit marginal effect is $1.2$ and its cross-unit effect is zero under this exact restriction. For the second covariate, direct multiplication gives

S_2 =\frac{(I+0.25W)(-0.8I+0.15W)}{0.9375} =\frac{-0.7625I-0.05W}{0.9375} = \begin{pmatrix} -0.8133&-0.05333\\ -0.05333&-0.8133 \end{pmatrix}.

The second covariate therefore has nonzero cross-unit impacts in this example and does not meet the common-factor restriction. This coefficient comparison is an algebraic check, not a statistical test of the restriction.

Common Pitfalls and Why the Distinction Matters

Reading $\beta_k$ as the full marginal effect. In the Durbin model, the response matrix is $(I-\rho W)^{-1}(\beta_k I+\theta_kW)$. Its diagonal and off-diagonal entries generally differ from the raw coefficients.
Calling any model with $Wy$ a spatial-error model. A spatial-error representation requires the common-factor relationship between $\theta$, $\rho$, and $\beta$; the spatial Durbin model allows parameters that do not satisfy it.
Ignoring the intercept transformation. When $W\mathbf{1}=\mathbf{1}$ and the intercept is unlagged, the equivalent intercept is $\alpha/(1-\rho)$. It is not generally the original $\alpha$.
Interpreting a modeled spillover as a causal effect automatically. The impact matrix gives model-implied marginal responses. A causal interpretation also requires assumptions that justify treating the covariates and spatial structure as part of a causal design.

The distinction matters because the spatial Durbin model separates spatial feedback in outcomes from spatially lagged covariates, while the spatial-error representation places the spatial dependence in the disturbance. The common-factor restriction identifies a precise overlap between those representations. Stating the restriction and checking the intercept convention makes clear what that overlap does—and does not—imply.

Check Your Understanding

Use the model definitions and results above to answer the following questions.

For covariate $k$, write the spatial Durbin marginal-effect matrix and explain what its $i,j$ entry represents.
In a model with $\rho=0.4$ and $\beta_k=1.5$, what value of $\theta_k$ satisfies the common-factor restriction?
Why does the common-factor restriction make the covariate terms factor as $(I-\rho W)X\beta$?
If $W\mathbf{1}=\mathbf{1}$, what happens to an unlagged intercept $\alpha$ in the equivalent spatial-error representation?
Does a coefficient mismatch with the common-factor restriction establish that one model is empirically preferable? Explain.