CSPATIALREG Procedure

Impact Estimation

For models in PROC CSPATIALREG that include endogenous interactions or exogenous interactions, the interpretation of marginal effects is not as straightforward as it is in the classical linear regression model. For example, consider the following SDM model:

bold y equals rho bold upper W bold y plus bold upper X bold-italic beta plus bold upper W bold upper X bold-italic theta plus bold-italic epsilon

where bold-italic epsilon tilde upper N left-parenthesis bold 0 comma sigma squared bold upper I right-parenthesis.

For an explanatory variable x Subscript p in the model,

StartFraction partial-differential y Subscript i Baseline Over partial-differential x Subscript j p Baseline EndFraction equals left-bracket bold upper A Superscript negative 1 Baseline left-parenthesis beta Subscript p Baseline bold upper I Subscript n Baseline plus theta Subscript p Baseline bold upper W right-parenthesis right-bracket Subscript i j

where bold upper A equals bold upper I minus rho bold upper W and left-bracket bold upper B right-bracket Subscript i j denotes the left-parenthesis i comma j right-parenthesisth element of the matrix bold upper B for i comma j equals 1 comma 2 comma ellipsis comma n.

According to StartFraction partial-differential y Subscript i Baseline Over partial-differential x Subscript j p Baseline EndFraction, you can see that the impact of changes in x Subscript p at a location j on the dependent variable differ from one location to another. To quantify the marginal effects of an explanatory variable x Subscript p in the model, LeSage and Pace (2009) propose the average direct impacts (upper M overbar left-parenthesis p right-parenthesis Subscript direct), the average total impacts (upper M overbar left-parenthesis p right-parenthesis Subscript total), and the average indirect impacts (upper M overbar left-parenthesis p right-parenthesis Subscript indirect) as follows:

upper M overbar left-parenthesis p right-parenthesis Subscript direct Baseline equals StartFraction 1 Over n EndFraction sigma-summation Underscript i equals 1 Overscript n Endscripts left-bracket bold upper A Superscript negative 1 Baseline left-parenthesis beta Subscript p Baseline bold upper I Subscript n Baseline plus theta Subscript p Baseline bold upper W right-parenthesis right-bracket Subscript i i
upper M overbar left-parenthesis p right-parenthesis Subscript total Baseline equals StartFraction 1 Over n EndFraction sigma-summation Underscript i equals 1 Overscript n Endscripts sigma-summation Underscript j equals 1 Overscript n Endscripts left-bracket bold upper A Superscript negative 1 Baseline left-parenthesis beta Subscript p Baseline bold upper I Subscript n Baseline plus theta Subscript p Baseline bold upper W right-parenthesis right-bracket Subscript i j
upper M overbar left-parenthesis p right-parenthesis Subscript indirect Baseline equals upper M overbar left-parenthesis p right-parenthesis Subscript total Baseline minus upper M overbar left-parenthesis p right-parenthesis Subscript direct

In addition, LeSage and Pace (2009) describe how to use Monte Carlo simulation to quantify uncertainty that is associated with the three impact estimates. First, the empirical distribution for parameters (the regression coefficients and, if applicable, the spatial dependence parameter rho) is assumed to be a multivariate normal distribution where the mean vector contains the point estimates and the covariance matrix is estimated based on the maximum likelihood estimation. Second, M samples are drawn from the empirical distribution and for each sample being drawn, upper M overbar left-parenthesis p right-parenthesis Subscript direct, upper M overbar left-parenthesis p right-parenthesis Subscript total, and upper M overbar left-parenthesis p right-parenthesis Subscript indirect are computed. Based on these random samples, the sample mean and standard deviation for each of three impact estimates can be computed.

When upper M overbar left-parenthesis p right-parenthesis Subscript direct, upper M overbar left-parenthesis p right-parenthesis Subscript total, and upper M overbar left-parenthesis p right-parenthesis Subscript indirect are computed, the inverse of bold upper A equals bold upper I Subscript n Baseline minus rho bold upper W needs to be computed for each random sample drawn. Since the exact inversion of bold upper A can be computationally intensive, approximation can be helpful when it is applicable. When the spectral radius of rho bold upper W is less than 1, bold upper A Superscript negative 1 can be approximated by the partial sums of Neumann series as

left-parenthesis bold upper I Subscript n Baseline minus rho bold upper W right-parenthesis Superscript negative 1 Baseline equals sigma-summation Underscript o equals 0 Overscript plus normal infinity Endscripts left-parenthesis rho bold upper W right-parenthesis Superscript o Baseline almost-equals bold upper I Subscript n Baseline plus rho bold upper W plus midline-horizontal-ellipsis plus rho Superscript upper O Baseline bold upper W Superscript upper O

The IMPACT option in the PROC CSPATIALREG statement computes upper M overbar left-parenthesis p right-parenthesis Subscript direct, upper M overbar left-parenthesis p right-parenthesis Subscript total, and upper M overbar left-parenthesis p right-parenthesis Subscript indirect by using the Monte Carlo simulation that is proposed by LeSage and Pace (2009). In particular, you specify NMC=M in the IMPACT option to request that M samples be drawn from the empirical distribution of the relevant model parameters. You use the ORDER=O suboption of the IMPACT option to specify the highest order of Neumann series to approximate bold upper A Superscript negative 1. You use the SEED= suboption of the IMPACT option to specify an integer seed for a random number generator to replicate your results.

Impact estimation is not supported for models that include any of the following: a classification variable, interactions between regressions, a BOUNDS statement, or a RESTRICT statement. In addition, Monte Carlo simulation is not needed to compute summary statistics for upper M overbar left-parenthesis p right-parenthesis Subscript direct, upper M overbar left-parenthesis p right-parenthesis Subscript total, and upper M overbar left-parenthesis p right-parenthesis Subscript indirect for linear, SLX, SEM, SDEM, SMA, and SDMA models because a closed-form solution is available (upper M overbar left-parenthesis p right-parenthesis Subscript direct Baseline equals beta Subscript p, upper M overbar left-parenthesis p right-parenthesis Subscript total Baseline equals beta Subscript p Baseline plus theta Subscript p, and upper M overbar left-parenthesis p right-parenthesis Subscript indirect Baseline equals theta Subscript p). In these models, any NMC=, ORDER=, and SEED= suboptions that you specify in the IMPACT option are ignored. When you specify the IMPACT option for SAR, SDM, SAC, SDAC, SARMA, and SDARMA models, the spatial weights matrix bold upper W must be row-standardized. Moreover, the restriction negative 1 less-than rho less-than 1 is imposed for SAR and SDM models and the restrictions negative 1 less-than rho less-than 1 and negative 1 less-than lamda less-than 1 are imposed for SAC, SDAC, SARMA, and SDARMA models. With the restriction negative 1 less-than rho less-than 1, the average total impacts upper M overbar left-parenthesis p right-parenthesis Subscript total can be simplified as

upper M overbar left-parenthesis p right-parenthesis Subscript total Baseline equals StartFraction 1 Over 1 minus rho EndFraction left-parenthesis beta Subscript p Baseline plus theta Subscript p Baseline right-parenthesis
Last updated: July 09, 2026