Why Spatial Weights Sensitivity Matters
A spatial-lag model depends on how the analyst represents which units influence which others. A border-contiguity matrix, a distance-based matrix, and a nearest-neighbour matrix can encode different defensible ideas about interaction. Since the spatial lag \(Wy\) is built from that choice, changing the matrix can change the estimated spatial-lag coefficient, model diagnostics, and estimated impacts. A result that appears stable under one arbitrary matrix is not, by itself, evidence that the spatial relationship is robust.
A sensitivity analysis compares a prespecified set of plausible matrices while holding the outcome, sample, covariates, and estimation method fixed. It also separates a meaningful change in relative connections from a harmless change of scale. This distinction matters because \(\rho\) is measured relative to the scale of \(W\), whereas the product \(\rho W\) governs the spatial feedback in the model.
Model and Conditions for a Meaningful Comparison
For a spatial-lag model, let \(y\) be an \(n\)-vector of outcomes, \(X\) the covariate matrix, \(W\) a specified \(n\times n\) weights matrix, and \(u\) an \(n\)-vector of disturbances. The model is
Here \(\rho\) is the spatial-lag coefficient and \(\beta\) is the vector of covariate coefficients. The reduced form requires \(I-\rho W\) to be invertible. The model also requires an appropriate specification for the conditional mean and disturbances, and the estimation method's assumptions must be credible. For sensitivity comparisons, use the same observations, outcome, covariates, and estimator in each fit; otherwise, a difference cannot be attributed to the weights matrix alone.
A candidate \(W\) should have a clear substantive rationale: for example, shared borders, a prespecified distance threshold, or a fixed number of nearest neighbours. Document how ties, isolated units, and diagonal elements are handled. In many applications the diagonal is set to zero, so a unit does not count itself as a neighbour. Do not choose a matrix simply because it produces a significant \(\hat{\rho}\) or a preferred substantive conclusion.
The numerical value of \(\hat{\rho}\) is not generally comparable across matrices with different scales. Its meaning depends on the units of \(W\). Compare it alongside the matrix definition and normalization, and give particular attention to fitted spatial dependence, diagnostics, and impacts. If matrices encode different relative connections, even impacts on the same outcome scale may differ substantively.
Worked Example: A Change of Scale Is Not a Change of Structure
Suppose \(W\) is row-standardized, and a fitted model has \(\hat{\rho}=0.30\). Now define \(W^*=2W\). This multiplies every weight by the same constant and preserves every relative connection. What parameter in the model using \(W^*\) gives exactly the same spatial feedback?
The spatial feedback term in the original model is \(\hat{\rho}W=0.30W\). For the new matrix, choose \(\hat{\rho}^*=0.15\):
Thus \(I-\hat{\rho}^*W^*=I-\hat{\rho}W\), so the reduced form, fitted values, and impacts are identical. Under an exactly equivalent reparameterization, the likelihood and diagnostics based on the fitted model are also unchanged. The coefficient and its standard error are expressed on a different scale: \(\hat{\rho}^*=\hat{\rho}/2=0.15\), and its standard error is half the original one. A Wald statistic for testing the equivalent zero restriction is unchanged because both estimate and standard error are halved.
The admissible parameter interval must also change with the scale. If \(W\) is row-stochastic, its largest eigenvalue in absolute value is 1, so a common stability restriction is \(|\rho|<1\). For \(W^*=2W\), the corresponding restriction is \(|\rho^*|<0.5\). Comparing the numerical values 0.30 and 0.15 as if they represented different strengths of feedback would therefore be misleading.
From Weights to Impact Measures
In a spatial-lag model, a change in one unit's covariate can affect its own outcome and, through feedback, outcomes at other units. For covariate \(k\), the matrix of marginal effects is
The average direct effect is the average of the diagonal elements of \(S_k(W)\). The average total effect is the average row sum, and the average indirect effect is the total effect minus the direct effect. The decomposition depends on \(W\) through the spatial multiplier, not just on \(\beta_k\). Consequently, a stable regression coefficient does not guarantee stable direct or indirect effects.
A useful special case clarifies the role of row standardization. If \(W\mathbf{1}=\mathbf{1}\), where \(\mathbf{1}\) is a vector of ones, then \((I-\rho W)^{-1}\mathbf{1}=\mathbf{1}/(1-\rho)\), provided the inverse exists. For a covariate that enters every unit's equation with coefficient \(\beta_k\), the average total effect is therefore \(\beta_k/(1-\rho)\). The average direct effect still depends on the diagonal of the spatial multiplier, which can vary across alternative network structures.
Worked Example: Comparing Impacts for Two Three-Unit Networks
Consider two row-standardized networks with three units. The first is a chain: unit 1 and unit 3 connect to unit 2, and unit 2 connects equally to units 1 and 3. The second is a complete network in which every unit connects equally to the other two. Suppose \(\rho=0.30\) and a covariate coefficient is \(\beta_k=1\). These values are hypothetical and are used to isolate the effect of network structure.
The chain matrix is
For this chain, the average diagonal of \((I-0.30W_C)^{-1}\) is
So the average direct effect is \(1.0659\). Since the matrix is row-standardized, the average total effect is
The average indirect effect is calculated from the unrounded quantities:
For the complete network, each row assigns weight \(0.5\) to each of the other two units. Its multiplier has average diagonal
The complete-network average direct effect is \(1.0559\); its average total effect remains \(1.428571\ldots\), and its average indirect effect is
The same coefficient and \(\rho\) produce the same total effect here, but different direct and indirect effects. The topology changes how feedback is distributed between own-unit and other-unit responses. In an empirical analysis, \(\rho\) is usually estimated separately for each matrix, so the total effect may also change. Report effects calculated from unrounded quantities; subtracting rounded values can give an incorrect final digit.
Compare Estimates and Diagnostics Systematically
A comparison should include more than a table of \(\hat{\rho}\). For each candidate matrix, record its substantive definition and normalization, the estimated spatial-lag coefficient and uncertainty, model fit under a common likelihood specification, relevant residual diagnostics, and the direct, indirect, and total effects for key covariates. A compact display makes patterns and departures visible.
Diagnostics are also conditional on a spatial structure. A residual spatial-dependence test evaluated with one matrix need not have the same result when evaluated with another. It is useful to state the diagnostic weights explicitly and, when substantively justified, assess residual dependence under more than one plausible diagnostic matrix. Do not treat a non-significant diagnostic as proof that the model is correct; it indicates only that the test did not detect the specified form of dependence at its chosen level and under its assumptions.
Likelihoods or information criteria can be compared only when the fits use the same outcome observations and a comparable likelihood and model specification. A better fit is evidence about relative statistical fit, not an automatic reason to prefer a matrix with no substantive justification. Inference should reflect the sample structure and estimator; especially with few spatial units or networks, routine large-sample \(p\)-values may be unreliable.
Worked Example: Reading a Hypothetical Sensitivity Table
Suppose an analyst fits the same spatial-lag specification to the same hypothetical sample using three defensible matrices. The following numbers are illustrative model output; the residual test in each column uses that column's matrix.
| Weights matrix | \(\hat{\rho}\) (SE) | Log likelihood | Residual Moran's I (p) |
|---|---|---|---|
| Row-standardized contiguity | 0.28 (0.11) | -41.2 | 0.03 (0.62) |
| Row-standardized distance band | 0.18 (0.12) | -40.8 | -0.04 (0.48) |
| Row-standardized 4-nearest neighbours | 0.25 (0.10) | -41.0 | 0.01 (0.81) |
The point estimates range from 0.18 to 0.28, so the estimated magnitude depends on how neighbours are defined. Their standard errors are substantial relative to the estimates, and this illustrative table alone does not establish a precise difference across matrices. Each reported residual test fails to reject its tested null at conventional levels, but the tests use different spatial structures; this is not proof that residual dependence is absent under every plausible structure.
The distance-band model has the largest log likelihood in this table, but the difference is modest and should not override the substantive rationale for the matrices. A useful next comparison would calculate the covariate impacts under all three fits. If the estimated \(\rho\) values differ but the impacts remain similar, the substantive conclusions may be comparatively robust. If a key indirect effect changes sign or magnitude materially, the conclusion is sensitive and should be reported as such. These data do not supply impact estimates, so none should be inferred from this table alone.
A Practical Sensitivity Workflow
A disciplined comparison makes the source of each change traceable. Keep a record of the matrices themselves or their construction rules, including any row standardization, symmetrization, or treatment of isolated units. The workflow below separates scale checks from substantive comparisons.
Choose matrices that represent plausible interaction mechanisms before examining which one gives the most favourable result.
Use the same units, sample, outcome, covariates, estimator, and inferential convention in each fit. Record normalization and any changes to the graph.
Check whether one matrix is a scalar multiple of another. If so, transform \(\rho\) accordingly; otherwise compare the matrices as substantively different structures.
Assess uncertainty and fit alongside residual diagnostics, then calculate direct, indirect, and total effects for each fitted model.
State what is stable, what changes, and which choices of \(W\) produce material differences. Do not conceal a defensible alternative because it weakens a preferred result.
Common Pitfall: Treating Every Change in \(\hat{\rho}\) as Substantive
A common mistake is to say that spatial dependence has weakened because \(\hat{\rho}\) falls after changing from a binary matrix to a row-standardized matrix. That conclusion may be wrong: row standardization can alter the scale, and, when units have different numbers of neighbours, it also changes relative weights. In the special case of a uniform scalar rescaling, the equivalent \(\rho\) changes inversely while \(\rho W\) and all model implications stay the same. When relative weights change, the models are genuinely different and their impacts and diagnostics require comparison.
The correction is to inspect the matrices and compare the implied spatial feedback, fitted results, and impact measures, not the bare coefficient alone. State clearly which conclusions survive plausible changes in the definition of neighbours. Sensitivity is not a defect to be hidden; it identifies which findings depend on a modelling choice.
Check Your Understanding
Use these questions to test whether you can distinguish normalization effects from structural sensitivity.
- If \(W^*=3W\), what value of \(\rho^*\) preserves the spatial feedback term for a model with \(\rho=0.24\)?
- Why can two row-standardized matrices yield different average direct effects even when \(\rho\) and \(\beta_k\) are held fixed?
- What must be held constant across model fits for changes in \(\hat{\rho}\) to be attributed to the weights matrix?
- Why does a non-significant residual spatial-dependence test not establish that the model is correctly specified?
- In a sensitivity table, why should impacts be computed from unrounded estimates rather than by subtracting displayed rounded values?