Constructing Spatial Weights Matrices

Why the Weights Matrix Is a Modeling Choice

In spatial and network econometrics, a weights matrix represents which units are connected and how strongly one unit is linked to another. The units might be regions, firms, schools, or people. A matrix is therefore not merely a convenient data-processing object: its entries encode a substantive assumption about the interaction structure. Two matrices built from the same locations can represent different assumptions if one uses shared borders and the other uses distance.

This tutorial focuses on constructing and interpreting that matrix. It does not determine which interaction structure is correct for a particular application. That decision requires a defensible account of how units could interact, together with a transparent description of the rule used to represent those interactions.

The row-receiver convention is common, but some software or authors use the transpose convention. Always check the indexing definition before interpreting a matrix. The unit ordering must also be identical in every row, column, and data vector to which \(W\) is applied.

Contiguity, Distance, and Network Rules

For areal units such as counties or grid cells, contiguity weights define neighbors using shared boundaries. Under rook contiguity, two regions are neighbors only if they share a boundary segment of positive length. Under queen contiguity, sharing either a boundary segment or a vertex is enough. These rules differ for regions that touch only at a corner. A binary contiguity matrix records a 1 for a pair classified as neighbors and a 0 otherwise.

A distance-based rule uses distances between locations, such as region centroids or known facility coordinates. One option sets a link when the distance is below a specified cutoff. Another assigns a positive weight that declines with distance. For example, an inverse-distance rule with exponent \(\rho>0\) can be written as \(d_{ij}^{-\rho}\) for distinct units within a chosen distance range. The cutoff and the definition of distance are part of the rule, not details that can be omitted.

For networks, an observed or hypothesized link supplies the basic connection. Examples include a directed referral, a road connection, or a reported communication tie. The resulting adjacency matrix may be symmetric or asymmetric. Symmetry is appropriate only when the link relation is mutual under the chosen definition; a directed referral from \(j\) to \(i\) does not automatically imply one from \(i\) to \(j\).

With a binary matrix, one linked unit contributes weight 1 regardless of how many other links the receiving unit has. The total weight in row \(i\) is its number of linked sources. With row standardization, each nonzero row instead has total weight 1, so the linked sources collectively receive a normalized weight allocation. If the initial weights are binary, a unit with \(m_i\) neighbors gives each one weight \(1/m_i\). If the initial weights are unequal, standardization preserves their relative proportions within the row.

Choosing and Calculating Distance Weights

Distance rules require more specification than the phrase “nearby units.” The locations used to calculate distance must be identified, as must the distance metric, the units of measurement, and whether a cutoff applies. A centroid-based distance, for instance, is not the same as the minimum distance between two boundaries. A cutoff chosen in kilometers cannot be applied directly to coordinates measured in degrees without a suitable conversion or distance calculation.

A distance-threshold rule can be written \(a_{ij}=1\) when \(0<d_{ij}\leq c\), and \(a_{ij}=0\) otherwise, where \(d_{ij}\) is the distance between distinct units and \(c\) is the specified cutoff. An inverse-distance rule could instead use \(a_{ij}=d_{ij}^{-\rho}\) for pairs within the cutoff and zero outside it. Here \(a_{ij}\) denotes the initial weight, before any row standardization. A positive exponent \(\rho\) makes the weight decline as distance grows. Units at identical coordinates require an explicit convention, since inverse distance is undefined at distance zero.

Network Weights and Direction

For a network, state what counts as a link and whether links are directed. If the relation is directed, \(w_{ij}\) and \(w_{ji}\) can differ. In the row-receiver convention used here, \(w_{ij}>0\) means that \(j\) is included as a source linked to receiver \(i\). For an undirected network, the binary adjacency matrix is symmetric. Whether weights represent link presence alone, link strength, frequency, or some other quantity must be specified separately.

Islands, Self-Links, and Normalization

An island is a unit with no links under the chosen rule, so its initial weights row sums to zero. This can arise in contiguity data when a region does not touch another unit, or in a threshold-distance construction when all other units are beyond the cutoff. It may also occur in a network with a unit that has no qualifying sources. Row standardization cannot turn a zero row into a row summing to one: division by zero is undefined.

There are several defensible ways to address islands, but they answer different substantive questions. One can retain the zero row and recognize that the unit has no linked sources under the rule. One can revise the rule, for example by increasing a distance cutoff or using a nearest-neighbor construction, if the original rule was not intended to leave units isolated. One can also exclude the unit from a particular analysis when exclusion is substantively and statistically justified. Each choice should be reported; silently inserting a link changes the assumed structure.

A self-link is a positive diagonal entry \(w_{ii}\). In many spatial weights constructions it is set to zero because the matrix is intended to describe links to other units, not a unit's relation to itself. A positive diagonal may be warranted in a specific network or application, but then it represents an explicit own-unit contribution. It should not be introduced merely to avoid an island: a self-link does not connect an island to another unit.

Row standardization is also not a neutral rescaling in every respect. It makes each non-island row sum to one, but it changes the relative scale of rows with different numbers or strengths of links. With binary weights, it gives each receiver equal total weight across its linked sources, regardless of its degree. The individual linked weights then depend on how many sources that receiver has. This is a different assumption from binary weights, which give every link the same individual weight and allow total row weight to vary.

A Short Justification: Why Row Standardization Gives Weighted Averages

Let \(x_j\) be a measured value for source unit \(j\), and suppose row \(i\) has nonnegative initial weights \(a_{ij}\) with positive row sum \(s_i=\sum_j a_{ij}\). Define the row-standardized weights as \(w_{ij}=a_{ij}/s_i\). Their sum is:

Consequently, \(\sum_j w_{ij}x_j=\sum_j a_{ij}x_j/s_i\) is a normalized weighted average of the linked source values, provided the weights are nonnegative. This explains the interpretation of a row-standardized row. For an island \(s_i=0\), the division used in this argument is not defined, which is why standard row standardization leaves the island unresolved rather than automatically assigning it a value.

Common Mistakes and Practical Checks

A weights matrix should be reported in enough detail for another analyst to reconstruct it. At minimum, describe the units and their ordering, the neighbor rule, the distance definition and units if relevant, the treatment of direction, the initial weights, the diagonal convention, the standardization rule, and the handling of islands.

• Assuming row standardization makes matrices equivalent. It does not. It preserves some within-row relative weights while changing row totals, and it cannot make different underlying link rules represent the same structure.
• Calling a cutoff matrix “distance weights” without a cutoff or metric. State how distance was calculated and the cutoff value with units. A binary threshold rule and inverse-distance weighting are not interchangeable.
• Standardizing an island by dividing by zero. A zero row has no positive sum to divide by. Keep it as zero or adopt and explain a substantively justified alternative.
• Adding diagonal ones to solve isolation. This creates self-links, not links to other units. Use a diagonal entry only when an own-unit link is part of the intended definition.
• Ignoring matrix orientation. Under the convention here, row \(i\) receives weights from column units \(j\). A transposed convention changes the interpretation for directed links.
• Treating a chosen matrix as proof of actual interaction. The matrix encodes a modeling assumption about possible interaction. A border, short distance, or network tie does not by itself establish that one unit causally affects another.