There is another type of problem which occurs when spatial weights are based on
a distance criterion, such that two units, i and j, are defined as neighbours if the
distance between them (or, in the case of units of area units, the distance between
their centroids) is less than a given critical value. If there is a high degree of
heterogeneity in the spatial distribution of points or the areas of regions, there may
not be a satisfactory critical distance. In those instances, a ‘small’ distance will tend
to yield numerous islands (or, unconnected observations). A distance chosen to
ensure that each unit has at least one neighbour could also result in an unacceptably
large number of neighbours for the smaller units. A common solution to this
problem is to constrain the neighbour structure to the k-nearest neighbours, thereby precluding islands and forcing each unit to have the same number of neighbours
(Anselin 2002).
A third issue that could arise occurs when weights are based on economic distance
or another general metric, derived perhaps from a social network structure. Care
must be taken to ensure that the resulting weights are meaningful, finite and nonnegative.
We also need to account for the ‘zero-distance’ problem. This problem
occurs when a distance measure, such as dij ¼ jzi . zjj, becomes zero as a result of
rounding or because two observations have identical socio-economic profiles. As a
result, inverse distance weights such as wij ¼ 1/dij, are undefined.