Economic Restructuring, Moehr
24
Since the spatial error model did not significantly change any of the model
parameters, it might be tempting to dismiss it as an unnecessary complication. The
overall improved model fit, however, argues in favor of including spatial heterogeneity in
an analysis of the socio-economic system (R
2
= .428 for net migration and R
2
= .335 for
poverty rates). Most importantly the spatial error model removes the high degree of
spatial autocorrelation among the residuals. The Moran’s I values within Table 2
demonstrate that in the non-spatial models the residuals tend to cluster together. On the
other hand, residuals of the spatial error model exhibit negative spatial autocorrelation
which means that similar residual values tend to be dispersed. Unfortunately, the
negative Moran’s I values are statistically significant (p < .01 for both net migration and
poverty rates). The goal of spatial error models is to create spatially random residuals,
but in this case the model estimation seems to have over corrected for the spatial
clustering. The degree of spatial patterning, however, is significantly less dramatic as can
be seen in Figure 4 for net migration rates and Figure 5 for change in poverty rates.
The spatial error model is the preferred model. Empirically, it creates the best
model fit and least amount of spatial autocorrelation among residuals, and, theoretically,
it incorporates both the reciprocal effects of migration and poverty and the effect of
geographically distributed economic change.
To crystallize my arguments about economic restructuring and geography, it may
be more useful to apply the statistical model to a few examples and illustrate how specific
places experienced the new economy. I choose Cleveland, Ohio and Pittsburgh,
Pennsylvania to represent places that were historically tied to old economy sectors such
as manufacturing. As can be seen in Table 3, both of the central city counties for each
Convention