|
POSITION STATEMENT
One area that I have done some work in and one where the explicit consideration of the spatial dimension of externalities has led to new insights is in the provision of public goods. When the geography is defined by political boundaries, like school districts, cities, states, and nations, many spending and taxing decisions by one political entity cause "spillovers" that impose costs or benefits in other entities. Failure to enforce environmental controls, the provision of goods with public characteristics (environmental quality, parks, golf courses, traffic flow), and the setting of tax rates are some of the obvious examples. Within this context, we wish to study the type of behavior exhibited by the political entities vis-�-vis one another; i.e., the collective action. Various economic models (Nash-type, collusive, cooperative, Stackelberg) generate simultaneous empirical models that are essentially spatial econometric models. Distinguishing between the models is quite interesting in a policy context (e.g., "race to the bottom" or Lindahl solutions).
Based on my experiences with these problems, I have a couple of general observations. First, the dependent variable is often a limited (participate or not in the provision), suggesting the need for estimators (and software) for limited dependent variable models with spatial dependence. Second, these problems frequently give insights into the nature of the dependence and, therefore, how to construct the spatial weight matrix. For example, determining the spillins of sulfur emissions to one European country from other nations requires calculation of their location relative to prevailing winds as well as distance. In the context of local communities, modeling the provision of (say) parks and recreation requires some consideration of accessibility (perhaps travel time) rather than just distance. In "traditional" spatial econometric models, the weights are formulated with geographic concepts like contiguity, relative distance, and nearness, and then calculated using GIS tools. A common example would be the use of pointdistance in Arc/Info to generate a distance matrix that could be used to make various spatial weight matrices. The examples above, however, suggest that some weight matrices may be based on the geographic data or features (points, lines, polygons) and some of the attributes of that data. Hence, construction of the matrices may require more interaction with a GIS than the standard problems.
More specifically, I have found that in estimating such models, relative to "regular" data analysis and econometric estimation (using a package like SAS or STATA, for example), the integration of GIS technologies (spatial data analysis) and spatial econometric estimation is somewhat lacking in a couple ways. First, in an ordinary problem, missing values can be handled almost seamlessly. In spatial problems, however, the simple introduction of a new attribute with missing values results in quite a bit of work to re-create just the spatial weight matrix. Additionally, there may be conceptual problems with missing values as they generate a "hole" in the space under consideration. A similar situation exits with respect to sampling. Regular econometric packages make sampling quite painless. Sampling in a spatial context, in my experience, is a different matter. Not only do I often not know how to sample (theoretically) but even when I do know, it seems difficult to implement the plan.
|