underlying this variable is that the more populated parishes are likely to be more
concerned about reducing water pollution. Hence, population density is expected to have
a negative sign1.
The error components, uit , can take different structures. The specification of error
components can depend solely on the cross section to which the observation belongs or
on both the cross section and time series. If the specification depends on the cross
section, then we have ut = vi + εit ; and if the specification is assumed to be dependent on
both cross section and time series, then the error components follow
uit = vi + et + εit. The term vi is intended to capture the heterogeneity across individual
parishes and the term et is to represent the heterogeneity over time. Furthermore, vi and
et can either be random or nonrandom, and εit is the classical error term with zero mean
and homoscedastic covariance matrix. The nature of the error structures leads to different
estimation procedures depending on the specification. For this study, we estimated the
models using one-way and two-way fixed and random effects models with F-tests and
Hausman tests used to evaluate the appropriateness of the model specifications.
Spatial panels
Cross sectional correlation can be an important factor in panel data model of parish level
pollution differences. Pollution and social capital relationship can be modeled using the
spatial correlation as well as the heterogeneity across parish using a spatial error
component regression model. The model is (Baltagi 2001):
1 Relationship between population density and water pollution may be positive or
negative depending on where the data come from. The hypothesis is open to an empirical
testing.
12