D-= [D1-,D2-,...,D1-2] take value 1. Therefore, ∂Lij /∂Xij =β when Xij is positive and
∂Li /∂Xi =γ otherwise.
ij ij
For the 4 lists noted earlier, we estimated equation (12) using OLS, FE and RE
procedures for each of the two years: 1997 and 2002. A number of specification tests and error-
structure analyses are conducted to choose the final specification which best fitted the data. Due
to space constraints, we do not report results of specification tests. 7 The first is the Lagrange
Multiplier test, which strongly rejected all four OLS specifications in favor of either FE or RE
regressions. Next, we rejected the restriction that the coefficient on the dissimilarity index is the
same regardless of its sign i.e., positive or negative dissimilarity, using a F test in most
specifications. To be consistent in reporting, we only report results from specifications with
asymmetric coefficients. The Hausman test was then employed to choose between FE and RE
estimators. In most cases, the FE effects specification is preferred over the random effects,
where the latter often assumes that the unobserved, random and state-specific effect is
independent of explanatory variables.8 In two cases, we failed to reject the null hypothesis of
Hausman test, i.e., the RE specification. However, the coefficients of RE and FE models are
qualitatively similar with some quantitative differences. Again, to be consistent in reporting, we
only present results from only FE models. Our error-structure analysis using a LM test indicated
the presence of groupwise (state-specific) heteroskedasticity. So, we utilize the feasible
generalized least squares estimator with fixed effects to estimate equation (12) for the 4 lists.
7A J-test and Cox test indicated that the share-based indexes (soil, land and water) better fit our model compared to
size-based indexes (Greene, 1997). Similarly, dollar shares of political contributions are preferred over volume-
based measures (e.g., number or share in total number of contributions).
8If the difference between the variance-covariance matrix of FE and RE model is not positive definite, the chi-
squared statistic of the Hausman test can take negative values. We obtained few negative values, where Greene
(1997) suggests setting the Hausman statistic to zero.
18