In this simple case, the problem is not difficult to remedy and we do so by allowing for a communist dummy in our estimations.21 In other cases such a simple fix is not available. Many of the candidate measures of trade frictions or “openness” may be contaminated by other more subtle country characteristics that jointly determine trade and environmental policy. For example, the trade intensity variable we employ reflects country type considerations such as proximity to markets, geographic size and natural resource endowments and in general tends to be highest for small countries within close proximity to their trading partners. Because our pollutant under study is well known to have serious transboundary effects, there may be a correlation between countries with large measured openness and SO2 regulation. The openness measure developed by Sachs and Warner (1995) and measures of the black-market exchange-rate premium also suffer from similar problems.
Panel-data methods offer different ways to deal with the possibility of country-specific and/or site-specific excluded variables. When such effects are viewed simply as parametric shifts of our regression function, a suitable estimation approach is the least-squares dummy-variable (i.e., fixed-effects) estimator that treats these effects as constants. This approach is appropriate when the model is viewed as applying only to the countries or observation sites in the sample but not to additional ones outside the sample.
If, however, the model is viewed as a random draw of countries or observation sites from a larger population, it is appropriate to use a random-effects estimator to capture the level effect through a random variable. Because this estimator treats the level effects as uncorrelated with the other regressors, it may suffer from inconsistency due to omitted variables. By comparison, the fixed-effects estimator does not suffer from this inconsistency problem, but it focuses exclusively on the variation over time in our data.
Acknowledging the strengths and weaknesses of both types of estimators, our strategy is to estimate both fixed and random effects versions of our model whenever possible. We also report results from the Hausman test comparing these two methods. Occasionally we are forced to rely on the random effects implementation alone because some of our regressors would not be identified in a fixed effects estimation. Both of these methods have been widely used in the literature. payday loans online direct lenders only