tire set of intermediate reduced equations--and, for the sake of simplicity, discuss only the final reduced equations. 14
We have three final comments to make. First, we noted a problem with skewed distributions on both independent and dependent variables that had the potential to adversely affect all statistical tests. To eliminate this problem, we normalized our data to the extent possible. We did this through a combination of recoding extreme values and, in some instances, using either a log linear or square root transformation (in accordance with procedures recommended by Tabachnick and Fiddell, 1989:68-89).
Second, we were interested in determining if conditional effects were present. To assess interactive effects, we created exhaustive sets of terms from our pool of independent variables and entered these terms one at a time. The term was retained if it achieved significance at the p < .05 level. Since this aspect of the analysis was exploratory--since interactive terms are usually not considered unless there is some prior theoretical justifications for their existence--we treat these results as advisory and recommend other researchers try to replicate our findings.
In order to test for conditional effects, we adopted the technique recommended primarily by R. J. Friedrich ( 1982) but repeated and explained elsewhere ( Aiken and West, 1991; Cohen and Cohen, 1983:301-350). This technique requires that all variables are standardized before construction of interactive terms and that the equations be generated on the basis of standardized data. Under these conditions, reported results are regression coefficients (b) based on the standardized data.
Finally, multicollinearity is a problem with the type of data and measures we are using. We minimize this problem, in part, through the use of factor scales to subsume highly correlated measures of the same concept (for example, our social class scale). However, the problem could not be completely eliminated, and the data being the data, we simply had to accept that uncorrelated measures of the concepts of interest were simply not available. In order to produce a complete picture, we reproduce the tolerance statistics for each of our equations. Tolerance scores, which range from 0.00 to 1.00, indicate the degree of multicollinearity for respective independent variables and are interpreted as the smaller the tolerance score, the higher the level of multicollinearity. We note that the vast majority of the literature on the subject of environmental injustice does not report this statistic.
In this chapter, we have articulated the model that guides the analysis, its concepts and measures, the hypotheses that we will examine at the state, county, and city levels, and the method of analysis employed. In Chapters