Notice the extremely high correlation between Positive and Image based press releases, as well as the correlation between Negative and Policy based releases. Such high correlations initially caused problems for the model, because we hoped to mix these groups of variables together. Statistically, however, this would be very unwise, as the amount of correlation would dramatically skew results.
Therefore, the models were altered to compare Positive to Negative and Image to Policy, as these were the only pairing without overly high correlation coefficients.