Separate tests were performed for each frequency, region and test, yielding a total of 20 tests. P values were adjusted using Holm's stepwise procedure55. Both predictors were mean-centered, to reduce multicollinearity and to facilitate interpretation of coefficients. Variance Inflation Factors were low (all <2), suggesting that results were not biased by multicollinearity. To test the reliability of the model in the presence of outliers and heteroscedasticity, whenever results were significant we performed additional robust regressions (with iteratively re-weighted least squares using the “lmrob” function from “robustbase” package in the R environment). We report results from ordinary least-squares regression which remained significant using robust regressions.
