Modeling--always log transform data (outlier reduction, variance stabilization).  Variance depends on absolute signal levels, even after log transform.  Consider incorporating this dependence into ANOVA models by fitting simple parametric models s = f(m).