Single-matrix models underestimate the level of compositional heterogeneity and overestimate the site-specific biochemical diversity of the informational gene alignments.

Predicted distributions (means and standard deviations) and observed value (in parenthesis) of the test statistic. All differences are significant at P 0.05. The models are those used in [18]; the LG model also fails all the tests for all four genes (see Table S1).