There is sometimes a bit of confusion among beginners as to what, exactly, we are doing when we carry out a comparative method, especially when testing for character correlations. Common language that comparative methods “control for phylogeny” or “remove the phylogeny from the data” is not necessarily enlightening or even always accurate. Another common suggestion is that species are not statistically independent and that we must account for that with comparative methods. While accurate, I still don’t think this statement fully captures the tree-thinking perspective enabled by comparative methods. In this section, I will use the particular example of correlated evolution to try to illustrate the power of comparative methods and how they differ from standard statistical approaches that do not use phylogenies.
In statistics, two variables can be correlated with one another. We might refer to this as a standard correlation. When two traits are correlated, it means that given the value of one trait – say, body size in mammals – one can predict the value of another – like home range area. Correlations can be positive (large values of x are associated with large values of y) or negative (large values of x are associated with small values of y). A surprisingly wide variety of hypotheses in biology can be tested by evaluating correlations between characters.
In comparative biology, we are often interested more specifically in evolutionary correlations. Evolutionary correlations occur when two traits tend to evolve together due to processes like mutation, genetic drift, or natural selection. If there is an evolutionary correlation between two characters, it means that we can predict the magnitude and direction of changes in one character given knowledge of evolutionary changes in another. Just like standard correlations, evolutionary correlations can be positive (increases in trait x are associated with increases in y) or negative (decreases in x are associated with increases in y).
We can now contrast standard correlations, testing the relationships between trait values across a set of species, with evolutionary correlations - where evolutionary changes in two traits are related to each other. This is a key distinction, because phylogenetic relatedness alone can lead to a relationship between two variables that are not, in fact, evolving together (Figure 5.1; also see Felsenstein 1985). In such cases, standard correlations will, correctly, tell us that one can predict the value of trait y by knowing the value of trait x, at least among extant species; but we would be misled if we tried to make any evolutionary causal inference from this pattern. In the example of Figure 5.1, we can only predict x from y because the value of trait x tells us which clade the species belongs to, which, in turn, allows reasonable prediction of y. In fact, this is a classical example of a case where correlation is not causation: the two variables are only correlated with one another because both are related to phylogeny.
If we want to test hypotheses about trait evolution, we should specifically test evolutionary correlations1. If we find a relationship among the independent contrasts for two characters, for example, then we can infer that changes in each character are related to changes in the other – an inference that is much closer to most biological hypotheses about why characters might be related. In this case, then, we can think of statistical comparative methods as focused on disentangling patterns due to phylogenetic relatedness from patterns due to traits evolving in a correlated manner.