Graphing the group selection problem

Group selection.

Steven Pinker calls it “a scientific dust bunny, a hairy blob in which anything having to do with ‘groups’ clings to anything having to do with ‘selection’.” It’s a popular view, if a bit unkind. 

But there are talented scientists who believe that group selection is the light and the way. And there are still others who see no controversy at all, because individual and group selection approaches are predictively equivalent.

Do any of these positions get it right? Certainly, not all things “group” need group selection, as Pinker suggests, but perhaps some of them might. And, certainly, individual and group selection methods tend to find the same result—both carve up the same metaphorical pie, after all—but predictive equivalence isn’t causal equivalence. So, perhaps group selection is a useful concept that, when applied appropriately, tells us something true. 

This is the message Samir Okasha has been pushing for years and, in light of his latest work, I think it will finally stick. In a paper currently in press, Okasha uses causal graphs (also known as path analysis) to show how we can properly distinguish individual selection from group selection.

Causal graphs are visual representations of cause-and-effect relationships. For example, the figure below depicts the hypothesis that ice on the road causes car tires to slip (the arrow running from “ice” to “slip”), snow on the road causes tires to slip (the arrow running from “snow” to “slip”), and that, whether caused by ice or by snow, slipping tires cause cars to crash (the arrow running from “slip” to “crash”). So ice and snow directly cause slipping and indirectly cause car crashes. 

The graph also tells us that the presence of ice and snow are correlated (the dashed line between “ice” and “snow”), and so it’s possible to mistakenly attribute a car crash to icy conditions when, in reality, the tires slipped on snow. By analogy, it’s possible to mistakenly attribute the evolution of a trait to group selection when it was actually caused by individual selection. 

As Okasha has shown, causal graphs make this distinction clear. So clear, in fact, that I’ve adopted the method for three new papers of my own.

In the first, Pat Barclay and I published a commentary on Richerson and colleagues’ fascinating review of the evidence that group selection has shaped human behavior. You can read both pieces, along with several other excellent commentaries, here. In it, we use a straightforward application of Okasha’s central argument: group selection happens when a group trait directly causes group fitness and only indirectly causes individual fitness. Conversely, individual selection happens when an individual trait directly causes individual fitness and indirectly causes group fitness. So the graph on the left, below, represents group selection and the graph on the right represents individual selection.  

If you’re wondering what it means, in a practical sense, for a trait to directly cause individual fitness versus directly cause group fitness, my second causal graph paper should help. In it, I review Okasha’s methods and give a few examples, the basic issue being whether individuals survive and reproduce independently of one another. However, I also extend the causal graph method to one of the stickiest wickets in the group selection field: what to do about frequency-dependent and “emergent” effects—that is, the effects of group traits on fitness. When people get annoyed by a zeal for group selection, emergence is often at the root of it. (I refer you back to the Pinker quotation, above.) We already know that individual selection methods can handle emergence, despite protests to the contrary, but is that the right way to think about them? The graphs I develop show that it can be: an individual selection approach is called for whenever the emergent or frequency dependent trait directly affects individual fitness, as in the graph below. 

As you can see, the group trait—the emergent property—directly affects individual fitness and only indirectly affects group fitness in this case. So, although there’s a group trait involved, we’re still dealing with a clear case of individual selection. Consequently, an emergent property will sometimes evolve by individual selection and will sometimes evolve by group selection. 

In the third paper, David Logue and I apply this logic to a particular problem: the effect of vocal duets (an emergent property) and their constituent, individual parts on fitness. Although much of the paper concerns group-level traits, we never rely on group selection for understanding, because duets are almost certainly the product of individual selection. This is because duets will tend to affect individual fitness directly and group fitness only indirectly. In our article, we present three different causal models of duet evolution, two of which assume that duets have evolved functions designed by individual selection. 

Anyone interested in the group selection debate should first get a handle on Okasha’s causal graph argument. My new paper in TREE provides, I hope, a nice summary of that work and a useful extension of it. My collaborators and I have also applied causal graph methods to group selection in humans and to duetting in songbirds. With these methods, it’s possible that we’ll finally be on the road to resolving this half century-old debate and learning something new in the process. 

Cover image: "The Incredulity of Saint Thomas" by Caravaggio (Public Domain), via Wikimedia Commons.