What is multilevel selection?

The arguments in the group selection debate at The Edge, as kicked off by Steven Pinker, contain some useful descriptions on what is meant by multilevel selection in a modern sense and how this varies from older formulations of group selection. Some of this is worth drawing out.

The old story of group selection might run as follows. There are two types of people - altruists and egoists. Altruists are willing to incur individual costs for the benefit of the group, while egoists shirk this responsibility for their own benefit. As a result, egoists have higher fitness than altruists within groups, while groups with higher proportions of altruists do better than groups with relatively more egoists. If altruistic groups have a large enough advantage over groups with more selfish individuals, and as a result grow faster and bud off new groups, it may be possible for altruists to increase in overall prevalence even though egoists have an advantage within groups.

Maynard Smith’s haystack model was one of the many early critiques of this picture. Maynard Smith argued that the conditions required for the evolution of an altruistic trait (in his haystack model, timidity compared to the dominant aggressive trait), were so limited that they were unlikely to be satisfied. These conditions included a limited level of migration between groups and large differences in relative group fitness. Group selection theory also suffered from criticism of a lack of preciseness about how the selection at various levels should be weighted.

Since then, multilevel selection theory has tightened a few of those issues up. First, from David Queller:

Modern group selection theory is as mathematically rigorous as individual selection or inclusive fitness theory. … They simply divide up fitness in slightly different ways – inclusive fitness into effects on self versus others, and multilevel selection into between-group and within-group parts – and a simple partition of fitness should not alter predictions. Inclusive fitness became popular, despite the head start enjoyed by multilevel selection thinking, because it successfully weighted the relative importance of its two fitness components, using genetic relatedness. Without a similar set of weights, group selection advantages could not be accurately judged, and their strength and importance was often overemphasized. … However, modern multilevel selection theory does have such weights, the between-group and within-group genetic variances, whose ratio happens to be relatedness of the actor to its groupmates (including itself). Once the proper weights are accounted for, the two approaches give essentially identical results.

As well as tightening up the mathematics, the definition of group was also tweaked. Rather than talking about competition between distinct populations, multilevel selection looks at competition of all levels of organisation, with groups formed within populations at various stages of the life-cycle. Pinker writes:

In most models of the new group selection, a group is defined as any subset of interacting individuals, that is, as organisms which interact with one another more intensely than they interact with organisms selected from the population at random. Two sisters who help each other, for example, or a pair of friends who trade favors, are dubbed “a group”.

It is this new grouping arrangement that is partitioned and weighted:

… Once a “group” is defined as a subset of interacting individuals, the variance in the fitness of individuals can be partitioned into two statistical components: how fit the individual is with respect to his groupmates, and how fit his group is with respect to other groups. … Examples include huddling for warmth, mobbing a predator, and Tooby’s example of pooling resources to get higher expected returns in a risky investment. In such cases one can separate the benefits that accrue to the entire group (including me) and whatever benefits or costs are assumed by me but no one else in the group.

This is not to say that everyone is on board with this statement that the two approaches are mathematically equivalent. However, the mathematical equivalence and greater flexibility about what constitutes a group provide group selection advocates an alternative argument about why it is useful.

Unfortunately, the new approach is often used as a trojan horse for the old group selection approach, particularly in popular discussions of human evolution. The responses to Pinker contain various varieties of this.

Pinker also makes some interesting points about the costs of the new approach. First, the flexibility in the definition of groups is multi-level selection’s weakness.

If the two theories really are equivalent, then any advantage of group selection (in this new sense) would have to come from the models’ being more convenient, elegant, simple, transparent, explanatory, or mathematically tractable. Yet by stretching the meaning of “group” beyond its ordinary sense, that’s just what they fail to be. …

[A] mathematical model that submerges the psychological forces that keep different “groups” together (such as genetic relatedness or mutual sensitivity to altruism), and requires theorists to dig them out from under the equations, is hardly a perspicuous way to analyze sociality (as Coyne points out). In gene-selectionist theories, the theoretical constructs that power the models turn out to be fantastically psychologically important, including sensitivity to kinship, scrutiny of individuals, moralistic emotions elicited by benefits conferred or withheld, and the psychological differentiation of relationships into discrete models corresponding to mutualism, kinship, reciprocity, and dominance.

Pinker also picks up what I consider to be the major reason multilevel selection has not been as broadly accepted as its proponents hope:

Mathematical biologists such as Alan Grafen, Stuart West, Ashleigh Griffin, and Andy Gardner have criticized this formulation because it obfuscates the fact that individuals are still maximizing their genetic fitness: “The fundamental point is that the spread of a gene is determined by its ‘fitness relative to others in the breeding population, and not to others with which it happens to interact.’ … Natural selection selects for a gene if it causes a behavior that leads to that gene increasing in frequency in the population, not some other arbitrarily defined scale such as social partners.”

This point can be drawn out through examining some of the similarities between the cooperation identified in a multilevel selection framework and the way economists look at economic exchange. I will offer some more thoughts on this in my next post on group selection.

*As an aside, I am deliberately avoiding the cultural group selection issue for the moment.