Why “The Spandrels of San Marco” isn’t a good paper

Stephen Jay Gould & Richard Lewontin’s 1979 article “The spandrels of San Marco and the Panglossian paradigm: a critique of the adaptationist programme” is one of the most widely-discussed papers in evolutionary biology over the last 40 years (cited over 3700 times according to Google Scholar). Probably every graduate student in evolution reads it, as well they should, both because of its historical importance and because it’s creative and intentionally provocative. Arguing about “Spandrels”, as the paper is colloquially known, is a good mental workout.*

Which isn’t to say it’s a good paper. Indeed, I actually think it’s a bad paper. It’s the kind of thing that might work as a deliberately-provocative blog post (and I certainly approve of those). But as a serious critique of adaptationist thinking, or even as a serious critique of bad adaptationist thinking, it just doesn’t hold up. And that’s a problem, because even deliberate provocation needs to have a solid foundation, lest the provocation just come off as empty rhetoric. Which is exactly what I think ends up happening in “Spandrels”.

A good critical review paper needs to do several things. It needs to describe the central questions driving the research it critiques. It needs to provide a thorough and comprehensive review of the approaches taken to address those questions, and the results obtained. Those elements provide the foundation for criticism. And it needs to lay out and argue for, in as much detail as possible, new and better approaches for addressing current questions, and new and better questions and approaches for addressing them. Often, it will be possible to cite recent work which exemplifies these new research directions.

“Spandrels” falls down on pretty much all those fronts.

1. The paper draws a now-famous analogy between putative adaptations of living organisms, and what Gould and Lewontin call “spandrels”**, which are an architectural feature of domed churches such as St. Mark’s Basilica in Venice:

What Gould and Lewontin call a spandrel is the roughly-triangular area at the center of this image, located at the junction of curved arches which support a domed ceiling. Gould and Lewontin say that this area is an inevitable consequence of putting a dome on top of four arches. Although the area displays a tilework image (here, an image of Christ), that’s not what it was designed to do–the surface actually isn’t a design feature of the building at all, it’s an architectural by-product of other features. Analogously, many features of organisms which appear to be adaptations to perform a certain function, actually aren’t adaptations at all. Gould and Lewontin use gaps between fans in the fan vaulted roof of Cambridge’s King’s College Chapel as a second architectural example of non-adaptations.

This is a very creative analogy. Too bad it’s flat-out wrong. As architectural historian and engineer Robert Marks (1996) points out, there is actually more than one way to support a dome on four arches–and the choice made by the builders of St. Mark’s Basilica was the only one known at the time to be sufficiently strong to support a dome as large as the one on St. Mark’s. In other words, the spandrels of San Marco are adaptive.*** Similar points hold for the gaps in the fan vaulting of King’s College Chapel.

The whole point of “Spandrels” is to critique unsupported “just so stories” about adaptation. But adaptations aren’t the only thing one can tell “just so stories” about. Ironically, Gould and Lewontin’s spandrels analogy is an unsupported “just so story” about non-adaptation.

There’s a deeper problem here as well. Maybe it’s just me, but I find the whole idea of an “architectural by-product” or an “architectural constraint” very fuzzy, and I doubt any example could make it clear. It just seems like a very poor analogy for any sort of constraint to which living organisms might be subject, as evidenced by the fact that subsequent thinking about “constraints” in evolution does not appeal to architectural analogies (Pigliucci and Kaplan 2000). This seems to be one of those cases where we’re better off just focusing on the system of interest (living, developing, evolving organisms) rather than on analogies to the system of interest. There certainly are times when a good analogy or metaphor can help to clarify concepts, but there are also times when a superficially-plausible analogy or metaphor only serves to mislead, or to hide unclear thinking.

2. Gould and Lewontin criticize adaptationist research for various other putative sins besides failure to recognize the possibility of “spandrels”. One of their other complaints is that adaptationists “atomize” organisms into aggregations of discrete traits, each of which is viewed as an independent adaptation to perform an independent function. This is inevitably distorting, say Gould and Lewontin; organisms are integrated wholes and selection acts on those wholes.

The problem here, which recurs throughout “Spandrels”, is not that this criticism is wrong, it’s that Gould and Lewontin don’t have much to say about what to do about it, as a good critical review paper ought to. “Spandrels” is a strikingly negative article; Gould and Lewontin are all about tearing down the “adaptationist programme”. But having torn down that programme, by emphasizing all the alternative hypotheses it purportedly ignores, they make only cursory and superficial remarks about how test those alternatives, or about how to pursue an alternative research programme built on different conceptual foundations. For instance, simply saying “organisms are integrated wholes, so don’t atomize them” is useless advice for researchers. We can’t do without atomization; the phenotype of an organism is too complex to summarize in one number. Fortunately, there are objective strategies for dealing with this problem, for instance by using correlation matrices to summarize which (possibly arbitrarily-individuated) traits are correlated with which others, and to what degree, and using this information to predict the direct and correlated consequences of selection on a given trait, such as chin size (Gould and Lewontin claim chins are spandrels, but neglect to note that one could certainly select on chin size, with predictable direct and correlated consequences for phenotypic evolution).

Nor can natural selection make due without atomization. IIRC (sorry, can’t find the reference), Lewontin himself once pointed out that an organism that wasn’t atomized into independent traits almost certainly couldn’t evolve by natural selection, because beneficial mutations would be all but impossible (any mutation would almost certainly just lead to a non-viable organism). That is, evolution by natural selection won’t work, or won’t work very well, unless the “genotype-phenotype map” has certain properties.

Of course, this doesn’t mean that natural selection “sees” individual traits. It is indeed true that, ultimately, natural selection only “sees” the whole organism. Specifically, all that natural selection “sees” is relative fitness, which is an integrated measure of how well the whole organism is adapted to its current environment (relative to other members of the population). Natural selection only acts indirectly on every other trait, however those traits are individuated and however they are intercorrelated. And you know who emphasized this point most keenly? Famous arch-selectionist R. A. Fisher, in The Genetical Theory of Natural Selection! So I have no idea why a “non-atomizing” approach to understanding organismal form and function couldn’t be thoroughly adaptationist.

3. Gould and Lewontin attack a few specific instances of what they claim are general tendencies in adaptationist research, but the reader is supposed to take their word for it that these few instances are representative. Sorry, but I don’t take anybody’s word for anything (which in retrospect seems to be wise policy when it comes to Gould’s pronouncements on the work of others). Gould and Lewontin were writing at or around the time that some cast-iron classic field studies of adaptation were underway–the work of Peter and Rosemary Grant on Darwin’s finches, and the work of John Endler on guppies, to name just two. In light of that, broad-brush characterizations of the entire field of evolutionary biology based on a few cherry-picked examples just won’t fly. Most of the space given over to deliberate provocations (discussion of architecture, invocations of Voltaire’s Dr. Pangloss, criticisms of A. R. Wallace and Herbert Spencer, cherry-picking supportive quotes from Darwin…) would’ve been better used in a more comprehensive review of the literature at the time. Exactly what alternative hypotheses typically are considered in studies of adaptation, and how often are those hypotheses tested and with what lines of evidence? This sort of review was common, even in that era. For instance (to pick examples from my own field of ecology), Connell 1983 and Schoener 1983 are examples from that time period of the sort of comprehensive review I’m thinking of (notably, both were hugely influential without needing to resort to deliberate provocation; sometimes substance trumps, or at least doesn’t need, style). Instead of doing a proper literature review, Gould and Lewontin offer up statements like this:

“At this point, some evolutionists will protest that we are caricaturing their view of adaptation. After all, do they not admit genetic drift, allometry, and a variety of reasons for non-adaptive evolution? They do, to be sure, but we make a different point. In natural history, all possible things happen sometimes; you generally do not support your favored phenomenon by declaring rivals impossible in theory. Rather, you acknowledge the rival but circumscribe its domain of action so narrowly that it cannot have any importance in the affairs of nature. Then, you often congratulate yourself for being such an undogmatic and ecumenical chap.”

Frankly, I think it’s pretty obvious who’s congratulating themselves on being undogmatic and ecumenical chaps here. In my view, this passage crosses the line from deliberately provocative into offensive; it’s an attack not on evolutionary biology, but evolutionary biologists. And not just a few isolated individuals, but evolutionary biologists “generally”. Which is complete and utter crap (wonder if “generally” includes Peter and Rosemary Grant?) Science is hard, scientists mostly do their best, and their best is mostly pretty good (that’s how we’ve learned so much about how the world works since the 17th century) Calling your fellow scientists self-congratulatory, willfully-biased dogmatists is pretty rich, especially coming from the same guys who can write “allometric patterns are as subject to selection as static morphology itself” (p. 157), and then immediately go back to touting allometry as a non-adaptive alternative to selection. Thanks guys–way to set a good example for the self-congratulatory, dogmatic masses by taking alternative hypotheses so seriously.****

4. Gould & Lewontin claim that developmental constraints and allometry may well be collectively so important that their effects should be regarded as the main story of evolution, with adaptation by natural selection being a relatively unimportant detail (p. 160). What they never bother to explain is how the heck we ever got complex, functioning organisms that could exhibit developmental constraints and allometry in the first place. Seriously, how in the name of Darwin was that supposed to happen, except via natural selection? See Peter Godfrey-Smith (e.g., this) for discussion of Gould and Lewontin’s tendency to run together empirical claims about the prevalence of adaptation with conceptual and methodological claims about the questions evolutionary biology ought to address and the methods by which it ought to address them.

5. Gould and Lewontin go on and on about how important forces besides natural selection are, but they completely miss all kinds of interesting questions that can be asked about the interplay between multiple evolutionary “forces”. They talk as if these other forces are alternatives to natural selection, when in fact it’s typically the case that all these forces co-occur and really surprising phenomena arise from their interplay (it’s not always that the “strongest” force just swamps all the others). For instance, developmental systems themselves vary and variation in developmental outcomes has fitness consequences, so developmental systems are subject to selection as well as constraining the phenotypes that an organism can build.  So how do developmental constraints themselves evolve? Can organisms evolve to become more “evolvable” (e.g., to have more atomized traits)? It’s not as if we’ve only learned in the last few years how to ask those kinds of questions. In population genetics, Sewell Wright’s shifting balance theory is all about the complex interplay of drift and selection, to pick just one famous example of this kind of thinking which predates “Spandrels”. It’s ironic that two Marxists like Gould and Lewontin would try to steer evolutionary biologists towards testing and rejecting alternative hypotheses, rather than pushing for evolutionary biologists to ask “synthetic” questions about the interplay of different, non-mutually exclusive causal factors which feed back on one another. Indeed, near the end of “Spandrels” Gould and Lewontin actually raise, and then dismiss as uninteresting, the fact that similar phenotypes can have different developmental underpinnings. Presumably, Gould and Lewontin are reluctant to even ask these kinds of questions because it would mean conceding ground to adaptationists. They prefer to cast their lot with old-school continental European developmental biology and insist (without any particularly convincing evidence) that Bauplans (“body plans”) constrain evolution so strongly that the constraints themselves are “much the most interesting aspect of evolution” (p. 160).

In conclusion, let’s step back from “Spandrels” itself and consider its legacy and influence. Certainly, evolutionary biologists today are much more rigorous and sophisticated about testing alternatives to adaptation than they were in 1979.***** Nielsen (2009) attributes a good deal of this increased rigor to “Spandrels”, and I suppose I should take his word for it because he’s an evolutionary biologist and I’m not. But honestly, I’m not at all convinced that “Spandrels” should get much credit. The bar always gets raised rather than lowered in science. Further, all areas of science have become much more quantitatively sophisticated since 1979 due in large part to improvements in computer hardware and software and associated rapid advances in statistical theory. Those advances let us develop more rigorous and sophisticated hypotheses, and test them in more rigorous and sophisticated ways. And evolutionary biology has been flooded with gene sequences and other molecular data which would’ve become available even if “Spandrels” had never been written.

Probably molecular data are the single most important factor driving the increased rigor of adaptationist research, but don’t give short shrift to the increasing importance and sophistication of mathematical modeling. Indeed, many of the questions evolutionary biologists ask about “constraints” these days are pretty much impossible to ask without mathematical modeling (and pretty much impossible to answer without molecular data) (Pigliucci and Kaplan 2000). Mathematical approaches to evolution of course have a long history–population genetics has always been mathematical–but recent conceptual progress in other areas, such as evolutionary developmental biology, has depended heavily on mathematical models such as those from Gunter Wagner. Today, theoretical and empirical research on evolutionary constraints and how they evolve is a hot area (review here). But it’s an area shot through with mathematical modeling and shows little sign of being influenced by the non-biological analogies and references to older, non-mathematical work which loom large in all of Gould’s thinking. Nor is this increasing prominence of mathematics a recent phenomenon; I’m not damning Gould and Lewontin for failing to push for work they couldn’t possibly have foreseen. For instance, Lande and Arnold (1983) developed a now-standard regression method for analyzing selection on possibly-intercorrelated traits just a few years after “Spandrels”.

This seems to me to highlight a significant limitation of Gould’s whole approach to science. Reasoning with mathematics is very different than verbal reasoning based on analogies and metaphors. Math forces you to be much more explicit and precise in your assumptions, and in the logic with which you derive your predictions or conclusions. Things like “developmental constraints” have very precise definitions in the context of mathematical models like those of Gunter Wagner, which not only facilitates empirical testing but makes metaphors and analogies superfluous. Michael Lynch is another good example of someone making heavy use of mathematical modeling to understand the interplay of selective and non-selective forces in evolution (in his case, in the context of genome evolution). Put it this way: which has advanced our understanding of the evolution of correlated characters more: the metaphor of the Bauplan and the analogy of “spandrels”–or the statistical methods developed by Lande and Arnold (1983)? And while mathematical models often are inspired by analogies between different systems, formalizing those analogies mathematically forces you to make them much more precise. And trying and failing to formalize an analogy is a very good way to learn that the analogy is actually a bad one. “Spandrels” generated a lot of heat–but the light currently being shed on adaptation (and non-adaptation) comes from sources that owe little to it.******

UPDATE: For contrasting takes on “Spandrels”, see here, here, and here. For a take in the same spirit as mine, see here. And for critiques that are if anything even more negative than mine, see here and here (I especially like Jeremy Yoder’s take, which points out that Gould’s fondness for flowery rhetoric was (ironically) accompanied by dislike of flowery rhetoric from others).

UDPATE #2: Commenters on this post point out Alasdair Houston’s very nice discussion of architectural analogies in evolution (concluding, like me, that “spandrels” are likely adaptive, and that architectural analogies are unhelpful), and to economist Paul Krugman’s sharply-worded assessment of Gould (“there’s no there there”!)

*fn 1: In my own field of ecology, Robert Peters’ A Critique for Ecology used to fill much the same pedagogical role as Spandrels–a deliberately provocative, contrarian, and highly problematic treatment of fundamental methodological issues. I’m not sure if Critique is still widely-assigned these days, though.

**fn 2: Gould and Lewontin’s terminology is not strictly accurate; see Robert Marks’ article for discussion.

***fn 3: This also undermines Dan Dennett’s (1995) claim, contra Gould and Lewontin, that spandrels are an “obligatory design opportunity”–a space which the builder is free to fill with whatever he wishes.

****fn 4: In “Spandrels” and elsewhere, Gould claims that it’s only low-sloped intraspecific allometries that are non-adaptive. AFAIK (correct me if I’m wrong), he never considers the alternative hypothesis, taken for granted in much recent work on the evolution of allometries, that intraspecific allometries only look different than (adaptive) interspecific ones because there’s less variation within species than among species. A basic fact about linear regression is that, if you shrink the range of variation on the X-axis, you lower your estimated slope (and broaden the confidence interval).

*****fn 5: I’m talking about evolutionary biologists here, not evolutionary psychologists.

*****fn 6: This isn’t to say things are perfect now. See Nielsen (2009) for a (productive and positive rather than provocative and negative) call for more rigorous testing of adaptive hypotheses in genomics.