Ackermann and Cheverud (2004) consider the pattern of selection necessary to change a nonrobust australopithecine cranium (i.e. Sts 5) into a robust australopithecine or early Homo cranium. To do this, they measure seven fossil specimens (KNM-ER 1470, KNM-ER 1813, KNM-ER 3733, KNM-ER 406, KNM-WT 17000, SK 48, and Sts 5), taking eight linear measurements on each (nasion-nasale, nasion-frontomalare, nasale-anterior nasal spine, anterior nasal spine-intradentale superior, anterior nasal spine-zygomaxillare superior, frontal-maxillary nasal suture point-zygomaxillare superior, and zygomaxillare superior-frontomalare). These are all facial measurements, so the question is the degree to which gross facial dimensions change over time.
They also have samples of humans, chimpanzees, and gorillas to provide a model of within-population facial variation. These comparative samples are used as follows:
Phenotypic within-population V/CV matrices for the facial variables from all three living primate populations were obtained by using the residual CV matrix from a multiple ANOVA with the eight traits as the dependent variables and subspecies as the independent variable, thus pooling the CV across subspecies, and were then simplified to their principal components (PCs). The PCs of the within-population V/CV matrix are ordered by their level of V (eigenvalues) and are uncorrelated with one another so that on the scale of the PCs, the within-population V/CV martix is a simple diagonal matrix with no CVs among components. PC scores are calculated for each fossil population by multiplying trait means by the standardized within-population PC loadings. The between-population V for each PC can then be calculated as the V among these population mean PC scores; these values are given in Table 2 along with the within-population Vs (eigenvalues) for each extant model (Ackermann and Cheverud 2004:17947).
I had to read that through a few times before I got it, and since I'll surely forget it, I quoted it verbatim.
From the values for each group, the amount of morphological change is estimated between the hominid populations. Random drift is expected to cause evolutionary divergence between populations in a way that is proportional to the within-population variation of traits. In other words, traits that are highly variable within populations should also vary highly between populations, and two traits that are correlated within populations should both vary in the same way between populations. By examining the PC scores within populations, they eliminate the correlations, and can consider whether the within and between-population differences are linearly related. Under the hypothesis of genetic drift, the slope of a regression between the two should be 1.0. Greater divergence time increases the expected between-population difference, but the relation with within-population variance of each trait should remain linear and with a slope of 1.0 (17948).
Since these are PC scores, it is hard for me to figure out quite what deviations from this model would mean. For example, there is undoubtedly variation about the regression line, but how much variation is too much to be consistent with the model? If the slope deviates from 1.0, that means that the high variance traits are either significantly more or less different than they should be, or the low variance traits are significantly higher in between-variation difference than they should be (they would seem unlikely to be lower than expected). And are the PC's drawn from living hominoid populations really applicable to the fossils? After all, the three contemporary species generate different PC's, and the fossil taxa are analyzed three different ways here as a result. Does the significant result in the robust australopithecines mean that they really were subject to selection, or that they fit the pattern of variation in living hominoids poorly? Since these variances are pooled into classes (australopiths, robusts, Homo), it is hard to tell.
The hominid data seem difficult to shoehorn into Lande's (1976) predictions about drift. The expectation for the degree of change over time for a trait within a population is no change. In an infinitely large set of populations, the degree of change in a single trait is expected to be distributed normally, with a variance determined by the within-population variance of the trait. What that means is that if there are a large set of populations that have diverged simultaneously from a common ancestor, the degree of variance among those populations should be predicted by the degree of variance within the populations. But for the early hominids we do not have a large range of populations; we have a set of pairwise contrasts. I should also mention that we do not have a sample of species means; we have a sample of individuals drawn from species at different times and places. For example, the "australopith" class equals the "robust" class plus Sts 5. Together, all this should mean that the scatter plot of between-population difference against within-population variance should be exactly that--a scatter.
For the significant results, Ackermann and Cheverud (2004) estimate the degree of selection necessary on each linear measurement to create the observed amount of change. These values are smoothed across a map of the face to create a graphic picture of selection intensity. This is not a real quantification of selection in terms of deaths, since each measure used here may really be linked to one or more correlated traits that affect mortality in early hominid populations.
Their results indicate that robust australopithecines emerged through a selective process operating on the shape of the upper and lower face, while early Homo did not exhibit a significant divergence from the predictions of genetic drift among the specimens sampled. They describe this result (17951):
[A]lthough the initial divergence of Homo from the australopiths may have involved selection, divergence after this time (at least in the facial characters analyzed) could have occurred through random processes alone. In other words, much of the facial diversity seen in the Homo lineage from ~2.5 million to 1 million years ago may result from random evolutionary processes, rather than adaptive evolution. Other studies have shown that craniofacial diversity in most populations of modern humans can be explained by random processes. Lynch (1990) suggests that the development of cultural inheritance could have released many of the morphological traits of humans from the pressures of stabilizing selection. This study supports that idea and supplies it with a temporal context, potentially providing direct biological evidence of a shift early on in this lineage toward nonbiological adaptation (i.e., culture) as early hominins increasingly relied on technology. Because drift tends to play a larger role in shaping diversity when populations are finite, these results also may reflect a demographic revolution toward increasingly isolated and widespread populations.
The time range cited here is probably overstating matters, considering their sample of 3 specimens. Indeed, the great size disparity among the early Homo fossils used here may have some effect on the results. I would posit that it is not time yet to accept the null hypothesis of drift, considering the failure to detect the fairly profound selection on size among the specimens. The fact that the variation among these specimens correlates linearly with the variation within the comparative samples does not prove drift; there are a number of ways that selection on facial shape might result in such a pattern. Consider for example that the gorilla comparative model (which has the greatest degree of within-sample variation and would best match the two-specimen distance of KNM-ER 1470 and KNM-ER 1813) yields a regression for the early Homo sample with an R2 of 0.01. In other words, the scatter is super-high, and doesn't really prove or disprove anything. The chimpanzee and human models are very close to a slope of 1.0, and have less scatter; but their variance is arguably a worse model for early Homo including H. habilis.
For the robust australopithecines, Wood and Lieberman (2001) found that the traits that vary most within the lineage are those that are related to masticatory strain; the fact that these characters are much less variable between robust taxa (indicated by the very low regression slope found in this study) is not especially surprising: this would seem to indicate strong stabilizing selection on variable characters within taxa rather than directional selection on robust facial morphology over time.
I think a stronger study would control better by sample, with larger samples compared to each other and specific evolutionary hypotheses being tested (for example, are all early Homo erectus specimens, including Dmanisi, significantly divergent from Homo habilis?). I won't say that more features should be added, because until there are more specimens, there is no point whatsoever in adding features from a statistical point of view. Indeed, more focused questions might require that variable number be reduced. A resampling test that could determine if a very small number of traits are consistent with drift would be useful here. Even if this were simply a test of the regression, it might be handy, but especially if it tested drift directly without recourse to a regression. In particular, if such a test could directly compare pairs of species samples instead of variance within multiple-species samples, it would be a more useful test.
Ackermann RR and Cheverud JM. 2004. Detecting genetic drift versus selection in human evolution. Proc Natl Acad Sci U S A 101:17946-17951.