How to Test an Evolutionary Explanation

In 1977 a drought on Daphne Major, a small island in the Galápagos, killed about 85 percent of its medium ground finches, and the birds that lived were not a random sample of the ones that died. That sentence makes three separate claims sound like one. Something sorted the birds. Whatever was sorted might or might not be passed on. And the island itself had changed. Each needs its own evidence, and none of them can be checked by rerunning 1977.

This lesson asks how you test a claim about evolution, much of which lies in the past. It sets out three kinds of evidence, uses the finches to keep the three claims apart, works one prediction with numbers, and ends on what any one dataset can and cannot carry.

Three kinds of evidence, and what each can carry

Think of a detective, a chemist and a historian. One notices that two things go together, one changes a single thing and watches, and one reconstructs an event from what it left behind. Evolutionary biology uses all three.

Observation of natural variation shows association. In the streams of Trinidad, guppy (Poecilia reticulata) populations that live with predators of adult fish differ in life history (age and size at maturity, number of young) from populations that live with a predator that mostly takes juveniles. That pattern fits the idea that predators shaped those traits. It cannot show it, because the sites also differ in stream size, food, light and history.

Experimental contrast changes the proposed cause on purpose. Guppies from a site with predators of adults were moved to a site whose predator preys on juveniles. Reznick, Bryga and Endler reported in 1990 that, 11 years later (30 to 60 generations), the moved population had changed its life history in the direction life-history theory predicted, the same direction as the natural comparison. A manipulation rules out explanations an observation leaves open, but it is still one stream and one introduction.

Historical inference tests rival explanations of a past event against traces that survive, choosing the traces that the rivals predict differently. The philosopher Carol Cleland has argued that this is a different method from experiment, not a weaker one. On Daphne Major, a male finch arrived and bred with a resident medium ground finch. Was he an unusual local bird or an immigrant? His genome answered it: Lamichhaney and colleagues placed him as a Geospiza conirostris from Española, more than 100 km away. The breeding was watched. His origin was inferred.

Daphne Major: selection, inheritance and a changing environment

The Grants and their colleagues caught, banded and measured nearly every finch on the island, year after year. That turned a drought into data. Here fitness means reproductive success, the number of surviving offspring an individual leaves in a given environment. Natural selection is a consistent difference in survival or reproduction among individuals that differ in a trait.

Claim 1: selection happened. Boag and Grant (1981) compared the 1977 survivors with the population before the drought. Larger birds, especially males with larger beaks, survived best, and the food left on the island was dominated by large hard seeds. Survivors differing from their starting population is selection within one generation. It says nothing yet about the next.

Claim 2: the trait is inherited. That needs offspring compared with parents. The natural shortcut is to take the birds feeding a chick as its parents. Keller, Grant, Grant and Petren (2001) tested that shortcut with eight microsatellite loci: 20 percent of the offspring in their sample were extra-pair young. Removing them raised father-offspring estimates of inheritance by 21 percent on average. Narrow-sense heritability, h^2, is the fraction of a trait’s variation in a population that is additive genetic variation, the part that makes offspring resemble parents. It belongs to one population in one environment, not to the trait everywhere.

Claim 3: the environment changed. The wet El Niño of 1982 and 1983 left a seed supply dominated by small soft seeds. In the 1985 drought, small birds with pointed beaks survived disproportionately well. The direction of selection reversed. Drought did not favour large beaks as such; the seeds that remained did. Over the whole record from 1972 to 2001, the Grants (2002) found selection on these populations that ranged from one-directional to oscillating, and from episodic to gradual.

Predicting the next generation, and why forecasts fail over decades

Selection and inheritance combine in the breeder’s equation, which complex traits and GWAS derives:

R = h^2 x S

S, the selection differential, is the mean of the selected parents minus the mean of the population they came from. R, the response, is the offspring mean minus that same original mean.

Worked problem. These numbers are illustrative, not the Grants’ measurements. Before a drought, adult mean beak depth in a finch population is 9.40 mm. The birds that survive and breed average 9.80 mm. Assume h^2 = 0.60, an additive model, no other trait correlated with beak depth under selection, and offspring reared in the same environment as their parents.

1. Selection differential. S = 9.80 - 9.40 = 0.40 mm.
2. Predicted response. R = 0.60 x 0.40 = 0.24 mm. Heritability has no units, so R stays in millimetres.
3. Predicted offspring mean. 9.40 + 0.24 = 9.64 mm.
4. A reversal. Suppose that in a later drought, after a wet spell, the survivors average 9.20 mm. Then S = 9.20 - 9.40 = -0.20 mm, R = 0.60 x (-0.20) = -0.12 mm, and the predicted offspring mean is 9.40 - 0.12 = 9.28 mm.
5. Check. 0.24 / 0.40 = 0.60 and -0.12 / -0.20 = 0.60, the heritability you put in.

S happens within a generation and R across generations. Only R is evolution. If h^2 were 0, R would be 0 however strong the selection: the survivors’ children would look like the population before the drought. The prediction is also only a prediction. An actual offspring sample is what tests it.

Survivors’ mean beak depth 9.80 mm Narrow-sense heritability h² 0.60

S, within one generation

R = h² × S, across generations

Predicted offspring mean

Check: R ÷ S

In this sketch: an additive model, no correlated trait under selection, and offspring reared in the same environment as their parents. Curve widths are drawn as a sketch, not to scale. The numbers are illustrative, not the Grants’.

S happens within a generation and R across generations; only R is evolution, and the prediction is only a prediction until offspring are measured.

On Daphne Major, predictions of this kind did well over a single episode. In 2006 the Grants reported that medium ground finches diverged in beak size from the large ground finch, Geospiza magnirostris, 22 years after it had arrived on the island, when the two species had severely depleted their shared food supply. The response was close to the value predicted from beak-size heritability.

Over decades, prediction failed for a different reason. The Grants (2002) concluded that evolution in these finches could be predicted in the short term from selection and inheritance, and that unpredictable fluctuations in the environment limited longer-term forecasts. The equation was fine. Nobody could forecast the rain.

A second limit lies inside the equation. The single-trait version assumes beak depth is not correlated with other traits that themselves affect fitness. Beak depth, beak length and body size go together in these birds, so selection on body size could show up as an apparent response in beak depth. Morrissey and colleagues (2010) pointed out that even the multivariate form is valid only if every correlated trait that affects fitness has been measured, and advocated the Robertson-Price identity as a more robust alternative. For the algebra of drift and selection on single genes, see population genetics.

What a dataset cannot show

Each kind of data answers its own question and leaves others open.

• Survival comparison (1977): selection on the measured traits in that year. It cannot show inheritance, and it does not prove beak size was the trait selected rather than something correlated with it.
• DNA-checked parent-offspring regression: how strongly the traits pass on in that population. It cannot identify what did the selecting, and it can still mistake a shared environment for inheritance, because parents rear their own chicks.
• The 30-year record: that selection varied in direction and pattern on one island, in two species, over those years. It does not tell you what happens on other islands or in other lineages. “Darwin’s finches show” is a broader claim than the data make.
• Genomes: Enbody and colleagues (2023) analysed 3,955 genomes from Daphne Major finches and found six loci of large effect that together explained 45 percent of beak-size variation in the medium ground finch. That leaves most of the variation unassigned, and an association with a locus is not a mechanism.
• The hybrid lineage (Lamichhaney and colleagues, 2018): reproductive isolation established within three generations in this case. It does not show that the lineage will persist or that it is a new species, and it gives no general timescale for speciation.

Uncertainty sits in different places in each: sample size in a survival count, pedigree errors and shared environments in a heritability, the next year’s weather in a forecast.

Design problem

The offspring of drought survivors have deeper beaks than the population before the drought. A colleague says that proves beak depth is inherited. Another says well-fed parents simply raise well-fed chicks. Design a study that tells these explanations apart, and state its limits.

One way to do it

Separate genes from rearing. Swap eggs between nests soon after laying: some eggs of deep-beaked pairs go to shallow-beaked foster parents, and the reverse. Leave other clutches with their own parents as a comparison. Measure every chick again as an adult, when beak growth is complete.

Get the pedigree right. Type chicks and all candidate parents at microsatellite loci. Keller and colleagues found 20 percent extra-pair young in their Daphne Major sample, so assign each chick its genetic father by DNA, not by who fed it.

State the prediction before the data come in. If beak depth is inherited, adult offspring will resemble their genetic parents more than their foster parents: the regression on genetic parents will have a clearly positive slope, and the regression on foster parents a slope near zero. If rearing explains the resemblance, the pattern will be the reverse.

Limits. This is one island, one species and one cohort, and the estimate belongs to that environment. Foster nests still differ in territory and food, and swapping happens after the egg is laid, so effects of the mother through the egg remain. A result in one year may not hold in a year with different seeds.