20.05.02 · philosophy / phil-of-biology

The unit of selection

draft3 tiersLean: nonepending prereqs

Anchor (Master): Okasha, *Evolution and the Levels of Selection* (Oxford, 2006); Wilson & Sober, *Unto Others: The Evolution and Psychology of Unselfish Behavior* (Harvard, 1998); Sober, *Evidence and Evolution* (Cambridge, 2008); primary lit (*Biology & Philosophy*, *Studies in History and Philosophy of Biological and Biomedical Sciences*, *Philosophy of Science*)

Intuition [Beginner]

Natural selection acts on something. When biologists say "the giraffes with longer necks left more descendants, and over time the population's necks got longer," there is a clear story: individual giraffes vary in neck length, longer-necked ones eat more leaves and have more babies, the babies inherit the necks, the population shifts. Selection acts on individual giraffes.

This is the textbook answer, and for most cases it is fine. The question of what selection acts on — the unit of selection — becomes a real puzzle only when standard cases stop working cleanly. Three kinds of case force the question.

First case: a sterile worker bee. She has no offspring at all. Her individual fitness, measured as the number of her own descendants, is zero. By the textbook account, selection should eliminate sterility in one generation. But sterile workers exist throughout the social insects — bees, ants, termites — and are the rule rather than the exception in those lineages. Something has gone wrong with the picture of selection acting on individuals.

Second case: the peacock's tail. A male peacock's enormous, conspicuous tail makes him slower, more visible to predators, and more expensive to grow. From the standpoint of survival, the tail is a liability. But peahens prefer males with more elaborate tails, so males with extravagant tails leave more offspring. Selection here acts on individuals — but the trait being selected for hurts the individual's survival while helping his reproduction.

Third case: a "selfish" stretch of DNA that copies itself faster than the rest of the genome, accumulating across generations, even if it does no good — or active harm — to the organism carrying it. Transposable elements ("jumping genes") do exactly this. The genome contains thousands of such sequences in most species. Selection here acts on the gene, not on the organism.

These three cases pull in different directions. The first invites a group answer: the worker bee's behaviour benefits the colony, and colonies with workers out-compete colonies without. The second invites the textbook individual answer with the wrinkle that "fitness" must mean reproductive success, not just survival. The third invites a gene answer: alleles compete with each other to be copied, and the organism is at best a temporary vehicle they ride in.

In 1976, Richard Dawkins published The Selfish Gene, which made the gene-level answer famous. His claim was strong: the gene is the only true unit of selection, because only genes have the fidelity of copying needed for selection to accumulate. An individual organism is shuffled away every generation through sex; a group is shuffled by migration; only a gene gets to persist long enough across generations for selection to do meaningful work. Organisms and groups are "vehicles"; genes are "replicators."

On the gene's-eye view, even cooperation is decoded as gene-level strategy. A copy of a gene in one body helps a copy of the same gene in a related body, because helping that other body is helping that other copy of the gene. This is kin selection, formalised by W. D. Hamilton in 1964.

Three years before Dawkins, V. C. Wynne-Edwards (1962) had defended a strong group selection picture: animals restrain their reproduction "for the good of the species," because populations whose members over-reproduce crash, while populations whose members self-regulate persist. George Williams (1966) and John Maynard Smith (1964) demolished this view: it requires groups to be sufficiently isolated and sufficiently short-lived for between-group differences to outpace within-group cheating, and the conditions are rarely met in nature. Group selection went into ecological exile.

Then it came back. Starting in the 1970s and accelerating after David Sloan Wilson and Elliott Sober's Unto Others (1998), a revised "multi-level selection" theory argued that selection can act at any level — gene, organism, group — simultaneously, and that the question of which level matters in a given case is empirical, not a matter of pre-emptive metaphysical choice. The mathematical machinery (the Price equation, due to George Price, 1970) makes this rigorous: any change in a trait's mean value in a population can be partitioned into a between-group component and a within-group component, and the two components measure selection at the two levels.

The unit-of-selection problem is whether this partitioning fixes a fact about what selection is acting on. Population-genetics dynamics (the formal framework that the Hardy-Weinberg equilibrium result anchors) is consistent with multiple ways of describing the same allele-frequency change. The Dawkins gene's-eye view and the Wilson-Sober group-selection view often agree on every measurable quantity. So is there a fact of the matter about which level selection is "really" acting on, or are these two equally good descriptions of one underlying process?

That is the question the rest of this unit develops. The Intermediate tier states the formal apparatus (Lewontin's three conditions, the replicator/interactor distinction, the Price equation, the MLS1/MLS2 split) that makes the question precise. The Master tier surveys the position space and engages a contemporary primary paper.

Visual [Beginner]

Picture three nested levels of biological organisation: gene copies inside cells, cells inside organisms, organisms inside groups. At each level there is variation, and at each level reproductive success differs across that variation. Now run a thought experiment.

Imagine a population of beetles. Some beetles are solitary (carrying allele $S$ ); others are cooperative (carrying allele $C$ ). Cooperative beetles share food, so the average $C$ -beetle has fewer of its own offspring than the average $S$ -beetle within any single group — within-group, cooperation is costly. But groups containing more cooperators have more total offspring, because cooperation increases the group's resource yield — between-group, cooperation is favoured. Within each group, selection acts against $C$ ; across groups, selection acts for $C$ . Which level wins depends on the balance of within-group and between-group variance.

The picture has three diagrams. On the gene's-eye view, each allele $C$ and $S$ is a player; their reproductive success is read directly from offspring count, and the population's allele frequencies shift according to whichever allele copies itself more on net. On the individual view, each beetle is a player; their fitness depends on the genotype combination in their group. On the group view, each group is a player; groups compete to reproduce more groups (by colonising new patches, say).

The image is the same allele-frequency change, decomposed three different ways. The Intermediate tier asks: which decomposition tells us what selection really acts on?

Worked example [Beginner]

A population of $1000$ beetles is split into $10$ groups of $100$ . In each group, beetles with the cooperative allele $C$ have a per-capita reproductive success that is $5%$ lower than beetles with the solitary allele $S$ , because $C$ -beetles share food. But the group total reproductive output is proportional to the fraction of $C$ -beetles in the group: a $100%$ - $C$ group produces $1.5$ times as many offspring as a $100%$ - $S$ group.

Three of the ten groups start at $80%$ $C$ , four at $50%$ $C$ , and three at $20%$ $C$ . The overall starting frequency of $C$ is $(3 \cdot 0.8 + 4 \cdot 0.5 + 3 \cdot 0.2) /10 = 0.5$ .

Step 1. Within-group dynamics. In the $80%$ - $C$ groups, $C$ loses ground: after reproduction, $C$ frequency drops by roughly $5% \cdot 0.8 \cdot 0.2 = 0.008$ , so the new $C$ frequency within those groups is about $0.792$ . Similar small drops happen in the other groups: the $50%$ groups go to $\sim 0.4875$ , the $20%$ groups to $\sim 0.192$ .

Step 2. Between-group dynamics. The $80%$ - $C$ groups produce more total offspring, because their group output is higher. If we weight each group's offspring contribution by $(1 + 0.5 \cdot group C -fraction)$ , the $80%$ - $C$ groups produce $1.4$ times the baseline output, the $50%$ groups $1.25$ , the $20%$ groups $1.1$ . The next generation's pool is biased toward the high- $C$ groups.

Step 3. Combine. The new overall frequency of $C$ comes from summing each group's post-selection frequency weighted by its offspring share. The arithmetic (full version in the Intermediate tier exercise) gives a new overall $C$ -frequency of about $0.515$ — a net increase of $0.015$ despite the within-group drop everywhere.

Step 4. The takeaway. $C$ went down in every single group. $C$ went up overall. This is Simpson's paradox for selection: a trait can decrease in every subpopulation while increasing in the union of subpopulations, when subpopulations differ in size. Group selection is real — but only in the sense that the between-group component of the Price equation is non-zero. Whether that between-group component is the unit of selection, or whether the same arithmetic is more honestly read as gene-level selection with group-structured fitness effects, is the open question.

Check your understanding [Beginner]

Exercise (easy, multiple choice).

Which one of the following best states the unit-of-selection problem?

A. Whether evolution is true. B. Whether natural selection acts on genes, individuals, groups, or some combination, and whether this question has a determinate answer. C. Whether populations always satisfy Hardy-Weinberg equilibrium. D. Whether sterile insect workers are evolutionarily stable.

Hint

The unit-of-selection problem is a question about selection, not about whether selection happens.

Answer

Option B.

The problem assumes selection happens and asks what entity selection acts on. A is question-begging (evolution is the agreed background). C confuses the null model for evolution with the question of what evolves. D is one downstream empirical phenomenon that bears on the question but is not the question itself — sterile workers are evidence in the debate, not the debate.

Exercise (easy, true/false).

Hamilton's rule $r b > c$ — where $r$ is the coefficient of relatedness between an actor and a recipient, $b$ is the benefit to the recipient, and $c$ is the cost to the actor — was historically taken to support gene-centric selectionism because it lets altruism be explained at the level of allele copies rather than at the level of group benefit.

Hint

Hamilton's reformulation rewrites "for the good of the group" as "for the good of allele copies shared with relatives."

Answer

True.

Hamilton's 1964 papers showed that an apparently altruistic behaviour — sacrificing personal reproductive success to help relatives — is favoured by selection at the gene level when an actor's lost reproductive output (cost $c$ ) is more than offset by the gain in copies of its alleles passed on through relatives (benefit $b$ weighted by relatedness $r$ ).

The reformulation made it possible to explain eusocial-insect altruism, alarm calls, and kin-directed grooming without invoking group selection: gene copies that promote help-your-kin behaviour are favoured because they are helping other copies of themselves carried by close relatives. This was the central technical argument used by Williams 1966 and Dawkins 1976 to make gene-level selectionism the consensus view from the late 1960s through the 1990s.

Exercise (easy, multiple choice).

Sterile worker bees pose a puzzle for which view of natural selection most directly?

A. Gene-level selectionism. B. Group-level selectionism. C. Individual-level selectionism. D. Multi-level selectionism.

Hint

The puzzle is that sterile workers have zero direct reproductive output. Which view of selection makes that fact most difficult to absorb without further apparatus?

Answer

Option C.

A sterile worker has zero direct offspring, so under a strict individual-fitness account she should be eliminated in one generation by selection against her sterility. The puzzle is most acute for individual-level selectionism in its naive form. Gene-level selectionism resolves the puzzle through kin selection — the worker's alleles are propagated by her queen sister's reproduction. Group-level selectionism resolves it by appeal to colony-level fitness. Multi-level selectionism resolves it by allowing both individual and colony levels to bear fitness simultaneously.

Exercise (medium, short answer).

Imagine a future technology that made every organism reproduce clonally — every offspring an exact genetic copy of its single parent, with no recombination or sex. List two ways this would change the unit-of-selection debate.

Hint

Think about (1) what makes a "replicator" in Dawkins's sense, and (2) what role recombination plays in keeping organism-level traits inheritable.

Answer

Two changes are immediate. First, under strict clonal reproduction the entire organismic genome would be inherited as a single unit; the gene-level granularity that Dawkins emphasises (where individual alleles persist across generations while the genome they sat in this generation is dispersed by meiosis) would collapse. The genome and the organism would become co-extensive as replicators, weakening the case for the gene as the uniquely fundamental level.

Second, the formal foundation for Hamilton's rule changes: relatedness $r$ between any parent-offspring or sibling-sibling pair is $1$ , not $1/2$ , so kin selection's distinctive arithmetic (the discounting of distant relatives) disappears. Both gene-selectionism's strongest case and group-selectionism's main competitor would have to be rebuilt. A weaker third effect: clonal reproduction makes whole-genome lineages directly visible to selection, which strengthens the case that the organism (now genome-identical with its lineage) is a unit in a way it is not under sexual reproduction.

Formal definition [Intermediate+]

The intuitive picture made precise. Lewontin's three conditions (Lewontin 1970): an ensemble of entities is subject to natural selection iff

phenotypic variation — the entities differ from each other in some heritable trait;
differential fitness — the trait correlates with reproductive success (broadly: differential survival, reproduction, or replication);
heritability — offspring resemble their parents in the trait above chance.

Any entity satisfying these three conditions is a unit of selection in Lewontin's sense. Critically, the conditions are agnostic about the entity type: genes, cells, organisms, colonies, species can each in principle satisfy them.

Replicators and interactors (Hull 1980; Dawkins 1976/1982). The Lewontin conditions conflate two roles: the entity whose differential survival drives selection (the interactor, in Hull's term — the vehicle, in Dawkins's), and the entity whose persistent identity across generations carries the selected variation (the replicator). For Dawkins the replicator is uniquely the gene-copy, because only at gene-copy scale does sufficient copying fidelity exist for selection to accumulate; sexual reproduction shuffles individual genomes every generation, and migration shuffles groups, so neither individuals nor groups satisfy the replicator role. The vehicle/interactor is, in Dawkins's view, whatever phenotypic structure most efficiently propagates the replicator — for sexually reproducing animals this is usually the organism, but for some traits the unit can be the group (eusocial colonies) or a slice of the genome (transposable element clusters).

The Price equation (Price 1970, 1972). Let a population consist of indexed entities $i = 1, \dots, N$ with phenotypic values $z_{i}$ and absolute fitnesses $w_{i}$ (the number of descendant entities at the next time step). Write $\overset{w}{ˉ} := N^{- 1} \sum_{i} w_{i}$ and $\overset{z}{ˉ} := N^{- 1} \sum_{i} z_{i}$ . Let $Δ \overset{z}{ˉ}$ denote the change in mean trait value from this generation to the next. Then

\overset{w}{ˉ} Δ \overset{z}{ˉ} = Cov (w_{i}, z_{i}) + E [w_{i} Δ z_{i}],

where $Cov$ is the population covariance, $E$ the population expectation, and $Δ z_{i}$ is the difference between the mean trait value of $i$ 's offspring and $z_{i}$ itself. The identity is exact (not a leading-order approximation): the first term measures the contribution of differential reproductive success at the entity level (entities with high $z$ have high $w$ ), the second term measures the contribution of transmission bias (offspring of $i$ deviate systematically from $i$ ).

The Price equation admits a recursive partition. If the entities $i$ are grouped into groups $G_{k}$ indexed by $k$ , then the same identity can be re-applied within each group, yielding

\overset{w}{ˉ} Δ \overset{z}{ˉ} = between-group component Cov_{k} (W_{k}, Z_{k}) + within-group component E_{k} [W_{k} Δ z_{k}],

where $W_{k}$ and $Z_{k}$ are the group-level fitness and trait means and $Δ z_{k}$ is the within-group change in $z$ . The between-group term measures selection between groups — groups with higher trait means reproduce more groups; the within-group term measures selection within groups — within each group, entities with higher trait values reproduce more (averaged over groups).

MLS1 versus MLS2 (Damuth & Heisler 1988). The recursive Price decomposition admits two interpretations that are formally distinct and, on careful analysis, semantically different:

MLS1: the group-level trait $Z_{k}$ is the average of the entity-level traits in the group ( $Z_{k} = mean (z_{i} : i \in G_{k})$ ), and the group-level fitness $W_{k}$ is the total entity-level fitness ( $W_{k} = \sum_{i} w_{i}$ for $i \in G_{k}$ ). What is being tracked across generations is the entity. Groups appear only as a useful re-grouping of entity-level events.
MLS2: the group has its own collective trait (e.g., colony size, division of labour) and its own collective fitness (e.g., number of daughter colonies produced). Groups are the entities being tracked across generations; the within-group composition matters only insofar as it affects the group's own properties.

Hardy-Weinberg [^hardy-weinberg-link] supplies the null against which any non-zero between-group covariance is read as group selection. If allele frequencies are at equilibrium and groups are sampled independently with no selection at any level, the between-group covariance $Cov_{k} (W_{k}, Z_{k})$ is zero in expectation; an observed non-zero value diagnoses selection somewhere in the recursive decomposition.

[^hardy-weinberg-link]: See unit 19.02.01 pending(19.02.01) for the formal statement of Hardy-Weinberg and the diagnosis of deviations.

Hamilton's rule (Hamilton 1964; Maynard Smith 1964). For social behaviour, write $c$ = the actor's cost (lost fitness), $b$ = the recipient's benefit (gained fitness), $r$ = the coefficient of relatedness between actor and recipient (probability that a randomly chosen allele in the actor is identical-by-descent to one in the recipient). Then a behaviour is favoured by selection iff

r b > c .

Equivalent reformulation: the inclusive fitness of an actor — its own fitness plus $r$ -weighted contributions to relatives' fitness — is maximised by behaviours satisfying $r b > c$ . Hamilton's rule and the multi-level Price equation are mathematically equivalent in their predictions over a wide class of cases (Hamilton himself derived inclusive fitness using a primitive version of the Price equation). Whether they say the same thing about what selection acts on is the open question. Sober (1984, 2010) argues yes: kin selection is one specific instance of group selection, with the "group" being the relevant kin-correlation neighbourhood. Dawkins (1982) argues no: kin selection is gene-level selection across vehicles, and group selection requires real group-level fitness in the MLS2 sense.

Counterexamples to common slips

Mathematical equivalence does not imply conceptual equivalence. Two formal decompositions of the same data can be related by an invertible transformation and still disagree on what causes what. Hamilton's rule and the multi-level Price equation are inter-translatable as formal identities; whether they pick out the same causal structure is a separate, partially independent question.
High copying fidelity is not the only replicator criterion. Dawkins's argument that only genes are replicators because only genes are copied with high fidelity is empirically contested. Genome-level inheritance in clonal organisms, cytoplasmic inheritance, and the supra-generational stability of some symbiotic associations all furnish candidate replicators above the gene level. The argument as Dawkins states it presupposes diploid sexual reproduction.
Group selection requires neither species benefit nor consciously cooperating individuals. Wynne-Edwards 1962 invoked species-level benefit and was rightly criticised; multi-level selection theory in the post-1980s sense does not. A group needs only to bear fitness as a group — to produce more groups when it has certain compositional properties — for the between-group component of the Price equation to be non-zero.
"The gene is the unit of selection" is ambiguous. It can mean (i) a particular allele at a particular locus is what natural selection acts on, in any context; or (ii) the appropriate level of accounting for selection is always the allele-frequency change, even when the causal action is at higher levels. Dawkins is committed to (ii) and sometimes to (i); the multi-level selection literature accepts (ii) as a calculational convenience but rejects (i).

Key theorem with proof [Intermediate+]

The closest analogue of a theorem in this domain is the Price equation as an exact accounting identity, together with its multi-level extension. Stating and proving it makes the formal apparatus that the entire debate rests on transparent.

Theorem (Price equation, finite-population version). Let a population of $N$ entities with traits $z_{i}$ and absolute fitnesses $w_{i}$ produce a next generation whose mean trait is $\overset{z}{ˉ}^{'}$ . Let $Δ z_{i}$ denote the difference between the mean trait value of $i$ 's offspring and $z_{i}$ , with the convention $Δ z_{i} := 0$ if $w_{i} = 0$ . Then

\overset{w}{ˉ} (\overset{z}{ˉ}^{'} - \overset{z}{ˉ}) = Cov (w_{i}, z_{i}) + E [w_{i} Δ z_{i}],

where $Cov$ and $E$ are the unweighted population covariance and expectation over $i = 1, \dots, N$ .

Proof. By definition of the next-generation mean trait,

\overset{z}{ˉ}^{'} = \frac{\sum _{i} w _{i} ( z _{i} + Δ z _{i} )}{\sum _{i} w _{i}} = \frac{\sum _{i} w _{i} z _{i} + \sum _{i} w _{i} Δ z _{i}}{N w ˉ} .

Multiply both sides by $\overset{w}{ˉ}$ and subtract $\overset{w}{ˉ} \overset{z}{ˉ}$ :

\overset{w}{ˉ} (\overset{z}{ˉ}^{'} - \overset{z}{ˉ}) = \frac{1}{N} (i \sum w_{i} z_{i} + i \sum w_{i} Δ z_{i}) - \overset{w}{ˉ} \overset{z}{ˉ} = (\frac{1}{N} i \sum w_{i} z_{i} - \overset{w}{ˉ} \overset{z}{ˉ}) + \frac{1}{N} i \sum w_{i} Δ z_{i} .

The first parenthesis is the population covariance $Cov (w_{i}, z_{i}) = E [w_{i} z_{i}] - E [w_{i}] E [z_{i}]$ . The second is the expectation $E [w_{i} Δ z_{i}]$ . Combining,

\overset{w}{ˉ} (\overset{z}{ˉ}^{'} - \overset{z}{ˉ}) = Cov (w_{i}, z_{i}) + E [w_{i} Δ z_{i}] . □

Multi-level extension. Suppose the entities are partitioned into groups $G_{1}, \dots, G_{M}$ , with group sizes $n_{k} = ∣ G_{k} ∣$ , group mean traits $Z_{k} = n_{k}^{- 1} \sum_{i \in G_{k}} z_{i}$ , group mean fitnesses $\overset{w}{ˉ}_{k} = n_{k}^{- 1} \sum_{i \in G_{k}} w_{i}$ , and within-group changes $Δ z_{k} = \overset{w}{ˉ}_{k}^{- 1} (\overset{w}{ˉ}_{k} Z_{k}^{'} - \overset{w}{ˉ}_{k} Z_{k})$ where $Z_{k}^{'}$ is the next-generation group mean. Then

\overset{w}{ˉ} (\overset{z}{ˉ}^{'} - \overset{z}{ˉ}) = Cov_{k} (\overset{w}{ˉ}_{k}, Z_{k}) + E_{k} [\overset{w}{ˉ}_{k} Δ z_{k}],

where $Cov_{k}, E_{k}$ are the size-weighted covariance and expectation over groups.

Proof. Apply the Price equation to each group $G_{k}$ treated as its own population:

\overset{w}{ˉ}_{k} (Z_{k}^{'} - Z_{k}) = Cov_{i \in G_{k}} (w_{i}, z_{i}) + E_{i \in G_{k}} [w_{i} Δ z_{i}] .

Take the size-weighted average over $k$ , using $\sum_{k} n_{k} = N$ and $\sum_{k} n_{k} \overset{w}{ˉ}_{k} = N \overset{w}{ˉ}$ . The size-weighted average of $\overset{w}{ˉ}_{k} (Z_{k}^{'} - Z_{k})$ is $\overset{w}{ˉ} (\overset{z}{ˉ}^{'} - \overset{z}{ˉ})$ minus a correction term that vanishes when group sizes are constant. By the law of total covariance,

Cov (w_{i}, z_{i}) = Cov_{k} (\overset{w}{ˉ}_{k}, Z_{k}) + E_{k} [Cov_{i \in G_{k}} (w_{i}, z_{i})],

and substituting into the single-level Price equation,

\overset{w}{ˉ} (\overset{z}{ˉ}^{'} - \overset{z}{ˉ}) = Cov_{k} (\overset{w}{ˉ}_{k}, Z_{k}) + E_{k} [Cov_{i \in G_{k}} (w_{i}, z_{i})] + E [w_{i} Δ z_{i}] = Cov_{k} (\overset{w}{ˉ}_{k}, Z_{k}) + E_{k} [\overset{w}{ˉ}_{k} Δ z_{k}],

where the last equality uses the within-group Price equation per $k$ . $□$

Interpretation. The between-group covariance $Cov_{k} (\overset{w}{ˉ}_{k}, Z_{k})$ is the contribution of selection at the group level — groups with high trait means reproduce more. The within-group expectation $E_{k} [\overset{w}{ˉ}_{k} Δ z_{k}]$ is the contribution of selection within groups — entities with high traits reproduce more inside each group. Neither term is privileged formally; the partition is an algebraic identity.

The unit-of-selection problem arises because the same allele-frequency change can be partitioned more than one way. A given $\overset{w}{ˉ} (\overset{z}{ˉ}^{'} - \overset{z}{ˉ})$ admits the gene-eye decomposition (where each gene-copy is one "entity" $i$ with its own $w_{i}$ ), the individual-level decomposition (entities are organisms, groups are kin-correlated neighbourhoods), the colony-level MLS1 decomposition (entities are organisms, groups are real spatial colonies), and the colony-level MLS2 decomposition (entities are colonies treated as wholes). Each gives a well-defined numerical answer to "how much selection is happening at each level," and the answers do not in general agree. Whether any of these decompositions is the correct one is the substantive question — the formalism does not answer it.

Exercises [Intermediate+]

Exercise 1 (easy, short answer).

State Lewontin's three conditions for natural selection. For each, identify one implicit assumption the condition leaves unstated.

Hint

The three conditions are variation, differential fitness, and heritability. Each one presupposes something about the entity type or the timescale that Lewontin does not make explicit.

Answer

(1) Phenotypic variation. Implicit: the entities are individuable — there is a fact of the matter about which entity is which, both within a generation and across generations. For genes this is uncontroversial (sequence identity), for organisms mostly so, for groups (especially spatially overlapping or socially fluid groups) far less so. (2) Differential fitness. Implicit: the "fitness" $w_{i}$ is a property of the entity itself, not jointly of the entity and its environment. In multi-level cases this assumption breaks: an individual's fitness depends on its group composition, so what counts as "the entity's fitness" already presupposes a partition. (3) Heritability. Implicit: there is a generational structure with a well-defined transmission relation from parents to offspring. For genes (vertical transmission, simple inheritance) this is unproblematic. For social groups (membership fluxes through migration; "offspring" groups inherit only some of the composition), it requires careful definition of the parent-offspring relation. Sober 1984 and Godfrey-Smith 2009 elaborate these implicit moves.

Exercise 2 (medium, short answer).

Produce a counterexample to the claim "only entities with high replication fidelity can be units of selection." Your counterexample should be a real biological case in which an entity with imperfect transmission is nevertheless plausibly subject to natural selection.

Hint

Cultural transmission, organelles with non-Mendelian inheritance, and the supra-organismic stability of host-microbe associations are three candidate domains.

Answer

Cultural traits in primates and humans are inherited with low fidelity by the standards Dawkins applies to genes — a tool-use technique copied by a juvenile from a parent has many opportunities to drift, be re-invented, or be lost. Yet the cultural-evolution literature (Boyd & Richerson 1985; Henrich 2015) shows that cultural traits are subject to selection-like dynamics: variants with higher payoff propagate, the population mean shifts, and the changes are heritable in the relevant transgenerational sense. The fidelity is far below allele-copying fidelity but well above zero, and the resulting dynamics satisfy Lewontin's three conditions. A second case: holobiont selection in coral-algal symbioses, where the holobiont's composition is partially heritable through vertical transmission of symbionts and partially through environmental acquisition; the inheritance fidelity is intermediate but the holobiont's fitness varies systematically with composition. In both cases, Dawkins's high-fidelity threshold is empirical rather than conceptual, and the empirical bar is not set as high as he supposed.

Exercise 3 (medium, open-ended).

Write a paragraph defending kin selection as gene-selection-in-disguise. Use Hamilton's rule and the inclusive-fitness reformulation as the central technical move.

Hint

The argument is: $r b > c$ is most naturally read as "an allele copy in the actor that promotes helping behaviour propagates because the r-weighted helped relative carries copies of the same allele." The level at which the selection is acting is the allele.

Answer

Kin selection is gene-selection-in-disguise because the conserved quantity Hamilton's rule tracks is the allele copy, not the organism. When an actor with allele $A$ at a behavioural locus performs a costly action helping a sibling, the actor's $A$ -copy loses fitness $c$ ; the sibling, by symmetry of meiosis, carries an $A$ -copy with probability $r = 1/2$ , and that $A$ -copy gains fitness $b$ . The condition for an increase in the frequency of $A$ in the next generation is $r b > c$ — the same condition Hamilton derived. Read this way, every "altruistic" behaviour is explained without invoking groups at all: it is allele copies (sometimes in different organisms) increasing the frequency of their own kind. The mathematical apparatus (inclusive fitness) is a re-bookkeeping of gene-level reproductive success, generalising the simple direct-fitness account to include the indirect propagation of allele copies through related vehicles. The organism remains a vehicle in Dawkins's sense; its perceived "interests" diverge from the gene's only because the organism is what we observe. The Hamilton's-rule arithmetic gives the gene's-eye view in formal dress: alleles select for their own propagation, including via copies in other bodies.

Exercise 4 (medium, open-ended).

Write the opposing paragraph: defend kin selection as group-selection-in-disguise. Use the multi-level Price decomposition as the central technical move.

Hint

The argument is: kin neighbourhoods are groups, in the sense the multi-level Price equation requires; the between-group covariance is the kin-selection term re-expressed. Sober 1984, 2010; Wilson & Sober 1998.

Answer

Kin selection is group-selection-in-disguise because the kin neighbourhood — the set of relatives sharing higher-than-baseline rates of identity-by-descent at the relevant locus — is exactly the kind of population substructure that produces a non-zero between-group covariance in the Price equation. When relatives interact preferentially, the population is partitioned (often non-spatially, sometimes spatially) into groups whose mean trait values vary; groups containing more help-givers have higher mean fitness because their members interact mostly with each other; the between-group component of the multi-level Price equation is non-zero. Hamilton's rule $r b > c$ is the threshold at which the between-group selection term outweighs the within-group selection term against altruism: $r$ corresponds formally to the between-group correlation coefficient, $b$ to the group-level fitness gain, $c$ to the within-group cost. The "kin" in "kin selection" is what makes the population structure that group selection requires. To insist on calling the same arithmetic gene-level rather than group-level is to privilege one re-partition of the same identity over the other on grounds the formalism does not supply. Hamilton himself, late in his career, accepted this reading; Sober and Wilson 1998 made it canonical in the multi-level-selection literature.

Exercise 5 (hard, symbolic).

Derive the single-level Price equation from first principles, starting from the definition $\overset{z}{ˉ}^{'} = (\sum_{i} w_{i} z_{i}^{'}) / (\sum_{i} w_{i})$ where $z_{i}^{'} = z_{i} + Δ z_{i}$ . Show every step.

Hint

Expand the numerator, subtract $\overset{w}{ˉ} \overset{z}{ˉ}$ , and recognise the covariance and expectation forms. The proof is short but every step must be explicit.

Answer

Start from $\overset{z}{ˉ}^{'} = \sum_{i} w_{i} z_{i}^{'} / \sum_{i} w_{i}$ , with $z_{i}^{'} = z_{i} + Δ z_{i}$ the mean trait of $i$ 's offspring (with convention $Δ z_{i} = 0$ if $w_{i} = 0$ , so the product $w_{i} Δ z_{i}$ is well-defined uniformly).

Step 1. Expand the numerator: $\sum_{i} w_{i} z_{i}^{'} = \sum_{i} w_{i} z_{i} + \sum_{i} w_{i} Δ z_{i}$ .

Step 2. Multiply both sides by $\sum_{i} w_{i} = N \overset{w}{ˉ}$ :

N \overset{w}{ˉ} \overset{z}{ˉ}^{'} = i \sum w_{i} z_{i} + i \sum w_{i} Δ z_{i} .

Step 3. Subtract $N \overset{w}{ˉ} \overset{z}{ˉ}$ from both sides:

N \overset{w}{ˉ} (\overset{z}{ˉ}^{'} - \overset{z}{ˉ}) = i \sum w_{i} z_{i} - N \overset{w}{ˉ} \overset{z}{ˉ} + i \sum w_{i} Δ z_{i} .

Step 4. Recognise the first two terms on the right as $N$ times a covariance:

i \sum w_{i} z_{i} - N \overset{w}{ˉ} \overset{z}{ˉ} = N (\frac{1}{N} i \sum w_{i} z_{i} - \overset{w}{ˉ} \overset{z}{ˉ}) = N Cov (w_{i}, z_{i}),

using the identity $Cov (X, Y) = E [X Y] - E [X] E [Y]$ .

Step 5. Recognise the third term as $N$ times an expectation:

i \sum w_{i} Δ z_{i} = N \cdot \frac{1}{N} i \sum w_{i} Δ z_{i} = N E [w_{i} Δ z_{i}] .

Step 6. Divide by $N$ :

\overset{w}{ˉ} (\overset{z}{ˉ}^{'} - \overset{z}{ˉ}) = Cov (w_{i}, z_{i}) + E [w_{i} Δ z_{i}] .

This is the Price equation as stated. The covariance term measures selection (high- $w$ entities have high $z$ or vice versa); the expectation term measures transmission bias (offspring of $i$ deviate from $z_{i}$ on average). Under perfect transmission $Δ z_{i} \equiv 0$ and only the covariance term remains; in cultural or epigenetic inheritance the expectation term carries the bulk of the dynamics.

Exercise 6 (hard, short answer).

Sober and Wilson (1998) argue that group selection is "real but rare" — formally well-defined, but operative only when between-group variance is large relative to within-group variance. State two empirical phenomena that have been cited as evidence of group selection in this sense, and one common objection to each.

Hint

Eusocial-insect colony selection and microbial-virulence evolution are the two canonical examples in Unto Others and follow-up literature. The objections are usually that the same data can be accounted for by kin selection or by frequency-dependent individual selection.

Answer

(1) Eusocial insect colonies. Honey-bee and ant colonies vary in average worker behaviour (hygienic vs. non-hygienic; aggressive vs. docile); colonies with hygienic workers have lower pathogen load and higher colony reproductive success. The between-colony fitness variance is large because the colony's reproduction (production of new queens and males) is the unit of inter-colony selection. Objection: every worker in a colony is closely related to every other worker ( $r$ values between 0.5 and 0.75 depending on the colony's mating structure), so Hamilton's rule with $r \approx 0.5$ predicts the same outcome — the colony-level selection signal and the gene-level kin-selection signal cannot be empirically distinguished. (2) Reduced virulence in horizontally transmitted parasites. Myxoma-virus-rabbit dynamics in Australia (Fenner 1965) show that virulence dropped over decades because virulent strains kill hosts before transmitting, while less-virulent strains transmit more, so between-host (between-group, where hosts are groups of replicating viruses) selection favours mildness over within-host fast-replication selection for virulence. Objection: the same dynamics can be described as kin selection between viruses within hosts (relatives sharing a host benefit from collective restraint) or as direct frequency-dependent selection on individual virus reproductive strategies in different epidemiological contexts; the multi-level decomposition adds little explanatory work beyond the within-host / between-host bookkeeping that virologists already do. The pattern across both cases — that the multi-level account is empirically equivalent to a kin-selection account — is the standing complaint against the post-1998 group-selection programme.

Exercise 7 (hard, short answer).

Hamilton's rule $r b > c$ is sometimes called the "fundamental theorem of social evolution." State two ways it fails or requires modification, with a brief example of each.

Hint

Think about (1) non-additive fitness interactions (synergy) and (2) population structure where $r$ is hard to define unambiguously (group-living species with weak kin structure but strong reciprocity).

Answer

(1) Non-additive interactions (synergy). The basic Hamilton's rule assumes that the actor's cost $c$ and the recipient's benefit $b$ combine additively with their other fitness components. When the helping behaviour produces synergistic gains — three cooperators together do far better than the sum of their pairwise contributions, as in cooperative hunting in lions or wolves — the additive Hamilton's rule under-predicts the threshold for cooperation's evolution. Queller (1985) extended Hamilton's rule to include a synergy term $d$ ; the modified condition is $r b + r^{'} d > c$ , where $r^{'}$ is the relatedness among the synergising group. Example: cooperative breeding in social vertebrates where helpers contribute to nest defence; the marginal benefit of an additional helper is not linear, and the basic Hamilton's rule misses by a measurable amount. (2) Reciprocity without kinship. When organisms in a group cooperate via repeated interactions without being related — vampire bats sharing blood meals, primate grooming exchanges — relatedness $r$ is near zero, but cooperation is still favoured by reciprocity (Trivers 1971) or by reputation-based partner choice. Hamilton's rule says no cooperation should evolve under $r \approx 0$ ; the data say otherwise, and the gap is filled by reciprocity theory rather than kin-selection theory. The two theories overlap when the same individuals are both relatives and repeated interactors, but in pure reciprocity cases Hamilton's rule does not apply at all.

Levels of organisation and the major-transitions framework [Master]

The unit-of-selection problem inherits its modern shape from a sequence of monographs running from Williams 1966 through Okasha 2006. The two empirically distinctive moves in this literature are (i) the partial rehabilitation of group selection after 1975 and (ii) the major-transitions argument of Maynard Smith and Szathmáry (1995), which reframes the problem as a question about how individuality itself evolves up the biological hierarchy.

Maynard Smith and Szathmáry identify eight major evolutionary transitions: replicating molecules to populations of molecules in compartments; independent replicators to chromosomes; RNA as gene and enzyme to DNA + protein; prokaryotes to eukaryotic cells; asexual clones to sexual populations; protists to multicellular organisms; solitary individuals to colonies with non-reproductive castes; primate societies to human societies with language. Each transition is a change in the unit of selection itself: entities that previously bore fitness as separate units come, after the transition, to function as a single higher-level unit whose components no longer have independent reproductive interests.

The transition framework converts the unit-of-selection problem into a developmental one. At the start of a transition the lower-level entities are units of selection in Lewontin's sense; at the end the higher-level entity is. The question of which level is the real unit is, on this view, locally answerable in any given empirical case but not globally fixed across time — the answer depends on where in a major transition the lineage in question sits.

This is the framing Okasha (2006) develops at book length, and it shapes the post-2006 phil-of-biology literature. Recent treatments include Bourrat (2015, 2021), Birch (2017), and Godfrey-Smith (2009, 2016). The continuing debate is whether the major-transitions framework dissolves the unit-of-selection problem (because the question has different right answers at different evolutionary moments) or whether it exposes a deeper failure of analytic resolution (because every supposedly distinct level turns out to be unstable across long enough timescales). Okasha tracks the position space carefully without endorsing a single resolution; the master-tier sections below survey the four major positions in the contemporary literature.

Gene-centric selectionism (Williams 1966; Dawkins 1976, 1982; with formal underpinning in Hamilton 1964 and Maynard Smith 1964)

The position: selection acts on alleles, full stop. Organisms are vehicles built by alleles to propagate themselves; groups are loose aggregations of vehicles. Kin selection, frequency-dependent selection, and reciprocity are all decoded as allele-level strategies playing out across vehicles. The gene is the unique replicator because only the gene has the inheritance fidelity required for selection to accumulate over evolutionary timescales — a gene-copy persists with high probability across generations, while organisms and groups are dispersed every generation by meiosis or migration.

The strongest argument for the position is the accounting argument: any evolutionary change must, in the end, register as a change in allele frequencies; whatever else happened along the way, the gene's-eye bookkeeping is the one that closes the books. The Hamilton-Maynard Smith inclusive-fitness apparatus shows that even apparently group-benefiting behaviours can be re-derived in this bookkeeping.

The strongest argument against the position, due to Sober 1984 and elaborated by Wilson and Sober 1998, is that gene-level accounting does not preserve causal information. Two scenarios with identical allele-frequency outcomes can have radically different causal structures — in one, the relevant selection is between groups containing different allele frequencies; in another, the same alleles undergo selection within a single panmictic population. The accounting cannot distinguish them; the causal structure can. Insisting on gene-level bookkeeping is therefore a methodological choice rather than a discovery about what selection acts on.

Hierarchical pluralism (Gould & Lewontin 1979; Sober 1984; Lloyd 1988, 2017)

The position: selection acts at multiple levels simultaneously and the levels are not reducible to each other. There is selection on alleles, selection on individuals, selection on species lineages (Stanley 1975's "species selection"); these are distinct causal processes contributing to evolutionary change. No level is privileged.

The strongest argument for the position is causal heterogeneity: the kinds of biological structure that drive selection are different at different levels. Allele-level selection works through differential replication of DNA sequences. Individual-level selection works through differential survival and reproduction of phenotypically variable organisms. Species-level selection works through differential origination and extinction rates among lineages. Each kind of selection has its own characteristic dynamics, its own characteristic counterfactuals, and its own characteristic stabilising or destabilising effects. To collapse them all to "gene-level" is to flatten a structured causal hierarchy.

The strongest argument against the position is the redundancy worry: when the multi-level decomposition is mathematically equivalent to a single-level account, what is added by the multi-level description? If every empirically distinguishable claim of the hierarchical pluralist can also be made in gene-eye language, the pluralism may be a notational preference rather than a substantive metaphysical thesis. Sober 1984 and Lloyd 2017 respond that the equivalence of formal predictions is consistent with the multi-level description being the causally honest one, while the gene-eye description is a derived accounting — but this requires an independent argument for what "causally honest" means, which the literature has not converged on.

Multi-level selection theory (Wilson 1975; Wilson & Sober 1994, 1998; Wilson 2015)

The position: group selection is quantitatively important in empirically identifiable cases. The Price equation's between-group covariance term is non-zero in many real biological systems, and where it is non-zero, group selection is doing causal work. The MLS1 / MLS2 distinction allows for two flavours: groups as collections of entities (MLS1, where what is tracked across generations is the entity but the group structures the entity-level fitness) and groups as themselves entities (MLS2, where what is tracked is the group). Both flavours are real; both contribute.

The strongest argument for the position is its account of altruism. Behaviours costly to the actor and beneficial to recipients are widespread in nature, and the kin-selection apparatus, while formally adequate, requires extending "kin" to cover any positively correlated population structure — at which point the kin-selection account is the multi-level account, just with relatedness $r$ standing in for between-group covariance. Wilson and Sober argue that calling this kin selection rather than group selection is historical contingency, not theoretical economy.

The strongest argument against is the equivalence-with-kin-selection objection (Maynard Smith 1976, 1998; West, Griffin & Gardner 2007, 2008). Every published case of "group selection" in the post-1980 sense can be re-described as kin selection at no informational loss; the choice between the two formulations is a matter of presentational preference, not a discovery about which level matters. West and colleagues argue that the multi-level vocabulary actively confuses applied biologists because it suggests a separate phenomenon when only a re-bookkeeping is on offer. The two camps in the field (kin-selection partisans and group-selection partisans) have failed to converge on whether this objection is decisive.

Causal-process interpretation (Sterelny 1996; Godfrey-Smith 2009; Bourrat 2015, 2021)

The position: a level of selection is real iff the causal processes operating at that level are minimally autonomous — they have their own characteristic timescales, their own characteristic feedback structures, and their own characteristic stability conditions. This is a coarse-graining criterion: levels are real if they admit a self-consistent dynamics at their own scale.

The strongest argument for the position is its handling of intermediate cases. A microbial mat is, on different measurements, a population of cells, a community of distinct strains, or a single quasi-organism with a coordinated metabolism. Each description captures real causal structure at a different scale. The causal-process view says all three are valid units of selection, for the dynamics that operate at their scales; the question of which is the true unit is malformed. Godfrey-Smith 2009 develops this through the notion of Darwinian populations admitting degrees: a population of entities is more or less Darwinian depending on how strongly each Lewontin condition is satisfied, and the unit-of-selection question reduces to which level has the highest Darwinian rating in a given empirical context.

The strongest argument against is that minimally autonomous is doing too much work without independent definition. What counts as a separate causal timescale or feedback structure depends on which abstractions one is willing to accept; reasonable disagreement here can produce reasonable disagreement about which levels count. The view is sometimes accused of pluralism-by-stipulation: any level whose causal coarse-graining is plausible counts, which can include arbitrarily many. Bourrat 2021 acknowledges the worry and proposes operational metrics for "minimal autonomy"; the metrics have not yet been independently validated.

Reading a contemporary primary paper [Master]

Pick Okasha's 2018 monograph Agents and Goals in Evolution (Oxford University Press), Ch. 3 (the chapter on type-2 agency and multi-level selection). Okasha asks whether collective agents — colonies, multi-cellular organisms — can be said to have goals in the same sense in which individual organisms can, and whether this affects the unit-of-selection picture.

The argument runs in three parts. First, Okasha distinguishes type-1 agency (an organism behaving as if to maximise its inclusive fitness) from type-2 agency (an organism behaving as if to maximise some other quantity, e.g., colony output, while sacrificing its own inclusive fitness). Standard inclusive-fitness theory predicts type-1 agency under generic conditions; type-2 agency is what would be evolutionarily unexpected on a strict gene-centric account.

Second, Okasha argues that the cellular slime mould Dictyostelium discoideum — a unicellular organism that, under starvation, forms a multi-cellular fruiting body in which roughly $20%$ of the cells die to form a sterile stalk supporting the remaining $80%$ that become reproductive spores — exhibits type-2 agency at the level of the individual amoeba. Each amoeba's "decision" to become a sterile stalk cell or a fertile spore cell is partially predicted by the cell's allelic state, but the decision-rule appears to be sensitive to colony-level outcomes in a way that suggests the colony, not the amoeba, is the agent whose fitness the behaviour serves.

Third, Okasha concludes that the unit-of-selection literature must accommodate cases in which the agent of selection (the entity whose fitness is causally efficacious) is at a different level from the replicators whose allele frequencies change. He proposes a four-fold matrix — replicators × interactors × agents × entities — and asks the field to specify, for each empirical case, which row and column apply. The 2018 book defends this multi-dimensional approach against both Dawkins-style gene-centrism and Wilson-Sober-style multi-level selection theory as inadequately fine-grained.

Original response. Okasha's framework gains expressive power at the cost of empirical traction: the four-fold matrix can in principle describe any case, but the cost is that no case forces a unique entry. The slime-mould example, taken seriously, illustrates this. Each amoeba's stalk-versus-spore decision is, on Okasha's reading, a type-2 agent decision serving the colony; on a strict kin-selection reading, it is a type-1 agent decision serving the amoeba's allele copies through the closely related spore cells. The two readings agree on every measurable quantity — the proportion of stalk versus spore cells, the inclusive-fitness arithmetic, the colony-level reproductive success — and disagree only on which entity to assign the goal to. The matrix records both readings as legitimate without supplying a procedure for choosing between them, which leaves the underlying problem (multiple decompositions of the same dynamics) untouched. The hierarchical-pluralist view (Sober, Lloyd) is closer to Okasha's framework in spirit but more honest about the residual indeterminacy: where Okasha catalogues, the pluralists name the catalogue as ineliminable. The hidden premise Okasha leaves unstated is that agency talk is doing genuine causal work — that there is a fact about which entity bears the goal, independent of how the bookkeeping is partitioned. Without an argument for that premise, the type-2 agency proposal is a fifth way of saying what Sober and Lloyd had already said three ways: the level question may not have a unique answer.

This is a place where machine-verifiability fails outright. The Price equation can be checked formally (and Mathlib could in principle verify it); the question of whether one decomposition is causally more honest than another decomposition admitting the same algebraic data is a substantive interpretive question with no formal counterpart. Per docs/plans/PHILOSOPHY_PLAN.md §6.3, master-tier philosophy of biology operates at the lowest machine-verifiability level in the curriculum; human review is the only correctness gate.

Connections [Master]

The unit-of-selection debate connects to:

19.02.01 pending(19.02.01) Hardy-Weinberg equilibrium. The null model against which any allele-frequency change is read; the Price equation's between-group component is identically zero in expectation when every level is at Hardy-Weinberg with no selection.
20.05.03 The species problem (pending). The question of what counts as a species is logically downstream of the question of what counts as a unit of selection: species selection (Stanley 1975) requires species to satisfy Lewontin's three conditions as entities.
20.05.04 Fitness (pending). What "fitness" measures and whether different levels can simultaneously bear fitness is a sub-problem of the unit-of-selection problem.
20.07.01 Explanation and theory choice in science (pending). The unit-of-selection debate is a paradigm case of empirical-theoretical underdetermination — two formal frameworks (gene-eye and multi-level) are observationally equivalent on a large class of cases, raising methodological questions about how to choose between them.
19.03.NN Natural selection (pending). The empirical surface on which the philosophical debate plays out; selection equations $Δ p = pq (p s_{1} - q s_{2}) / \overset{w}{ˉ}$ derived in the bio-side unit are the formal target the levels-of-selection debate decomposes.
20.essays.05 Practice (existing). The major-transitions argument's account of how individuality emerges across evolutionary time — separate replicating molecules consolidating into chromosomes, single cells into multicellular organisms — connects to the practice-essay's treatment of how distinctions stabilise into apparent objects through repeated coordination.

Historical & philosophical context [Master]

The contemporary unit-of-selection literature begins with two near-simultaneous papers in 1964. W. D. Hamilton's two-part paper in Journal of Theoretical Biology (Hamilton 1964) introduced inclusive fitness and the apparatus that became Hamilton's rule; in the same year, John Maynard Smith's Nature paper (Maynard Smith 1964) introduced the term "kin selection" and criticised V. C. Wynne-Edwards's 1962 monograph Animal Dispersion in Relation to Social Behaviour, which had defended species-benefit reasoning as the standard explanation for self-limiting reproduction in animal populations. George Williams's 1966 monograph Adaptation and Natural Selection extended the critique into a general programme: group-level adaptations require either the groups themselves to function as units of selection (which the population-genetics conditions rarely permit) or the appearance of group benefit to be a side-effect of individual-level selection.

Williams 1966 set the consensus for the next decade: selection acts on individuals, occasionally on alleles, and only in highly restrictive circumstances on groups. Richard Dawkins's 1976 The Selfish Gene shifted the consensus further toward the gene-level view, both technically (through the replicator-vehicle distinction) and rhetorically (through the gene's-eye view metaphor). David Hull's 1980 paper Individuality and Selection in Annual Review of Ecology and Systematics refined the technical apparatus by introducing the replicator/interactor terminology that most of the subsequent literature uses.

George Price's 1970 Nature paper Selection and covariance introduced the Price equation, the exact accounting identity that has structured every subsequent technical discussion. Price's contribution was partly unrecognised in his lifetime; the equation's centrality became clear only in the 1980s through the work of W. D. Hamilton, who used Price's identity to re-derive his 1964 inclusive-fitness results, and through Steven Frank's Foundations of Social Evolution (Princeton, 1998), which made the Price equation the organising tool of social-evolution theory.

The group-selection revival begins with David Sloan Wilson's 1975 paper A theory of group selection (Proceedings of the National Academy of Sciences 72, 143–146), which proposed a "trait-group" model rescuing group selection from the strict-isolation requirements that Williams 1966 had used to dismiss it. Wilson and Elliott Sober's 1998 monograph Unto Others: The Evolution and Psychology of Unselfish Behavior (Harvard University Press) consolidated the post-1975 multi-level selection theory and made it a serious competitor to the gene-centric view, both in evolutionary biology and in philosophy of biology.

John Damuth and I. L. Heisler's 1988 paper in Biology and Philosophy introduced the MLS1 / MLS2 distinction that Okasha 2006 made standard. John Maynard Smith and Eörs Szathmáry's 1995 monograph The Major Transitions in Evolution (Freeman) reframed the unit-of-selection problem in terms of how individuality itself evolves up the biological hierarchy.

Samir Okasha's 2006 Evolution and the Levels of Selection (Oxford University Press) is the canonical contemporary survey and the master-tier anchor for this unit. Recent contributions include Pierrick Bourrat (2015 Studies in History and Philosophy of Biological and Biomedical Sciences; 2021 Facts, Conventions, and the Levels of Selection, Cambridge), Jonathan Birch (2017 The Philosophy of Social Evolution, Oxford), and Peter Godfrey-Smith's continuing work on Darwinian populations (2009, 2016). The continuing debate in the journals Biology and Philosophy and Philosophy of Science has not converged.

Bibliography [Master]

Hamilton, W. D. (1964). The genetical evolution of social behaviour, I and II. Journal of Theoretical Biology 7, 1–16 and 17–52.
Maynard Smith, J. (1964). Group selection and kin selection. Nature 201, 1145–1147.
Williams, G. C. (1966). Adaptation and Natural Selection. Princeton University Press.
Wynne-Edwards, V. C. (1962). Animal Dispersion in Relation to Social Behaviour. Oliver & Boyd.
Lewontin, R. C. (1970). The units of selection. Annual Review of Ecology and Systematics 1, 1–18.
Price, G. R. (1970). Selection and covariance. Nature 227, 520–521.
Price, G. R. (1972). Extension of covariance selection mathematics. Annals of Human Genetics 35, 485–490.
Dawkins, R. (1976). The Selfish Gene. Oxford University Press. (30th-anniversary ed. 2006.)
Dawkins, R. (1982). The Extended Phenotype. Oxford University Press.
Hull, D. L. (1980). Individuality and selection. Annual Review of Ecology and Systematics 11, 311–332.
Wilson, D. S. (1975). A theory of group selection. Proceedings of the National Academy of Sciences 72, 143–146.
Gould, S. J. & Lewontin, R. C. (1979). The spandrels of San Marco and the Panglossian paradigm. Proceedings of the Royal Society of London B 205, 581–598.
Sober, E. (1984). The Nature of Selection. MIT Press.
Damuth, J. & Heisler, I. L. (1988). Alternative formulations of multilevel selection. Biology and Philosophy 3, 407–430.
Maynard Smith, J. & Szathmáry, E. (1995). The Major Transitions in Evolution. Freeman.
Wilson, D. S. & Sober, E. (1998). Unto Others: The Evolution and Psychology of Unselfish Behavior. Harvard University Press.
Sterelny, K. & Griffiths, P. E. (1999). Sex and Death: An Introduction to the Philosophy of Biology. University of Chicago Press.
Sterelny, K. (1996). The return of the group. Philosophy of Science 63, 562–584.
Lloyd, E. A. (1988). The Structure and Confirmation of Evolutionary Theory. Princeton University Press.
Okasha, S. (2006). Evolution and the Levels of Selection. Oxford University Press.
Frank, S. A. (1998). Foundations of Social Evolution. Princeton University Press.
West, S. A., Griffin, A. S. & Gardner, A. (2007). Social semantics: altruism, cooperation, mutualism, strong reciprocity and group selection. Journal of Evolutionary Biology 20, 415–432.
Godfrey-Smith, P. (2009). Darwinian Populations and Natural Selection. Oxford University Press.
Bourrat, P. (2015). Levels of selection are artefacts of different fitness temporal measures. Ratio 28, 40–50.
Birch, J. (2017). The Philosophy of Social Evolution. Oxford University Press.
Okasha, S. (2018). Agents and Goals in Evolution. Oxford University Press.
Bourrat, P. (2021). Facts, Conventions, and the Levels of Selection. Cambridge University Press.

Prerequisites

19.02.01 pending

Tier anchors

beginner: Sober, *Philosophy of Biology*, 2nd ed. (Westview, 2000), Ch. 4 (the units of selection problem at intro level); Godfrey-Smith, *Philosophy of Biology* (Princeton Foundations of Contemporary Philosophy, 2014), Ch. 3
intermediate: Sober, *Philosophy of Biology* (full); Sober, *The Nature of Selection: Evolutionary Theory in Philosophical Focus* (MIT Press, 1984); Godfrey-Smith, *Darwinian Populations and Natural Selection* (Oxford, 2009), Ch. 5–6
master: Okasha, *Evolution and the Levels of Selection* (Oxford, 2006); Wilson & Sober, *Unto Others: The Evolution and Psychology of Unselfish Behavior* (Harvard, 1998); Sober, *Evidence and Evolution* (Cambridge, 2008); primary lit (*Biology & Philosophy*, *Studies in History and Philosophy of Biological and Biomedical Sciences*, *Philosophy of Science*)

References

TODO_REF
Hamilton, W. D. — *Journal of Theoretical Biology* 7, 1–16 and 17–52 (1964) · Two-part originator paper, *The genetical evolution of social behaviour* — inclusive fitness, Hamilton's rule $rb > c$
TODO_REF
Maynard Smith, J. — *Nature* 201, 1145–1147 (1964) · Originator paper introducing the term *kin selection* and the haystack-model critique of group selection
TODO_REF
Williams, G. C. — *Adaptation and Natural Selection* (Princeton University Press, 1966) · Canonical anti-group-selection monograph; defends the gene/individual locus of adaptation
TODO_REF
Wynne-Edwards, V. C. — *Animal Dispersion in Relation to Social Behaviour* (Oliver & Boyd, 1962) · The high-water mark of naive group selection that Williams 1966 and Maynard Smith 1964 dismantled
TODO_REF
Lewontin, R. C. — *Annual Review of Ecology and Systematics* 1, 1–18 (1970) · *The units of selection* — the three-conditions (variation / differential fitness / heritability) framework for any unit of selection
TODO_REF
Dawkins, R. — *The Selfish Gene* (Oxford University Press, 1976; 30th-anniversary ed. 2006) · Popularisation and defence of gene-replicator selectionism; vehicle-replicator distinction at Ch. 11–13
TODO_REF
Dawkins, R. — *The Extended Phenotype* (Oxford University Press, 1982) · Technical follow-up; replicator-vehicle distinction formalised; case against organism-as-fundamental-vehicle
TODO_REF
Hull, D. L. — *Annual Review of Ecology and Systematics* 11, 311–332 (1980) · *Individuality and selection* — replicator/interactor terminology adopted by most of the subsequent literature
TODO_REF
Sober, E. — *The Nature of Selection: Evolutionary Theory in Philosophical Focus* (MIT Press, 1984) · Canonical analytic-phil treatment; pluralism on levels; causal-decomposition reading of selection
TODO_REF
Wilson, D. S. & Sober, E. — *Unto Others: The Evolution and Psychology of Unselfish Behavior* (Harvard University Press, 1998) · Multi-level-selection-theory monograph; defence of group selection as causally real
TODO_REF
Okasha, S. — *Evolution and the Levels of Selection* (Oxford University Press, 2006) · Master text; MLS1/MLS2 distinction; systematic treatment of the unit-of-selection literature 1966–2006
TODO_REF
Maynard Smith, J. & Szathmáry, E. — *The Major Transitions in Evolution* (Freeman, 1995) · The hierarchy-of-individuality programme; eight major transitions from RNA to language
TODO_REF
Damuth, J. & Heisler, I. L. — *Biology and Philosophy* 3, 407–430 (1988) · Originator paper of the MLS1 / MLS2 terminology adopted by Okasha 2006
TODO_REF
Price, G. R. — *Nature* 227, 520–521 (1970); *Annals of Human Genetics* 35, 485–490 (1972) · The Price equation as exact accounting identity for evolutionary change across levels
TODO_REF
Sterelny, K. & Griffiths, P. E. — *Sex and Death: An Introduction to the Philosophy of Biology* (University of Chicago Press, 1999) · Standard graduate-level survey; replicator pluralism; causal interpretation of selectionism
TODO_REF
Godfrey-Smith, P. — *Darwinian Populations and Natural Selection* (Oxford University Press, 2009) · Population-thinking re-grounding of the unit-of-selection debate; gradations of Darwinian populations
TODO_REF
Sober, E. — *Evidence and Evolution: The Logic Behind the Science* (Cambridge University Press, 2008) · Likelihoodist treatment of selection inference; the unit-of-selection problem as a model-selection question
TODO_REF
Okasha, S. — *Agents and Goals in Evolution* (Oxford University Press, 2018) · Follow-up monograph; type-2 agency, multi-level fitness, persistent ambiguity of MLS decomposition

Reviewer

Tyler (pending external phil-of-biology reviewer per PHILOSOPHY_PLAN §9 — top recruitment priority; phil-of-biology specialist with population-genetics fluency needed for master-tier sign-off)

Estimated time

beginner: 16m
intermediate: 38m
master: 60m