NOTE-184: SEP — Units and levels of selection

Contribution

The entry states in full, as revised in 2024, Lloyd’s “anatomy” of the units-of-selection problem. It separates four research questions. Two come from Hull 1980: the interactor and the replicator, now Griesemer’s reproducer. Two are Lloyd’s own additions from 1988: the manifestor of adaptation and the beneficiary. She pairs the anatomy with a second distinction she calls “extremely influential but usually unrecognized” (§2.5). One school is the Wrightian “Evolutionary Change” school: quantitative genetics, contextual analysis, interest in process. The other is the Fisherian “Adaptationist/KS” school: ESS, inclusive fitness, optimality, interest in engineering adaptation. Six debates are then dissected with both tools: group selection, species selection, genic selectionism, genic pluralism, ETI, and holobionts/demibionts.

Key insight

“Unit of selection” names four roles, not one thing. Being an interactor, an entity that interacts “as a cohesive whole” with its environment so that reproduction is differential (Hull 1980: 318), is “adaptation-neutral” (§2.3). It requires only a higher-level character whose relation to fitness cannot be reduced to lower-level characters. It does not require that the level have accumulated “engineering” adaptations. Much of the literature against group, species or holobiont selection is really an argument that those levels lack engineering adaptations, presented as an argument that they are not units at all.

Assumptions

• A premise of the method: scientific theories are families of formal and informal “model-types”, structures with some parameters left unspecified (§3, n. 30; semantic view, Lloyd 1988 [1994]; Beatty 1980; Griesemer 1984).
• Interactors are identified empirically and statistically. The standard is contextual analysis, a multilevel partial regression of individual fitness on individual and group characters. Earnshaw’s summary is w Δz = a Var(Z) + b Var(z) (§4.1). The additivity approach and screening off are named alternatives (n. 50).
• Adaptation has two readings. A “selection-product” adaptation is any trait that is a direct result of selection. An “engineering” adaptation is an accumulated, design-like change beyond the ancestral range of variation (§2.3). The engineering reading is taken to meet both the etiological and the systems-analysis accounts of function (Allen 2002).
• The Evolutionary Change school’s empirical record is taken as established: group-selection experiments, IGEs, field contextual analyses, Muir’s hens, Wade & Breden on cheaters. It is cited, not re-examined.

Key results

The entry argues rather than proves. Its main theses:

• Four RQs (§2, §5). What is the interactor? What is the reproducer (or the narrower replicator)? What manifests accumulated engineering adaptations at a level? What is the long-run beneficiary? They are “distinct and separable”. One entity may fill several roles, but the questions are not interchangeable (§2.5).
• Two schools (§3). They ask different RQs and make different modelling assumptions: “universal epistasis and pleiotropy vs. additive genetic effects”, small subdivided vs large panmictic populations (Wade & Goodnight 1998: 1547, quoted §3.3). The Adaptationist/KS research question has two parts. The first identifies an optimal higher-level engineering adaptation. The second finds a KS or inclusive-fitness process that could produce it (§3.2).
• Group selection (§4.1). G. C. Williams’s 1966 attack was on group benefit. Its critics “never considered groups as interactors, only as interactors-cum-manifestors”. Williams later accepted the interactor-only criterion (1990: 504; 1992). The claimed formal equivalence of kin and group selection models rests heavily on Dugatkin & Reeve 1994. Lloyd calls that argument parametrically and dynamically insufficient, because allelic parameters depend on genotypic fitnesses. On cheaters, a selfish mutant spreads only if (1/2k) b_Total < c, which is the reverse of Hamilton’s rule (Wade & Breden 1981, via Wade 2016: 146–147). Cheaters are therefore held at a mutation–selection-like balance and are not a “relentless threat”.
• Price equation (§4.1). In the multilevel form ΔZ = Cov(Ω_k, Z_k) + E(Ω_k ΔZ_k), the single-level transmission-bias term becomes a “within-collective selection term”. Following Griesemer, this is either a conflation of inheritance with lower-level selection or an “entanglement”. That may be a reason to abandon the Price formalism for multilevel selection. Lloyd leaves it as “an open question concerning best methods”.
• Species selection (§4.2). The paleontologists’ early demand for “emergent properties” (species-level engineering adaptations) fused the interactor and manifestor roles. Lloyd holds that the current accepted definition of species selection is an interactor definition, with geographic range as the standard trait (Jablonski).
• Genic selectionism (§4.3). Dawkins’s question is the beneficiary question, answered by definition, combined with the manifestor question. Dawkins accepts that vehicles are the “proximal subjects of selection” (1982a: 47). His argument against vehicles targets only the organism as inclusive-fitness maximiser, the “Central Theorem”.
• Genic pluralism (§4.4). Waters, Sterelny and Kitcher rebuild genic models by renaming interactors as “genic environments”. They need the interactor analysis to do so, so their models are derivative and introduce “no new allelic causes”. Lloyd argues their denial that “targets of selection” exist (Sterelny & Kitcher 1988: 359) goes beyond anything Dawkins or Williams held.
• ETI (§4.5). Maynard Smith & Szathmáry are committed to a gene-centred approach (1995: 8). Godfrey-Smith’s “Darwinian populations” deny that the unit question is ambiguous (2009: 111). Lloyd’s reply is that his analysis still uses all four roles (n. 117). The ETI model-type involves reproduction at the higher level and suppression of the lower level, and she holds it does not generalise to multilevel evolution.
• Holobionts (§4.6). Opponents apply the ETI model-type with cheater dynamics, and demand vertical transmission. Community-genetics models incorporating IGEs and epistasis (Drown & Wade 2014; Lloyd & Wade 2019) are said to show that strict vertical transmission is not required, and that conflict can be self-limiting while mutualism is self-accelerating. Demibionts are asymmetric partial dependences on a continuum from independence to full holobiont.

Claims

id	claim	strength	support
C1	”Unit of selection” covers four distinct research questions: interactor, reproducer, manifestor of adaptation and beneficiary	moderate	conceptual analysis (§2), with the historical uses traced in §4; Maynard Smith 2001 is quoted agreeing on the diagnosis
C2	The interactor role is adaptation-neutral, and taking an interactor to be a manifestor is “the most common error” in the debates	moderate	informal argument plus case histories (§2.1, §§4.1–4.6); “most common” is asserted, not counted
C3	Williams’s 1966 critique concerned group benefit, not group selection, and Williams later accepted the interactor criterion	moderate	textual citation (Williams 1966; 1990: 504; 1992: 27)
C4	Claims that kin and group selection models are formally equivalent rest on an inadequate argument (Dugatkin & Reeve 1994)	weak–moderate	argument (parametric and dynamic insufficiency) and citations (Lloyd & Feldman 2002; Bijma & Wade 2008); the opposing technical literature (Marshall 2011; Gardner et al.) is quoted but not refuted in detail
C5	The “cheater problem” for group-selected altruism and for mutualisms is a fallacy	weak–moderate	a result reported secondhand (Wade & Breden 1981, via Wade 2016) and one empirical survey (Jones et al. 2015: no well-studied mutualism “cheater” meets a fitness-based criterion); partisan framing
C6	Genic pluralism’s genic models are derivative renamings of multilevel models and introduce no new causes	moderate	informal argument (§4.4, n. 103: the pluralists use the same interactor-identifying tools)
C7	The multilevel Price equation entangles inheritance with lower-level selection, which may be a reason to abandon it for multilevel selection	weak	informal argument credited to Griesemer (personal correspondence, n. 63); explicitly left open
C8	Models of holobiont evolution with IGEs and epistasis show that vertical transmission is not required and that mutualism tends to self-accelerate	weak (as reported here)	citation of Drown & Wade 2014, Lloyd & Wade 2019 and Koskella & Bergelson 2020; the models are not presented
C9	The research programmes of Godfrey-Smith, Doolittle & Booth, Inkpen & Doolittle, Bourrat and Okasha reintroduce the four roles while denying or neglecting them	weak	assertion with selective quotation (§2.5, §4.6, n. 80, n. 117)

Method

Lloyd’s “logic of research questions” (Lloyd 2015). For each debate she states the research question, or the combination of questions, each party is answering. She then states the possible and responsive answers within that party’s model-types. Finally she shows where a disagreement is really a difference of question or a fusing of roles. She pairs this with a sociological sort into schools (§3). In §4 she applies both to six debates.

Concepts

• interactor — an entity that “interacts, as a cohesive whole, directly with its environment in such a way that replication is differential” (Hull 1980: 318). It is identified through “emergent” fitness–character relations at a level, for example by contextual analysis.
• replicator / reproducer — replicator: “an entity that passes on its structure directly in replication” (Hull 1980: 318). Reproducer: Griesemer’s generalisation, “multiplication of material overlapping propagules that confer the capacity to develop” (2000c: 74). It brings in development and epigenetic inheritance. The reconstitutor, which needs no material overlap, is mentioned (n. 15).
• manifestor of adaptation — the entity at whose level adaptations accumulate. In the entry’s usage the adaptations meant are engineering adaptations.
• beneficiary — who benefits from selection in the long run. For Dawkins this is the replicator, understood as an agent that initiates causal chains.
• selection-product vs engineering adaptation — any trait resulting directly from selection, versus accumulated design beyond the ancestral range of variation (§2.3).
• Evolutionary Change school — the Wrightian, quantitative-genetic multilevel tradition. It is interested in process, epistasis and IGEs (§3.1).
• Adaptationist/KS school — the Fisherian, game-theoretic and inclusive-fitness tradition. It is interested in optimal engineering adaptations (§3.2).
• emergent fitness — a higher-level character’s relation to fitness that the associations of lower-level characters cannot account for. Lloyd puts it in place of “emergent properties” (§4.2, n. 88–89).
• demibiont — a host–symbiont combination in which one lineage has adapted to the other but not the reverse (§4.6, n. 141).

Connections

Of the held works, the most directly engaged is Damuth & Heisler (LIT-201, Deferred, seeded from its abstract only). Lloyd relies on their contextual analysis and quotes their empirical criterion. She also reads them as conflating interactor with manifestor (§4.1, n. 113), and that bears on how the MLS1/MLS2 distinction is summarised in LIT-201.

Bourrat’s earlier work (2016–2022) is cited and criticised on four points: contextual analysis, Price-equation-based unit definitions, the use of equivalence claims, and reinventing group-trait reproduction (n. 4, 62, 71, 109, 140). The held Bourrat papers continue that programme. LIT-157 builds on Griesemer’s reproducer, the concept Lloyd champions in §2.2. LIT-049 and LIT-005 are information-theoretic and coarse-graining criteria for ETI. A reader of those should know that this entry is where the opposing view is set out.

Griesemer is Lloyd’s closest ally in the entry: the reproducer (§2.2, §4.5) and the “entanglement” reading of the Price equation (§4.1). Griesemer & Shavit (LIT-188) is not cited, but it extends Griesemer’s programme towards the “community” level that §4.6 argues for. Shavit 2005 is cited once, as a general criticism of pluralism about group selection (§4.1).

The entry sends readers to the SEP “Biological Individuals” entry (LIT-168) for individuality and keeps the two topics apart on purpose. Where LIT-168 maps Darwinian vs physiological individuals, this entry objects to “Darwinian individual” as a fusion of roles (n. 80, n. 115). Baedke (LIT-167) holds that holobionts and colonies also meet the “persister” criterion. That is compatible with Lloyd’s view of holobionts as interactors, but neither work cites the other. O’Malley & Parke (LIT-210) cite this entry for the unresolved question of whether host and microbes are one unit of selection (their §5.1). Lloyd cites O’Malley’s work with Dupré as the “common fate” and “polygenomic consortia” account of holobionts (§4.6).

On the identity tag: the entry supplies no account of what an individual is. The view it presupposes is functional-role pluralism. What counts as “the unit” is relative to which of four roles is in question, an entity may fill several roles, and the roles are “abstract and functional” and “uncommitted to any particular biological hierarchy” (n. 137, citing Griesemer 2005). That is a thesis about how to individuate the relata of selection, not about persistence or identity over time. The entry states that it is separate from the individuality question (§4.3, n. 99).

Bearing on the record

It carries nothing for ML practice. The contrast between contextual analysis and the Price equation is a statistical-methods argument inside evolutionary genetics, not a recommendation that transfers.

For this record the entry is the hub the dossier took it to be, with two qualifications. It is partisan. And it is critical of several works in the cluster: Damuth & Heisler, Okasha, Godfrey-Smith and Bourrat. A THEORY document or LIT summary that cites this entry for “units of selection is four questions” is on firm ground. One that cites it as a neutral statement of consensus (on kin/group equivalence, on the cheater problem, or on the Price equation) is not, and should attribute the claim to Lloyd and the Evolutionary Change school.

LIT-201’s summary says the MLS1/MLS2 distinction answers different questions. That agrees with Lloyd’s reading, but her further charge that Damuth & Heisler conflate interactor and manifestor should be checked when LIT-201 is read.

Limitations

• It is one participant’s view, written in the first person. The opposing school’s replies are quoted briefly and often called “false claims” or “confusion” (§4.1, §4.4) rather than answered in technical detail.
• No model or data is presented in full. The empirical support (Muir’s hens, up to 60% egg-laying improvement in 6 generations; Jones et al. 2015; the IGE literature) is cited, not examined.
• “Most common error” and “nearly universally” are asserted without a survey (§2.1, §4.1).
• Several arguments depend on personal communications (Hamilton 1989, n. 66; Williams 1989, n. 20; Griesemer 2022, n. 63).
• The prose is loose in places. A few sentences are incomplete (§2.5 “When applying this…”), and an editor’s placeholder is left in §4.3. “Thies” in §4.6’s opening list and “Thies et al. 2016” (n. 139) misspell Theis: the bibliography has Theis et al. 2016. “Thies & Watson 2021” is a different, correctly spelled author.
• The revision predates 2024, so it cannot speak to Bourrat 2023–2025 or Griesemer & Shavit 2023.

Open questions

• Whether the multilevel Price equation should be abandoned in favour of contextual analysis and causal-graph methods for multilevel selection (§4.1: “remains an open question”).
• Whether kin-selection and group-selection model-types are equivalent formally, empirically, conceptually or explanatorily (§4.1, RQ (1)). Settling it needs a comparison of the dynamics and the parameters, not only of predicted allele frequencies.
• Whether group selection leads to group adaptation (§4.1, RQ (2)), in the engineering sense.
• Which model-type, ETI or ordinary multilevel, should represent holobiont evolution (§4.6), and whether demibionts count as units under any role.

Corrections to the seeded skim

• The dossier asks how far the four-question framework is Lloyd’s own programme. It is entirely hers. §1 attributes the two added questions to “one philosopher” (Lloyd 1988 [1994]). §2.3 and §5 are in the first person (“I call this…”, “I offer a framework”). The method is her “logic of research questions” (Lloyd 2015, n. 25). The entry also says it is “not meant to resolve” which question matters most (§1), yet it repeatedly rules on disputes. Examples: genic pluralists “fundamentally over-reached” (§4.4); the cheater problem is a “fallacy” (§4.1, quoting Wade 2016); applying ETI model-types to holobionts “constitutes a fundamental confusion” (§4.6). The dossier’s word “clarifies” understates this. The entry is advocacy as well as a map.
• The dossier’s summary (“apparent disagreements are answers to different questions”) is right about §5, but it misses the entry’s main diagnosis. The error Lloyd finds most often, and names “the most common error” (§2.1), is taking an interactor at a level to be a manifestor of adaptation at that level. She traces it through group selection (§4.1), species selection (§4.2), ETI (§4.5) and holobionts (§4.6). She also says that the ambiguity between “selection-product” and “engineering” adaptation (§2.3, peppered moth vs Darwin’s finches) drives it.
• The dossier’s check of §4.2’s “emergent properties” against the emergence literature: the emergence at issue is not metaphysical emergence. For the paleontologists (Vrba 1984, Eldredge 1985), “emergent properties” meant species-level engineering adaptations. Lloyd’s replacement is “emergent fitnesses”: a statistical relation between higher-level characters and fitness, detected by contextual analysis or non-additivity. G. C. Williams endorsed it (1992: 27, quoted n. 88). The links to LIT-156 and LIT-150 are therefore at most terminological.
• The dossier places Damuth & Heisler (LIT-201) as the base that the entry builds on. The entry does credit their contextual analysis, and it quotes their claim that whether a unit is functioning as one is an empirical question (§1, §2.1, n. 4, n. 52). It also says twice that Heisler & Damuth, like Okasha 2006, conflate interactor with manifestor. They cited Sober’s “common fate” definition as equivalent to their approach (§4.1, Heisler & Damuth 1987: 588), and the MLS1/MLS2 reading of transitions inherits the conflation (§4.5, n. 113).
• On Bourrat, whom the dossier lists as part of the cluster: the entry is hostile rather than allied. It names Bourrat 2016–2022 among the works that “reintroduced” her roles under new names while neglecting her analysis (§2.5). It rejects his demand that statistical tools individuate populations (“to mistake methods for metaphysics”, n. 109; also n. 4, n. 62, n. 71, n. 140). The held Bourrat works (LIT-157, LIT-049, LIT-005) postdate the revision and are not cited.
• Omitted by the dossier: §2.2 says the replicator role has been “rejected… as misconceived, and too narrow” and replaced by Griesemer’s material “reproducer”. Every “replicator” in the entry is read in that light. §4.6 introduces the “demibiont” (Lloyd & Wade 2019; Suárez & Stencel 2020, n. 141) alongside holobionts. §4.1 argues that the multilevel Price equation turns the single-level transmission term into a within-collective selection term. That relates to the Luque paper (LIT-205), which the entry does not cite.
• An editing artefact is published in the live text. §4.3 ends with “[[please add Links listed in NOTES 99 and 100 for this page]]”.
• identity tag: justified, as a secondary tag. The entry says expressly that questions about units of selection must not be conflated with questions about biological individuality (§4.3 and n. 99). Its demibiont continuum “leaves untouched” what a biological individual is (n. 141). Even so, §4.5 (evolutionary transitions in individuality) and §4.6 (holobionts, and Darwinian individuals as reproducers) are central material for anyone browsing individuality. The entry also cross-links the SEP “Biological Individuals” entry (LIT-168) throughout.
• primary topic: philosophy-of-science (keep).