Multilevel selection (MLS) occurs when natural selection simultaneously acts at two or more levels of biological organization (Damuth and Heisler, 1988; Okasha, 2006; Wilson and Wilson, 2007; Marín, 2024). Specifically, MLS occurs when there is differential reproduction of groups in addition to reproduction of individual entities within them, or when the differential reproduction of individuals is based on their group composition or characteristics, e.g., the social environment (Goodnight et al., 1992) (see key definitions in Table 1). Goodnight et al. (1992) have defined MLS as: “variation in the fitness of individuals that is due to both properties of the individuals and properties of the group or groups of which they are members” (p. 745). Goodnight et al. (1992) definition incorporates models that explicitly include differential extinction of entire groups (Levins, 1970), trait-group models (Wilson, 1975), and Wright’s (1945) definition of interdemic selection, which does not require group extinction.

The MLS framework has been useful, even essential in studying the central dogma in molecular biology (Takeuchi and Kaneko, 2019), horizontal gene transfer in bacteria (Lee et al., 2022), the origin of multicellularity (Bozdag et al., 2023), cancer (Aktipis et al., 2015), disease/virus evolution (Blackstone et al., 2020), animal (Craig and Muir, 1996) and plant breeding (Zhu et al., 2019b), as well as economics (Wilson et al., 2020) and cultural institutions (Wilson et al., 2023). The clear value of an MLS approach, whether related to the selection of particular traits, or to the discovery of what affects fitness in a given system/organism, is its focus on identifying both the direction and strength of selection from multiple sources. Despite this, criticisms and skepticism persist among biologists (Eldakar and Wilson, 2011) – albeit anthropologists seem to favor an MLS framework, according to a survey by Yaworsky et al. (2015). Marı́n (2024) has identified three main arguments in favor of MLS: first, the term “unit of selection” has a polysemic nature, with at least three different meanings: “interactors”, “replicators or reproducers or reconstitutors”, and “manifestors of accumulated adaptations” (Suárez and Lloyd, 2023; Lloyd, 2024). The second is the fact that biological entities as complex as organisms or genes must-at least- have evolved from less complex entities (Okasha, 2006). And third, there is vast empirical evidence for this theory both in laboratory and natural populations. Sound literature reviews of such empirical evidence of MLS can be found in: Wilson and Sober (1994), Goodnight and Stevens (1997), Eldakar and Wilson (2011), Goodnight (2015), Marín (2015, 2016, 2024), and Hertler et al. (2020). Despite these clear reviews and a diversity of empirical studies across a range of taxa, the misconception that MLS lacks empirical support persists (Harms et al., 2023). Here we address this misconception head on, by revealing an abundance (not a paucity) of examples of MLS in a diversity of taxa and biological systems, levels of biological organization, and type of research topics and tools.

In evolutionary biology, the evolution of altruism or prosociality has been a main focus of MLS debates for decades, but altruism is just one trait that can evolve via MLS. On the one hand, the classic example of the evolution of altruism considers groups within which selfish individuals out-compete altruists, while groups with more altruists contribute more offspring to the next generation than groups comprised of more selfish individuals (Darwin, 1871; Wilson and Wilson, 2007). On the other hand, MLS also occurs when emergent group traits (e.g., social network structure, density, collective colony personality, among other descriptors) have significant effects on the reproductive success of a focal individual (Damuth and Heisler, 1988; Goodnight et al., 1992; Philson et al., 2025). Such effects of emergent or contextual traits have been amply demonstrated, for example in studies of epistasis (Burch et al., 2024) and indirect genetic effects (IGEs) (Linksvayer et al., 2009; Buttery et al., 2010; Bijma, 2014; Baud et al., 2021; Santostefano et al., 2025), and using techniques such as contextual analysis (Marín, 2016; Suárez and Lloyd, 2023; Lloyd, 2024; Philson et al., 2025).

We conducted a bibliometric review of the scientific literature to identify the breadth and depth of empirical evidence of MLS across levels of biological organization. In addition, we also focused on phenotypic selection studies that use contextual analysis (Heisler and Damuth, 1987) to decompose the strength and direction of selection at different levels (individual organisms and groups of organisms). We then organized the literature on the basis of study systems (i.e., eusocial insects, humans, wild plants and algae, crops, etc.), levels of biological organization assessed (demes, communities, cells, etc.), and type of research (i.e. in situ studies of natural populations, artificial selection experiments, breeding through group selection, etc.). The focus of this review is to provide an introduction, accounting, and organization of the vast empirical support of MLS and its utility to understand the natural world. In this review, “support” means only that selection at different levels was explicitly measured, not that higher levels or “group” selection outweighed lower levels. For example, there were some studies in which lower-level selection or individual properties were shown to be more important than higher-level selection or properties in explaining focal individual fitness (Tsuji, 1995; Donohue, 2003; Fisher et al., 2017; Fisher et al., 2021). Thus, our bibliometric review captured the full spectrum of studies, whether or not individual or higher-level selection is the prevailing force – something perfectly consistent with MLS theory, as MLS should be evaluated in a case-by-case basis (Wilson and Wilson, 2007). In addition, while discussions of alternative and complementary frameworks (such as inclusive fitness theory or direct reciprocity) and mechanisms that partition variation within and between groups (e.g., conditional dispersal, kinship, and kin groups) are of general interest (for detailed discussions, see Bijma and Wade, 2008; Goodnight, 2013; Frank, 2025; Marín and Wade, 2025), the consideration of such topics are beyond the scope of this review. However, in Appendix 1 we discuss why MLS is not mathematically equivalent to other frameworks such as inclusive fitness theory.

The identification phase of the Scopus search yielded a total of 2,950 articles (after deleting duplicates) (Figure 1). A total of 1,829 papers remained after exclusion of all articles not related to biology, cooperation, and social behavior in general (screening phase). From these, only 414 papers included empirical studies and thus persisted in the eligibility phase (Figure 1). Finally, 166 articles indicating group selection or MLS as possible or plausible mechanism but not ensuring it as an explanation, were also excluded, resulting in a total of 248 papers providing empirical support for MLS found with Scopus. The additional search with Google Scholar, which was restricted to Web of Science and Scopus-indexed articles, added 32 articles to this list, leading to a total of 280 scientific articles providing empirical support for MLS (Figure 1).

These articles spanned 1976–2024, and 180 consisted of laboratory-controlled experiments, while the remaining 100 consisted of in situ (field) measurements and/or experiments (Figure 2). Only years 2019, 2021, and 2023, yielded 20 or more MLS empirical papers, with a peak of 22 studies in 2019 (Figure 2). Only 81 studies were published during the first 35 years of MLS empirical research (1976–2011), while the remaining 199 have been published since 2012, showing a marked increase in research in the last 12 years, both on MLS in situ and experimental studies (Figure 2).

Regarding the taxa or study systems, systems like farm animals, eusocial insects, “other insects” (this means non-eusocial insects such as beetles, spiders, water striders, among others), and humans, together compose approximately 65% and 55% of experimental and MLS in situ studies, respectively (Figure 3). However, in MLS in situ studies, systems like wild plants and algae, wild mammals, and wild birds also make up an important proportion of studies, while this is the case for microbial populations and fungi in MLS experimental studies (Figure 3). Many other study systems or taxa have been empirically investigated under a MLS framework: other invertebrates (such as tunicates and polychaetes), crops, algae, fish and amphibians, microbiomes, etc. (Figure 3).

Regarding the levels of biological organization investigated, 90.4% of MLS empirical studies focused on individual organisms and groups of organisms. In particular, 71% of studies (198 papers) focused on demes, while 19.4% of studies focused on tighter organismal groups: 24 studies were conducted at the “aggregate” level (aggregates of bacteria, amoebas, algae, and yeast) and 31 studies investigated colony-level selection (mostly in eusocial insects but also including spider and Caenorhabditis elegans colony studies). A 9.6% of MLS empirical studies focused on organization levels above or below organisms/groups of organisms: four studies were conducted at the cell level (this include horizontal gene transfer or RNA viruses, for example); three studies were conducted at the genetic element level (specifically investigating selfish genetic elements or gene transfer agents using an MLS framework); 13 studies investigated community-level selection (mostly microbiomes, but also including beetles, ants, and arthropod communities); three studies with either algae or sea-grass investigated clonal or module-level selection, i.e. selection acting at the clonal level; two studies with fungi investigated natural selection at the nuclei level, as some fungal taxa can contain thousands of nuclei on a single spore (Jany and Pawlowska, 2010); and finally, two studies investigated natural selection at the species level.

Both in situ and experimental MLS empirical evidence comes from many different sources, types of study, and taxa or study systems, to the point that our nine main categories were sub-categorized into 70 sub-categories (Figure 3; Appendix 2). More than half (n=54) of MLS in situ studies used either IGEs measurements or contextual analysis, with categories such as group effects (n=18) and dataset analyses (n=18) also having important numbers (Figure 3). Similarly, 107 of the 180 MLS experimental studies were laboratory experiments of different types, with the group selection treatments on wild animals sub-category in particular having 18 studies (Appendix 2). Other MLS experimental studies categories also had an important number of articles, including breeding through group selection (n=32), artificial selection (n=22), and psychology experiment (n=19) (Figure 3). A brief summary of nine representative studies of each one of the main categories is given in Table 2 for MLS in situ studies and in Table 3 for MLS experimental studies.

Finally, regarding the 26 studies that implemented contextual analysis in natural populations, it was not possible to extract the regression coefficient information from five of them (Appendix 2; Appendix 3). Thus, Figure 4 shows the regression coefficients from 21 studies spanning 1995–2025, which were conducted in a plethora of study systems: from plants and water striders to chipmunks and humans. In Figure 4, the effects of individual (“size”) and group (average “size” of the neighborhood individuals) traits on focal individuals’ fitness is shown with the Beta (β) regression coefficients. In some studies (Stevens et al., 1995), group selection is stronger and goes in an opposite direction than individual selection, while in other studies (Donohue, 2004) the strength and direction of individual and group selection are similar, and in other studies (Bolstad et al., 2012), individual selection is significantly stronger than group selection (Figure 4). In summary, there is a variety of selection outcomes across the 21 studies as revealed by contextual analysis, with some showing selection at different levels acting in concert while others show selection acting in opposition (Figure 4).

In this review, we catalogued a vast body of empirical evidence for multilevel selection (MLS), from in situ observations to experimental studies, spanning five decades. Such evidence encompasses a broad spectrum of study systems and taxa, albeit systems like farm animals, eusocial and non-eusocial insects, and humans have been the main focus of MLS research. Similarly, and likely due to the organismal focus of most biologists, but also due to methodological feasibility, individual organisms and groups of organisms (demes, colonies, aggregates) have been the most investigated levels of selection in the MLS empirical research literature. With our analysis we can conclude that there is not a single or majority way to investigate MLS in situ or experimentally. Rather, multiple tools or ways of empirically investigating MLS have been used through the decades, which respond to the specificity of each study system or taxa, level of organization, topic, and/or tool. Further, our bibliometric screening shows that from 1,829 articles that deal in some way with MLS or social evolution, 1,415 articles (77%) consisted of mathematical and conceptual models (Figure 1), opinion pieces, debates, reviews, simulations, and so on. These are important in their own right but are excluded here because we are concerned with the realized utility of the MLS framework in empirical research. Group selection was initially rejected, not due to lack of evidence, but due to its supposed theoretically implausibility (Maynard Smith, 1964; Williams, 1966). The large number of models demonstrating the theoretical plausibility of MLS therefore complements our review of the empirical literature.

The debate on the units of selection has gone on for too long. It is high time to move on and focus on the empirical evidence and data (Marín, 2015; Marín, 2016; Marín, 2024). Because there is a plurality of levels of selection investigated, from selfish genetic elements and nuclei to microbiomes, not all tools or experiments will work the same. For example, it would be quite challenging to apply contextual analysis in artificial selection experiments dealing with multicellularity evolution. Following our comprehensive MLS definition (Table 1), here we focused on compiling an extensive list of studies showing either differential reproduction of entire groups, or the differential reproduction of individuals being affected by their group composition or characteristics.

The plurality of study types and systems involves a variety of different methods to assess MLS in situ. For example, some in situ studies infer MLS by using molecular sequencing tools such as microsatellites, fingerprinting, and genome-wide association studies, among other tools. More than half of MLS in situ studies implemented either IGE assessment or contextual analysis (Figure 3), finding quite significant effects of the neighborhood traits or emergent traits on focal individuals’ fitness (and individual trait variation). Such neighborhood/emergent effects are a core feature of MLS (Table 1), with IGEs, contextual analysis, and group effects measurements, representing different ways in which they are calculated. It is not within the scope of this article to compare such mechanisms of assessment, as this has been done plenty in the literature (Bijma and Wade, 2008; Goodnight, 2013). In particular, Bijma and Wade (2008) have shown the relationships between kin selection (inclusive fitness theory), MLS, and IGEs. In Appendix 1 you can find an expanded Discussion on these relationships and on the non-equivalence between MLS and inclusive fitness theory. Rather, here we show that when group composition or characteristics or average/emergent traits are considered in the response to selection models’ (in addition to individual traits), focal individuals’ fitness are affected by such group composition or characteristics. This is supported by recent meta-analyses by Santostefano et al. (2025) and Burch et al. (2024), which respectively showed that IGE and epistasis are ubiquitous across the Tree of Life.

The comprehensive definition of MLS that we employ here (Table 1) falls into the disambiguating project of the units of selection literature (Suárez and Lloyd, 2023). In the disambiguating project (Suárez and Lloyd, 2023; Lloyd, 2024), the “units of selection” at any given level in the biological hierarchy can have one or more of three functional roles in the process of natural selection, which must be distinguished from each other. These roles are: interactors (phenotypic variation and differential proliferation), replicators (or reproducers or reconstitutors; inheritance), and manifestation of accumulated adaptations (Suárez and Lloyd, 2023). Thus, for a biological entity to be considered an interactor, two minimal things are required: phenotypic variation and differential proliferation. Furthermore, Suárez and Lloyd (2023, p. 17) have defined natural selection as a “process in which the differential proliferation of interactors causes the differential replication of replicators” [or the differential reproduction of reproducers or the differential reconstitution of reconstitutors]. This clarification is necessary, as many of the historical (Williams, 1966) and current-day (Harms et al., 2023) critiques of MLS confound the roles of the different units of selection (Gould and Lloyd, 1999), requiring phenotypic variation, differential proliferation, and inheritance at the same level of biological organization to be considered as a unit of selection. This is not necessarily the case. For example, although typically genes constitute replicators, in specific cases such as selfish genetic elements, genes might also be considered as interactors (Gitschlag et al., 2020).

The comprehensive definition of MLS (Table 1) employed here captures instances in which entire groups constitute the inheritance unit (replicator/reproducer/reconstitutor) and instances in which entire groups constitute the interactor but inheritance occurs at a lower level of biological organization (most typically, the individual organism or its genetic material). The latter cases are typically detectable with techniques such as IGE measurements, social network analysis, the Price equation, and contextual analysis, among others (Marín and Wade, 2025), as mentioned above. In summary, MLS occurs when natural selection operates simultaneously among two or more different levels of a nested biological hierarchy, which either causes differential reproduction of entire groups (i.e., the group is also the replicator/reproducer/reconstitutor) or when the differential reproduction of individuals is influenced by their group composition or its characteristics (i.e., lower-level entities are the replicator/reproducer/reconstitutor).

The pioneering study by Wade (1976) (Table 3) was the starting point of laboratory studies on which group selection was imposed as a treatment. Several dozen similar studies (imposed group selection in laboratory populations) were conducted through the decades, generally finding rapid responses to the group selection treatments after a few generations. Further, such imposed group selection studies found that selection sometimes acts in concert and sometimes in opposition at the individual and group levels, also with varying strength. Interestingly, the same pattern is found when analyzing contextual analysis studies (Figure 4): natural selection sometimes acts at the same and sometimes at different directions and strengths across levels of biological organization. As such, no generalization can be made about MLS and it should be investigated on a case by case manner (Wilson and Wilson, 2007; Eldakar and Wilson, 2011). However, ecological constraints can help predict responses to selection. For example, when in 2017 the category 4 Hurricane Maria almost totally destroyed a Puerto Rican island inhabited by rhesus macaques, shade became a very scarce resource. As a response, there was a marked increase in tolerance and decrease in aggression among macaques (Testard et al., 2024), with the most tolerant animals having the highest survival. Similarly, in plant-mycorrhizal associations it has long been known that under scarcity of nutrients (particularly nitrogen and phosphorous), this symbiotic association becomes more mutualistic while under “luxury” conditions (excess of nutrients), the usually benign mycorrhizal fungal microbiomes can behave as nutritional parasites (Johnson et al., 1997; Johnson and Marín, 2025).

Several other influential MLS experimental studies include Craig and Muir (1996), Swenson et al. (2000), Ratcliff et al. (2012), and Bozdag et al. (2023) (Table 3). Craig and Muir (1996) and several dozen more studies (a total of 32 studies; Figure 3) have shown that MLS is a very useful framework for breeding programs of farm animals and crops. Furthermore, when farm animals or crops are bred through group selection treatments (i.e., selecting group traits) or when IGEs are considered in breeding programs, the outcome is almost always the desired for the farmer: higher yields or more production. Even MLS sceptics recognize the value of MLS-focused breeding programs in wheat cultivars (Zhu et al., 2019b; Zhu et al., 2019a; Zhu et al., 2022). Empirical evidence showing the success of wheat breeding for higher yields over the past 100 years in north-western China has been argued to result in part from “unconscious group selection on root traits” (Zhu et al., 2019b), which results in smaller, less branched, and deeper roots.

Swenson et al. (2000) pioneered the framework of artificial ecosystem selection as a way of selecting communities of soil microorganisms based on plant performance. This implies exposing multiple generations of plants to particular selection pressures, selecting the microbiomes that increase plant fitness (or selected traits) to the next generation, while the genetic basis of the host remains the same. This approach has been successfully used to engineer belowground communities that increase plant tolerance to drought (Lau and Lennon, 2012; Jochum et al., 2019) and salinity (Mueller et al., 2021), or that increase leaf greenness (Jacquiod et al., 2022), among others (reviewed in Sánchez et al., 2021; Sanchez et al., 2023; Yu et al., 2023). On their part, Ratcliff et al. (2012) and Bozdag et al. (2023), implementing artificial selection regimes in yeast aggregates, have shown some of the most visually stunning examples of experimental MLS: they shown de novo evolution of macroscopic multicellularity just after one year and 600 rounds of selection (Bozdag et al., 2023). In particular, in an anaerobic treatment, yeast evolved to be macroscopic, becoming 2 x 10⁴ times larger than at the beginning, while maintaining a clonal multicellular life cycle (Bozdag et al., 2023).

A MLS framework has long been used to investigate human culture (Soltis et al., 1995), originating a whole sub-discipline, deemed “cultural multilevel selection” (Wilson et al., 2020; Wilson et al., 2023). In our review, a total of 30 MLS empirical studies were centered around humans: 19 consisted of psychological experiments, 10 assessed or inferred cultural MLS in situ, and one implemented contextual analysis over 55 years of polygyny and polyandry data, based on the Utah Population Database (Moorad, 2013). MLS seems to explain the most important cultural macro-evolutionary patterns and historical trends, including competition and warfare but also exchange and selective imitation (Turchin and Gavrilets, 2021) (Table 2). The utility of MLS has been recognized in anthropology: a survey to 175 evolutionary anthropologists (faculty members of graduate programs) finds that 78.7% of them regard cultural MLS as “important”, while 64.9% disagree with the statement “Group selection has no useful role to play in social science” (Yaworsky et al., 2015). Whether a similar acceptance rate of MLS by evolutionary biologists not working with humans is yet to be analyzed/surveyed. However, it is worth noticing that two recent analyses of biology, evolution, and behavior undergraduate textbooks show that MLS theory is generally dismissed as unimportant when compared to individual selection (Greene et al., 2025), with teaching on evolutionary transitions in individuality also being absent (La et al., 2025).

Our findings showing a marked increase in MLS research in the last 12 years (Figure 2), with 199 MLS studies for 2012–2024, indicates both that MLS is becoming more accepted as a conceptual framework and that many studies are using adequate sample sizes to ask questions across levels of biological organization. With the marked increase since 2012 and expanding acceptance of MLS as an conceptual evolutionary framework, many more ground-breaking studies are to come in the next few decades.

There are some caveats to our findings that the evidence for MLS is vast. First, we expect a publication bias towards studies finding positive outcomes, by which we mean that some studies where no selection at a higher level was found, were likely not captured. Despite this, our database does include studies in which higher-level selection or group properties were not important in explaining focal individual trait variation and fitness (Philson and Blumstein, 2023a; Philson and Blumstein, 2023b). Further, in several of the contextual analysis studies (Tsuji, 1995; Donohue, 2003; Weinig et al., 2007; Boege, 2010; Eldakar et al., 2010; Formica et al., 2011; Bolstad et al., 2012; Laiolo and Obeso, 2012; Fisher et al., 2017; Fisher et al., 2021) (Figure 4), the magnitude of selection was stronger at the individual than at the group level. Similarly, direct genetic effects are also usually stronger than indirect genetic effects, as shown by the meta-analysis of Santostefano et al. (2025) and through our database (but see Santostefano et al., 2021). However, because MLS should be evaluated in a case-by-case basis (Wilson and Wilson, 2007), this is not problematic for our framework: depending on the environmental context, case, and traits, it is expected that there will be cases in which there are no group effects or they are not as important as individual-level effects. It is, in fact, a central point that, unless a MLS perspective is applied, one would not know if or how strongly individual selection was causally related to trait evolution (Wade, 2026). Secondly, in order to have a distinct cut-off, we excluded MLS empirical evidence produced after 2024, thus missing new studies such as Lipowska et al. (2025), showing how bank vole holobionts selected for herbivorous capability evolved distinct and robust gut bacterial communities.

In general, we were quite strict in our search. For example, a study classically cited by some as the first MLS empirical study (Lewontin, 1962) was excluded, because, although it is based on real lab mice population data, the conclusions (about interdemic selection) are based on Monte Carlo simulations. Similarly, studies arguing that MLS is a “likely” (Dyer et al., 2005) or “possible” (Aviles, 1986) explanation were also excluded. Thus our total of 280 articles obtained is an underestimate of the evidence and conceptual use, because many more studies that clearly show results consistent with the MLS framework (Pope, 1992; Heinsohn and Packer, 1995; Ingvarsson, 2000; Papkou et al., 2023; Barnett et al., 2025), have historically avoiding using the term (Eldakar and Wilson, 2011; Greene et al., 2025). For example results based on Wright’s fitness landscapes (Papkou et al., 2023) or on evolvability (Barnett et al., 2025), explicitly require a MLS perspective to understand them, but avoid the terminology. Although a MLS framework may not be explicitly mentioned by name, and in some cases may be avoided due to historical misconceptions (Eldakar and Wilson, 2011), it is implicit in experimental design and rationale.

In summary, a thorough search of the literature shows that contrary to common misconceptions which plagued the field since the 1960s, there is vast empirical evidence of selection acting at multiple levels and of the utility of assessing multilevel selection (MLS) both in situ and via experimental studies. We found 280 papers providing empirical support for MLS: 100 in situ and 180 laboratory experiments. The studies span many taxa and research methodologies, meaning MLS is not situational or an exception: MLS is a powerful evolutionary force in nature. Disregarding MLS will continue to hold the field of evolutionary biology back and prevent us from more fully understanding life on earth.