The co-evolution of individual behaviors and social institutions.

Abstract

We present agent-based simulations of a model of a deme-structured population in which group differences in social institutions are culturally transmitted and individual behaviors are genetically transmitted. We use a standard extended fitness accounting framework to identify the parameter space for which this co-evolutionary process generates high levels of group-beneficial behaviors. We show that intergroup conflicts may explain the evolutionary success of both: (a) altruistic forms of human sociality towards unrelated members of one's group; and (b) group-level institutional structures such as food sharing which have emerged and diffused repeatedly in a wide variety of ecologies during the course of human history. Group-beneficial behaviors may evolve if (a) they inflict sufficient fitness costs on outgroup individuals and (b) group-level institutions limit the individual fitness costs of these behaviors and thereby attenuate within-group selection against these behaviors. Thus, the evolutionary success of individually costly but group-beneficial behaviors in the relevant environments during the first 90,000 years of anatomically modern human existence may have been a consequence of distinctive human capacities in social institution building.

Full text

Page 1

Journal of Theoretical Biology 223 (2003) 135–147
The co-evolution of individual behaviors and social institutions
Samuel Bowlesa,b,*, Jung-Kyoo Choic, Astrid Hopfensitzd
aSanta Fe Institute, 1399 Hyde Park Road, Santa Fe, NM 87501, USA
bFaculty of Economics, University of Siena, Siena 53100, Italy
cSanta Fe Institute, Department of Economics, University of Massachusetts at Amherst 01002, USA
dCenter for Research in Experimental Economics and Decision-Making, University of Amsterdam, 1018 WB Amsterdam, The Netherlands
Received 5 March 2001; received in revised form 10 October 2002; accepted 15 January 2003
Abstract
We present agent-based simulations of a model of a deme-structured population in which group differences in social institutions
are culturally transmitted and individual behaviors are genetically transmitted. We use a standard extended fitness accounting
framework to identify the parameter space for which this co-evolutionary process generates high levels of group-beneficial
behaviors. We show that intergroup conflicts may explain the evolutionary success of both: (a) altruistic forms of human sociality
towards unrelated members of one’s group; and (b) group-level institutional structures such as food sharing which have emerged
and diffused repeatedly in a wide variety of ecologies during the course of human history. Group-beneficial behaviors may evolve if
(a) they inflict sufficient fitness costs on outgroup individuals and (b) group-level institutions limit the individual fitness costs of these
behaviors and thereby attenuate within-group selection against these behaviors. Thus, the evolutionary success of individually costly
but group-beneficial behaviors in the relevant environments during the first 90,000 years of anatomically modern human existence
may have been a consequence of distinctive human capacities in social institution building.
r 2003 Elsevier Science Ltd. All rights reserved.
Keywords: Human cooperation; Multi-level selection; Intergroup conflicts
1. Introduction
Is the remarkable level of cooperation among
unrelated humans a result of the distinctive capacities
of humans to construct institutional environments
which limit competition and reduce phenotypic varia-
tion within groups, allowing individually costly but
group-beneficial behaviors to co-evolve with these
supporting environments through a process of inter-
demic group selection? We use simulations of a standard
extended fitness accounting framework to investigate
this question, identifying the parameter space for which
this co-evolutionary process generates high levels of
group-beneficial behaviors.
The idea that the suppression of within-group
competition may be a strong influence on evolutionary
dynamics has been widely recognized in eusocial insects
and other species (Smith and Szathmary, 1995; Frank,
1995; Michod, 1996; Buss, 1987; Ratnieks, 1988).
Boehm (1982) and Irenaus Eibl-Eibesfeldt (1982) first
applied this reasoning to human evolution, exploring the
role of culturally transmitted practices which reduce
phenotypic variation within groups. Examples of such
variance-reducing practices are leveling institutions,
such as monogamy and food sharing among non-kin,
namely those which reduce within-group differences in
reproductive fitness or material well-being. Monoga-
mous or polygamous mating systems, distinct systems of
resource sharing, and the like may be termed institu-
tions, by which we mean a uniformity in the structure of
human interactions, that is characteristic of a group but
may differ among groups. Such structures may have
attenuated within-group selection operating against
individually costly but group-beneficial practices, result-
ing in higher group average fitness or material success. If
so, groups adopting these variance-reducing institutions
would have had advantages in coping with climatic
adversity, intergroup conflicts and other threats. A
group’s institutions thus constitute a niche, that is, a
modified environment capable of imparting distinctive
ARTICLE IN PRESS
*Corresponding author. Tel.: +1-505-984-8800; fax: +505-982-
0565.
E-mail address: bowles@santafe.edu (S. Bowles).
0022-5193/03/$-see front matter r 2003 Elsevier Science Ltd. All rights reserved.
doi:10.1016/S0022-5193(03)00060-2

Page 2

direction and pace of evolutionary change (Laland et al.,
2000; Bowles, 2000). According to this view, the
evolutionary success of variance-reducing social institu-
tions may be explained by the fact that they retard
selection pressures working against in-group-beneficial
individual traits coupled with the fact that high
frequencies of bearers of these traits reduces the
likelihood of group extinctions (or increases the like-
lihood of a group’s expanding and propagating new
groups).
Theevolutionary mechanisms
account are multi-level selection processes with the
novel features (adapted from Bowles, 2001) that both
genetically transmitted influences on individual beha-
viors as well as culturally transmitted group-level
institutional characteristics are subject to selection, with
inter-group conflicts playing a decisive role in group-
level selection. The model is thus an example of
a gene-culture evolutionary process (Cavalli-Sforza
and Feldman, 1981; Boyd and Richerson, 1985;
Durham, 1991).
It has been long recognized that in populations
composed of groups characterized by a markedly
higher level of interaction among members than with
outsiders, evolutionary processes may be decomposed
into between-group and within-group selection effects
(Lewontin, 1965; Price, 1972; Crow and Kimura, 1970;
Uyenoyama and Feldman, 1980). Where the rate of
replication of a trait depends on the frequency of
the trait in the group and where group differences in
trait frequencies are substantial and persistent, group
selection contributes to the pace and direction of
evolutionary change. But most who have modeled
evolutionary processes under the joint influence of
group and individual selection have concluded that the
group selection pressures cannot override individual-
level selection except where special circumstances (e.g.
small group size, limited migration) heighten and sustain
differences between groups relative to within-group
differences (Eshel, 1972; Boorman and Levitt, 1973;
Smith, 1976).
Beginning with Darwin (for example Darwin, 1873;
p. 156 and other passages), a number of evolutionary
theorists have suggested that human evolution might
provide an exception to this negative assessment of the
force of multi-level selection. Haldane (1932) suggested
that in a population of small endogamous ‘‘tribes’’, an
altruistic trait might evolve because the ‘‘tribe splitting’’
which occurs when successful groups reach a certain size
would create a few successor groups with a very high
frequency of altruists, reducing within-group variance
and increasing between-group variance, a process very
similar to that simulated in this paper (pp.210ff).
Hamilton (1975) took up Haldane’s suggestion, adding
that if the allocation of members to successor groups
following tribe splitting was not random but was rather
involved inthis
what he called ‘‘associative’’ (p.137), group selection
pressures would be further enhanced. Subsequently, a
number of writers have pointed out that group selection
may be of considerably greater importance among
humans than among other animals given the advanced
level of human cognitive and linguistic capabilities, the
consequent capacity to maintain group boundaries and
to formulate general rules of behavior for large groups,
and the resulting substantial influence of cultural
inheritance on human behavior (Alexander, 1987;
Cavalli-Sforza and Feldman, 1973; Boyd and Richer-
son, 1985, 1990; Sober and Wilson, 1994; Boehm, 1997;
Gintis, 2000).
It is now widely accepted that the distribution of
culturally transmitted traits (i.e. learned behaviors) may
be strongly influenced by group selection pressures
(Boyd and Richerson, 2002; Soltis et al., 1995). But
many doubt the importance of group selection for traits
governed by genetic transmission. Whether they are
right is an empirical question: could a genetically
transmitted altruistic trait evolve under the influence
of group selection pressures in an environment approx-
imating past human social and ecological interactions?
This is the question our simulations seek to answer.
Among the distinctive human characteristics which
may enhance group selection effects on genetic variation
is our capacity for the suppression of within-group
phenotypic differences in reproductive or material
success, our patterns of social differentiation supporting
positive assortation (non-random pairing), and the
frequency of intergroup conflict. Thus, the two key
features of our model will be intergroup conflicts and
culturally transmitted group differences in institutional
structure. We stress intergroup conflicts for empirical
reasons: the central role of war and the extinction or
reduced fitness of loser populations in the spread of
behavioral traits. The institutions we model are the
commonly observed human practices of resource shar-
ing among group members including non-kin and
patterns of residence and social differentiation that
result in a greater likelihood of like types interacting
(positive assortation). Our model could easily be
extended to study other group level institutions that,
like resource sharing, reduce the within group variance
of material and hence reproductive success. Included are
information sharing, consensus decision making, and
monogamy.
Group differences in institutional structure persist
over long periods of time due to the nature of
institutions as conventions. A convention is a common
practice that is adhered to by virtually all group
members because the relevant behaviors—for example
sharing meat, or not engaging in extra-pair copula-
tions—are mutual best responses conditional on the
expectation of similar behaviors by most others (Young,
1995). We do not here model the reasons why the
ARTICLE IN PRESS
S. Bowles et al. / Journal of Theoretical Biology 223 (2003) 135–147
136

Page 3

behavior prescribed by the institution is a mutual best
response, but plausible accounts are not difficult to
provide. Those violating sharing norms may bear fitness
costs of ostracism, for example (Boehm, 1993). The
conventional nature of institutions accounts for their
long-term persistence and also their occasional rapid
demise under the influence of shocks. We study
institutional evolution in ways analogous to the evolu-
tion of individual traits. Just as the individuals in our
model are the bearers of genes, groups are the bearers of
institutions, and a successful institution produces many
replicas, while unsuccessful ones are eliminated. The
inheritance of group-level institutions results from a
cultural transmission process based on learned beha-
viors: as new members of the population mature or
immigrate, they adhere to the existing institutions, not
due to any conformist predisposition, but because this
is a best response as long as most others do the same.
The resulting behavioral uniformity in adherence to a
group’s institutions permits us to treat the institution as
a group-level characteristic.
By contrast, the group beneficial individual traits in
our model are replicated by a standard fitness-based
mechanism in which the above pressures for uniformity
are absent. We consider a single individually costly but
group-beneficial trait relevant to dyadic interactions
among group members. Other formally altruistic traits
could be modeled in a similar manner. Included are
individual contributions in an n-person public good
interaction (common defense, insurance as modeled in
Bowles and Gintis, 2003), or the punishment of those
who fail to contribute in such situations, and other ways
that cheating is sanctioned (e.g. Clutton-Brock and
Parker, 1995; Boyd et al., 2003). Our simulations seek
to determine the environments in which such a trait, if
initially rare, can proliferate in the population.
2. Social institutions and multi-level selection
The causal importance of social institutions in our
simulations will be illuminated by a simple model of
multi-level selection in which between-group conflicts
are absent. (We introduce these presently). Consider a
single trait, which may be absent or present in each
individual in a large population, the members of which
each belong to one of a number of groups. For
concreteness, consider an altruistic behavior which costs
the individual c and confers a benefit of b (both
measured in units of some material resource) on a
randomly paired (single) other member of the group. To
simplify this example, suppose that differences in
material payoffs are expressed directly as differences in
fitness. This means that a member in a group composed
entirely of A’s (that is, altruists) has material payoffs
exceeding those of a member in another group with no
altruists by the amount b–c, with fitness correspondingly
higher as well. As we assume b?c>0, altruism is group-
beneficial. But in any mixed group, the expected payoffs
and fitness of altruists will be lower than that of the N’s
(the non-altruists).1So within-group selection will work
against the altruists.
Let pij¼ 1 indicate that individual i in group j has the
trait, with pij¼ 0 otherwise (those without the trait are
N’s). Using a discrete time framework, let p and p0
represent the fraction of the population with the trait
during a given time period and the subsequent period,
respectively, and Dp ? p0? p:2Price (1972) showed that
Dp can be partitioned into group and individual effects.
Define wijas the number of offspring, next period, of an
individual of type i in group j: Let wijdepend additively
on type i’s own trait and on the frequency of the trait in
the group ðpjA½0;1?Þ according to:
wij¼ b0þ pjbgþ pijbi;
ð1Þ
where bgand biare the partial effects on wij of the
frequency of the trait in the group and the presence of
the trait in the individual, respectively (the subscripts
refer to group and individual effects) and b0; a constant,
captures other influences on fitness. Define bG? bgþ bi
as the effect on the group average number of replicas of
the frequency of the trait in the group (the difference in
the number of offspring of an individual in a group
composed entirely of those with the trait and a group
entirely without is bG). Thus using the definitions above,
bi¼ ?c;bg¼ b and bG¼ b ? c: Then following Price
(1972), and taking the expected value of Dp as an
ARTICLE IN PRESS
A
Ab-c -c
N0
N
The Altruism Game:
Row's Payoffs
b
1For groups of infinite size the expected difference in payoffs is c. As
Pepper (2000) shows, the ‘‘large group’’ assumption is required for this
approximation because in a q-person group with qAoq A’s, the
probability that an A will be paired with another A under random
pairing is not pj? qA=q; but ðqA? 1Þ=ðq ? 1Þopj: The discrepancy
arises because the individual A-member cannot be paired with itself.
As a result, for finite group size, the difference in average payoffs of the
A’s and the N’s is not c but rather c þ b=ðq ? 1Þ: For large populations
the discrepancy may be ignored, as it is in order to simplify the
presentation in this section. In the simulations below, the ‘‘large
group’’ assumption is dropped and the relevant probabilities reflect the
actual group size.
2We thus do not provide an explicit treatment of the mating system
and genotype–phenotype relationship but instead assume that the
behavior in question is the expression of a single gene, and that
reproduction is asexual.
S. Bowles et al. / Journal of Theoretical Biology 223 (2003) 135–147
137

Page 4

adequate approximation of Dp due to the large
population size assumed, we have
wDp ¼ varðpjÞbGþ EfvarðpijÞgbi
ð2Þ
or
wDp ¼ varðpjÞðb ? cÞ ? EfvarðpijÞgc;
where w is the population-wide average of the number of
offspring (which we normalize to unity) and the
expectation operator Efg indicates a weighted average
over groups (the weights being relative group size). The
first term captures the group-selection effect (which is
positive), while the second represents the effect of
individual selection, which is negative (a simple deriva-
tion of this decomposition is in Bowles (2001)). Setting
aside degenerate cases such as zero variances, it follows
that an interior frequency of the trait will be stationary
where these two terms are of equal absolute magnitude
(assuming that the b’s and variances making up these
terms are themselves stationary). Because the second
term is negative, the frequency of the A-trait within all
surviving groups will fall over time. But as bGis positive,
this tendency will be offset by the continual extinction
of groups with disproportionately low frequencies of
the trait and their replacement by ‘‘new’’ groups with
disproportionately high frequencies.
Then rearranging the stationarity condition for p (2)
we see that Dp ¼ 0 when
c=b ¼ varðpjÞ=½EfvarðpijÞg þ varðpjÞ?ð3Þ
with
Dp > 0 for c=bovarðpjÞ=½EfvarðpijÞg þ varðpjÞ?
Dpo0 for c=b > varðpjÞ=½EfvarðpijÞg þ varðpjÞ?:
The left-hand term is the benefit-to-cost ratio of the
altruistic trait. The right-hand term is the ratio of
between-group to the within-group plus the between-
group variance of the trait. It is easily shown (Crow and
Kimura, 1970) that this ratio measures the difference
between the probabilities that an altruist will be paired
with an altruist, PðAjAÞ; and that a non-altruist be
paired with an altruist, PðAjNÞ: Thus
r ? varðpjÞ=½EfvarðpijÞg þ varðpjÞ? ¼ PðAjAÞ ? PðAjNÞ:
The variance ratio, r; is thus a population-wide measure
of the degree of non-randomness resulting not because
of non-random pairing within groups, but because the
population is group-structured. Eq. (3) shows that in
order for an altruistic trait to proliferate in a population,
the more costly (relative to the benefits) is the trait, the
greater must be the between-group variance (relative to
the within-group variance).
When the variance among group means is zero, A’s
no longer have the advantage of being in groups with
disproportionally many A’s. In this case group selection
is inoperative, so only a costless form of group benefit
could proliferate. By contrast when varðpijÞ ¼ 0 8j;
groups are either all A or all N, and one meets only
one’s own type, independently of the composition of the
total population. In this case, within-group selection is
absent and between-group selection is the only selective
force at work.
Thus the force of group selection will depend on the
magnitude of the group benefit relative to the individual
cost (b and c in the example) and the degree to which
groups differ in their frequency of the trait, relative to
the degree of within-group variance of the trait.
Rewriting Eq. (3) as rb ? c ¼ 0 we see that the
stationarity condition for p in a group-structured
population is just another version of Hamilton’s rule
for the degree of positive assortation permitting an
altruistic trait to proliferate when rare. In this respect,
multi-level selection works by the same processes as
other evolutionary processes based on non-random
pairing.
Fig. 1 shows how the group structure of the
population overcomes the disadvantage of bearing the
costs of altruistic behaviors. While the expected payoff
to the non-altruist ðpNÞ exceeds that to an altruist ðpAÞ
when they both have the same probability of being
paired with an altruist, the difference in the probability
of meeting an altruist conditional on one’s type may
overcome this disadvantage. The figure illustrates a
value of the variance ratio r (that is, the difference
PðAjAÞ?PðAjNÞ) that is just sufficient to equate the
expected payoffs of the two types and thus to maintain a
stationary value of p. How large this difference must be
depends, as we have seen and as the figure makes clear,
on the payoff differences between the bearers of the two
traits.
Group level social institutions may reduce these
within group payoff (and hence fitness) differences
between the A’s and the N’s. To explore these effects
ARTICLE IN PRESS
-c
b
πN
πA
1
0
P(A|A)P(A|N)
The Probability of Being Paired with an A
b-c
Fig. 1. The evolution of an altruistic trait in a group-structured
population. If the population structure’s variance ratio is such that the
difference in the conditional probabilities of being paired with an A
(P(A|A)?P(A|N)) is as shown (r ¼ r?), p is stationary, because the
expected payoffs of the two types pAand pNare equal.
S. Bowles et al. / Journal of Theoretical Biology 223 (2003) 135–147
138

Page 5

of institutions we need to model the process of
differential replication. Suppose that in the absence of
the group-level institutions (to be introduced presently)
the selection process within a group is modeled (for
group j) by the standard replicator dynamic equation
Dpj¼ pjð1 ? pjÞðpA? pNÞ ¼ pjð1 ? pjÞð?cÞ:
ð4Þ
Now imagine that the group has adopted the practice,
common among foragers and other human groups, of
within-group resource sharing. Some fraction of the
resources an individual acquires—perhaps specific kinds
of food as among the Ache (Kaplan and Hill, 1985)—is
deposited in a common pot to be shared equally among
all group members. This sharing institution may be
modeled as a linear tax, tA½0;1Þ; collected from the
members payoffs with the proceeds distributed equally
to all members of the population. The effect is to reduce
payoff differences between A’s and N’s, that is: pA? pN
¼ ?ð1 ? tÞc: Fig. 2 shows the effect of resource sharing
on the payoff differences of the two types. The
difference in the probability of meeting an A (condi-
tional on ones own type) that equalizes expected payoffs
is no longer PðAjAÞ?PðAjNÞ ¼r? as shown in Fig. 1, but
is now PTðAjAÞ?PTðAjNÞ ¼ rTwith rTor?. Comparing
the two figures one sees that r? ¼ c=b while rT¼
cð1 ? tÞ=b: As a result, were the population structure
as in Fig. 1 ðr?Þ and the sharing institution in place
(t>0), then pA> pN; so p would increase.
Suppose that in addition to the institution of resource
sharing, groups are also segmented, so that in the
pairing process within groups, A’s are more likely to
interact with A’s and N’s with N’s than would occur by
random matching. Suppose that the probability that an
A-member of group j is matched with an A is not pjbut
sjþ ð1 ? sjÞpjXpjand the probability that a N-member
of group j is matched with an A is ð1 ? sjÞpjppj. Then
we define sjX0 as the degree of segmentation in group,
or the difference in the conditional probability of an A
meeting an A and an N meting an A in the within-group
pairing. A transparent interpretation of sj is that it
represents the fraction of interactions that take place
with one’s own type for reasons such as common
residence and the like, the remaining ð1 ? sjÞ pairings
being random. But it could arise for other reasons,
deliberate but imperfect attempts by the A’s to avoid
interactions with the N’s, for example. Then ignoring
the subscripts, and abstracting from the tax: pA? pN¼
sb ? c: Segmentation reduces the expected payoff
disadvantage of altruists because within a given group
they are disproportionately likely to meet other altruists,
while N’s are disproportionately likely to meet other
N’s. If s > c=b; A’s will on average do better than N’s
within every group and as a result the A’s will proliferate
as a result of both within and between-group selection.
Thus, both terms in the Price equation will be positive.
To pose the classical group-selection problem, we
assume soc=b; so the A’s will only proliferate if
group-selection pressures are strong enough. Like
resource sharing, segmentation is a convention and is
passed on culturally.
Taking account of both segmentation and resource
sharing, thedifferences
received by N’s and A’s within a group will now be
ð1 ? tjÞðsjb ? cÞ so we have
Dpj¼ pjð1 ? pjÞð1 ? tÞðsb ? cÞ
intheexpected payoffs
ð5Þ
from which it is clear (comparing (5) and (4)) that both
institutions retard the within-group selection against the
A’s. This can be seen by noting that
@Dpj=@t ¼ ?pjð1 ? pjÞðsb ? cÞ;
@Dpj=@s ¼ pjð1 ? pjÞð1 ? tÞb:
ð6Þ
For pjAð0;1Þ both expressions are positive, meaning
that both segmentation and resource sharing attenuate
the negative selection against the A’s. Note that the
effect of each institution is greater when pj is close to
one half, and when the other institution is at a low
level. Thus, in terms of their benefits in retarding
selection against the A’s, for a given level of pj, the
institutions are substitutes, not complements: their
beneficial effects are enhanced the lesser is the presence
of the other.
The model shows how group level institutions may
retard individual level selection and thus facilitate the
proliferation of an otherwise unviable trait by means of
group selection. But the analysis is incomplete. The
Price equation gives the stationarity condition for p, but
it does not account for the movement of the variances
upon which the movement in p is based. For most
species, the between-group variance-enhancing mechan-
isms (mutation, genetic drift) are weak and tend to be
swamped by the homogenizing effects of selection itself,
along with migration among groups. This is the reason
whygroup-selectionpressuresamongnon-human
ARTICLE IN PRESS
-c
b
πN
1
πA
0
PT(A|N)P(A|N)
-(1-t)c
b-ct
b-c
The Probability of Being Paired with an A
PT(A|A) P(A|A)
Fig. 2. Resource sharing increases the relative importance of popula-
tion structure in the evolution of an altruistic trait. The dashed payoff
functions indicate the effect of within group resource sharing; the
altruistic trait will proliferate if r=r*.
S. Bowles et al. / Journal of Theoretical Biology 223 (2003) 135–147
139

Page 6

animals are thought to be weak. However, among
humans, where effective group size is small (e.g. the
members of a foraging band) and where groups
frequently divide either in response to increased size or
to interpersonal tensions within the group, a process of
even random (rather than Hamilton’s ‘‘associative’’)
division will increase between-group variance.
Thus small group size and frequent group division
coupled with social institutions that attenuate the
within-group selection against the A-trait constitute an
environmentfavorable for
pressures to support the evolution of the A-trait. For
any model even minimally faithful to the empirical
circumstances of human evolution, the only practical
way to determine if these between group variance-
enhancing effects and within group attenuation of
individual level selection are strong enough to make
group selection an important influence on evolution is to
simulate a group-structured population under reason-
able parameter values.
multi-levelselection
3. An agent-based model of multi-level selection
We simulated an artificial population living in 20
groups. For each simulation, total population size is
given and group size is approximately constant,
modified only by random migration among groups
and by the outcomes of group conflict, as explained
below. In the model above, groups with a high
frequency of A’s produce more offspring and thus grow
in size. In the simulations to follow, a group’s size
is restricted by its site, and a high frequency of A’s
contributes to the group’s success in intergroup con-
flicts, allowing it to occupy a new site and thus to
increase in size.
Reflecting the effect of payoffs on fitness, an
individual’s expected share of the group’s next gener-
ation’s offspring is equal to the individual’s share of the
group’s total payoffs. We assume that each individual
has access to material resources from sources other than
the interaction we are modeling and set these ‘‘baseline
payoffs’’ at 10 units. Because offspring are produced in
proportion to the individual’s share of the group’s total
material payoffs and the expected difference in payoffs
is approximately c ¼ 1 (in the absence of segmentation
and resource sharing), the N’s produce 10% more
offspring than the A’s. Individual replication is subject
to mutations, such that with a small probability, e, the
offspring of an A will be an N or an A with equal
probability and conversely.
The institutions represented by s and t differ among
groups, and they also evolve. When conflict occurs
between groups, the group with the higher total payoff
wins. The losing group’s members die and the winning
group populates the site occupied by the losers with
replicas of themselves.3The new inhabitants of the site
adopt the institutions of the group from which they
descended. Institutions are also subject to stochastic
variation, increasing or lowering t and s by chance each
period. Both segmentation and resource sharing impose
costs on the groups adopting them. More segmented
groups may fail to capture the benefits of diversity or of
economies of scale, and resource sharing may reduce
incentives to acquire the resources to be shared. Neither
of these costs are modeled formally, but to capture their
impact, group average benefits are reduced by an
amount that is rising and convex in both s and t. Unlike
many institutions, both s and t may be introduced at low
levels, so the initial emergence of resource sharing and
segmentation could readily take place through the
extension to an initially small number of unrelated
individuals of the practice of within-family resource
sharing or a preference for interaction with individuals
sharing common traits, proximity, or other similarities.
The benchmark values of the parameters in the
simulations, and the range of alternative values that
we explored appear in Table 1. The structure of
our simulation is described in Fig. 3 and its notes.
(Additional details are available at http://www.santafe.
edu/Bbowles/artificial history).
The key parameters concern the rate of (random)
migration among groups, group size, and the probability
in any period that a group will engage in a between-
group conflicts. Because our group conflicts are lethal
for the losers, we have chosen a benchmark probability
of conflict giving an expected frequency of a single war
every four generations. Of course group conflicts more
commonly result in fitness differentials between winners
and losers without group extinctions. Our benchmark
likelihood of an extinction is chosen to reflect the long
term consequences of plausible values of differential
reproductive success between adjacent stronger and
ARTICLE IN PRESS
Table 1
Key parameters for the simulations
Benchmark valuesRange explored
Mean group size (n/g)
Migration rate (m)
Probability of conflict (k)
Mutation rate (e)
20
0.2
0.25
0.001
7–47
0.1–0.3
0.18–0.4
0.01–0.000001
Note: Total population size is n, and there are g groups; m, k and e are
per generation. Other Parameters: benefit (b): 2; cost (c): 1; baseline
payoffs: 10. We varied group size by varying n. For reasons explained
in the text, we restricted s to not exceed 1/2 while tA½0;1?: The costs
imposed on the group by these institutions are ðs2þ t2Þ=2:
3An alternative formulation would have the losing group survive as
a subordinated people with less access to resources and hence reduced
fitness. We have modeled group conflict in this way elsewhere (Bowles
and Choi, 2003a) but will not pursue it here.
S. Bowles et al. / Journal of Theoretical Biology 223 (2003) 135–147
140

Page 7

weaker groups engaged in on going conflict. The other
benchmark values were also chosen on grounds of
empirical plausibility, the evidence for which we review
in the penultimate section.
We initiated each simulation with altruists and
institutions absent at time zero, to see if both the
individual A-trait and the group level institutions would
proliferate if initially rare (the individual and institu-
tional mutation process will introduce some variability
in the population). To explore the effects of varying
parameter values, we ran at least 10 simulations of at
least 10,000 generations for each parameter set investi-
gated, as indicated in the notes to Fig. 6.
The early generations of a typical simulation appear
in Fig. 4. The rise in p is supported by the chance
increase in both s and t (between periods 100 and 150).
ARTICLE IN PRESS
(1)
(1)
(2)
(1)
(3)
(2)
(0)
(1)
(0)
(0)
(0)
group i
1) pairing and interacting
2) pay off determines the
number of offspring of
each player (in parenthesis)
3) new generation and
mutation:
4) migration:
5) competition
between groups:
6) winning group
repopulates the site of
losing group and splits to
two newgroups
7) new group
Agents playing N
Agents playing A
Agents switching by chance
emigrating to group x
immigrating from group y
losing group j
group j'
Go To Step (1)
Go To Step (1)
group i'
winning group i
temporarily enlarged
winning group i
group i
group i
group i
Fig. 3. Individual and group-level selection in the simulation. Notes. We assign n individuals to g groups. At t=0 all are N. 1. Pairing. In each period,
each member of a group is randomly paired to play the PD game once, with another member with payoffs given in the text (in some runs modified by
the resource-sharing rule). With segmentation, the member interacts with a similar type with probability s and is paired randomly with probability
1?s. 2. Reproduction. Replicas of the current generation constitute the next generation. They are produced by drawing (with replacement) from the
current group membership with the probability that any member will be drawn equal to that member’s share of the total payoffs of the group. 3.
Mutation. With probability e a member of the next generation is not a replica of its parent, but is A or N with equal probability. 4. Migration. With
probability m each member of the new generation relocates to a group randomly selected from the other groups. 5. Group competition. With
probability k each group is selected and among those selected competition takes place between randomly paired groups. The winning group is that
with the highest total payoff (net of the costs of sharing and segmentation, if any). 6. Repopulation and fission. The members of the losing group are
replaced by replicas of the members of the winning group, and the resulting (temporarily enlarged) winning group splits with members assigned
randomly to two new groups. (In simulations with resource sharing or segmentation, the two new groups adopt the institutions of the winning
group.)
S. Bowles et al. / Journal of Theoretical Biology 223 (2003) 135–147
141

Page 8

When p reaches high levels (periods 532–588, for
example) both s and t decline, typically leading to a
sharp decline in p. The subsequent rise in s or t occurs by
chance. This pattern emerges for the following reason.
When the population is evenly divided between A’s and
N’s, many groups are also approximately evenly
divided. From Eq. (6), we know that the beneficial
effectsofinstitutions—the
selection gained by higher levels of t or s—are
maximized in this region. When p is well above 0.5,
the benefits of the protection of A’s offered by the
institutions is of less value. But the institutions are costly
to bear, so when p is high, groups with substantial levels
of segmentation or resource sharing are likely to lose
conflicts with other groups, and the sites they occupied
are then populated by the descendants of winners, who
typically bear lower levels of these institutional vari-
ables. As a result, both s and t fall.
To explore further the impact of institutions on the
updating process we estimated the Price equation
statistically, exploring the effect of institutions (that is,
constraining s, t, both, or neither to zero). Using data
from four simulations of 10,000 generation each, we
regressed the observed Dp on the previous period’s
values for varðpjÞ and EfvarðpijÞg; where the second term
is the mean across all groups of the within-group
variances. The coefficients of these variables are
estimates of bGand bifrom Eq (2), respectively. As
Table 2 shows, the combined effect of resource sharing
and segmentation is to reduce by half the extent of
within-group selection against the altruists, that is, the
estimate of biis ?0.102 without institutions and ?0.055
with both institutions. Note that with no institutions the
estimate of bi(0.102) is very close to the expected value
given that the baseline fitness is 10 (so N’s have a 10%
advantage in fitness). The estimate of the between-group
retarded within-group
effect, bG; varies little in response to which institutions
are allowed to evolve, and is in all cases more than four
times as large as the within-group effect. The mean
within-group variance is correspondingly much larger
than the between-group variance.
Note that we can rewrite Eq. (2), the condition for
Dp ¼ 0; as
?Xi=XG¼ varðpjÞ=EfvarðpijÞg ¼ R
ð20Þ
with Dp > 0 if the variance ratio, R, exceeds the ratio of
within- to between-group effects, and conversely. Do we
observe this in our simulations? Using the econometric
estimates of the within and between-group effects
described in Table 2 as well as the mean variance ratios
ARTICLE IN PRESS
Fig. 4. The dynamic interaction between group institutions and individual behaviors. The figure presents a 1000 period history of a run using the
benchmark parameters from Table 1. The population average frequency of altruists is p, while t and s give the average across the 20 groups of the
level of resource sharing and segmentation. Altruism and both group-level institutions are initially rare. The particular time frame shown in Fig. 4
was selected because it clearly reveals this dynamic, which is observed over long periods in many runs.
Table 2
Institutions retard within-group selection against altruists
Institutions
bi
?t
None
Resource sharing
Segmentation
Both
?0.102
?0.080
?0.063
?0.055
8.5
16.6
4
11.2
Note: Column bigives the ordinary least-squares estimate of the
coefficient of the group mean value of pj(1?pj) as a predictor of Dpj
(the other regressor is the between-group variance, i.e var(pj)). The last
column is the negative of the t-statistic for the estimate.
Table 3
An estimate of the price equation
Institutions Effects ratioVariance ratio
p
None
Both
0.252
0.127
0.134
0.132
0.063
0.516
Note: The second column is the ratio ?bi=bG; estimated as described in
Table 2, while the third column is the mean of var(pj)/E{var(pij)} over
the same simulations; p is the average fraction of A’s in the population
for these runs.
S. Bowles et al. / Journal of Theoretical Biology 223 (2003) 135–147
142

Page 9

observed in the same simulations, we have the results in
Table 3. With both institutions constrained to zero, the
ratio of the within-group selection effect to the between-
group selection effect, ?bi=bG; is almost twice the ratio
of between to within-group variances. Thus, were the
population at these mean and estimated values, Dp
would be negative. Thus it is no surprise to find that in
the simulations on which these estimates are based, the
mean value of p is 0.06. However, with both institutions
unconstrained, the variance ratio is approximately equal
to the effects ratio, meaning that the within-group
effects operating against the A’s are exactly offset by the
between-group effects supporting their proliferation. In
the simulation on which these estimates are based the
mean value of p is 0.516.
Between-group conflicts play a key role in supporting
both group-level institutions and individual-level altru-
ism. In the simulations reported, the expected frequency
of conflict was 1/k, where k is the probability that a
group is drawn for a conflict in every generation. It
seems likely that over long historical periods, the
frequency of conflict varied considerably, perhaps in
response to the need to migrate in times of climatic
variability. To explore the sensitivity of the simulations
to the frequency of conflicts, we varied k stochastically
using the auto-regressive system described in the notes
to Fig. 5. During periods in which conflict was frequent
(e.g. around the 21,000th generation), high levels of
altruism were sustained, but periodic outbreaks of
relative peace among the groups (around the 25,300th,
27,000th and 29,600th generations) led to sharp
reductions in the fraction of A’s in the population.
The 500-generation period following generation 28,500
illustrates the strong path dependency in the model. The
high level of p induced by the sharp rise in the frequency
of intergroup conflict around 28,500 persists even as the
frequency of conflict sharply declines in subsequent
generations. But the ‘‘lock-in’’ is not permanent: when k
remains below 0.2 for a number of periods, p crashes.
We sought to answer two other questions as well.
Could altruism have evolved had group level institutions
not co-evolved with individual level altruism? And how
sensitive are our simulations to variations in the key
parameters? To answer these two questions, we varied
group size from 7 to 47, and for each size ran 10
simulations of 50,000 generations, with the other
parameters at their baseline values. We did this with
both institutions constrained to not evolve, with each
singly constrained to not evolve, and with neither
constrained. We performed the same operation for
variations in the migration rate from 0.1 to 0.3, and the
probability of conflict (k) from 0.18 to 0.51. The results
appear in Fig. 6.4
The top panel shows that with both institutions
constrained not to evolve, a group size of 7 supports
high levels of altruism, but group sizes greater than 8
result in a frequency of altruists of less than 0.3. Taking
as a benchmark the group size for which p>0.5, we see
that with no institutions, the critical size is 8, while with
both institutions, p>0.5 for all group sizes less than 22.
The results for the migration rate are similar. Without
institutions, sustaining p>0.5 requires a (per genera-
tion) migration rate of 0.13, but with both institutions
free to evolve, the critical migration rate is 0.21. The
bottom panel shows that institutions also allow the
evolution of substantial frequencies of altruism with
significantly fewer between-group conflicts. A ‘‘vertical’’
reading of the figure is also illuminating: for example,
ARTICLE IN PRESS
0
20000
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
2100022000 23000 24000 250002600027000 28000 29000
generation
average p
k
Fig. 5. High frequencies of group conflict favor altruism. The figure shows a thousand generation period from a run in which both institutions
evolved endogenously, and in which k, the frequency of between-group conflict varies over time according to kt¼ k0þ rkt?1þ stwhere r ¼ 0:99; st
is randomly drawn from the uniform distribution [?0.02, 0.02], and k0is selected so that the mean of ktis the same as the baseline k;namely, 0.25.
4We also investigated whether the institutions would evolve if p is
constrained to zero. They do not, because institutions are costly and
where there are no altruists in the population they perform no group-
beneficial function, thus leading groups that by chance adopt a high
level of sharing or segmentation to lose conflicts in which they are
involved.
S. Bowles et al. / Journal of Theoretical Biology 223 (2003) 135–147
143

Page 10

the bottom panel shows that for k =0.3, p is less than
0.2 without institutions, but is greater than 0.8 with both
institutions free to evolve.5
Experiments with mutation rates ranging from 10?2
to 10?5gave similar results to those shown. Without
institutions, p remains low, while with both institutions,
ARTICLE IN PRESS
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
710
13
16 1922 2528
3134
4043
46
seg
tax
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.10.11 0.12 0.13 0.14 0.15 0.16 0.17 0.18 0.190.20.21 0.22 0.23 0.24 0.25 0.26 0.27 0.28 0.29
Migration rate (m)
seg
tax
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.180.210.240.270.3 0.330.36 0.39 0.42 0.450.480.51
tax
seg
Frequency of Group Conflicts
Probability of Group Conflicts (k)
both institutions
no institutions
Group Size
Group Size (n/g)
Migration Rate
both institutions
no institutions
no institutions
both institutions
Pop. Frequency of Altruist
Pop. Frequency of Altruists
Pop. Frequency of Altruists
37
Fig. 6. Group-level institutions increase the size of the parameter space for which altruistic behaviors are common. Notes. Each data point is the
average frequency of altruists in the entire population over 10 runs of 50,000 periods each for the parameter value indicated on the horizontal axis. In
each panel the other parameters are the benchmark values shown in Table 1. Each run began with p, t, and s set equal to zero. The curve labeled
‘‘none’’ gives the results for runs in which t and s were constrained to zero; the other curves indicate runs in which one or both of the institutions were
free to evolve. (‘‘Tax’’ refers to resource sharing.) The horizontal distance between the curves indicates the enlargement of the parameter space made
possible by group level institutions. The vertical distance between the curves shows the impact of institutions on average p.
5Fig. 6 and Table 2 suggest that segregation is a more powerful
influence than resource sharing: the segmentation alone has a larger
effect that resource sharing alone both in retarding within-group
selection against the A’s and in broadening the parameter space for
which the A’s constitute large fractions of the population. This is
artifact of our modeling choices. The cost functions for s and t are
identical but s has a greater impact on within-group updating, as can
be seen from Eq. (6). Comparing the effect of s when t ¼ 0 with the
(footnote continued)
effect of t when s ¼ 0; we see that the former is b=c times the latter and
b > c because the altruistic behavior is group-beneficial. (In our
simulations, b ¼ 2 and c ¼ 1 so the s-effect is twice the t-effect.) Also,
note that from Eq. (5), if sj¼ c=b ¼ 1=2;Dpj¼ ð1 ? pjÞð1 ? tjÞðsjb ? cÞ
¼ 0; but the value of t required to halt within-group selection against
the A’s is 1. (In the quadratic cost function we used, the costs of t ¼ 1
are four times the cost of s ¼ 1=2:)
S. Bowles et al. / Journal of Theoretical Biology 223 (2003) 135–147
144

Page 11

the average of p in five simulations of 100,000
generations each (for mutation rates of 10?2,10?3,
10?4, and 10?5) exceeds one half. The average p for
the five simulations with a mutation rate of 10?5ranged
from 0.75 to 0.83, in each case a sharp rise in p occurred
between the 17,150th and 25,855th generation, and high
levels of p were sustained throughout the rest of the
simulation. The waiting time before a take-off depends
on the time it takes for a single group to accumulate a
significant number of altruists. This waiting time would
be shortened considerably where there are more than 20
groups. Because we set p=0 at the initial generation,
very low rates of mutation (less than 10?5) sustain low
levels of p over very long periods; when, as very
occasionally occurs, p rises to high levels it is sustained
over very long periods.
4. Early human environments
Does this model illuminate the process by which
human group-beneficial behaviors and group level
institutions might have evolved? The answer must
depend on whether the parameter space in which this
co-evolutionary process occurs in our simulations
approximates the relevant environment, namely the first
50,000 or 100,000 years of modern human existence,
prior to the dramatic transformation of social structure
accompanying the advent of agriculture around 11,000
years ago.
Little is known about the relevant late Pleistocene
environments, and the difficulty in making inferences
about the social organization of human groups during
this period on the basis of contemporary simple societies
is well-known (Foley, 1987, Kelly, 1995). We can say
with some confidence, however, that during much of this
period climate was exceptionally variable and that small
mobile foraging bands composed of both kin and non-
kin, and lacking complex political organization were a
common form of social organization (Richerson et al.,
2001.)
Our benchmark value for group size, 20, is based on
an approximation of the median of the 235 hunter
gather groups recorded in Binford (2001), namely, 19.
Our handling of group size is not entirely realistic,
however. Recall that small size contributes to group-
selection pressures by increasing the between-group
variance arising when groups that win conflicts double
in size and divide. In reality, group fissioning is not by a
random draw, but rather appears to be a highly political
conflict-resolving process in which kin and coalitions are
likely to remain together. Thus fissioning contributes to
between-group variance and to reduce within group
variance in ways which our model does not capture. A
study of fissioning among Amazonian peoples (Neves,
1995, p. 198) reports that ‘‘fissioning y keeps close kin
together but separates them from more distant kinythe
potential line of cleavage is furnished by the division in
patrilineages.’’6
As the bearers of the group-beneficial trait are likely
to be numerically and socially dominant in the winning
group, they may practice what Hamilton termed
associative division by segregating bearers of the ‘‘other’’
trait insofar as recognition of traits or characteristics
correlated with traits allows. Were this the case, it seems
likely that much larger group sizes would sustain the
evolutionary processes indicated above though we have
not simulated an associative fissioning process.
Very little is know about group conflict during early
human history. We do know that deaths due to warfare
constitute a substantial fraction of all deaths in many of
the pre-state societies in the ethnographic and arche-
ological record. The average reported by Keeley (1996)
for ethnographic studies of pre-state societies is 0.19,
and for pre-state societies studied by archeologists is
0.16. This compares with estimates well below 0.1 for
Europe and the US in the 20th century, 0.03 for 19th
century France and 0.02 for Western Europe in the 17th
century. A 50-year record of 200 wars among the Mae-
Enga in New Guinea, for example, took 800 lives from a
population of about 5000, resulting in an annual death
rate from warfare (0.0032 deaths per year per head),
double that of Germany and Russia in the 20th century
but well below the average of the pre-state societies on
record (Keeley, 1996, p. 195). Whether these extra-
ordinarily lethal episodes were common during the Late
Pleistocene is difficult to say. But some speculations are
possible based on what we know about climate change
and the migrations that this induced. Boehm (2000,
p. 19) writes:
.. towards the end of the Pleistocene as anatomically
modern humans began to emerge, group extinction
rates could have risen dramatically as needy bands of
well armed hunters, strangers lacking established
patterns of political interaction frequently collided,
either locally or in the course of long distance
migration.
Ember (1978) collected data on the frequency of
warfare among 50 foraging groups in the present or
recent past. Excluding those who practice some herding
or sedentary agriculture, 64%of the groups had warfare
every 2 years or more frequently. Even excluding those
more bellicose groups who either had horses or relied on
fishing, warfare is described as ‘‘rare’’ in only 12% of the
groups.
While movement between ethno-linguistic units was
probably quite rare, it seems likely that substantial rates
ARTICLE IN PRESS
6Chagnon (1983, pp. 141–143) studied a Yanomamo village that
subdivided and found that average relatedness in the prefission village
was lower than either of the newly formed units.
S. Bowles et al. / Journal of Theoretical Biology 223 (2003) 135–147
145

Page 12

of migration among the bands making up these units
occurred. Migration rates for the 13 societies surveyed
by Rogers (1990) averaged 22% percent a generation
with the maximum (the !Kung) less than one half. As
Rogers’ data refer to somewhat larger than band-sized
groups, these data may understate the rate of migration
somewhat.
5. Conclusion
We have described a process whereby institutions
such as resource-sharing and segmentation provide an
environment within which an individually costly group-
beneficial trait may evolve, and in which these institu-
tions proliferate in the population because of their
contribution to the evolutionary success of the group-
beneficial trait. Our simulations have shown that if
group-level institutions implementing resource sharing
or positive assortation within groups are free to evolve,
group-level selection processes support the co-evolution
of group-beneficial individual behaviors along with
these institutions, even where these institutions impose
significant costs on the groups adopting them. In the
absence of these group-level institutions, however,
group-selection pressures support the evolution of
group-beneficial traits only when intergroup conflicts
are very frequent, groups are small, and migration rates
are low.
Notwithstanding the highly speculative nature of
these inferences, it seems possible that the social and
physical environments of the late Pleistocene may fall
within the parameter space supporting the co-evolu-
tionary trajectories illustrated in Fig. 6. If so, the multi-
level selection model with endogenous institutions may
provide at least a partial account of the evolution of
individually altruistic behaviors as well as group level
resource sharing, segmentation and perhaps other
institutions during this critical period of human evolu-
tion. An implication, explored elsewhere is that altruistic
behaviors and warfare as a group practice may have co-
evolved, the frequency of warfare contributing to the
evolutionary success of altruism, and the presence of a
significant fraction of altruists in a group contributing to
a group’s war-making capacity.7
The main causal mechanisms of the model—institu-
tionalized resource sharing among non-kin and inter-
group conflict in particular—suggest a central role for
uniquely human cognitive, linguistic and other capa-
cities in this process, perhaps helping to account for the
distinctivelevelsof cooperationamongnon-kin
practiced by humans. The same observation suggests
the limited applicability of the model and simulations to
most other animals. However, for species in which
neighboring groups including unrelated members com-
pete for resources or in which group extinctions are
common, a similar model might apply. In these cases
individually costly group-beneficial behaviors may con-
tribute via group size or in other ways to the success of
the group in avoiding extinctions or in gaining resources
from neighboring groups. Examples include social
mammals such as the cooperative mongoose Suricata
suricatta, for which group extinction rates are inversely
correlated with group size and in some years exceed half
the groups under observation (Clutton-Brock et al.,
1999). Similarly, fire ants (Solenopsis invicta) and a large
number of other ant species form breeding groups with
multiple unrelated queens and practice brood raiding
and other forms of hostility toward neighboring groups,
with success positively related to group size (Bernasconi
and Strassmann, 1999). Whether the levels of coopera-
tion observed in these and other species might be
explained in part by the causal mechanisms at work in
our model is an interesting question which we have not
explored.
Acknowledgements
We thank Tim Clutton-Brock, Christopher Boehm,
Robert Boyd, Martin Daly, Marcus Feldman, Steven
Frank, Herbert Gintis, Alan Grafen, Peter Hammer-
stein, Guy Hoelzer, Alex Kacelnik, Hillard Kaplan,
John Pepper, Peter Richerson, Eors Szathmary, David
Sloan Wilson, Margo Wilson, Elisabeth Wood, two
anonymous referees for this Journal, and seminar
participants at the Santa Fe Institute, Department of
Economics, MIT, Faculty of Zoology, Oxford, the
Collegium Budapest, Department of Anthropology,
UCLA, the Dahlem Conference, Berlin, and Depart-
ment of Sociology, University of Washington for
contributions to this paper, Bridget Longridge and Bae
Smith for research assistance and the Santa Fe institute
and the MacArthur Foundation for financial support.
Authors’ coordinates: bowles@santafe.edu, jkchoi@
santafe.edu, and astridh@fee.uva.nl.
References
Alexander, R.D., 1987. The Biology of Moral Systems. Adine de
Gruyter, New York.
Bernasconi, G., Strassmann, J.E., 1999. Cooperation among unrelated
individuals: the ant foundress case. Trends Ecol. Evol. 14 (12),
477–482.
Binford, L., 2001. Constructing Frames of Reference: An Analytical
Method for Archeological Theory Using Hunter–Gatherer and
Environmental Data Sets. University of California Press, Berkeley.
ARTICLE IN PRESS
7A simulation similar to that reported in Fig. 5, for example, shows
that if low values of p made conflicts unlikely and conversely, the
population spends most of the time (over a very long run) with either
high levels of both p and k, or low levels of both (Bowles and Choi,
2003b).
S. Bowles et al. / Journal of Theoretical Biology 223 (2003) 135–147
146

Page 13

Boehm, C., 1982. The evolutionary development of morality as an
effect of dominance behavior and conflict interference. J. Social
Biol. Struct. 5, 413–421.
Boehm, C., 1993. Egalitarian behavior and reverse dominance
hierarchy. Curr. Anthropol. 34 (3), 227–254.
Boehm, C., 1997. Impact of the human egalitarian syndrome on
darwinian selection mechanics. Am. Nat. 150 (suppl.), S100–S121.
Boehm,C.,2000.Groupselection
J. Consciousness Stud. 7 (1-2), 211–219.
Boorman, S., Levitt, P.R., 1973. Group selection on the boundary of a
stable population. Theor. Popul. Biol. 4, 85–128.
Bowles, S., 2000. Economic institutions as ecological niches. Behav.
Brain Sci, 23.
Bowles, S., 2001. Individual interactions group conflicts and the
evolution of preferences. In: Durlauf, S., Young, P. (Eds.), Social
Dynamics. MIT Press, Cambridge, MA, pp. 155–190.
Bowles, S., Choi, J.-K., 2003a. The first property rights revolution.
Santa Fe Institute (working paper).
Bowles, S., Choi, J.-K., 2003b. The coevolution of love and hate. Santa
Fe Institute (working paper).
Bowles, S., Gintis, H., 2003. The evolution of strong reciprocity. Santa
Fe Institute, unpublished manuscript.
Boyd, R., Gintis, H., Bowles, S., Richerson, P., 2003. The evolution of
altruistic punishment. Proc. Natl. Acad. Sci. (in press).
Boyd, R., Richerson, P.J., 1985. Culture and the Evolutionary Process.
University of Chicago Press, Chicago.
Boyd, R., Richerson, P.J., 1990. Group selection among alternative
evolutionarily stable strategies. J. Theor. Biol. 145, 331–342.
Boyd, R., Richerson, P., 2002. Group beneficial noms can spread
rapidly in a structured population. J. Theor. Biol. 215 (3), 287–296.
Buss, L.W., 1987. The Evolution of Individuality. Princeton University
Press, Princeton.
Cavalli-Sforza, L., Feldman, M.W., 1973. Models for cultural
inheritance: group mean and within group variation. Theor. Popul.
Biol. 4 (42), 42–55.
Cavalli-Sforza, L.L., Feldman, M.W., 1981. Cultural Transmission
and Evolution: A Quantitative Approach. Princeton University
Press, Princeton, N.J.
Chagnon, N.A., 1983. Yanomamo: The Fierce People. Holt, Rhine-
hart and Winston, New York.
Clutton-Brock, T.H., Gaynor, D., McIlrath, G.M., Maccoll, A.D.C.,
Kansky, R., Chadwick, P., Manser, M., Skinner, J.D., Brotherton,
P.N.M., 1999. Predation group size and mortality in a cooperative
mongoose, suricata suricatta. J. Anim. Ecol. 68, 672–683.
Clutton-Brock, T.H., Parker, G.A., 1995. Punishment in animal
societies. Nature 373, 209–216.
Crow, J.F., Kimura, M., 1970. An Introduction to Population Genetic
Theory. Harper & Row, New York.
Darwin, C., 1873. The Descent of Man. D. Appleton and Company,
New York.
Durham, W.H., 1991. Coevolution: Genes, Culture, and Human
Diversity. Stanford University Press, Stanford.
Eibl-Eibesfeldt, I., 1982. Warfare man’s indoctrinability and group
selection. J. Comparative Ethol. 60 (3), 177–198.
Ember, C., 1978. Myths about hunter-gatherers. Ethnology 17 (4),
439–449.
intheupperpaleolithic.
Eshel, I., 1972. On the neighbor effect and the evolution of altruistic
traits. Theor. Popul. Biol. 3, 258–277.
Foley, R., 1987. Another Unique Species: Patterns in Human
Evolutionary Ecology. Wiley, New York.
Frank, S.A., 1995. Mutual policing and repression of competition in
the evolution of cooperative groups. Nature 377, 520–522.
Gintis, H., 2000. Strong reciprocity and human sociality. J. Theor.
Biol. 206, 169–179.
Haldane, J.B.S., 1932. The Causes of Evolution. Harper and Brothers,
New York.
Hamilton, W.D., 1975. Innate social aptitudes of man: an approach
from evolutionary genetics. In: Fox, R. (Ed.), Biosocial Anthro-
pology. Wiley, New York, pp. 133–155.
Kaplan, H., Kim, H., 1985. Food sharing among ache foragers: tests of
explanatory hypotheses. Curr. Anthropol. 26 (2), 223–246.
Keeley, L., 1996. War Before Civilization. Oxford University Press,
New York.
Kelly, R.L., 1995. The Foraging Spectrum: Diversity in Hunter-
Gatherer Lifeways. Smithsonian Institution Press, Washington.
Laland, K., Odling-Smee, F.J., Feldman, M., 2000. Group selection:
a niche construction perspective. J. Conciousness Stud. 7 (1/2),
221–224.
Lewontin, R.C., 1965. Selection in and of populations. In: Moore, J.
(Ed.) Ideas in Modern Biology. The Natural History Press,
New York.
Michod, R., 1996. Cooperation and conflict in the evolution of
individuality, 2 Conflict Mediation. Proc. Roy. Soc. London B.
263, 813–822.
Neves, E.G., 1995. Village fissioning in amazonia. Rev. Museo.
Arqueol. Ethnol. (S. Paulo) 5, 195–209.
Pepper, J., 2000. Relatedness in trait group models of social evolution.
J. Theor. Biol. 206, 355–368.
Price, G.R., 1972. Extension of covariance selection mathematics.
Ann. Hum. Genet. 35, 485–490.
Ratnieks, F., 1988. Reproductive harmony via mutual policing by
workers in eusocial hymenoptera. Am. Nat. 132 (2), 217–236.
Richerson, P., Robert,, B, Bettinger, R.L., 2001. Was agriculture
impossible during the pleistocene but mandatory during the
holocene? A climate change hypothesis. Am. Antiquity 66 (3),
387–411.
Rogers, A.R., 1990. Group selection by selective emigration: the effects
of migration and kin structure. Am. Nat. 135 (3), 398–413.
Smith, J.M., 1976. Group selection. Quart. Rev. of Biol. 51,
277–283.
Smith, J.M., Szthmary, E., 1995. The Major Transitions in Evolution.
Oxford University Press, Oxford.
Sober, E., Wilson, D.S., 1994. Reintroducing group selection to the
human behavioral sciences. Behav. Brain. Sci. 17, 585–654.
Soltis, J., Boyd, R., Richerson, P.J., 1995. Can group-functional
behaviors evolve by cultural group selection: an empirical test.
Curr. Anthropol. 36 (3), 473–483.
Uyenoyama, M., Feldman, M., 1980. Theories of kin and group
selection: a population genetics approach. Theor. Popul. Biol. 17,
380–414.
Young, P., 1995. Economics of conventions. J. Econ. Perspectives 10
(2), 105–122.
ARTICLE IN PRESS
S. Bowles et al. / Journal of Theoretical Biology 223 (2003) 135–147
147