Best Friends: Alliances, Friend Ranking, and the
MySpace Social Network
∗
Peter DeScioli
Economic Science Institute
Chapman University
Robert Kurzban
Department of Psychology
University of Pennsylvania
Elizabeth N. Koch
†
Department of Computer Science
University of Minnesota
David Liben-Nowell
Department of Computer Science
Carleton College
Abstract
Like many topics of psychological research, the explanation for friendship is at once
apparently intuitive and yet simultaneously difficult to address empirically. The stan-
dard difficulties worsen when one seeks, as we do, to go beyond “obvious” explanations
(“humans are social creatures”) to ask deeper questions such as: what is the evolved
function of human friendship? In recent years, however, a new window into human
behavior has opened as a growing fraction of people’s social activity has moved online,
leaving a wealth of digital traces behind. One example is a feature of the MySpace
social network that allows millions of users to rank their “Top Friends.” Here we
collect over 10 million people’s friendship decisions from MySpace to test predictions
made by hypotheses about human friendship. We find particular support for the al-
liance hypothesis, which holds that human friendship is caused by cognitive systems
that function to create alliances for potential disputes. Because an ally’s support can
be undermined by a stronger outside relationship, the alliance model predicts that
people will prefer partners who rank them above other friends. Consistent with the
alliance model, we find that an individual’s choice of best friend in MySpace is strongly
predicted by how partners rank that individual.
∗
The authoritative version of this paper appears in Perspectives on Psychological Science, 6(1):6–8, Jan-
uary 2011. This work was supported in part by NSF grant C CF-0728779, by an NSF Graduate Research
Fellowship to the third author, and by a predoctoral fellowship to the first author from the International
Foundation for Research in Experimental Economics (IFREE). We thank Carol Drysdale and HongDa Tang
for helpful comments.
†
Work performed in part at Carleton College.
1
In recent years, hundreds of millions of people have left behind digital records of their
social behavior as they interact via online communities like Facebook, MySpace, and Twitter.
These digital traces form a massive naturalistic data source that can address a wide variety
of psychological questions: Alice is now in a relationship. Bob has 150 friends. Charlie
has 12 new followers. Alice is now single; Bob likes this. Our interes ts in this paper lie
in explanations of human friendship, and MySpace has a feature that makes it particularly
useful for research on c lose relationships. MySpace profiles contain a “Top Friends” list, in
which individuals designate a subset of their friends as “Top Friends” and organize these
friends in a ranked order. Through this feature, MySpace users have created a vast network
of ranked friendships. This data source can be used to test fine-grained predictions about
people’s rankings of friends—and theories about why friendships exist at all.
Humans are unusual in that they form long-term, dyadic relationships with non-relatives,
and the evolved function of friendship behavior remains unclear (Silk, 2003). The traditional
theory is that friendship is a trade relationship in which people exchange goods and services
to reap gains in trade (Trivers, 1971). Cognitive systems designed for trade should closely
monitor benefits given and received (Cosmides and Tooby, 1992; Trivers, 1971). However,
substantial evidence shows that friends cooperate without closely monitoring contributions
(reviewed by Silk (2003)). For friend rankings, trade models predict that people will favor
friends who generate more trade surplus than others, without special concern about friends’
other friendships.
An alternative theory is the alliance hypothesis, which holds that friendship is caused
by cognitive systems designed to assemble a group of allies for potential disputes (DeScioli
and Kurzban, 2009). The value of an ally crucially depends on the partner’s alliances with
others because an individual cannot count on a partner for support when the partner has a
stronger alliance with the individual’s opponent (Snyder, 1997). This point leads to a key
prediction of the alliance hypothesis: an individual will favor those friends who rank fewer
others above the individual. A partner who ranks the individual first, as the best friend, is
particularly valuable because that partner’s alliance support cannot be undermined by an
outside relationship.
We wrote software to collect a sample of ≈11M MySpace profiles, including demo-
graphic data (age, sex, geographic location; Figure 1) and each individual’s rank-ordered
“Top Friends.” We focused on the best-friend network defined by the connections listed in
the first-ranked position in each of these ≈11M lists. Of them, ≈3.5M named best friends
also in the sample. In the best-friend network, we computed the most common connected
components, subsets of individuals connected through best-friend links (Figure 2). Broadly,
the most common component was a mutual pair of best friends, whereas stars, paths, and
other network structures occurred much less frequently (see Supporting Analysis, p. 11).
A central prediction of the alliance hypothesis is that people will be very concerned about
how their friends rank them among other friends. We tested this idea by looking at people’s
decisions about whom they rank first versus second in their “Top Friends .” Specifically,
we tested whether relative rank, defined as alter’s rank of ego (relative to other alters’
ranks of ego, DeScioli and Kurzban (2009)), predicts first-ranked friendship. We also looked
2
Figure 1: The geographic distribution of our sample from the MySpace network. We were
able to determine the locations of 41.7% of crawled profiles in a United States Geological
Survey (USGS) database of longitudes and latitudes of cities in the contiguous United States.
For each longitude–latitude pair x, a circle centered at x with area proportional to the number
of crawled MySpace users declaring their location as x is shown. For scale, circles of sizes
corresponding to 500, 5000, and 50,000 people are shown on the left.
3
Figure 2: Distribution of connected components in the best-friend network. A person is
represented by a circle, and an arrow leads from each person to that person’s best friend. The
figure shows the ten most frequent components, the number of occurrences, the percentage
of components that are of this type (%c), and the percentage of all individuals who occur
in this type of component (%n). Shaded nodes are part of the component’s cycle (see
text). Components without a cycle contain one person whose best friend was uncrawled
or a non-person, indicated by an arrow that do es not point to a circle. The ten depicted
components account for 91.85% of 1,585,561 total components; 70.17% of the 3,445,329 total
best friendships; and 75.38% of the 4,495,696 nonisolated people (individuals who list a best
friend or are listed as a best friend).
4
Ego’s first-ranked friend . . . All friends Same-sex friends
. . . ranks ego better 881,909 68.85 391,804 66.32
. . . is geographically closer 114,266 56.41 45,747 53.85
. . . is opposite sex of ego 534,587 56.85 —
. . . is closer in age 1,151,751 50.71 457,005 50.93
. . . is older 1,199,391 51.36 475,208 51.99
. . . is ranked more often 1,354,710
ns
49.99 562,938 51.51
. . . is ranked #1 more often 914,703 51.83 381,867 53.36
Table 1: Predictors of ego’s first- versus second-ranked friend. Percentages indicate the
proportion of individuals whose first-ranked friends had a better predictor value than the
individual’s second-ranked friend. All percentages differ from chance (50%) at the p < .001
level unless labeled “ns.” In the same-sex sample, egos are included only if ego, ego’s first-
ranked alter, and ego’s second-ranked alter are all male or all female. In all cases, n denotes
the number of individuals who had in-sample first- and second-ranked friends who differed
in the predictor variable.
at demographic predictors based on the alter’s age, sex, and geographic distance from ego.
Finally, we tested two predictors based on network popularity: the number of other members
(excluding ego) who declared the alter as their first-ranked friend, and the total number of
people who ranked the alter somewhere in their “Top Friends” list.
For each predictor variable, we found the subset of egos for which the first- and second-
ranked alters differed in the value of the variable (114,266 < all ns < 1,199,391). We analyzed
whether these differences predict first- and second-ranked friendships (Table 1). Individuals
showed a weak tendency to choose best friends who were older (51%, p < .001, binomial
test) and closer in age (51%, p < .001). Individuals tended to choose best friends who
were opposite sex (57%, p < .001), w hich might reflect romantic partners; however, when
we repeated all comparisons using only same-sex friendships, we found the same qualitative
patterns for all other variables in the same-sex-only sample (Table 1, Figure 3, Figure 4).
The two measures of popularity showed little predictive power: the number of best-frienders
had a significant but small effect (52%, p < .001), and the number of appearances in a “Top
Friends” list was not significant (50%, p = .76).
Consistent with the importance of physical proximity (Festinger et al., 1950; Liben-
Nowell et al., 2005), individuals chose geographically closer individuals as best friends (56%,
p < .001). This result extends previous research by showing that geographic distance predicts
not only the presence or absence of friendships but also fine-grained distinctions among
individuals’ highest-ranked friends.
Finally, the key alliance variable, alter’s rank of ego, was the best predictor of first-ranked
friendship: 69% of egos (p < .001) selected the alter who ranked ego better. This shows that
knowing only how alters rank ego, researchers can distinguish first- and second-ranked friends
with 69% accuracy. The relative rank variable predicted a significantly greater proportion
of best friendships than the next best predictor, geographic proximity (z = 84.44, p < .001).
We conducted the same comparisons between first-ranked versus third- through eighth-
5
