REPORTS

Acknowledgments: We thank the Deutsche Forschungsgemein- with the graphics, and A. Holzapfel for editorial work Figs. S1 to S5 schaft (grants WE2039/8-1, RI525/17-1, KU683/9-1, and (all from Univ. of Cologne). Tables S1 and S2 KU683/12-1) and the U.S. NSF Paleoclimate Program for References financial support, the Alfred Wegener Institute for Polar and Supporting Online Material Marine Research for logistic support, J. Rethemeyer for www.sciencemag.org/cgi/content/full/334/6060/1265/DC1 conducting part of the AMS 14C measurements, B. Weninger Materials and Methods 3 June 2011; accepted 18 October 2011 for discussing the 14C data, D. Sprenk for assisting SOM Text 10.1126/science.1209299

obese individuals both receiving less exposure An Experimental Study of Homophily to valuable health innovations and having fewer sources of positive social influence, resulting in in the Adoption of Health Behavior lower levels of adoption (8, 9, 16, 18). By con- trast, populations in which obese individuals are Damon Centola better socially connected to healthier individuals should provide greater access to health innova- How does the composition of a population affect the adoption of health behaviors and innovations? tions (5, 6, 11, 20) and more social support for Homophily—similarity of social contacts—can increase dyadic-level influence, but it can also force less adopting new behaviors (7, 8, 16, 21), thereby healthy individuals to interact primarily with one another, thereby excluding them from interactions increasing both overall levels of adoption and the with healthier, more influential, early adopters. As a result, an important network-level effect of use of health innovations among the obese homophilyisthatthepeoplewhoaremostinneed of a health innovation may be among the least population (8, 9, 21, 22). likely to adopt it. Despite the importance of this thesis, confounding factors in observational data have An empirical test of these individual and made it difficult to test empirically. We report results from a controlled experimental study on the network-level effects of homophily has proven spread of a health innovation through fixed social networks in which the level of homophily was difficult because homophily in observational data independently varied. We found that homophily significantly increased overall adoption of a new health is usually confounded with other relevant factors behavior, especially among those most in need of it. such as the topological structure of social net- works (5, 11, 23), interpersonal affect in relation- ocial networks are a primary channel for alters than by homophilous ties to similarly low- ships (12, 24), and shared history and frequency the spread of health behaviors (1–3). How- status individuals (12, 15). of interaction among connected individuals Sever, it is not just the existence of social Although these accounts of homophily may (11, 12, 24). Moreover, individual-level factors ties between individuals that matters for diffu- be in tension with one another at the dyadic level, can be difficult to distinguish from relational sion. Just as important are the demographic com- at the network level both views support the the- ones: Are obese individuals simply less likely position of the population and the distribution of sis that homophily will reduce overall adoption, to adopt a health behavior, or is it their social individual characteristics throughout the social thereby increasing health inequalities across di- environment that reduces their likelihood of network (4–6). Homophily—the tendency of so- verse populations (12–14, 16). This network- adoption (25, 26)? Addressing these difficulties cial contacts to be similar to one another—can level effect emerges from the fact that homophily requires the ability to independently control the affect the extent of a behavior’s adoption in a can result in less healthy individuals having fewer degree of homophily in a while population (7–12). At the dyadic level, research social ties to healthier early adopters, which lim- simultaneously holding the topology of the net- on diffusion has suggested that homophilous ties its their level of exposure to health innovations work, and the distribution of individual- and can promote the spread of behavior between (3, 7, 11, 12). Moreover, fewer ties to healthy indi- population-level parameters, constant. individuals (11–13). This is because actors are viduals also means that the exposure that the less We addressed these issues by studying the more likely to be influenced by alters who are healthy individuals do receive is less likely to come effects of homophily on adoption experimen- similar to themselves. However, research on so- from healthier members of the population—who tally. We conducted an Internet-based social net- cial influence has also suggested that the effects may be more effective at influencing others to work experiment (27) in which we manipulated of status can interact with those of homophily adopt new behaviors (7–11)—thereby reducing the the level of homophily in people’ssocialnet- (12, 14, 15). Homophily among high-status likelihood of adoption among those less healthy works. Our study focused on the spread of a individuals may help to promote diffusion, but individuals who are ultimately exposed to the in- health behavior through a World Wide Web social low-status individuals may be more likely to be novation (15, 17). networking environment composed of 710 par- influenced by heterophilous ties to high-status This network thesis has important implica- ticipants, all of whom were recruited from an tions for obese members of a population because online fitness program (28). We constructed two Sloan School, Massachusetts Institute of Technology, the homophilous “clustering” of health character- types of online social networking communities: Cambridge, MA 02142, USA. E-mail: [email protected] istics in social networks (1, 3, 18, 19) can result in (i) homophilously structured populations, in which

AB C DE 6 6 6 6 6 5 5 5 5 5 4 4 4 4 4 3 3 3 3 3 2 2 2 2 2 1 1 1 1 1 0 0 0 0 0

Number of Adopters Number of 1 2 3 4 5 6 7 1 2 3 4 5 6 7 1 2 3 4 5 6 7 1 2 3 4 5 6 7 1 2 3 4 5 6 7 Weeks Fig. 1. (A to E) Time series showing the number of adopters in each of the five trials. Adoption levels are shown for all homophilous (solid circles) and unstructured (open triangles) networks.

www.sciencemag.org SCIENCE VOL 334 2 DECEMBER 2011 1269 REPORTS

individual traits [gender, age, and body mass in- the notification allowed these neighbors to sign of adopting it, Fig. 2 compares adoption levels dex (BMI)] were “clustered” in the social network up. If they completed the sign-up procedure, this among the “obese” (i.e., BMI ≥ 30) and “non- (28, 29), and (ii) unstructured populations, which would result in each of the adopters’ neighbors obese” (BMI < 30) members of the population. had fully “integrated” neighborhoods in which also receiving notifications about the diet diary. Within the homophilous condition, a significantly participants were mixed at random regardless of At the beginning of week 1, we simultaneous- greater number of non-obese than obese individ- their individual characteristics. ly activated seed nodes in each of the 10 networks. uals adopted the behavior (P < 0.05, Mann-Whitney Each participant in the study created an anon- We then observed the spread of adoption for a U test) (Fig. 2A). This is expected because of the ymous online profile, which included the par- total of 7 weeks. high ratio of non-obese to obese individuals in ticipant’s gender, age, BMI, fitness level, diet The results (Fig. 1) show that homophilously the population (6:1) (28) (table S1). However, the preferences, and favorite exercise (28). Subjects structured social networks exhibited significantly fraction of obese adopters was significantly greater were then matched with other participants in more adoption than unstructured networks. De- than the fraction of non-obese adopters (P <0.05) the study—referred to as “health buddies”—as spite low overall levels of adoption, by the third (Fig. 2B), indicating that relative to their popula- members of an online health community. Each week of the study there was already a noticeable tion sizes, homophilous networks promoted greater participant was provided with a personalized difference between conditions: In each of the five uptake of the behavior among obese individuals online “health dashboard” that displayed real-time trials, there were greater numbers of adopters in than among non-obese individuals. By contrast, health information (e.g., daily exercise minutes) the homophilous condition than in the unstruc- in the unstructured condition, both the number as well as basic profile information (as listed tured condition. By week 4, adoption levels were (Fig. 2A) and the fraction (Fig. 2B) of non-obese above) for each participant and his or her health significantly different across experimental con- adopters was significantly greater than that of buddies (28). The health dashboards also dis- ditions (P < 0.05, Mann-Whitney U test). This obese adopters (P < 0.05). Not one obese in- played a record of any health behaviors that the difference increased in week 5 (P < 0.01) and dividual signed up for the diet diary in the un- participants and their health buddies adopted. maintained this level of significance in weeks 6 structured networks, which suggests that obese Participants arriving to the study were random- and 7. At the end of week 7, the total number of members of the population were very reluctant ized to one of the two experimental conditions—a adopters across all homophilous networks was to adopt the behavior. homophilous population condition and an unstruc- more than 3 times the total number of adopters The comparison across conditions shows that tured population condition—that were distinguished in the unstructured networks. Additional analyses homophily significantly increased adoption both only by the clustering of health characteristics in examining the effects of conditions on subjects’ among the obese (P < 0.01) and non-obese (P < the social network neighborhoods. The occupants participation in the study found no significant dif- 0.05) members of the community (Mann-Whitney of the immediately adjacent nodes in the network ferences in subjects’ use of the environment across U test). Among non-obese individuals, adoption (i.e., the network neighbors) constituted a partic- experimental conditions (28) (fig. S9). levels more than doubled in the homophilous net- ipant’s health buddies. All networks in the study To evaluate the spread of the behavior to works. Among obese individuals, the average had the same size (N = 72), neighborhood struc- the members of the population most “in need” fraction of adopters increased from zero in the ture (hexagonal lattice network, clustering coeffi- Z A B cient = 0.4), and degree distribution ( =6),which 4 0.16 ensured that every individual within and between 0.14 conditions had an identical number of health bud- 3 0.12 28 dies ( ). Random assignment of subjects to ex- 0.1 perimental conditions ensured that each network 2 0.08 in the study had equivalent distributions of individ- 0.06 ual characteristics. Consequently, at the start of the 1 0.04 Number of Adopters study, the only difference between conditions was Fraction of Adopters 0.02 the level of homophily within the social networks. 0 0 Unstructured Homophilous Unstructured Homophilous Unstructured Homophilous Unstructured Homophilous We report results from five independent trials Obese Individuals Non-Obese Individuals Obese Individuals Non-Obese Individuals of this experimental design. Each trial consisted of two social networks, one from each experi- C D 20 0.35 mental condition. All five trials of the study ran 0.3 concurrently for 7 weeks. 15 Participants in the study made decisions about 0.25 0.2 whether to adopt an Internet-based diet diary 10 (28). The diet diary was not known, or accessible, 0.15 to anyone except participants in the experiment. 5 0.1 This ensured that the only way a participant could 0.05

learn about or adopt the behavior was to receive a 0 0 Fraction of Exposed Individuals

Number of Exposed Individuals Unstructured Homophilous Unstructured Homophilous Unstructured Homophilous Unstructured Homophilous health dashboard notification from one of his or Obese Individuals Non-Obese Individuals Obese Individuals Non-Obese Individuals her health buddies. “ Adoption was initiated using a healthy seed Fig. 2. (A to E) Exposure and adoption for obese E ” — 0.5 node in each network amemberoftheonline and non-obese individuals in both experimental community with above average fitness, high ex- conditions at the end of week 7. Means and SEs 0.4 ercise minutes, and a low BMI (28). This en- are shown for all trials. Homophilous networks sured that in every network the behavior originated improved (A) the number and (B) the fraction of 0.3 from a “healthy” individual (28, 30). The study individuals who adopted the behavior, as well as 0.2 began by triggering the seed nodes to adopt the increasing (C) the number and (D) the fraction of behavior. Once each seed was triggered, a notifi- individuals who were exposed to the behavior, 0.1 cation appeared on the health dashboards of each and (E) the percentage of exposed individuals ’ who adopted the behavior. 0 of the seed s neighbors, indicating that their health Unstructured Homophilous Unstructured Homophilous buddy had started using the diet diary. Clicking on Percent Exposed Who Adopted Obese Individuals Non-Obese Individuals

1270 2 DECEMBER 2011 VOL 334 SCIENCE www.sciencemag.org REPORTS unstructured networks to more than 12% of Homophily thus not only significantly increased individual choices for homophilous tie forma- the obese population in the homophilous net- obese individuals’ access to the health innova- tion (4, 11, 31, 32), having a shared history or works (Fig. 2B). Remarkably, the number of tion, but also significantly increased their likeli- context (5, 6, 33), or meeting through mutual obese adopters in the homophilous networks was hood of adoption once they were exposed to it. friendships (6, 11, 33)]. Rather, we find that the equal to the number of non-obese adopters in the Our experimental design did not permit caus- simple fact of homophilous relations can pro- unstructured networks (Fig. 2A). Homophily thus al identification at the individual level; however, vide a significant foundation for social influence. did not restrict adoption of the health behavior to dyadic-level analyses of the correlations between Second, these findings distinguish the ef- only the more fit individuals, but instead signif- homophilous ties and the likelihood of behavior fects of choice homophily in the dynamics of icantly increased uptake by the less fit members spread indicate that partial overlap on traits be- tie formation from the role of observed homo- of the population. tween neighbors may be sufficient to significant- phily in the dynamics of behavioral adoption. This Although these results demonstrate a signif- ly increase the likelihood of transmission (28) was accomplished through our experimental de- icant effect of homophily on adoption, they do (fig. S11). Although these correlations do not sign, which permitted manipulation of the level not identify the mechanism responsible for the imply causal effects of specific traits, they do of homophily while holding network structure success of the homophilous networks. Did obese suggest that a minimal level of overlapping char- constant. individuals have higher adoption rates because acteristics between social contacts may improve Third, our use of the lattice network substrate more obese individuals were exposed to the be- the spread of behaviors through social networks for this study provides two important advantages havior, or because homophilous networks in- (28). To formalize this intuition, we develop a over observational network data: (i) The uniform creased the likelihood of adoption among those simple model of homophily and behavioral influ- structure of the lattice network allowed us to who were exposed? To address this question, Fig. ence in which we assume that actors will influ- identify the individual-level effects of homophily 2, C and D, shows the levels of exposure among ence each other if they have a sufficient number on adoption without the confounding effects of obese and non-obese individuals in each condi- of their traits in common (28). The results (Fig. 3) heterogeneous neighborhood structure on indi- tion. Within both conditions, there were no sig- indicate that increased influence of homophilous vidual adoption behavior; and (ii) holding the nificant differences in the fraction of exposed ties—based on relational similarity (6, 12), rather lattice network constant across conditions allowed obese and non-obese individuals (Fig. 2D). than based on individual characteristics (4, 15)— us to identify the aggregate effects of homophily Across conditions, homophily significantly im- offers a good approximation of the empirically ob- independent of topology. In particular, recent proved obese individuals’ access to the health in- served adoption dynamics (28). studies on the diffusion of behavior have shown novation. A significantly greater number (P <0.01) Our findings suggest two primary conclu- that the lack of “wide bridges” in social networks and fraction (P < 0.05) of obese individuals were sions. First, homophily significantly improves the can significantly inhibit the level of adoption in exposed to the behavior in the homophilous net- adoption of health behavior. Homophily can al- a population (23, 27). By keeping bridge width works than in the unstructured networks. Ex- low a behavior to spread more successfully across constant among all of the neighborhoods in all of posure increased from an average of 17% in the a heterogeneous population, providing greater lev- the networks, our design prevented these topo- unstructured networks to an average of 33% of els of exposure to individuals with diverse health logical effects from confounding the effects of the obese population in the homophilous net- characteristics. Second, homophily can significant- homophily. works (Fig. 2D). These findings are particularly ly increase the likelihood of adoption across dy- To test the robustness of our experimental striking given that, by construction, obese indi- adic ties, in particular among obese individuals. findings for alternative network structures, we viduals initially had greater exposure to the health Our findings suggest that obese individuals may replicated our formal model of homophily (Fig. innovation in the unstructured networks than in be more dependent than healthier individuals 3) on more complex topologies (34) (fig. S12). the homophilous networks (28) (Fig. S10). on the composition of their social networks for We found that the effects of homophily on dif- Putting these results together, Fig. 2E reports making decisions about adopting health behav- fusion were equally pronounced in heteroge- the effect of homophily on the likelihood that iors. This indicates that low adoption levels of neously structured social networks (fig. S13). exposed individuals would ultimately adopt. Ho- health innovations among less healthy individuals Finally, at the dyadic level, our findings sug- mophily had no significant effect on the likeli- (7, 8, 16, 20) may be a function of social environ- gest that some minimal level of similarity may be hood of adoption among non-obese individuals ment rather than a baseline reluctance for adop- necessary for alters to influence one another, exposed to the behavior. However, homophily tion (25, 26). indicating a possible “threshold” effect of homo- significantly increased the likelihood of adop- More generally, the experimental approach de- phily on adoption. Our results indicate that if this tion among exposed obese individuals (P < 0.01, veloped here provides four principal advances in minimal level of similarity is satisfied, partial Mann-Whitney U test). In the homophilous net- our understanding of homophily. First, our re- overlapintraitsmaybesufficienttosignificantly works, an average of more than 40% of exposed sults indicate that the positive effects of homophily increase the likelihood of behavioral influence. obese individuals adopted the behavior, as com- on adoption do not depend on the mechanisms This suggests that “hemiphilous” ties between paredwithzerointheunstructurednetworks. that generate homophilous relationships [e.g., partially similar individuals, who bridge diverse social clusters, may be effective in transmitting Fig. 3. Comparison of the average 6 behavior across heterogeneous populations. number of adopters in the homophi- In sum, we found not only that exposure and lous (open circles) and unstructured 5 adoption levels were greatest in homophilous (open triangles) conditions, across all networks, but that the most effective social en- 4 10 experimental trials, with the aver- vironment for increasing the “willingness” of age adoption levels for the model 3 obese individuals to adopt the behavior was the given in equation S1 (28), using homo- one in which they interacted with others with philous (solid circles) and unstructured 2 similar health characteristics. With respect to the (solid triangles) social networks (aver- design of interventions and the promotion of new Number of Adopters 1 aged over 200 realizations). Error bars health innovations, our results indicate that indicate SE. 0 homophilous ties can increase access to health innovations better than providing direct access to

1 2 3 4 5 6 7 diverse individuals across a community. These Weeks findings may be particularly important for the

www.sciencemag.org SCIENCE VOL 334 2 DECEMBER 2011 1271 REPORTS

design of online health communities, where the 15. S. J. Correll, C. L. Ridgeway, in Handbook of Social 31. S. Aral, L. Muchnik, A. Sundararajan, Proc. Natl. Acad. structure of social relations can be explicitly Psychology, J. Delamater, Ed. (Kluwer Academic, Sci. U.S.A. 106, 21544 (2009). ’ New York, 2003), pp. 29–51. 32. D. Centola, J. C. Gonzalez-Avella, V. M. Eguiluz, determined on the basis of individuals health 16. M. Rostila, Sociol. Health Illn. 32, 382 (2010). M. San Miguel, J. Conflict Resolut. 51, 905 (2007). characteristics. 17. D. Strang, S. Soule, Annu. Rev. Sociol. 24, 265 (1998). 33. G. Kossinets, D. J. Watts, Am. J. Sociol. 115, 405 18. R. Crosnoe, K. Frank, A. S. Mueller, Soc. Forces 86, (2009). References and Notes 1189 (2008). 34. J. P. Onnela et al., Proc. Natl. Acad. Sci. U.S.A. 104, 1. N. A. Christakis, J. H. Fowler, N. Engl. J. Med. 357, 19. T. Valente, K. Fujimoto, C.-P. Chou, D. Spruijt-Metz, 7332 (2007). 370 (2007). J. Adol. Health 45, 202 (2009). 2. T. W. Valente, D. Vlahov, Am.J.Pub.Health91, 406 (2001). 20. D. A. Luke, J. K. Harris, Annu. Rev. Public Health 28, Acknowledgments: I thank A. van de Rijt for helpful 3. K. Smith, N. Christakis, Annu. Rev. Sociol. 34, 405 69 (2007). discussion, and N. Christakis, S. Levin, J. Freese, (2008). 21. D. Umberson, R. Crosnoe, C. Reczek, Annu. Rev. Sociol. R. Reagans, E. Zuckerman, J. Fowler, F. Garip, and 4. P. Lazarsfeld, R. K. Merton, in Freedom and Control in 36, 139 (2010). P. DiMaggio for useful comments and suggestions; Modern Society, M. Berger, T. Abel, C. H. Page, Eds. 22. J. Freese, K. Lutfey, in The Handbook of the of W. Pan for programming assistance; A. Wagner, (Van Nostrand, New York, 1954), pp. 18–66. Health, Illness, and Healing, B. Pescosolido, J. K. Martin, T. Groves, and A. Michal for website development; and 5. P. Blau, J. Schwartz, Crosscutting Social Circles J. McLeod, A. Rogers, Eds. (Springer-Verlag, Heidelberg, K. Schive, M. Kirkbride, and MIT Medical for assistance with (Transaction, New Brunswick, NJ, 1984). 2011), pp. 67–84. site design and participant recruitment. Supported by the 6. M. McPherson, L. Smith-Lovin, Am. Sociol. Rev. 52, 23. D. Centola, M. Macy, Am. J. Sociol. 113, 702 (2007). James S. McDonnell Foundation, and by time in residence 370 (1987). 24. M. Granovetter, Am. J. Sociol. 78, 1360 (1973). at the Center for the Study of Social Organization at Princeton 7.F.C.Pampel,P.M.Krueger,J.T.Denney,Annu. Rev. Sociol. 25. L. F. Berkman, I. Kawachi, Social Epidemiology (Oxford University, and the Institute for Research in the Social 36, 349 (2010). Univ. Press, New York, 2000). Sciences at . Data on population 8. J. C. Phelan, B. G. Link, P. Tehranifar, J. Health Soc. 26. I. Kawachi, L. F. Berkman, Neighborhoods and Health characteristics and adoption behavior are provided in (28). Behav. 51 (suppl.), S28 (2010). (Oxford Univ. Press, New York, 2003). 9. V. W. Chang, D. S. Lauderdale, J. Health Soc. Behav. 27. D. Centola, Science 329, 1194 (2010). Supporting Online Material 50, 245 (2009). 28. See supporting material on Science Online. www.sciencemag.org/cgi/content/full/334/6060/1269/DC1 10. F. C. Pampel, Soc. Probl. 49, 35 (2002). 29. We calculated the age and BMI homophily “score” using Materials and Methods 11. M. McPherson, L. Smith-Lovin, J. Cook, Annu. Rev. Sociol. the average difference between an actor and his or her SOM Text 27, 415 (2001). network neighbors. We then minimized this difference, Figs. S1 to S13 12. E. M. Rogers, Diffusion of Innovations (Free Press, while maximizing gender similarity, across all Tables S1 to S4 New York, 1995). neighborhoods in the network. References 13. A. G. Rao, E. M. Rogers, S. N. Singh, Ind. J. Ext. Edu. 30. The results include a subsequent seeding in both 16, 19 (1980). conditions of trial 1 that tests the robustness of our 15 April 2011; accepted 17 October 2011 14. P. DiMaggio, F. Garip, Am. J. Sociol. 116, 1887 (2011). design for alternative seeding strategies. 10.1126/science.1207055

Paper wasps are a good system for examining Specialized Face Learning Is the evolution of face specialization because closely related wasp species differ in their ability Associated with Individual to individually recognize conspecific faces. The paper wasp, Polistes fuscatus, has variable facial features that are used to recognize individual Recognition in Paper Wasps conspecifics (10, 11). Visual recognition is pos- sible in Polistes wasps because they have acute Michael J. Sheehan* and Elizabeth A. Tibbetts vision (12) and live in well-lit nests. P. fuscatus We demonstrate that the evolution of facial recognition in wasps is associated with specialized nests are often initiated by groups of cooperating face-learning abilities. Polistes fuscatus can differentiate among normal wasp face images more queens, in which relative reproduction is deter- rapidly and accurately than nonface images or manipulated faces. A close relative lacking facial mined by a strict linear dominance hierarchy recognition, Polistes metricus, however, lacks specialized face learning. Similar specializations for (13, 14); individual recognition stabilizes social face learning are found in primates and other mammals, although P. fuscatus represents an interactions and reduces aggression within these independent evolution of specialization. Convergence toward face specialization in distant taxa as cooperative groups (15). Some wasp species, such well as divergence among closely related taxa with different recognition behavior suggests that as Polistes metricus, typically nest alone (16)and specialized cognition is surprisingly labile and may be adaptively shaped by species-specific therefore lack competition among queens. Soli- selective pressures such as face recognition. tary nest founding is associated with a lack of facial pattern variability (17), and experiments he cognitive mechanisms underlying learn- interactions, and our brains process the images of have shown that P. metricus does not recognize ing abilities are surprisingly similar across normal conspecific faces differently than any other conspecifics as individuals (18). Ttaxa as diverse as mammals, birds, insects, images (6). Further, individual recognition is a We tested the adaptive evolution of spe- and mollusks (1). Although the mechanisms that type of complex social behavior that could favor cialized face learning by comparing face spe- underlie learning are broadly generalized across specialized cognition (7) because it requires flex- cialization in P. fuscatus and P. metricus.We animals, there is increasing evidence that learning ible learning and memory and has the potential to predicted that P. fuscatus will learn normal face abilities are adaptively shaped by species’ ecology dramatically increase cognitive demands. How- images faster and more accurately than nonface and can be highly specialized (2). One of the most ever, the claim that face specialization is an images or manipulated faces (Fig. 1), whereas striking examples of specialized cognition is spe- adaptation to facilitate individual recognition has P. metricus will not. Comparing learning of normal cialized face learning found in some mammals, been contentious, in part because it is unclear and manipulated face images (Fig. 1) provides including humans (3–5). Individual face recog- whether face learning is based on conserved mech- a particularly good test of face specialization be- nition is an important aspect of human social anisms or has evolved independently in multiple cause manipulated faces are composed of the same mammalian lineages (8, 9). If face specializa- colors and patterns as those of normal faces (table Department of Ecology and Evolutionary Biology, University of tion is an adaptation to facilitate face recognition, S1), but alteration may prevent the perceptual Michigan, Ann Arbor, MI 48109, USA. we predict that specialization will be associated system from identifying the stimuli as faces. We *To whom correspondence should be addressed. E-mail: with the evolution of facial individual recognition tested learning by training wasps to discriminate [email protected] across distant taxa. between two images using a negatively reinforced

1272 2 DECEMBER 2011 VOL 334 SCIENCE www.sciencemag.org