Inferences of Competence from Faces Predict Election Outcomes

Abstract

We show that inferences of competence based solely on facial appearance predicted the outcomes of U.S. congressional elections
better than chance (e.g., 68.8% of the Senate races in 2004) and also were linearly related to the margin of victory. These
inferences were specific to competence and occurred within a 1-second exposure to the faces of the candidates. The findings
suggest that rapid, unreflective trait inferences can contribute to voting choices, which are widely assumed to be based primarily
on rational and deliberative considerations.

Full text

DOI: 10.1126/science.1110589
, 1623 (2005); 308Science
et al.Alexander Todorov,
Election Outcomes
Inferences of Competence from Faces Predict
www.sciencemag.org (this information is current as of July 30, 2008 ):
The following resources related to this article are available online at
http://www.sciencemag.org/cgi/content/full/308/5728/1623
version of this article at:
including high-resolution figures, can be found in the onlineUpdated information and services,
http://www.sciencemag.org/cgi/content/full/308/5728/1623/DC1
can be found at: Supporting Online Material
found at:
can berelated to this articleA list of selected additional articles on the Science Web sites
http://www.sciencemag.org/cgi/content/full/308/5728/1623#related-content
http://www.sciencemag.org/cgi/content/full/308/5728/1623#otherarticles
, 2 of which can be accessed for free: cites 16 articlesThis article
25 article(s) on the ISI Web of Science. cited byThis article has been
http://www.sciencemag.org/cgi/content/full/308/5728/1623#otherarticles
11 articles hosted by HighWire Press; see: cited byThis article has been
http://www.sciencemag.org/cgi/collection/psychology
Psychology
: subject collectionsThis article appears in the following
http://www.sciencemag.org/about/permissions.dtl
in whole or in part can be found at: this article
permission to reproduce of this article or about obtaining reprintsInformation about obtaining
registered trademark of AAAS.
is aScience2005 by the American Association for the Advancement of Science; all rights reserved. The title
CopyrightAmerican Association for the Advancement of Science, 1200 New York Avenue NW, Washington, DC 20005.
(print ISSN 0036-8075; online ISSN 1095-9203) is published weekly, except the last week in December, by theScience
on July 30, 2008 www.sciencemag.orgDownloaded from

We suspect that this system is not unique.
Several cod stocks, inhabiting similar ocean-
ographic regimes (north of 44-N latitude) in
the northwest Atlantic where they were the
dominant predators, collapsed in the early
1990s (decline by 995% of maximum histori-
cal biomass) and failed to respond to complete
cessation of fishing Ethere was one exception-
al stock (table S1)^. For example, the current
biomass of these stocks has increased only
slightly, ranging from 0.4 to 7.0% during the
past 10þ years (table S1). Reciprocal rela-
tionships between macroinvertebrate biomass
and cod abundance in these areas (12) suggest
that the processes that we document for the
Scotian Shelf may have occurred there. On
the other hand, the three major cod stocks
resident south of 44- N, though reaching his-
torical minimum levels at about the same time
as the northerly stocks and experiencing simi-
lar intensive fishing pressure, declined by only
50 to 70%; current biomass has increased from
10 to 44% of historical minimum levels. These
stocks inhabit different oceanographic regimes
with respect to temperature and stratification
and do not show the inverse relationship be-
tween the biomass of macroinvertebrates and
cod found by Worm and Myers (12). These
geographic differences in cod population dy-
namics merit additional study.
The changes in top-predator abundance
and the cascading effects on lower trophic
levels that we report reflect a major pertur-
bation of the eastern Scotian Shelf ecosys-
tem. This perturbation has produced a new
fishery regime in which the inflation-adjusted,
monetary value of the combined shrimp and
crab landings alone now far exceed that of the
groundfish fishery it replaced (13). From an
economicperspective,thismaybeamoreat-
tractive situation. However, one cannot ignore
the fundamental importance of biological and
functional diversity as a stabilizing force in
ecosystems, and indeed in individual popula-
tions (20), in the face of possible future per-
turbations (whether natural or human-made).
One must acknowledge the ecological risks
inherent in Bfishing down the food web[
(21), as is currently occurring on the Scotian
Shelf, or the ramifications associated with
indirect effects reverberating across levels
throughout the food web, such as altered pri-
mary production and nutrient cycling.
References and Notes
1. M. L. Pace, J. J. Cole, S. R. Carpenter, J. F. Kitchell,
Trends Ecol. Evol. 14, 483 (1999).
2. G. A. Polis, A. L. W. Sears, G. R. Huxel, D. R. Strong, J.
Maron, Trends Ecol. Evol. 15, 473 (2000).
3. J. B. C. Jackson, E. Sala, Sci. Mar. 65, 273 (2001).
4. M.Scheffer,S.Carpenter,J.A.Foley,C.Folke,B.Walker,
Nature 413, 591 (2001).
5. P. C. Reid, E. J. V. Battle, S. D. Batten, K. M. Brander,
ICES J. Mar. Sci. 57, 495 (2000).
6. D. R. Strong, Ecology 73, 747 (1992).
7. J. B. Shurin et al., Ecol. Lett. 5, 785 (2002).
8. J. Terborgh et al., Science 294, 1923 (2001).
9. J. A. Estes, M. T. Tinker, T. M. Williams, D. F. Doak,
Science 282, 473 (1998).
10. J. H. Steele, J. S. Collie, in The Global Coastal Ocean:
Multiscale Interdisciplinary Processes, A. R. Robinson,
K. Brink, Eds. (Harvard Univ. Press, Cambridge, MA,
2004), vol. 13, chap. 21.
11. F. Micheli, Science 285, 1396 (1999).
12. B. Worm, R. A. Myers, Ecology 84, 162 (2003).
13. Materials and methods are available as supporting
material on Science Online.
14. J. B. Jackson et al., Science 293, 629 (2001).
15. L. P. Fanning, R. K. Mohn, W. J. MacEachern, Canadian
Science Advisory Secretariat Research Document 27
(2003).
16. W. D. Bowen, J. McMillan, R. Mohn, ICES J. Mar. Sci.
60, 1265 (2003).
17. J. S. Choi, K. T. Frank, B. D. Petrie, W. C. Leggett,
Oceanogr. Mar. Biol. Annu. Rev. 43, 47 (2005).
18. K. T. Frank, N. L. Shackell, J. E. Simon, ICES J. Mar. Sci.
57, 1023 (2000).
19. J. S. Choi, K. T. Frank, W. C. Leggett, K. Drinkwater,
Can. J. Fish. Aquat. Sci. 61, 505 (2004).
20. K. T. Frank, D. Brickman, Can. J. Fish. Aquat. Sci. 57,
513 (2000).
21. D. Pauly, V. Christensen, J. Dalsgaard, R. Froese, F. C.
Torres Jr., Science 279, 860 (1998).
22. We thank the Department of Fisheries and Oceans
staff who collected and maintained the data with
care and thoroughness, and M. Pace, N. L. Shackell, J. E.
Carscadden, and two anonymous reviewers for helpful
criticisms. This research was supported by Fisheries
and Oceans Canada and a grant from the Natural
Sciences and Engineering Research Council of Canada
Discovery (to K.T.F. and W.C.L.).
Supporting Online Material
www.sciencemag.org/cgi/content/full/308/5728/1621/
DC1
Materials and Methods
SOM Text
Table S1
References
4 April 2005; accepted 7 April 2005
10.1126/science.1113075
Inferences of Competence from
Faces Predict Election Outcomes
Alexander Todorov,
1,2
*
Anesu N. Mandisodza,
1
. Amir Goren,
1
Crystal C. Hall
1
We show that inferences of competence based solely on facial appearance
predicted the outcomes of U.S. congressional elections better than chance
(e.g., 68.8% of the Senate races in 2004) and also were linearly related to the
margin of victory. These inferences were specific to competence and occurred
within a 1-second exposure to the faces of the candidates. The findings sug-
gest that rapid, unreflective trait inferences can contribute to voting choices,
which are widely assumed to be based primarily on rational and deliberative
considerations.
Faces are a major source of information
about other people. The rapid recognition of
familiar individuals and communication cues
(such as expressions of emotion) is critical
for successful social interaction (1). Howev-
er, people go beyond the inferences afforded
by a person_s facial appearance to make
inferences about personal dispositions (2, 3).
Here, we argue that rapid, unreflective trait
inferences from faces influence consequential
decisions. Specifically, we show that infer-
ences of competence, based solely on the
facial appearance of political candidates and
with no prior knowledge about the person,
predict the outcomes of elections for the U.S.
Congress.
In each election cycle, millions of dollars
are spent on campaigns to disseminate infor-
mation about candidates for the U.S. House
of Representatives and Senate and to con-
vince citizens to vote for these candidates. Is
it possible that quick, unreflective judgments
based solely on facial appearance can predict
the outcomes of these elections? There are
many reasons why inferences from facial ap-
pearance should not play an important role in
voting decisions. From a rational perspec-
tive, information about the candidates should
override any fleeting initial impressions. From
an ideological perspective, party affiliation
should sway such impressions. Party affilia-
tion is one of the most important predictors of
voting decisions in congressional elections (4).
From a voter_s subjective perspective, voting
decisions are justified not in terms of the can-
didate_s looks but in terms of the candidate_s
position on issues important to the voter.
Yet, from a psychological perspective,
rapid automatic inferences from the facial
appearance of political candidates can influ-
ence processing of subsequent information
about these candidates. Recent models of
social cognition and decision-making (5, 6)
posit a qualitative distinction between fast,
unreflective, effortless Bsystem 1[ processes
and slow, deliberate, effortful Bsystem 2[ pro-
cesses. Many inferences about other people,
including inferences from facial appearance,
1
Department of Psychology,
2
Woodrow Wilson
School of Public and International Affairs, Princeton
University, Princeton, NJ 08544, USA.
*To whom correspondence should be addressed.
E-mail: atodorov@princeton.edu
.Present address: Department of Psychology, New
York University, New York, NY 10003, USA.
www.sciencemag.org SCIENCE VOL 308 10 JUNE 2005
1623
R
EPORTS
on July 30, 2008 www.sciencemag.orgDownloaded from

can be characterized as system 1 processes
(7, 8). The implications of the dual-process
perspective are that person impressions can
be formed Bon-line[ in the very first encoun-
ter with the person and can have subtle and
often subjectively unrecognized effects on sub-
sequent deliberate judgments.
Competence emerges as one of the most
important trait attributes on which people eval-
uate politicians (9–11). If voters evaluate po-
litical candidates on competence, inferences
of competence from facial appearance could
influence their voting decisions. To test this
hypothesis, we asked naBve participants to
evaluate candidates for the U.S. Senate (2000,
2002, and 2004) and House (2002 and 2004)
on competence (12). In all studies, participants
were presented with pairs of black-and-white
head-shot photographs of the winners and the
runners-up (Fig. 1A) from the election races.
If participants recognized any of the faces in
a race pair, the data for this pair were not used
in subsequent analyses. Thus, all findings are
based on judgments derived from facial ap-
pearance in the absence of prior knowledge
about the person.
As shown in Table 1, the candidate who
was perceived as more competent won in
71.6% of the Senate races and in 66.8% of
the House races (13). Although the data for
the 2004 elections were collected before the
actual elections (14), there were no differ-
ences between the accuracy of the prospec-
tive predictions for these elections and the
accuracy of the retrospective predictions for
the 2000 and 2002 elections (15). Inferences
of competence not only predicted the winner
but also were linearly related to the margin
of victory. To model the relation between
inferred competence and actual votes, we
computed for each race the difference in the
proportion of votes (16). As shown in Fig. 1B,
competence judgments were positively corre-
lated with the differences in votes between
the candidates for Senate Er(95) 0 0.44, P G
0.001^ (17, 18). Similarly, the correlation was
0.37 (P G 0.001) for the 2002 House races
and 0.44 (P G 0.001) for the 2004 races.
Across 2002 and 2004, the correlation was
0.40 (P G 0.001).
In the previous studies, there were no
time constraints on the participants_ judgments.
However, system 1 processes are fast and
efficient. Thus, minimal time exposure to the
faces should be sufficient for participants to
make inferences of competence. We con-
ducted an experiment in which 40 partic-
ipants (19) were exposed to the faces of the
candidates for 1 s (per pair of faces) and
were then asked to make a competence judg-
ment. The average response time for the
judgment was about 1 s (mean 0 1051.60
ms, SD 0 135.59). These rapid judgments
based on minimal time exposure to faces pre-
dicted 67.6% of the actual Senate races (P G
0.004) (20). The correlation between compe-
tence judgments and differences in votes was
0.46 (P G 0.001).
The findings show that 1-s judgments of
competence suffice to predict the outcomes
of actual elections, but perhaps people are
making global inferences of likability rather
than specific inferences of competence. To
address this alternative hypothesis, we asked
participants to make judgments on seven dif-
ferent trait dimensions: competence, intelli-
gence, leadership, honesty, trustworthiness,
charisma, and likability (21). From a simple
halo-effect perspective (22), participants
should evaluate the candidates in the same
manner across traits. However, the trait judg-
ments were highly differentiated. Factor anal-
ysis showed that the judgments clustered in
three distinctive factors: competence (compe-
tence, intelligence, leadership), trust (honesty,
trustworthiness), and likability (charisma, lik-
ability), each accounting for more than 30%
of the variance in the data (table S1). More
important, only the judgments forming the
competence factor predicted the outcomes of
the elections. The correlation between the
mean score across the three judgments (com-
petence, intelligence, leadership) and differ-
ences in votes was 0.58 (P G 0.001). In contrast
to competence-related inferences, neither the
trust-related inferences (r 0 –0.09, P 0 0.65)
nor the likability-related inferences (r 0 –0.17,
P 0 0.38) predicted differences in votes. The
correlation between the competence judgment
Fig. 1. (A) An example
of a pair of faces used
in the experiments: the
2004 U.S. Senate race
in Wisconsin. In all ex-
periments, the positions
of the faces were coun-
terbalanced. (B)Scatter-
plot of differences in
proportions of votes be-
tween the winner and
the runner-up in races
for the Senate as a
function of inferred
competence from fa-
cial appearance. The
upper right and lower
left quadrants indicate
the correctly predicted
races. Each point rep-
resents a Senate race
from 2000, 2002, or
2004. The competence
score on the x axis
rangesfrom0to1and
represents the pro-
portion of participants
judging the candidate
on the right to be more
competent than the one
on the left. The midpoint score of 0.50 indicates that the candidates were judged as equally
competent. The difference in votes on the y axis ranges from –1 to þ1 [(votes of candidate on the
right – votes of candidate on the left)/(sum of votes)]. Scores below 0 indicate that the candidate on
the left won the election; scores above 0 indicate that the candidate on the right won the election.
[Photos in (A): Capitol Advantage]
B
A
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
Inferred competence from faces
Differences in votes
Which person is the more competent?
Table 1. Percentage of correctly predicted races for the U.S. Senate and House of Representatives as a
function of the perceived competence of the candidates. The percentages indicate the races in which the
candidate who was perceived as more competent won the race. The c
2
statistic tests the proportion of
correctly predicted races against the chance level of 50%.
Election Correctly predicted c
2
U.S. Senate
2000 (n 0 30) 73.3% 6.53 (P G 0.011)
2002 (n 0 33) 72.7% 6.82 (P G 0.009)
2004 (n 0 32) 68.8% 4.50 (P G 0.034)
Total (n 0 95) 71.6% 17.70 (P G 0.001)
U.S. House of Representatives
2002 (n 0 321) 66.0% 33.05 (P G 0.001)
2004 (n 0 279) 67.7% 35.13 (P G 0.001)
Total (n 0 600) 66.8% 68.01 (P G 0.001)
10 JUNE 2005 VOL 308 SCIENCE www.sciencemag.org
1624
R
EPORTS
on July 30, 2008 www.sciencemag.orgDownloaded from

alone and differences in votes was 0.55 (P G
0.002), and this judgment correctly predicted
70% of the Senate races (P G 0.028). These
findings show that people make highly differ-
entiated trait inferences from facial appearance
and that these inferences have selective effects
on decisions.
We also ruled out the possibility that the
age, attractiveness, and/or familiarity with
the faces of the candidates could account for
the relation between inferences of competence
and election outcomes. For example, older can-
didates can be judged as more competent (23)
and be more likely to win. Similarly, more
attractive candidates can be judged more fa-
vorably and be more likely to win (24). In the
case of face familiarity, though unrecognized
by our participants, incumbents might be more
familiar than challengers, and participants
might have misattributed this familiarity to
competence (25). However, a regression anal-
ysis controlling for all judgments showed that
the only significant predictor of differences in
votes was competence (Table 2). Competence
alone accounted for 30.2% of the variance for
the analyses of all Senate races and 45.0%
of the variance for the races in which can-
didates were of the same sex and ethnicity.
Thus, all other judgments combined contrib-
uted only 4.7% of the variance in the former
analysis and less than 1.0% in the latter
analysis.
Actual voting decisions are certainly based
on multiple sources of information other than
inferences from facial appearance. Voters can
use this additional information to modify initial
impressions of political candidates. However,
from a dual-system perspective, correction of
intuitive system 1 judgments is a prerogative of
system 2 processes that are attention-dependent
and are often anchored on intuitive system 1
judgments. Thus, correction of initial impres-
sions may be insufficient (26). In the case of
voting decisions, these decisions can be an-
chored on initial inferences of competence
from facial appearance. From this perspec-
tive, in the absence of any other information,
voting preferences should be closely related
to such inferences. In real-life voting deci-
sions, additional information may weaken the
relation between inferences from faces and
decisions but may not change the nature of
the relation.
To test this hypothesis, we conducted sim-
ulated voting studies in which participants
were asked to choose the person they would
have voted for in a political election (27). If
votingpreferencesbasedonfacialappearance
derive from inferences of competence, the re-
vealed preferences should be highly correlated
with competence judgments. As shown in Fig.
2, the correlation was 0.83 (P G 0.001) (28).
By comparison, the correlation between com-
petence judgments and actual differences in
votes was 0.56 (P G 0.001). These findings
suggest that the additional information that
voters had about the candidates diluted the
effect of initial impressions on voting deci-
sions. The simulated votes were also correlated
with the actual votes Er(63) 0 0.46, P G 0.001^
(29, 30). However, when controlling for infer-
ences of competence, this correlation dropped
to 0.01 (P 0 0.95), which suggests that both
simulated and actual voting preferences were
anchored on inferences of competence from
facial appearance.
Our findings have challenging implica-
tions for the rationality of voting preferences,
adding to other findings that consequential
decisions can be more Bshallow[ than we
would like to believe (31, 32). Of course, if
trait inferences from facial appearance are
correlated with the underlying traits, the ef-
fects of facial appearance on voting de-
cisions can be normatively justified. This is
certainly an empirical question that needs
to be addressed. Although research has
shown that inferences from thin slices of
nonverbal behaviors can be surprisingly ac-
curate (33), there is no good evidence that
trait inferences from facial appearance are
accurate (34–39). As Darwin recollected in
his autobiography (40), he was almost de-
niedthechancetotakethehistoricBeagle
voyage—the one that enabled the main ob-
servations of his theory of evolution—on
account of his nose. Apparently, the captain
did not believe that a person with such a
nose would Bpossess sufficient energy and
determination.[
References and Notes
1. J. V. Haxby, E. A. Hoffman, M. I. Gobbini, Trends Cognit.
Sci. 4, 223 (2000).
2. R. Hassin, Y. Trope, J. Pers. Soc. Psychol. 78, 837
(2000).
3. L. A. Zebrowitz, Reading Faces: Window to the Soul?
(Westview, Boulder, CO, 1999).
4. L. M. Bartels, Am. J. Polit. Sci. 44, 35 (2000).
5. S. Chaiken, Y. Trope, Eds., Dual Process Theories in
Social Psychology (Guilford, New York, 1999).
6. D. Kahneman, Am. Psychol. 58, 697 (2003).
7. A. Todorov, J. S. Uleman, J. Exp. Soc. Psychol. 39, 549
(2003).
8. J. S. Winston, B. A. Strange, J. O’Doherty, R. J. Dolan,
Nat. Neurosci. 5, 277 (2002).
9. D. R. Kinder, M. D. Peters, R. P. Abelson, S. T. Fiske,
Polit. Behav. 2, 315 (1980).
10. In one of our studies, 143 participants were asked to
rate the importance of 13 different traits in consid-
ering a person for public office. These traits included
competence, trustworthiness, likability, and 10 addi-
tional traits mapping into five trait dimensions that
are generally believed by personality psychologists to
explain the structure of personality: extraversion, neu-
roticism, conscientiousness, agreeableness, and open-
ness to experience (11). Competence was rated as
the most important trait. The mean importance as-
signed to competence was 6.65 (SD 0 0.69) on a scale
Table 2. Standardized regression coefficients of competence, age, attractiveness, and face familiarity
judgments as predictors of differences in proportions of votes between the winner and the runner-up in
races for the U.S. Senate in 2000 and 2002. Matched races are those in which both candidates were of
the same sex and ethnicity.
Predictor
Differences in votes between winner and runner-up
All races Matched races
Competence judgments 0.49 (P G 0.002) 0.58 (P G 0.002)
Age judgments 0.26 (P G 0.061) 0.07 (P 0 0.62)
Attractiveness judgments 0.07 (P 0 0.63) 0.08 (P 0 0.62)
Face familiarity judgments –0.05 (P 0 0.76) 0.03 (P 0 0.86)
Accounted variance (R
2
) 34.9% 45.8%
Number of races 63 47
Fig. 2. Scatterplot of
simulated voting pref-
erences as a function
of inferred compe-
tence from facial ap-
pearance. Each point
represents a U.S. Sen-
ate race from 2000 or
2002. One group of par-
ticipants was asked to
cast hypothetical votes
and another group was
asked to judge the
competence of candi-
dates. Both the com-
petence score and the
voting preference score
range from 0 to 1. The
competence score represents the proportion of participants judging the candidate on the right to
be more competent than the one on the left. The preference score represents the proportion of
participants choosing the candidate on the right over the one on the left. The midpoint score of
0.50 on the x axis indicates that the candidates were judged as equally competent. The midpoint
score of 0.50 on the y axis indicates lack of preference for either of the candidates.
0
0.2
0.4
0.6
0.8
1
0 0.2 0.4 0.6 0.8 1
Inferred competence from faces
Simulated voting preference
www.sciencemag.org SCIENCE VOL 308 10 JUNE 2005
1625
R
EPORTS
on July 30, 2008 www.sciencemag.orgDownloaded from

ranging from 1 (not at all important) to 7 (extremely
important). The importance assigned to competence
was significantly higher than the importance assigned
to any of the other 12 traits (Ps G 0.005).
11. S. D. Gosling, P. J. Rentfrow, W. B. Swan Jr., J. Res. Pers.
37, 504 (2003).
12. See supporting data on Science Online.
13. For the House races in 2002, we were able to obtain
pictures of both the winner and the runner-up for
321 of the 435 races. For the House races in 2004,
we were able to obtain pictures for 279 of the 435
races (12).
14. In the studies involving these races, we used photographs
of the Democratic and Republican candidates (12).
15. In addition, the accuracy of the predictions was not
affected by the race and sex of the candidates. This is
important because participants might have used race
and sex stereotypes to make competence judgments
for contests in which the candidates were of different
sexes and races. For example, in such contests
Caucasian male candidates were more likely to win.
However, if anything, competence judgments pre-
dicted the outcomes of elections in which the
candidates were of the same sex and race (73.1%
for the Senate and 68.5% for the House) more
accurately than elections in which they were of
different sexes and races (67.9% and 64.3%,
respectively). This difference possibly reflects partic-
ipants’ social desirability concerns when judging
people of different race and sex.
16. For races with more than two candidates, we stan-
dardized this difference so that it was comparable to
the difference in races with two candidates. Specif-
ically, the difference between the votes of the win-
ner and those of the runner-up was divided by the
sum of their votes.
17. From the scatterplot showing the relation between
competence judgments and votes for Senate (Fig. 1B),
seven races (three in the lower right quadrant and four
in the upper left quadrant) could be identified as de-
viating from the linear trend. It is a well-known fact
that incumbents have an advantage in U.S. elections
(18). In six of the seven races, the incumbent won
but was judged as less competent. In the seventh
race (Illinois, 2004) there was no incumbent, but the
person who won, Barack Obama, was the favorite
long before the election. Excluding these seven races,
the correlation between competence judgments and
differences in votes increased to 0.64 (P G 0.001). Al-
though incumbent status seemed to affect the strength
of the linear relation between inferences of compe-
tence and the margin of victory, it did not affect the
prediction of the outcome. Competence judgments
predicted the outcome in 72.9% of the races in which
the incumbent won, in 66.7% of the races in which the
incumbent lost, and in 68.8% of the cases in which
there was no incumbent (c
2
G 1.0 for the difference
between these percentages; P 0 0.89).
18. A. D. Cover, Am. J. Polit. Sci. 21, 523 (1977).
19. A bootstrapping data simulation showed that increasing
the sample size to more than 40 participants does not
improve the accuracy of prediction substantially (12)
(fig. S1).
20. Given the time constraints in this study, to avoid
judgments based on salient differences such as race
and sex, we used only Senate races (2000, 2002, and
2004) in which the candidates were of the same sex
and race.
21. For this study, we used the 2002 Senate races. The
judgments in this and the subsequent studies were
performed in the absence of time constraints (12).
22. H. H. Kelley, J. Pers. 18, 431 (1950).
23. J. M. Montepare, L. A. Zebrowitz, Adv. Exp. Soc. Psychol.
30, 93 (1998).
24. T. L. Budesheim, S. J. DePaola, Pers.Soc.Psychol.Bull.
20, 339 (1994).
25. C. M. Kelley, L. L. Jacoby, Acta Psychol. (Amsterdam)
98, 127 (1998).
26. D. T. Gilbert, in Unintended Thought,J.S.Uleman,J.A.
Bargh, Eds. (Prentice-Hall, Englewood Cliffs, NJ, 1989),
pp. 189–211.
27. For these studies, we used the 2000 and 2002 Senate
races (12).
28. An additional analysis from a study in which par-
ticipants made judgments of the candidates for the
Senate (2000 and 2002) on 13 different traits [see
(10) for the list of traits] provided additional evi-
dence that inferences of competence were the key
determinants of voting preferences in this situation.
We regressed voting preferences on the 13 trait judg-
ments. The only significant predictor of these prefer-
ences was the judgment of competence [b 0 0.67,
t(49) 0 4.46, P G 0.001].
29. A similar finding was obtained in an early study con-
ducted in Australia (30). Hypothetical votes based on
newspaper photographs of 11 politicians were closely
related to the actual votes in a local government
election. Moreover, both hypothetical and actual votes
correlated with inferences of competence.
30. D. S. Martin, Aust. J. Psychol. 30, 255 (1978).
31. G. A. Quattrone, A. Tversky, Am. Polit. Sci. Rev. 82,
719 (1988).
32. J. R. Zaller, The Nature and Origins of Mass Opinion
(Cambridge Univ. Press, New York, 1992).
33. N. Ambady, F. J. Bernieri, J. A. Richeson, Adv. Exp. Soc.
Psychol. 32, 201 (2000).
34. There is some evidence that judgments of intelli-
gence from facial appearance correlate modestly with
IQ scores (35). However, these correlations tend to be
small [e.g., G0.18 in (35)], they seem to be limited to
judgments of people from specific age groups (e.g.,
puberty), and the correlation is accounted for by the
judges’ reliance on physical attractiveness. That is,
attractive people are perceived as more intelligent,
and physical attractiveness is modestly correlated with
IQ scores.
35. L. A. Zebrowitz, J. A. Hall, N. A. Murphy, G. Rhodes,
Pers. Soc. Psychol. Bull. 28, 238 (2002).
36. Mueller and Mazur (37) found that judgments of dom-
inance from facial appearance of cadets predicted
military rank attainment. However, these judgments
did not correlate with a relatively objective measure
of performance based on academic grades, peer and
instructor ratings of leadership, military aptitude, and
physical education grades.
37. U. Mueller, A. Mazur, Soc. Forces 74, 823 (1996).
38. There is evidence that trait inferences from facial
appearance can be wrong. Collins and Zebrowitz
[cited in (23), p. 136] showed that baby-faced in-
dividuals who are judged as less competent than
mature-faced individuals actually tend to be more
intelligent. There is also evidence that subtle
alterations of facial features can influence the
trait impressions of highly familiar presidents such
as Reagan and Clinton (39).
39. C. F. Keating, D. Randall, T. Kendrick, Polit. Psychol.
20, 593 (1999).
40. F. Darwin, Ed., Charles Darwin’s Autobiography (Henry
Schuman,NewYork,1950),p.36.
41. Supported by the Department of Psychology and the
Woodrow Wilson School of Public and International
Affairs at Princeton University. We thank M. Savard,
R. Hackell, M. Gerbasi, E. Smith, B. Padilla, M. Pakrashi,
J. Wey, and R. G.-L. Tan for their help with this project
and E. Shafir, D. Prentice, S. Fiske, A. Conway, L. Bartels,
M. Prior, D. Lewis, and two anonymous reviewers for
their comments on previous drafts of this paper.
Supporting Online Material
www.sciencemag.org/cgi/content/full/308/5728/1623/
DC1
Materials and Methods
SOM Text
Fig. S1
Table S1
References
2 February 2005; accepted 7 April 2005
10.1126/science.1110589
TLR11 Activation of
Dendritic Cells by a Protozoan
Profilin-Like Protein
Felix Yarovinsky,
1
*
Dekai Zhang,
3
John F. Andersen,
2
Gerard L. Bannenberg,
4
. Charles N. Serhan,
4
Matthew S. Hayden,
3
Sara Hieny,
1
Fayyaz S. Sutterwala,
3
Richard A. Flavell,
3
Sankar Ghosh,
3
Alan Sher
1
*
Mammalian Toll-like receptors (TLRs) play an important role in the innate
recognition of pathogens by dendritic cells (DCs). Although TLRs are clearly
involved in the detection of bacteria and viruses, relatively little is known
about their function in the innate response to eukaryotic microorganisms. Here
we identify a profilin-like molecule from the protozoan parasite Toxoplasma
gondii that generates a potent interleukin-12 (IL-12) response in murine DCs
that is dependent on myeloid differentiation factor 88. T. gondii profilin ac-
tivates DCs through TLR11 and is the first chemically defined ligand for this
TLR. Moreover, TLR11 is required in vivo for parasite-induced IL-12 production
and optimal resistance to infection, thereby establishing a role for the recep-
tor in host recognition of protozoan pathogens.
Mammalian Toll-like receptors (TLRs) play
a fundamental role in the initiation of im-
mune responses to infectious agents through
their recognition of conserved microbial mo-
lecular patterns (1). TLR signaling in antigen-
presenting cells, such as dendritic cells (DCs),
results in the production of cytokines and
costimulatory molecules that are required for
initiation of the adaptive immune response
(2, 3). Human and mouse TLR family mem-
bers have been shown to have distinct ligand
specificities, recognizing molecular structures
such as lipopeptide (TLR2) (4), lipopolysaccha-
ride (TLR4) (5, 6), flagellin (TLR5) (7), double-
and single-stranded RNA (TLR3 and TLR7)
(8–11), and CpG motifs of DNA (TLR9) (12).
Although several TLRs have been shown to be
important for immune responses to microbial
products in vitro, their role in host resistance
to infection appears to be complex and not
10 JUNE 2005 VOL 308 SCIENCE www.sciencemag.org
1626
R
EPORTS
on July 30, 2008 www.sciencemag.orgDownloaded from

1
Materials and Methods
Participants
Participants were undergraduate or graduate students at Princeton University.
They participated for partial fulfillment of course credit or for payment. The total number
of participants across all studies was 843.
Selection of photographs
All initial photographs were obtained from the website of the Cable News
Network (CNN). There are 33 to 34 races for the US Senate and 435 races for the US
House of Representatives every two years. For the Senate races in 2000 and 2002, and
the House races in 2002, we obtained photographs of the winner and the runner-up in the
race. For the Senate and House races in 2004, we obtained photographs of the Republican
and Democratic candidates. If the quality of a photograph was poor, a web search for a
new photograph was undertaken and if a better quality photograph of the candidate was
found, it replaced the old photograph.
Because we were interested in inferences from facial appearance, uncontaminated
by prior knowledge, races involving highly familiar individuals were excluded from the
studies. For the Senate elections in 2000, we excluded the races for New York (Hillary
Clinton), Connecticut (Joe Lieberman), New Jersey (John Corzine), and Missouri (John
Aschcroft). The New Jersey race was excluded because all studies were conducted in
Princeton, NJ, and John Corzine’s campaign was extensively covered in the media. For
the Senate elections in 2002, we excluded the race for Massachusetts (John Kerry). For
the Senate elections in 2004, we excluded the race for Arizona (John McCain). The race

2
for Idaho was excluded too, because Senator Mike Crapo ran unopposed. Thus, we used
30 pairs of faces for the 2000 Senate races, 33 pairs for the 2002 races, and 32 pairs for
the 2004 races.
For many of the House of Representatives races, the photograph of one of the two
major candidates was missing and we were unable to include these races. We also
excluded races that were uncontested (i.e., there was only one candidate). Finally, we
excluded races involving highly familiar individuals. For the 2002 elections, these were
the race for Ohio (Dennis Kucinich) and the race for Missouri (Dick Gephardt), both
candidates highly familiar from the Democratic presidential primary. In total, we were
able to obtain photographs of the winner and the runner-up for 321 House races in 2002,
and photographs of the Republican and Democratic candidates for 279 House races in
2004.
All photographs were transformed to black-and-white bitmap files and
standardized in size. Any conspicuous background (e.g., the Capital or a U.S. flag) was
removed and replaced with gray background. For the photographs of candidates for the
2004 election, we created a standard gray background. Then, we removed all faces from
their original background and superimposed them on the standard gray background. In all
studies, each critical trial consisted of a pair of standardized black-and-white photographs
of the major contenders for a congressional race (Fig. 1a). The photographs were
presented either in a questionnaire format or on a computer screen.
General experimental procedures

3
In all studies, the position of the photographs was counterbalanced across
participants. We created two versions for each pair of faces - one in which the photograph
of the winner was positioned on the right side, and another in which the same photograph
was positioned on the left side. For the election in 2004, because the data were collected
before the outcome of the election was known, we used the party affiliation of the
candidates to counterbalance the positions of the photographs. That is, in one version the
photograph of the Democratic candidate was positioned on the right side, and in another
it was positioned on the left side. To avoid confounding the position of the photograph
and the election outcome, within each election – Senate 2000, Senate 2002, and House
2002 – for half of the races the photograph of the winner was positioned on the right side
and for the other half it was positioned on the left side. Within the 2004 elections, for half
of the races the photograph of the Democratic candidate was positioned on the right side
and for the other half it was positioned on the left side. Given the counterbalancing of the
positions of the photographs, two main experimental versions were created for each
study. Participants were randomly assigned to one of the versions. We also randomized
the order of presentation of the races. This manipulation is described below.
Procedures for questionnaire studies
Most of the studies, involving the elections for the Senate, were administered in
questionnaire sessions. Participants were asked to respond to a number of different and
unrelated questionnaires. The questionnaire with the Senate races was embedded in this
set of questionnaires. Participants worked individually and were paid $8 for their
participation in the questionnaire session. In all studies, participants were kept naïve with

4
respect to the objectives of the study. The studies were described as studies on face
perception and participants were encouraged to work as quickly as possible and to rely on
their “gut instincts” when responding. In addition to the main person judgments,
participants were always asked whether they recognized any of the faces. For all analyses
reported in the paper, judgments for races in which the participant recognized any of the
faces were excluded.
We also manipulated the order of presentation of the pairs of faces. For each
election – Senate 2000, Senate 2002, and Senate 2004 – we generated four different
random orders. With the counterbalancing of the positions of the pictures, this created
eight versions of the questionnaires: 2 (position) X 4 (random order). Participants were
randomly assigned to one of the eight versions.
Competence and other trait judgments. In the first study, conducted in May 2003,
we used only the Senate races from 2002. The races were divided into two sets, and 114
participants were randomly assigned to one of the sets. For each pair of faces, participants
were asked to make seven trait judgments: competence, trustworthiness, honesty,
leadership, intelligence, charisma, and likeability. Specifically, they selected the person
who they perceived as better on the respective trait (e.g., the more competent person).
We collected competence judgments for the Senate races in both 2000 and 2002
in three different waves. The first wave of data was collected in Fall 2003. One hundred
participants were presented with the photographs of the candidates for the Senate and
asked to select the person who was more competent. Participants made competence
judgments either for the races in 2000 (n = 50) or for the races in 2002 (n = 50).

5
Participants recognized very few of the politicians. The maximum number of
recognitions was 8 (out of 50 responses) for the Massachusetts race in 2000 (Senator
Kennedy was the incumbent). The mean recognition per race was 1.35 (SD = 1.59).
Because individual competence judgments for races in which participants recognized any
of the faces were excluded, the aggregated competence judgment for each race was based
on 42 to 50 individual judgments. In the first wave of data collection, another 100
participants were asked to make attractiveness, face familiarity, or age judgments. These
judgments are described below.
The second wave of data was collected in the beginning of 2004. In this study,
127 participants were asked to make thirteen trait judgments per pair of faces. The first
judgment was the competence judgment. In addition to this judgment, participants were
asked to decide who was more 2) honest, trustworthy; 3) likeable; 4) extraverted,
enthusiastic; 5) reserved, quiet; 6) calm, emotionally stable; 7) anxious, easily upset; 8)
dependable, self-disciplined; 9) disorganized, careless; 10) sympathetic, warm; 11)
critical, quarrelsome; 12) open to new experience, complex; and 13) conventional,
uncreative. Competence, trustworthiness, and likeability represented the three main traits
identified in the factor solution of the first study with multiple trait judgments per pair of
faces. There is a general consensus among personality psychologists that personality can
be explained in terms of five global factors: extraversion, neuroticism, conscientiousness,
agreeableness, and openness to experience (S1). We used a validated scale of 10 traits to
measure these five trait dimensions (S2): 4 and 5 (reversely scored) for extraversion, 6
(reversely scored) and 7 for neuroticism, 8 and 9 (reversely scored) for conscientiousness,

6
10 and 11 (reversely scored) for agreeableness, 12 and 13 (reversely scored) for openness
to experience.
To reduce the response burden on participants, within each Senate election – 2000
and 2002 - the races were divided into two sets. Thus, we created 4 sets of pairs of faces:
two sets for the 2000 races with 15 pairs of faces per set, and two sets for 2002 with 16
and 17 pairs per set respectively. Participants were randomly assigned to one of the four
sets (32, 31, 31, and 33 participants per set respectively). The maximum number of
recognitions per race was 6 and the mean recognition was 0.56 (SD = 1.33). The
aggregated competence judgment for each race was based on 25 to 33 individual
judgments.
The third wave of data was collected in May 2004. Seventy-four participants were
presented with all races for the Senate in 2000 and 2002 and asked to make a single
competence judgment for each of the 63 pairs of faces. The maximum number of
recognitions per race was 9 and the mean recognition was 2.79 (SD = 2.50). The
aggregated competence judgment for each race was based on 65 to 74 individual
judgments. Another 73 participants were asked to express a voting preference for each of
the 63 pairs of faces. These judgments are described below.
The competence judgments across the three waves of data collection were highly
correlated. The pair-wise correlations were greater than .76, p < .001, and a measure of
the internal consistency of the judgments, Cronbach’s alpha, indicated high reliability of
the combined competence judgment, α = .91. For these reasons, we computed a mean
competence judgment weighted according to the number of individual competence

7
judgments per pair of faces for each wave of data collection. The predictions for the
outcomes of the Senate races in 2000 and 2002 (Table 1) are based on this mean
competence judgment. However, it should be noted that each of the three competence
judgments predicted the outcomes of the elections for the Senate. Both the judgments
from wave 1 and wave 2 data collection predicted the outcomes of 66.7% of the races, χ
2
= 7.00, p < .008. The competence judgment from wave 3 predicted 74.6% of the races, χ
2
= 15.25, p < .001. The correlation of the judgment with the differences in votes between
the candidates was .50 for wave 1, .49 for wave 2, and .55 for wave 3, all ps < .001.
The competence judgments for the 2004 Senate elections were collected two
weeks before the elections on November 2, 2004. One hundred and twenty-seven
participants were asked to make single competence judgments for each of the 32 pairs of
faces. The maximum number of recognitions per race was 36 for the Illinois race. The
mean recognition was 7.13 (SD = 7.05). The aggregated competence judgment for each
race was based on 91 to 127 individual judgments.
Attractiveness, age, and face familiarity judgments. In addition to trait judgments
of the candidates for the Senate in 2000 and 2002, we collected judgments of age,
attractiveness, and face familiarity. A total of 100 participants were asked to make these
judgments. Participants were randomly assigned either to the races in 2000 or the races in
2002. For each pair of faces, 34 participants decided who was more attractive, 24
participants decided who was older, and 42 participants decided whose face was more
familiar.

8
Simulated voting preferences. Another 73 participants were asked to cast
hypothetical votes for each of the Senate races in 2000 and 2002. Participants were asked
to imagine that this was a political election and to choose the person for whom they
would vote. As with all other judgments, if participants recognized any of the candidates
in a race, the data for this race were not used in the analyses.
Procedures for computer studies
The studies for the House of Representatives races involved a large number of
faces (321 pairs of faces for 2000 and 279 for 2002) and the studies were computerized.
We also conducted a computerized study for the Senate races in order to control the rate
of presentation of faces and to measure response times for competence judgments. In all
three studies, participants were asked to make single competence judgments. As with the
questionnaire studies, the position of the faces was counterbalanced and participants were
randomly assigned to one of the two experimental versions. In all three studies, the order
of the pairs of faces was randomized for each participant by the computer. Participants
were kept naïve with respect to the objective of the experiment. The studies were
described as studies on face perception and how people assess personality from faces
alone. At no point in the study was any mention made of candidates or elections.
Participants worked in individual computer booths.
Forty-seven participants made judgments of the candidates competing for the
House in 2002, and 41 participants made judgments of the candidates competing in 2004.
Each experimental trial consisted of a pair of faces. The photographs had a standard size
of 3.2 cm (width) X 4.5 cm (height). The faces were positioned at the center of the screen

9
with a distance of 4 cm between them (2 cm to the central point of the screen). Each pair
of faces was presented on the screen until the participant selected the face that they
perceived as more competent. The next trial was presented immediately after the
participant’s response. At the end of the computer experiment, the experimenter showed a
printout of all photographs to the participant and asked them to circle faces that they
recognized.
Forty participants participated in the Senate study. In this study, we used all races
from 2000, 2002, and 2004 in which the candidates were of the same gender and race.
The total number of these races was 68 (24 from 2000, 24 from 2002, and 20 from 2004).
Each experimental trial started with a fixation point presented for 500 ms at the center of
the screen. Then, the pair of faces was presented for 1 second. The presentation was
immediately followed by the participant’s judgment of competence. After the
competence judgment, participants were asked whether they recognized any of the faces
in the pair. The inter-trial interval was 1 second.
Analyses
The unit of analysis in all studies was the individual election race – a state race in
the case of the elections for the Senate and a district race in the case of the elections for
the House of Representatives. For each race, the individual judgments were aggregated,
excluding judgments where participants recognized any of the faces for the respective
race. The final judgment could range from 0 to 1, reflecting the proportion of participants
who expressed a preference for one of the candidates. For example, if 45 out of 50
participants perceived one of the candidates as more competent, the competence score

10
was .90, assuming that none of the participants recognized any of the candidates. If more
than 50% of participants perceived the winner of the race as more competent, the race
was recorded as correctly predicted.
The competence score was also used in correlation and regression analyses as a
predictor of the differences in votes between the candidates in a given race. Although
only two candidates competed in most of the races, in many races there were more than
two candidates. To create a standardized difference score, we used only the votes for the
winner and the runner-up. This score was computed as the difference between the votes
for the person positioned on the right side in the studies and the votes for the other person
divided by the sum of the votes. Because the winner was presented on the right side for
half of the trials and on the left side for the other half, the difference score could range
from –1 to 1.
The response times for the competence judgments in the Senate study described
above were positively skewed. To remove outliers, we used a liberal procedure of not
using response times that were more than 10 standard deviations above the mean. The
mean response time across participants was 1181.56 milliseconds with a standard
deviation of 284.71 ms. Excluding response times longer than 4028.66 ms (mean + 10
SD) removed response times for only 2.02% of the trials.

11
Supporting text: Simulation of accuracy of prediction
In order to estimate how the accuracy of prediction of the outcomes of the Senate
elections changes as a function of the number of participants making competence
judgments from facial appearance, we performed bootstrapping data simulations using
two different samples of participants. The first sample consisted of 74 participants who
were asked to make single competence judgments for each of the 63 races for the Senate
in 2000 and 2002. The second sample consisted of 127 participants who made single
competence judgments for each of the 32 Senate races in 2004. In both cases, we
repeatedly drew random samples of participants with a fixed size. Starting with a sample
size of 10, we drew 50 random samples. For each of the 50 samples, we recorded the
percentage of correctly predicted Senate races. We repeated this procedure for larger
sample sizes, increasing the sample size by 10 participants at each step (Fig. S1).
The average individual accuracy was better than the chance prediction of 50%,
(M = 59%, SD = 7%), t(73) = 10.58, p < .001, for the predictions for the 2000 and 2002
Senate races, and (M = 53%, SD = 10%), t(126) = 3.58, p < .001, for the predictions for
the 2004 races. As shown in Fig. S1, increasing the sample size to 40 participants
substantially increased the accuracy of prediction over the average individual accuracy,
t(49) = 26.29, p < .001, for the 2000 and 2002 races, and t(49) = 19.46, p < .001, for the
2004 races. However, the benefit of increased sample size diminished after this point. In
fact, for both simulations – 2000/02 and 2004 races - the accuracy of prediction was
significantly improved with the increase of the sample size from 30 to 40 participants,
t(49) = 2.82, p < .007, and t(49) = 2.09, p < .042, respectively; but did not improve

12
significantly with the increase of the sample size from 40 to 50 participants, t = 1.05, p =
.30, and t < 1 respectively.
To capture the diminishing benefit of the increased sample size for the accuracy
of prediction, we modeled the accuracy as an inverse function of the sample size. For the
2000 and 2002 data, this function (1) accounted for 98.9% of the variance. For the 2004
data, the function (2) accounted for 94.1% of the variance.
(1) Accuracy (%) = 74.65 – 140.28 / n
(2) Accuracy (%) = 67.17 – 145.14 / n
Although the coefficients were slightly different, in both models increasing the sample
size from 10 to 40 participants increases the accuracy by more than 10%. At the same
time, increasing the sample size from 40 to 100 participants increases the accuracy by
slightly more than 2%.

13
Figure S1. Data simulation of accuracy of prediction of outcomes of the races for the US
Senate as a function of sample size for A) 2000 and 2002 Senate races and B) 2004
Senate races. The plots show the means of the 50 sample means for each sample size with
their corresponding standard errors.

14
Table S1. Factor loadings of trait judgments of candidates competing for the US Senate
in 2002 on factors identified in a principal components analysis with a Varimax rotation.
The factor analysis was performed on the aggregated judgments for each trait at the level
of the Senate races.
Factor solution
Traits
Competence
Trustworthiness
Likeability
Competence
.90
.33
-.20
Intelligence
.89
.21
-.26
Leadership
.82
-.34
.36
Honesty
.05
.98
.03
Trustworthiness
.16
.96
-.01
Charisma
-.05
-.14
.98
Likeability
-.12
.17
.96
Accounted variance
33.0%
31.4%
30.3%

15
References
S1. O. P. John, S. Srivastava, in Handbook of Personality: Theory and Research, L.
A. Pervin, O. P. John, Eds. (Guilford, New York, 1999), pp. 102-138.
S2. S. D. Gosling, P. J. Rentfrow, W. B. Swan, Jr., J. Res. Pers., 37, 504 (2003).