Gender Differences in Mate Selection
Gender Differences in Mate Selection
Gender Differences in Mate Selection
I. INTRODUCTION
The choice of a marriage partner is one of the most serious
decisions people face. In contemporary Western societies, this
decision usually follows a long learning period during which
people engage in more informal and often polygamous relationships, i.e., dating, which is the topic of this paper. In particular,
we analyze gender differences in dating preferences. As in all
matching markets, determining dating preferences from equilibrium outcomes is difficult because a given correlation of attributes across partners is often consistent with various preference structures. We overcome this problem by studying dating
behavior using an experimental Speed Dating market. In our
experimental paradigm, subjects meet a number of potential
mates (between 9 and 21, a number determined by the experimenters) for four minutes each, and have the opportunity to
accept or reject each partner.1 If both parties desire a future
673
674
675
676
AND
DATA DESCRIPTION
677
reported their names and e-mail addresses and completed a preevent survey.
SettingThe Speed Dating events were conducted in an enclosed room within a popular bar/restaurant near the campus.
The table arrangement, lighting, and type and volume of music
played were held constant across events. Rows of small square
tables were arranged with one chair on either side of each table.
ProcedureThe events were conducted over weekday evenings during 20022004; data from fourteen of these sessions are
utilized in this study.4 In general, two sessions were run in a
given evening, with participants randomly distributed between
them. Participants were not aware of the number of partners they
would be meeting at the Speed Dating event. The number of
participants and dates of each session are listed in Table I.5
Upon checking in, each participant was given a clipboard, a
pen, and a nametag on which only his or her ID number was
written. Each clipboard included a scorecard with a cover over it
so that participants responses would remain confidential. The
scorecard was divided into columns in which participants indicated the ID number of each person they met. Participants would
then circle yes or no under the ID number to indicate whether
they would like to see the other person again. Beneath the Yes/No
decision was a listing of the six attributes on which the participant was to rate his or her partner: Attractive, Sincere; Intelligent; Fun; Ambitious; Shared Interests.6
After all participants had arrived, two hosts instructed the
participants to sit at the two-person tables. The females were told
to sit on one side of the tables, while the males were seated across
from them. Males were instructed to rotate from table to table, so
4. We ran a total of 21 sessions. Seven have been omitted: one because we
imposed a maximum number of acceptances, two because we were unable to
attract sufficient participants, and four because they involved an experimental
intervention where participants were asked to bring their favorite book. These
four sessions were run specifically to study how decision weights and selectivity
would be affected by an intervention designed to shift subjects attention away
from superficial physical attributes. The inclusion of these four sessions does not
alter the results reported below; they are omitted so that the only experimental
difference across sessions is group size.
5. The number of males who attended the sessions was 202, as indicated by
Table I. However, two of these subjects did not indicate Yes or No in any of their
meetings, so their decisions cannot be included in the analyses. Nonetheless, they
were observed and rated by their partners, and we include these two males in our
analyses of female subjects choices.
6. A number of other responses, which we do not utilize in this paper, were
also elicited from the subjects. For the complete survey, please see http://
www2.gsb.columbia.edu/faculty/rfisman/Dating_Survey.pdf
678
NUMBER
OF
TABLE I
PARTICIPANTS IN EACH SPEED DATING SESSION
Round #
Women
Men
1
2
3
4
5
6
7
8
9
10
11
12
13
14
10
16
10
18
10
16
10
20
9
21
9
18
19
14
10
19
10
18
10
16
10
20
9
21
10
20
18
10
that by the end of the dating event they had rotated to all of the
tables, meeting all of the females.7 Each rotation consisted of four
minutes during which the participants engaged in conversation.
After the four minutes the Speed Dating hosts instructed the
participants to take one minute to fill out their scorecards for the
person with whom they were just speaking. In some events there
were slightly unequal numbers of males and females, so that
some subjects in these events experienced empty five-minute
intervals.
The morning after the Speed Dating event, participants were
sent an e-mail requesting that they complete the follow-up online
questionnaire. Ninety-one percent (51 percent female, 49 percent
male) of the Speed Dating participants completed this follow-up
questionnaire in order to obtain their matches. Upon receipt of
their follow-up questionnaire responses, participants were sent
an e-mail informing them of their match results.
Data DescriptionThe main variable of interest is the
Yes/No decision of subject i with respect to a partner j, which we
denote by Decisionij . Since our focus will be differential gender
7. This was the only asymmetry in the experimental treatment of men and
women. While we would have preferred to have men and women alternate in
rotating, we were advised against this by the owners of HurryDate. We believe
that this experimental asymmetry is unlikely to account for the observed gender
differences.
679
effects, we define an indicator variable Malei . In examining subjects decision weights, we use the ratings filled in after each
round. We limit ourselves to three of the six characteristics:
attractiveness, intelligence, and ambition. We omit the remaining
three characteristics primarily for brevity: our main interest is in
the gender differences in preferences, and we found very similar
attribute weights on the omitted characteristics. None of the
results are substantively affected by the inclusion of these additional covariates. Our notation for these ratings is given by Ratingijc , which is is rating, on a 10-point scale, of j on attribute c
{Attractiveness, Intelligence, Ambition}. Observations for which
at least one of these ratings is missing will necessarily be omitted
from the regressions. We also include results based on the average ratings of all other participants that rated j, which we denote
by Ratingijc , where the overbar represents that this is an average of ratings and the negative subscript indicates that i is
excluded from this average.
The pre-event survey additionally provides us with the information on partners undergraduate institution and the ZIP code
where the partner grew up. We match these variables to the log
of the median SAT score in the partners undergraduate institution, log(SAT) i , the log of the median income in 1990 in the ZIP
code where the partner grew up, log(Income)i , and the log population density in 1990 in that ZIP code, log(Density) i . 8
For subjects (as opposed to partners) attributes, we primarily rely on the self-ratings from the pre-event survey. These
self-ratings are denoted by Selfic , which is is self-rating on attribute c. As an alternative, we consider the consensus view of all
partners who rated i, which we denote by Othersic . Finally, we
use the pre-event survey to construct, for each subject-partner
pair, dummy variables for whether the pair has the same field of
study, SameFieldij , whether they are from the same part of the
world, SameRegionij , and whether they are of the same race,
SameRaceij .
For our individual-level regressions, we define the number of
acceptances that subject i gave as Yesesi . The variable NumberOfPartnersi denotes the number of meetings that i had during
the event. The fraction of partners that subject i accepted is
YesRatei Yesesi /NumberOfPartnersi . YesRatei will be our main
dependent variable for the selectivity analysis.
8. Our subjects had a median age of eleven in 1990.
680
A. Field of study
Business
Law
Service
Academic
Total
B. Race
White
Black
Hispanic
Asian
Total
C. Region of Origin
North America
Western Europe
Eastern Europe
Central Asia
Middle East
South Asia
East Asia
Latin America
Africa
Total
Percentage
Columbia graduate
population
Percentage
101
44
80
169
394
25.63%
11.17%
20.30%
42.89%
1925
1530
2161
4953
10569
18.21%
14.48%
20.45%
46.86%
228
22
31
67
348
65.52%
6.32%
8.91%
19.25%
3978
424
416
975
5793
68.67%
7.32%
7.18%
16.83%
287
32
7
6
6
10
29
14
1
392
73.21%
8.16%
1.79%
1.53%
1.53%
2.55%
7.40%
3.57%
0.26%
Statistics for the Columbia graduate student population reflect total (part-time and full-time) enrollment, and are taken from the Statistical Abstract of Columbia University 2004, available at http://www.columbia.edu/cu/opir/abstract/enrollment_fte_2004.html. No data are available on students countries of origin.
681
TABLE IIb
SUMMARY STATISTICS
Decision
Same Race
Same Field
Same Region
SAT
Income
Density
Mean
Std. dev.
Min
Max
Obs
0.43
0.49
0.35
0.55
1290.89
46056.30
13822.22
0.49
0.50
0.48
0.50
126.04
17661.54
26696.63
0.00
0.00
0.00
0.00
990
8607
5.89
1.00
1.00
1.00
1.00
1490
109031
122193.90
6276
4942
6102
6024
117
272
272
Decision is an indicator variable that takes on a value of one if a subject desired contact information for
a partner. SAT is the median SAT score in 2003 of the partners undergraduate institution. Income is the
median income of the partners ZIP code in 1990, measured in dollars, based on United States census data.
Density is the population density of the partners ZIP code in 1990, measured in people per square mile, based
on United States census data. Same Region is an indicator variable denoting that the subject and partner
were born in the same region of the world (North America, Western Europe, Eastern Europe, Central Asia,
Middle East, South Asia, East Asia, Latin America, or Africa). Same Field is an indicator variable denoting
that the subject and partner are in the same graduate school (aggregated to Business, Law, Service, and
Academic). Same Race is an indicator variable denoting that the subject and partner are of the same race. For
Same Race, Same Field, and Same Region, the level of observation is a subject-partner meeting. For SAT,
Income, and Density, the level of observation is the partner.
and professional students. Finally, the majority (nearly threequarters) of our subjects grew up in North America (i.e., the
United States and Canada).
Table IIb reports summary statistics on the subjects decision, the median SAT score in the partners undergraduate institution, and the median income and the population density in 1990
in the ZIP code where the partner grew up (for partners from the
United States). Of all meetings, 49 percent were between individuals of the same race, 35 percent between individuals in the
same field of study, and 55 percent between individuals from the
same region.
Our primary interest is understanding the desirability of
these attributes, which is the goal of the next section.
III. RESULTS
ON
ATTRIBUTE DEMAND
In this section we investigate the demand for partners attributes. Throughout the analysis we assume that the subjects
engage in straightforward behavior, i.e., that subjects are more
likely to accept partners whom they like better. In the theory
section, Proposition 1 lays out the conditions that are sufficient to
justify this assumption.
682
GENDER DIFFERENCES
(1)
Ambition
0.013**
(0.007)
Attractiveness 0.119***
(0.005)
Intelligence
0.045***
(0.007)
Ambition
*Male
Attractiveness
*Male
Intelligence
*Male
Subjects gender
Female
Rating measure
Observations
R2
TABLE III
SUBJECTIVE ATTRIBUTE WEIGHTS
IN
(2)
(3)
(4)
(5)
0.013*
(0.007)
0.140***
(0.005)
0.023***
(0.008)
0.013**
(0.007)
0.119***
(0.005)
0.045***
(0.007)
0.001
(0.009)
0.020***
(0.007)
0.022**
(0.011)
0.003
(0.021)
0.136***
(0.008)
0.044**
(0.019)
0.020
(0.020)
0.159***
(0.010)
0.005
(0.022)
Male
Both
Female
Male
OwnRatings
2655
0.52
2712
0.53
(6)
0.003
(0.021)
0.136***
(0.008)
0.044**
(0.019)
0.016
(0.029)
0.023*
(0.013)
0.039
(0.029)
Both
Consensus
5367
0.53
3128
0.38
3128
0.41
6256
0.40
Linear probability model; robust standard errors in parentheses, clustered by partner. The level of
observation is a subject-partner meeting. The dependent variable in all regressions is Decision, an indicator
variable that takes on a value of one if a subject desired contact information for a partner. In columns (1)(3)
the independent variables are the ratings that the subject assigned to the partner for Ambition, Attractiveness, and Intelligence. The ratings were on a 110 Likert Scale. In columns (4)(6) the independent variables
are the average ratings of a partner by all other subjects. Male is an indicator variable denoting whether the
subject is male. All regressions include subject fixed effects, and all observations are weighted by the inverse
of the number of observation per subject. * significant at 10 percent; ** significant at 5 percent; *** significant
at 1 percent.
Rating
c
ijc
ij ,
cC
683
tiveness than females do, while females put more weight on intelligence. This is consistent with the predictions of both the evolutionary and social structure theories of mate selection described
in the introduction.
The magnitudes of these differences are large. Each additional attractiveness point (on a 10-point scale) increases male
likelihood of saying Yes by 2.1 percentage points more than it
increases the female likelihood of saying Yes. This implies that
the effect of physical attractiveness is 18 percent higher for
males. The implied effect of intelligence on the probability of
Yes is 4.6 percentage points for women compared with 2.3
percentage points for men. We look at the statistical significance of these differences in column (3), where we pool all
subjects and include an interaction term RatingMale for each
attribute; for both attractiveness and intelligence, the interaction term is significant at the 5 percent level. We do not observe
any difference across genders in the importance of ambition.
When we repeat the same exercise using the average of all
subjects other than i, i.e., Ratingijc, as the measure of partner
attributes, we obtain qualitatively similar results (reported in
columns (4)(6) of Table III).10 Hence, the results are not
driven by idiosyncratic assessments of the attributes.
We next consider the influence of subjects own attributes
on the demand for particular partners. Specifically, we examine whether subjects are averse to choosing partners who are
superior to them on gender stereotypical attributes, as suggested by social structure theory [Eagly and Wood 1999].11 We
define an indicator variable, (Ratingijc Selfic), that takes on a
value of one if subject is rating of partner j on attribute c is
greater than the subjects own attribute rating. Our specification examines the impact of the difference in the attributes
(Ratingijc Selfic), allowing for the effect to change when the
difference is positive:
684
EFFECT
OF
TABLE IV
OWN ATTRIBUTES ON SUBJECTIVE ATTRIBUTE WEIGHTS
(1)
(2)
(3)
(4)
Ambition
0.009
0.031***
0.020**
0.030***
(0.008)
(0.008)
(0.010)
(0.009)
Ambition (Ambition Own
0.012
0.058*** 0.012
0.047***
Ambition)
(0.014)
(0.013)
(0.016)
(0.016)
Attractiveness
0.113***
0.134***
0.097***
0.136***
(0.006)
(0.007)
(0.008)
(0.009)
Attractiveness (Attractiveness 0.023
0.014
0.060***
0.006
Own Attractiveness)
(0.015)
(0.013)
(0.015)
(0.014)
Intelligence
0.049***
0.030***
0.041***
0.044***
(0.009)
(0.009)
(0.011)
(0.010)
Intelligence (Intelligence 0.007
0.043**
0.007
0.064***
Own Intelligence)
(0.019)
(0.018)
(0.018)
(0.020)
Subjects gender
Female
Male
Female
Self-rating
Observations
R2
2985
0.47
2978
0.50
Male
Partnerconsensus
3031
0.33
3016
0.50
Linear probability model; robust standard errors in parentheses, clustered by partner. The level of
observation is a subject-partner meeting. The dependent variable in all regressions is Decision, an indicator
variable that takes on a value of one if a subject desired contact information for a partner. Ambition,
Attractiveness, and Intelligence are the ratings that a subject assigned to a partner for each of the attributes.
(Ambition Own Ambition) is an indicator variable that takes on a value of one if the subjects rating of the
partners ambition is greater than the subjects own ambition. In columns (1) and (2) the measure of subject
ambition is a subjects pre-experiment self-rating; in columns (3) and (4) the measure of subject ambition is
the average rating by all partners that the subject met. Similar pairs of independent variables are defined for
Attractiveness and Intelligence. All regressions include subject fixed effects, and all observations are
weighted by the inverse of the number of observation per subject. * significant at 10 percent; ** significant
at 5 percent; *** significant at 1 percent.
Decision ij i
Rating
c0
ijc
Selfic
cC
Rating
c1
ijc
cC
The results are reported in Table IV, columns (1) and (2). For
attractiveness, the interaction term is insignificant for both men
and women. For ambition, however, the interaction term is insignificant for females but is significantly negative ( p 0.01) for
males. Furthermore, the effect of an increase in ambition above a
mans own level, given by the sum of the direct effect and the
interaction term, is negative. In other words, men strictly prefer
women with their own level of ambition to women more ambitious
685
686
may force women to directly confront the need to align choices with
a desired self-image.13
This concern about reason-based choice does not apply to our
next set of results,14 where we examine attribute valuations
based on objective characteristics that subjects were never asked
to rate. We first look at the impact of intelligence (measured by
the median SAT score in subjects college), economic background
(measured by the median income in 1990 in the ZIP code where
the subject grew up), and rusticity (measured by the population
density in 1990 in that ZIP code). Unfortunately, we obtained
subjects undergraduate institution only in the final six rounds,
and we possess income and population density information only
for subjects who grew up in the United States (and who responded to the survey question on ZIP code). As a result, including all variables simultaneously drastically reduces the sample
size. Hence, we run two separate regressions, one that relies on
college information and one that uses the ZIP codes:
Decision ij 0 logSATj ij ,
Decision ij 0 logIncomej 1 logDensityj ij .
We run these regressions separately for men and women and on
the full sample with a gender interaction term. Table V reports
the results. The first three columns support our previous finding
on the effect of partners intelligence: women put a greater emphasis on the SAT scores. This difference, as embodied in the
interaction term in column (3), is significant at the 10 percent
level. Further, in columns (4) through (6) we find that the coefficient on log(Income) is significant only for women. While the
difference between coefficients is large (0.08 versus 0.014), the
interaction log(Income)Male in column (6) is not significant ( pvalue 0.31). For both men and women, the coefficient on log(Density) is significantly negative, while the interaction with
Male is not significantly different from zero. Thus, we find tentative evidence that women prefer partners from more affluent
neighborhoods, and both men and women are more likely to accept
a partner who grew up in a less densely populated area.
687
log(SAT)
(1)
(2)
(3)
0.681**
(0.293)
0.101
(0.289)
0.681**
(0.288)
log(Income)
SUBJECTS DECISIONS
(4)
0.088*
(0.053)
0.020*
(0.011)
log(Density)
(5)
(6)
0.014
0.088*
(0.052)
(0.052)
0.022** 0.020*
(0.010)
(0.011)
0.782*
(0.409)
log(SAT)
*Male
log(Income)
*Male
log(Density)
*Male
Subjects gender
Observations
R2
AND
0.074
(0.074)
0.001
(0.015)
Female
794
0.32
Male
1120
0.27
Both
1914
0.29
Female
1915
0.28
Male
2410
0.30
Both
4325
0.30
Linear probability model; robust standard errors in parentheses, clustered by partner. The level of
observation is a subject-partner meeting. The dependent variable in all regressions is Decision, an indicator
variable that takes on a value of one if a subject desired contact information for a partner. Log(SAT) is the
logarithm of the median SAT score in 2003 of the partners undergraduate institution. Log(Income) is the
logarithm of median income of the partners ZIP code in 1990, measured in dollars, based on United States
census data. Log(Density) is the logarithm of the population density of the partners ZIP code in 1990,
measured in people per square mile, based on United States census data. Male is an indicator variable
denoting whether a subject is male. All regressions include subject fixed effects, and all observations are
weighted by the inverse of the number of observation per subject. * significant at 10 percent; ** significant
at 5 percent; *** significant at 1 percent.
688
Same Race
Same Field
Same Region
(1)
(2)
(3)
0.143***
(0.024)
0.002
(0.028)
0.075**
(0.033)
0.053
(0.032)
0.035
(0.026)
0.096**
(0.043)
0.143***
(0.024)
0.002
(0.028)
0.075**
(0.032)
0.090**
(0.040)
0.033
(0.038)
0.021
(0.054)
Female
2417
0.26
Male
2417
0.27
Both
4834
0.28
Same Race
*Male
Same Field
*Male
Same Region
*Male
Subjects gender
Observations
R2
Linear probability model; robust standard errors in parentheses, clustered by partner. The level of
observation is a subject-partner meeting. The dependent variable in all regressions is Decision, an indicator
variable that takes on a value of one if a subject desired contact information for a partner. Same Region is an
indicator variable denoting that the subject and partner were born in the same region of the world (North
America, Western Europe, Eastern Europe, Central Asia, Middle East, South Asia, East Asia, Latin America,
or Africa). Same Field is an indicator variable denoting that the subject and partner are in the same area of
study (aggregated to Business, Law, Service, and Academic). Same Race is an indicator variable denoting that
the subject and partner are of the same race. Male is an indicator variable denoting whether the subject is
male. All regressions include subject fixed effects, and all observations are weighted by the inverse of the
number of observation per subject. * significant at 10 percent; ** significant at 5 percent; *** significant at
1 percent.
689
ON
SELECTIVITY
690
EFFECT
Group size
OF
TABLE VII
GROUP SIZE ON SELECTIVITY
(1)
(2)
(3)
0.013**
(0.005)
0.003
(0.005)
0.036
(0.036)
0.166*
(0.092)
0.018**
(0.007)
Female
No
Male
No
Both
Yes
200
0.05
200
0.00
400
0.11
Male
Group size*
Male
Subjects gender
Round FE
Observations
R2
Robust standard errors are in parentheses. Regressions are at the subject level. The dependent variable
in all regressions is the fraction of partners for whom the subject desired contact information. Group size is
the number of meetings experienced by a subject. Male is an indicator variable denoting whether the subject
is male. * significant at 10 percent; ** significant at 5 percent; *** significant at 1 percent.
691
group size indicates that they may have linear benefit and cost
functions.19 Possible reasons why womens costs of going on dates
may be convex are many. Great attention paid to each date,
coupled with a finite time budget, is a plausible source of convexity. Social stigma may be another. Explanations for why women
could have more concave benefit functions may stem from different motivations in the dating search process for the two genders.
Alternatively, the gender difference in the relationship between selectivity and group size may stem from asymmetric costs
of reneging on a match, rather than asymmetric utility functions
over dates. Specifically, the observed difference may arise from
the social roles of men and women in setting up dates. That is, a
man reneges on a match by not contacting a woman (omission),
while a woman reneges by actively saying no to a man (commission).20 This may credibly create an asymmetric cost of reneging
on a match. However, such cost asymmetry alone merely predicts
that the level of selectivity will be higher for women than for men:
if the cost of declining dates is linear, the elasticity of selectivity
with respect to group size would be the same for men and women
even with gender-specific reneging costs. There is no clear intuition that suggests that this cost should be convex, and it could
very easily be the opposite (one grows inured to the psychological
trauma of rejecting others). That said, with the data we have, it
is not possible to completely rule out this possibility.
A second alternative is the potential gender difference in fear of
rejection discussed in Section III. As we note there, one benefit of
using the Speed Dating format is that there is no direct personal
feedback, which likely mitigates any psychic cost of rejection. Additionally, we may control to some degree for fear of rejection by
including the average rejection rate for each subject as a control
variable in our regressions on group size. We find that its inclusion
has little effect on the coefficient on NumberOfPartners.
Finally, since sessions with more partners are necessarily
longer, we investigated whether the probability that a subject
19. One may think that the Speed Dating sessions have a minimal impact on
each subjects overall dating opportunities, so that the number of partners ought
to be irrelevant for selectivity regardless of the second derivative properties of the
utility function. Our surveys reveal, however, that the matches formed at the
experimental sessions constitute a major component of the subjects dating experiences. More importantly, as the results make clear, group size does in fact
significantly influence selectivity of our female subjects.
20. Our follow-up survey indicates that roughly half of the matches resulted
in actual dates. Further, 64 percent of the men initiated contact with at least one
match compared with only 17 percent of the women.
692
says Yes changes over the course of the Speed Dating session. The
only time trend in the data is that women are significantly more
likely to say Yes in their first meeting, both in the small and in
the large groups. We have no explanation for this, and believe it
may be a statistical artifact.21
While we acknowledge the range of possible explanations for
the robust finding on group size, we favor the explanation based
on gender-specific dating utilities for the reasons given above. We
now provide a more structured theoretical discussion of the Speed
Dating game.
V. THEORETICAL ANALYSIS
Our theoretical framework is designed to identify the conditions that are sufficient to justify our interpretations of the results in the previous two sections. Many issues that are undeniably important in dating markets we ignore in order to preserve
tractability. Most importantly, we do not explicitly model matching with learning. As Das and Kamenica [2005] observe, matching with learning is a two-sided n-armed bandit problem so solving for equilibria can be prohibitively difficult. Therefore, we
present a simplified model more directly tailored to our empirical
analysis.
Let {1, 2, . . . , N } be the set of men and {N
1, . . . , 2N } be the set of women present on a particular evening.
v ij is the value of individual j to individual i. For each person i,
the v ij s are drawn from some distribution F i . An implicit
assumption in many models of marriage, from Becker [1973] to
Smith [2002], is sexwide homogeneity of preferences: v ij v j
@i @j and v ij v j @i @j . 22 We refrain from
this assumption since we are interested in how a persons taste
for a particular characteristic varies with that persons own characteristics (as in Tables IV and V). In particular, we take these
values to be functions of certain attributes. Let each individual
have K attributes, and let a ijk denote individual is appreciation
21. Ordering also does not affect our attribute weight results. All of our
reported results are robust to the inclusion of fixed effects for each meeting
number. Similarly, we find no significant interactions of attributes with time in
predicting decisions.
22. We borrow this terminology, as well as our basic notation, from Das and
Kamenica [2005].
693
u V
X
U iv ij,v ji jX
CX,
ijk
k1
694
utility is the sum of the values of his or her dates, with decreasing
value of lower-ranked partners, minus a cost that is increasing in
the total number of dates. The key assumptions underlying this
formulation are (i) the cost of saying Yes derives purely from the
cardinality of the number of dates; (ii) the benefit function is
separable across individuals; and (iii) utility is derived from dating alone. Utility functions of this form trivially lead to straightforward behavior.
X
uk(vijk) C(X),
PROPOSITION 1. Suppose that Ui({vij, vji}jX) k1
k
k1
where { j1, . . . , jX} X, vijk vijk1, u u ,C 0 and (u k )
0 @k. Then, there is an equilibrium where i engages
in straightforward behavior.
695
VI. CONCLUSION
Our paper serves as an important starting point in understanding the preferences underlying the search for a mate. Prior
work in economics has emphasized final matches, but the theory
and empirics have not been well-suited to the study of how these
matches are actually formed. In this paper we utilize an experimental design that allows us to directly observe individual decisions and develop a formal model that establishes the conditions
under which a regression framework is appropriate for estimating preferences over partner attributes.
There are a number of ways that our work may be improved
upon to generate more refined measures of preferences for romantic partners. Most notably, a similar methodology could be employed on a broader set of subject populations to examine the
extent to which our gender asymmetry results generalize. Such
experiments could also be used to better understand the differences across communities in dating preferences. A second important extension will be to develop data sets that similarly identify
individual preferences, but focus on longer run outcomes and
relationship formation.
Extensions to our theoretical analysis would also be useful
for a better understanding of dating markets. In particular, incorporating uncertainty and learning, which are especially relevant in the longer run, is an important next step for theory. More
ambitiously, we hope to develop models that incorporate strategic
behavior into dating decisions.
A deeper understanding of dating preferences is an essential
input for examining a number of social issues. Topics ranging
from race relations to intergenerational mobility strongly depend
on outcomes in the marriage market. The equilibrium household
formation in turn derives from the underlying preferences for
romantic partners. Prescriptively, a more complete and better
calibrated theory of dating preferences, built on results such as
those we report above, may eventually allow us to evaluate different institutional arrangements for generating matches.
APPENDIX
Proof of Proposition 1. Let Y be the set of individuals that say
Yes to i. Then, regardless of the behavior of other individuals of is
gender, a best response for i is to say Yes to { j1 . . . , jm} Y, subject
696
REFERENCES
Becker, Gary S., A Theory of Marriage: Part I, Journal of Political Economy,
LXXXI (1973), 813 846.
Bisin, Alberto, Giorgio Topa, and Thierry Verdier, Religious Intermarriage and
Socialization in the United States, Journal of Political Economy, CXII
(2004), 615 664.
Burdett, Kenneth, and Melvyn Coles, Marriage and Class, Quarterly Journal of
Economics, CXII (1997), 141168.
Buss, David M., Toward an Evolutionary Psychology of Human Mating, Behavioral and Brain Sciences, XII (1989), 39 49.
, The Strategies of Human Mating, American Scientist, LXXXII (1994),
238 249.
Buss, David M., and David P. Schmitt, Sexual Strategies Theory: An Evolutionary Perspective on Human Mating, Psychological Review, C (1993), 204 232.
Buss, David M., and Douglas T. Kenrick, Evolutionary Social Psychology, in
Daniel T. Gilbert, ed., The Handbook of Social Psychology, 4th ed., Vol. 2
(Boston, MA: McGraw-Hill, 1998), pp. 9821026.
Clark, Russell D., and Elaine Hatfield, Gender Differences in Receptivity to
Sexual Offers, Journal of Psychology and Human Sexuality, II (1989), 39 55.
Das, Sanmay, and Emir Kamenica, Two-Sided Bandits and the Dating Market,
Proceedings of the Nineteenth International Joint Conferences on Artificial
Intelligence (2005), 947952.
Dowd, Maureen, Men Just Want Mommy, The New York Times, January 13,
2005.
Eagly, Alice H., Sex Differences in Social Behavior: A Social-Role Interpretation
(Hillsdale, NJ: Erlbaum, 1987).
Eagly, Alice H., and Wendy Wood, The Origins of Sex Differences in Human
Behavior: Evolved Dispositions versus Social Roles, American Psychologist,
LIV (1999), 408 423.
Fisman, Raymond, Sheena Iyengar, Emir Kamenica, and Itamar Simonson, Racial Preferences in Dating: Evidence from a Speed Dating Experiment,
Working Paper, 2005.
Hitsch, Gunter, Ali Hortacsu, and Dan Ariely, What Makes You Click: An
Empirical Analysis of Online Dating, Working Paper, 2004.
Kenrick, Douglas T., and Richard C. Keefe, Age Preferences in Mates Reflect Sex
Differences in Human Reproductive Strategies, Behavioral and Brain Sciences, XV (1992), 7591.
Kerckhoff, Alan C., and Keith E. Davis, Value Consensus and Need Complementarity in Mate Selection, American Sociological Review, XXVII (1962),
295303.
Mortensen, Dale T., and Christopher A. Pissarides, New Developments in Models
of Search in the Labor Market, in O. Ashenfelter and D. Card, eds., Hand-
697