ArticlePDF Available

Understanding Infidelity: Correlates in a National Random Sample

Authors:

Abstract

Infidelity is a common phenomenon in marriages but is poorly understood. The current study examined variables related to extramarital sex using data from the 1991-1996 General Social Surveys. Predictor variables were entered into a logistic regression with presence of extramarital sex as the dependent variable. Results demonstrated that divorce, education, age when first married, and 2 "opportunity" variables--respondent's income and work status--significantly affected the likelihood of having engaged in infidelity. Also, there were 3 significant interactions related to infidelity: (a) between age and gender, (b) between marital satisfaction and religious behavior, and (c) between past divorce and educational level. Implications of these findings and directions for future research are discussed.
Journal
of
Family Psychology
2001,
Vol. 15, No.
4,
735-749
Copyright 2001
by the
American Psychological Association,
Inc.
O893-32OO/O1/S5.OO
DOI:
1O.1O37//O893-32OO.
15.4.735
Understanding Infidelity: Correlates
in a
National
Random Sample
David
C.
Atkins
University of Washington
Donald
H.
Baucom
University of North Carolina at Chapel Hill
Neil
S.
Jacobson
University of Washington
Infidelity
is a
common phenomenon
in
marriages
but is
poorly understood.
The
current study examined variables related
to
extramarital
sex
using data from
the
1991-1996 General Social Surveys. Predictor variables were entered into
a
logistic
regression with presence
of
extramarital
sex as the
dependent variable. Results
demonstrated that divorce, education,
age
when first married,
and 2
"opportunity"
variables—respondent's income
and
work status—significantly affected
the
likeli-
hood
of
having engaged
in
infidelity. Also, there were
3
significant interactions
related
to
infidelity:
(a)
between
age and
gender,
(b)
between marital satisfaction
and religious behavior,
and
(c) between past divorce
and
educational level. Impli-
cations
of
these findings
and
directions
for
future research
are
discussed.
Infidelity
1
occurs in a reliable minority of
American marriages. Conservative estimates
from recent national surveys indicate that be-
tween 20% and 25% of all Americans will have
sex with someone other than their spouse while
they are married (Greeley, 1994; Laumann,
Gagnon, Michael, & Michaels, 1994; Wieder-
man, 1997). Moreover, there is some evidence
that infidelity is also detrimental to many rela-
tionships. A survey of couple therapists demon-
David
C.
Atkins
and
Neil
S.
Jacobson, Center
for
Clinical Research, Department
of
Psychology,
Uni-
versity
of
Washington; Donald
H.
Baucom, Depart-
ment
of
Psychology, University
of
North Carolina
at
Chapel Hill.
While
the
manuscript
was in
preparation, Neil
S.
Jacobson passed away
on
June
2, 1999. He was a
great advisor, mentor,
and
friend
and is
deeply
missed.
Some
of the
analyses from this article were
pre-
sented
at the
32nd Annual Convention
of the
Asso-
ciation
for the
Advancement
of
Behavior Therapy,
Washington, D.C., November 1998.
Correspondence concerning this article should
be
addressed
to
David
C.
Atkins, Center
for
Clinical
Research,
1107
N.E. 45th Street, Suite 310, Seattle,
Washington 98105-4631. Electronic mail
may be
sent
to
datkins@u.washington.edu.
strated that clinicians view infidelity as one of
the most difficult problems to treat in therapy
and one of the most damaging issues for a
relationship (Whisman, Dixon, & Johnson,
1997).
In addition, some couple therapists have
estimated that
50%-65%
of couples in their
clinical practices are in therapy as the result of
infidelity (Glass & Wright, 1988; Humphrey,
1983,
as cited in Glass & Wright, 1988).
Yet, research on infidelity has not been com-
mensurate with its prevalence and impact on
American relationships. In fact, before the
1990s,
there were no reliable estimates of the
frequency of infidelity in American marriages.
Most of the research that has been conducted
has explored variables that might be related to
and influence infidelity. However, much of this
1
A
wide variety
of
terms have been used
in the
research literature to refer to infidelity. Depending on
the specific features
of
their study, researchers have
used such varying terms
as
nonmonogamy, extrady-
adic involvement, extramarital coitus,
and
poly-
amory.
For the
present article,
we
used
the
terms
infidelity
and
extramarital
sex to
refer
to sex
with
someone other than one's spouse while
one is mar-
ried. When
it
seemed appropriate
in
describing other
research, we used the terms that the study authors had
used.
735
736
ATKINS, BAUCOM, AND JACOBSON
literature has suffered from serious method-
ological problems. Many researchers have
failed to include samples in which infidelity has
actually taken place, either using a vignette pre-
sented to college students (e.g., Kitzinger &
Powell, 1995rMongeau & Schulz, 1997;
Parker, 1997; Sprecher, Regan, & McKinney,
1998) or having participants rate their future
likelihood of engaging in infidelity (e.g., Buss
& Shackelford, 1997; Shackelford & Buss,
1997).
Of the researchers who have included
participants who have engaged in infidelity,
several have not used random samples, calling
into question the generalizability of their find-
ings (e.g., Blumstein & Schwartz, 1983; Glass
& Wright, 1985; Spanier & Margolis, 1983).
Finally, researchers who have used random
samples with cases of infidelity have not used a
multivariable modeling strategy in predicting
infidelity (e.g., Greeley, 1994; Laumann et al.,
1994;
Wiederman, 1997; for two recent excep-
tions,
see Trseeh & Stigum, 1998, and Treas &
Giesen, 2000). When the relationship between
potential predictors and infidelity is not as-
sessed in a single model, important issues such
as interactions among predictors, multicol-
linearity, and the combined influence of multi-
ple predictors cannot be assessed.
The present study addressed each of these
methodological concerns. The data used in our
analyses were derived from a nationally repre-
sentative sample involving a percentage of
individuals who engaged in infidelity. Further-
more, the data were analyzed in a single regres-
sion model, allowing interactions among pre-
dictors and controlling for the joint influence of
multiple predictors. Thus, the present study
attempted to replicate some of the findings
from the previous infidelity literature using a
more rigorous method. The variables consid-
ered did not represent an exhaustive list of
variables from the previous literature, but they
did represent some of the significant domains
demographic, relationship, and environmental
that previous research has explored.
Keeping in mind the limitations of previous
research just noted, a number of factors have
demonstrated a relationship to infidelity. These
include age (Greeley, 1994; Traeen & Stigum,
1998),
education (Amato & Rogers, 1997;
Traeen & Stigum, 1998), history of divorce
(Wiederman, 1997), religiosity (Amato & Rog-
ers,
1997; Bell, Turner, & Rosen, 1975; Ed-
wards & Booth, 1976; Hurlburt, 1992), and
length of relationship (Bell et al., 1975; Glass &
Wright, 1977; Spanier & Margolis, 1983). In
addition, gender, relationship satisfaction, and
opportunity are particularly notable for the
attention—and, to some extent, the empirical
support—they have received in the literature
(Blumstein & Schwartz, 1983; Greeley, 1994;
Laumann et al., 1994; Traeen & Stigum, 1998;
Treas & Giesen, 2000).
Gender has been the most commonly studied
variable in extramarital behavior. The typical
finding has been that more men than women
have engaged in infidelity (Greeley,
1994:
Lau-
mann et al., 1994; Wiederman, 1997). Further-
more, men report a greater number of liaisons
(Lawson & Samson, 1988; Spanier & Margolis,
1983) and express greater interest in infidelity
(Buunk & Bakker, 1995; Seal, Agostinelli, &
Hannett, 1994; Thompson, 1984).
Some researchers have found evidence that
men and women engage in different types of
infidelity. For example, Spanier and Margolis
(1983) investigated the experiences of recently
separated and divorced individuals in which
either the research participant or the ex-spouse
had had an affair. Women who had been in-
volved in an affair tended to be more emotion-
ally involved with their affair partners than the
men who had been involved in an affair
(40.5%
of women vs. 11.5% of men reported that their
most recent extramarital relationship was a
"more long-term love relationship"), and the
involved women also reported greater feelings
of guilt. Similarly, Glass and Wright (1985)
showed that men are more likely than women to
have "sexual-only" affairs, and women are
more likely than men to have "emotional-only"
affairs.
However, recent research suggests that the
differences between the sexes in rates of infi-
delity may be decreasing (Greeley, 1994; Lau-
mann et al., 1994; Thompson, 1983; Wieder-
man, 1997; see Parker, 1997, for a recent study
that failed to reveal gender differences). Oliver
and Hyde (1993) conducted a meta-analysis in
which they examined gender differences in sex-
uality. In summarizing research exploring gen-
der differences in infidelity, they reported a
significant trend involving year of data collec-
tion, suggesting that the differing rates of infi-
delity for men and women are becoming more
similar in younger cohorts. In separate research
UNDERSTANDING INFIDELITY
737
involving survey data, Wiederman (1997) re-
ported that men and women less than 40 years
of age showed no differences in their reported
infidelity. Taking these findings together, it is
clear that gender plays a central role in infidelity
but that the conjoint impact of age and gender
on infidelity should be examined.
Another possible correlate of infidelity that
researchers have explored is the quality of the
primary relationship. In reviewing 10 studies of
infidelity, Thompson (1983) proposed a "deficit
model" to explain infidelity in which deficien-
cies in the primary relationship play a central
role in precipitating and sustaining infidelity.
Only
1
study of
the
10 that he reviewed failed to
reveal a significant relationship between marital
satisfaction and infidelity. Thompson estimated
that characteristics of the marriage (e.g., low
satisfaction and low sexual frequency) reliably
account for 25% of the variance in infidelity.
Analogue studies with young, married couples
have shown that marital conflict may make a
couple more susceptible to an extramarital affair
(Buss & Shackelford, 1997); relationship dissat-
isfaction may increase the desire to become
involved in infidelity (Prins, Buunk, & Van-
Yperen, 1993); and partners believe that low
marital satisfaction will lead to an affair (Wie-
derman & Allgeier, 1996).
Nonetheless, not all studies have shown a
relationship between infidelity and relationship
dissatisfaction. In their large-sample survey of
American couples, Blumstein and Schwartz
(1983) failed to find a relationship between in-
fidelity and marital satisfaction, sexual satisfac-
tion, or sexual frequency. In addition, Spanier
and Margolis (1983) found that quality of mar-
ital sex was unrelated to occurrence of extra-
marital sex (EMS) in their sample of recently
separated and divorced respondents. Moreover,
some couple therapists also support the idea that
infidelity does not automatically imply a. poor
primary relationship (Elbaum, 1981; Finzi,
1989).
There is some evidence that other variables
may moderate the association between relation-
ship dissatisfaction and infidelity. Glass and
Wright (1985) showed that "sexual-only" af-
fairs are less likely to be related to marital
dissatisfaction than "combined-type" affairs
that include both sexual and emotional compo-
nents.
These researchers found that women
were more likely to have emotional-only or
combined-type affairs. Consequently, there was
a stronger association between relationship sat-
isfaction and infidelity for women than for men.
Glass and Wright (1985) suggested that there
are different reasons for affairs, and thus rela-
tionship distress may be important in some
cases but not others.
Whereas gender and marital satisfaction have
been the most commonly studied factors in re-
search on infidelity, another variable that has
been shown to be related to infidelity in more
than one study is opportunity (Blumstein &
Schwartz, 1983; Greeley, 1994; Maykovich,
1976;
Traeen & Stigum, 1998). Opportunity is a
construct reflecting individuals' variability in
access and desirability to other people. Per-
ceived opportunity has been used to explain the
gender difference in rates of infidelity; men
have historically been in the workforce in
greater numbers than women, leading to more
contact with other people and an increased po-
tential for infidelity (Greeley, 1994; Schwartz &
Rutter, 1998). This view of opportunity may
explain the closing gender gap in rates of infi-
delity as greater numbers of women enter the
workplace.
However, Thompson (1983) noted that there
are some methodological problems in studies
measuring opportunity. Some studies have sim-
ply inquired whether or not the individual has
had the opportunity to engage in infidelity
(Johnson, 1970). This "either-or" question is
biased in that individuals who have engaged in
infidelity will necessarily answer
yes.
This strat-
egy does not explore degree of opportunity, and
it is difficult to see opportunity for infidelity as
being either present or absent. It is also not clear
that separate studies are referring to the same
construct when they discuss opportunity. Inves-
tigators have operationalized opportunity as
number of days traveling in a year (Traeen &
Stigum, 1998); liberalism, religiosity, and em-
ployment (Maykovich, 1976); occupational sta-
tus and use of prostitutes (Greeley, 1994);
strong views on personal independence (Blum-
stein & Schwartz, 1983); and early sexual ex-
perience, religious attendance, and work re-
quirements (Treas & Giesen, 2000). It is not
surprising that "opportunity," so variably de-
fined, has not consistently received empirical
support.
As noted earlier, the present study had two
primary goals. First, we sought to improve on
738
ATKINS, BAUCOM,
AND
JACOBSON
the methodological shortcomings of past studies
by using a national, random sample and a mul-
tivariable modeling strategy. Second, we at-
tempted to replicate findings from the previous
literature. In particular, we explored the associ-
ation of genderTmarital satisfaction, and oppor-
tunity with infidelity in a model containing
other demographic variables (i.e., age, length of
marriage, previous divorce, religious behavior,
and education). Related to our primary variables
of interest, we expected that men would report
more EMS than women but that this association
would be weaker among younger cohorts. Also,
the likelihood of EMS was expected to increase
with greater opportunity and lower marital sat-
isfaction. On the basis of previous research on
infidelity, we also had several secondary hy-
potheses. Specifically, we expected to find that
infidelity would be more likely among (a) those
in longer marriages, (b) less religious couples,
(c) more highly educated individuals, and (d)
individuals previously divorced.
Method
Participants
The data
for
this study were drawn from
the
Gen-
eral Social Surveys (GSSs) conducted
by the Na-
tional Opinion Research Center
at the
University
of
Chicago.
The
surveys were begun
in
1972 and, since
1994,
have been conducted every other year with
approximately 3,000 participants. Surveys
are
based
on structured interviews that last approximately
90
min. Each survey involves
a
national, cross-sectional
sample
of
noninstitutionalized, English-speaking
in-
dividuals
18
years
of
age
or
older
in the
continental
United States.
The
four surveys conducted from
1991-1996 included
a
question regarding EMS.
Spe-
cifically, respondents were asked whether they
had
"ever had
sex
with someone other than your husband
or wife while you were married?" We included in our
analyses all participants who were married at the time
of the interview
and
answered this question, yielding
a sample size
of
4,118.
Of
these participants, 544,
or
13.3%,
2
reported having
had EMS, and
3,574,
or
86.7%,
reported never having
had sex
with another
person while they were married. Survey response
rates
for
1991,1993,1994,
and
1996 were
78%,
82%,
78%,
and 76%, respectively (for more information
on
the sampling characteristics
of
the GSS, see Davis
&
Smith, 1996).
Measures
All
of
the variables used
in
the analyses were taken
from questions included
in the 1991 to 1996
GSSs
(Davis
&
Smith, 1996).
Marital satisfaction. Marital satisfaction
was
measured through
a
single item: "Taking
all
things
together,
how
would
you
describe your marriage?
Would
you say
that your marriage
is
very happy,
pretty happy,
or not too
happy?" Whereas
it
would
have been preferable
to
have
a
multi-item, well-
standardized measure
of
marital adjustment, such
measures
are not
feasible
in
large, multipurpose
sur-
veys.
However, there
is
reason
to
believe that even
a
single-item assessment
of
marital happiness
can
pro-
vide meaningful information about
the
respondent's
overall feelings about
the
relationship.
For
instance,
the frequently used Dyadic Adjustment Scale (DAS;
Spanier,
1976)
includes
a
single item rating happi-
ness with
the
relationship. Previous investigations
have demonstrated that this
one
item
is
highly
cor-
related with overall DAS scores (Goodwin, 1992).
Opportunity.
For the
purpose
of
this study,
we
operationalized opportunity
as a set of
variables that
might increase contact with other people
or
material
means that might facilitate infidelity.
We
attempted
to address opportunity through
two
variables
in-
cluded
in the
GSS. The first opportunity variable
we
used
was the
respondent's income. There were
two
reasons
for
thinking that
an
individual's income may
reflect
his or her
opportunity
for
infidelity.
In
some
instances, carrying
on a
clandestine relationship
re-
quires some financial means, and thus greater income
may facilitate EMS. Second, money
is
often equated
with power,
and
wealthy individuals
may be
more
appealing to potential extramarital partners. The GSS
measured respondents' income using
21
income
brackets ranging from $0-$ 1,000
to
more than
$75,000.
The
original scale included many more
brackets
for
lower incomes.
We
collapsed
the
origi-
nal scale into
12
income brackets
of
approximately
$5,000 increments.
The second opportunity variable
we
used
in our
analysis was
a
composite measure
of
the work status
of
the
survey respondent
and his or her
spouse.
In
general, individuals
in the
workforce have greater
contact with other people,
and
several researchers
have viewed people's work status
as a
measure
of
their opportunity
for
infidelity (Greeley, 1994;
May-
kovich, 1976). Furthermore, examining the combina-
tion
of
individuals'
and
their spouse's work status
is
one manner
of
assessing power
in the
relationship.
Situations
in
which
one
partner
is
working
and the
other partner
is not
may reflect unequal power
in the
relationship
and
greater opportunity
for
infidelity
for
the working partner.
In the
original survey, partici-
2
The reason
for the
somewhat lower rate
of
infi-
delity
in the
present sample
as
compared with other
national surveys (Greeley,
1994;
Laumann
et al.,
1994;
Wiederman,
1997) is
that
our
analyses were
restricted to individuals who were married
at
the time
of the survey.
UNDERSTANDING INFIDELITY 739
pants'
and
spouses' work status
was
coded into
one
of eight categories: working full time, working part
time,
temporarily unemployed, unemployed, retired,
in school, homemaker,
or
other.
We
collapsed these
eight categories into employed (full or part time)
3
and
not employed
(all
other categories)
and
then
com-
bined respondent
and
spouse scores. This yielded
a
categorical work status variable with four
levels:
both
respondent
and
spouse employed, respondent
em-
ployed
but
spouse
not
employed, spouse employed
but respondent not employed, and neither respondent
nor spouse employed.
Other variables.
The GSS has
used more than
100 separate questions through
the
years
to try to
understand Americans' religious beliefs
and
behav-
iors.
We
were interested
in a
measure
of
religious
behavior that might serve
as a
general marker
for the
importance
of
religion
to the
individual.
In
accor-
dance with this purpose, we used the following ques-
tion from
the
GSS:
"How
often
do you
attend reli-
gious services?" There were nine possible answers
ranging from "Never"
to
"More than once
a
week."
Several demographic variables were entered
in the
analysis, including gender, age, whether
the
respon-
dent
had
ever been divorced,
and the
respondent's
educational level (e.g., less than high school, high
school,
junior college, college, graduate). To evaluate
whether the length of the relationship might affect the
probability
of
infidelity,
we
also included
in our
analysis participants'
age
when they were first
married.
Table
1
Percentages of Missing Data by Predictor
Variable
Results
Missing Data
The GSS involves a core set of questions that
are included in every survey and asked of every
respondent. Other "rotating" questions are
asked of two thirds of respondents in a given
year, and questions change over the years as
modules of rotating questions are included and
dropped. Participants' age when they were first
married was a rotating question, thus creating a
sizable portion of missing data for that variable.
Moreover, some respondents failed to answer
all of the questions. Percentages of missing data
for each variable and for the total analysis are
displayed in Table 1.
We addressed the potential impact of missing
data through several means. First, following the
recommendation of Cohen and Cohen (1983),
we created a dummy variable indicating, for
each participant, whether she or he had any
missing data. This variable can be used to test
whether individuals with missing data are more
likely to report infidelity. If there is an associ-
Variable
Sex
Work status
Age
Educational degree
Previous divorce
Marital satisfaction
Religious behavior
Income
Age when first married
Total
Missing data
(%)
0.0
0.0
o!i
0.4
0.5
1.7
31.6
47.2
9.1
ation between missing data and the dependent
variable, this is a particularly difficult problem,
and analyses of the data are almost certainly
biased (Greenland & Finkle, 1995). In the
present study, there was not an association be-
tween missing data and the dependent variable,
^(1) = 1.51, p = .22, indicating that a data
imputation procedure might be appropriate. In
addition to the multiple imputation procedure
described subsequently, we analyzed our final
model using only cases without any missing
data (N =
1,413).
The results were very similar
to those that we found using multiple imputa-
tion, with somewhat reduced power because of
the smaller sample size.
Although there is no universally accepted
method for handling missing data, multiple im-
putation is a procedure that has been demon-
3
Some readers
may
wonder
why
full-time
em-
ployment and part-time employment were combined,
in that these
two
categories could
be
viewed
as dif-
ferent levels
of
opportunity. We chose to combine the
categories
for
two reasons. One issue
is
parsimony.
If
each spouse
is
classified
in
one
of
three ways (work-
ing full time, working part time,
or
not working), this
would yield nine combinations;
the
findings from
such
a
variable would
be
confusing
to
interpret.
An-
other issue
is the
wording
of the
question. Survey
respondents were asked
to
identify themselves
as
working either full time
or
part time. There
is
prob-
ably
a
great deal
of
heterogeneity within each cate-
gory
in
the number
of
hours
of
work and contact with
other people. Because
the GSS
does
not
contain
a
direct measure
of
contact with others resulting from
work
(or
otherwise),
we
believed that
it was
more
conservative
to
combine full-time
and
part-time
workers
to
indicate individuals who were working
at
all
as
compared with those
who
were
not.
740
ATKINS, BAUCOM, AND JACOBSON
strated to be superior to many alternatives (for
reviews, see Greenland & Finkle, 1995; Lit-
tle,
1992; Schafer & Olsen, 1998). Using only
cases with complete data (the default in many
statistical packages) can be very inefficient
because participants missing a single piece of
information are excluded from the entire anal-
ysis.
However, replacing missing values with
means or regression estimates underestimates
the true variability in the population, leading
to an inflated Type I error rate (see Greenland
& Finkle, 1995; Little, 1992). Multiple impu-
tation creates n ^ 2 complete data sets, using
the existing data and replacing missing values
with estimates sampled from distributions
based on the variables with missing data. Sta-
tistical analyses are then conducted on the n
data sets and averaged for a final estimate (for
information on how analyses are combined
across the n data sets, see Schafer & Olsen,
1998).
In simulation studies and practical ap-
plications, multiple imputation has been
shown to produce unbiased parameter esti-
mates from the missing data (Greenland &
Finkle, 1995; Schafer & Olsen, 1998).
In applying multiple imputation, a statisti-
cal assumption must be made about why the
data are missing. There are three potential
assumptions that can be made about the miss-
ing data. Data can be considered (a) missing
completely at random, in which the missing
data can be considered a random subset of the
sample; (b) missing at random or ignorable, in
which missing values may depend on other
variables in the analysis; or (c) nonignorable,
in which the missing data depend on the de-
pendent variable or do not depend on any
variables in the analysis. For practical pur-
poses,
the missing at random assumption is
usually reasonable; essentially, the assump-
tion is that missing values have some rela-
tionship with other variables in the analysis
(for a review of multiple imputation assump-
tions,
see Greenland & Finkle, 1995, or Scha-
fer & Olsen, 1998). We assumed that our
missing data were missing at random and
created 10 data sets on which to run our
analyses. For all of our statistical analyses,
including multiple imputation, we used the
Hmisc and Design libraries of functions in
S-PLUS Version 4.5, Release 2 for Windows
(Harrell, 1999a, 1999b; MathSoft, 1997).
Previous Divorce and Timing of the
Infidelity
One of the difficulties in using the GSS to
explore correlates of infidelity is the phrasing of
the question about EMS. Because the question
asks whether participants have "ever" had sex
with someone other than their spouse, it is un-
clear when the EMS occurred; it may be ongo-
ing, or it may have occurred many years ago.
Particularly worrisome for the present study is
the possibility that the EMS happened during a
previous marriage, which would make interpre-
tation of the marital satisfaction variable con-
fusing. Thus, it is possible that an unknown
percentage of divorced participants engaged in
infidelity during a previous marriage.
Ideally, we would like to know the exact
percentage; however, the GSS does not provide
any data about when the infidelity occurred. It
may be that a great percentage or relatively few
of the divorced participants engaged in infidel-
ity before their current relationship; this is im-
possible to determine conclusively. However, it
is possible to test whether the effects of our
model depend on participants' divorce status.
Moreover, for the present study, the crucial
point regarding the timing of the EMS is
whether divorced individuals whose EMS may
have occurred in a previous marriage demon-
strate radically different effects in regard to our
predictor variables—in particular, the marital
satisfaction variable—or whether they are more
or less similar to individuals whose EMS oc-
curred during their present marriage.
To investigate this possibility, we conducted
an analysis of our variables, including an inter-
action term between divorce and every other
term in our model. These divorce interaction
terms would indicate whether the effects of the
other predictors depended on participants' di-
vorce status. There was a single significant in-
teraction between divorce and educational de-
gree,
likelihood ratio /(I) = 10.20, p < .005.
The remainder of the divorce interaction terms
were not significant, including the Divorce X
Marital Satisfaction term. Thus, although there
was probably an unknown percentage of di-
vorced individuals who had EMS during a pre-
vious marriage, they did not appear to substan-
tially alter the other effects. The Divorce X
Education interaction term was included in the
main analyses so that it would not bias any of
the other findings.
UNDERSTANDING INFIDELITY
741
Logistic Regression
We modeled the relationship between EMS
and our predictor variables using multivariable
logistic regression. Because we were not testing
specific theories regarding EMS, we used a
model-building approach in the data analysis.
Initially, we entered all main effects and all
two-way interactions to create a full model and
then assessed the fit of various reduced models
using the log-likelihood chi-square, removing
nonsignificant terms (Agresti, 1996; Hosmer &
Lemeshow, 1989). Because previous literature
suggested that age may have a nonlinear asso-
ciation with EMS (Greeley, 1994; Wiederman,
1997),
we allowed our continuous variables
(age,
age when first married, religious behavior,
and income) to have curvilinear associations
with EMS. Researchers typically use polynomi-
als (i.e., quadratic, cubic, and so forth) to model
curvilinear responses. However, the "fit" of
polynomials has been shown to depend on de-
gree and type of curvilinearity (Seber & Wild,
1989).
An alternative way to fit nonlinear associa-
tions is through the use of spline functions (Har-
rell,
1997; Harrell, Lee, & Pollock, 1988).
Splines are smoothing functions that do not
require a priori specifications about shape. Sev-
eral "knots" are specified over the range of the
predictor variable, and the splines are smooth
polynomials between knots. There must be a
minimum of three knots to allow for nonlinear-
ity, and by reducing the number of knots, non-
linearity can be tested through a log-likelihood
chi-square with degrees of freedom equal to the
number of reduced knots (Harrell, 1997). For
the current analyses, we modeled continuous
variables using restricted cubic spline fits with
five knots placed evenly across the range of the
predictor variables. We entered categorical vari-
ables (gender, previous divorce, educational
level, marital satisfaction, and work status) into
the regression equation using treatment con-
trasts in which subsequent levels of the variable
were compared with a reference category.
All main effects of predictor variables were
significant in the range of p < .05 top <
.0001.
Two interactions in addition to the Divorce X
Education term were retained, one between age
and gender (p < .0001) and another between
marital satisfaction and religious behavior (p <
.05).
Residual analyses and tests for linearity
using the spline fits demonstrated that age and
age when first married had curvilinear relation-
ships with the probability of EMS. Income
showed a two-stage relationship: Among partic-
ipants with incomes up to $30,000, there was no
association between income and EMS; among
participants with incomes greater than $30,000,
there was a positive, linear association between
income and EMS. Religious behavior and edu-
cational level, an ordinal variable, demonstrated
linear relationships with the probability of
EMS.
Table 2 reports odds ratios, confidence
intervals for odds ratios, and significance levels
for a model with main effects and a model
including interactions.
Typically, in multiple regression analyses,
the model R
2
value is reported to indicate the
total amount of variance that the model ex-
plains. However, because the dependent vari-
able in logistic regression is dichotomous, R
2
values can depend heavily on the percentages of
the two categories in the dependent variable and
will tend to underrepresent the true proportion
of explained variance (see Agresti, 1996; Chris-
tensen, 1997; Mittlbock & Schemper, 1996). An
index that is similar to R
2
and can be computed
in logistic regression is the probability of con-
cordance index c, which has been shown to
represent the area under a receiver operating
characteristic curve (Harrell, 1997). Here this
index was a measure of how well the model
discriminated between those who had had EMS
and those who had not had EMS; a value of .5
would represent no ability to discriminate, and a
value of
1
would represent perfect discrimina-
tion. The c value for our model was .745, rep-
resenting a moderate ability of our model to
discriminate between survey participants on the
basis of their reported infidelity.
Similar to past research, we found that men
reported more EMS than women but that this
relationship was strongly dependent on the age
of the individual. Men 55-65 years of age at the
time of
the
survey were the most likely to report
infidelity, and men older and younger than this
cohort were less likely. It is important to note
that these findings indicate that this cohort of
men was more likely to have engaged in infi-
delity at any time in their lives, not necessarily
between 55 and 65 years of age. Similarly,
women 40-45 years old at the time of the
survey were the most likely to have ever had
EMS,
with women older and younger than this
cohort less likely to have had EMS. Further-
742
ATKINS, BAUCOM, AND JACOBSON
Table 2
Logistic Regression Analysis of Variables Associated With Extramarital Sex (N = 4,118)
Variable
Sex (female vs. male)
Age
Educational level
Religious behavior
Previous divorce (no vs. yes)
Income
Work status
Respondent works-spouse home
vs. both work
Respondent works-spouse home
vs. spouse works-respondent
home
Respondent works-spouse home
vs. neither work
Marital satisfaction
Very happy vs. pretty happy
Very happy vs. not too happy
Age when first married
Sex X Age
Marital Satisfaction X Religious
Behavior
Previous Divorce X Education
Level
Main
OR
2.64****
1.41*
1.32**
0.71***
1.95****
1.32**
LR*
2
0.81
0.86
0.51
LR^C
1.88
3.97
0.52****
effects model
95%
CI
for OR
2.09-3.34
1.06-1.86
1.09-1.61
0.59-0.85
1.58-2.42
1.05-1.66
(3) =
11.21**
0.62-1.05
0.61-1.21
0.34-0.76
2) = 63.76****
1.54-2.29
2.49-6.33
0.38-0.70
Model with
OR
1.83****
0.66****
1.09**
0.56****
1.63****
1.32**
LR x*(3)
=
0.76
0.79
0.52
LR ^(2) =
2.00
4.40
0.52****
LR ^(4) =
LR ^(2)
LR ^(1)
interactions
95%
CI
for OR
1.24-2.70
0.42-1.04
0.87-1.38
0.43-0.72
1.29-2.05
1.05-1.67
= 10.18**
0.58-1.00
0.56-1.12
0.35-0.79
69.91****
1.63-2.46
2.71-7.12
0.38-0.71
27.85****
= 7.42*
= 11.23*
Note. For continuous variables, odds ratios are based on the interquartile range of the variable. For
categorical variables, the odds ratio is the odds of extramarital sex of the second category listed relative to the
first category listed. The likelihood ratio chi-square is listed on the first line of categorical variables. OR =
odds ratio; CI = confidence intervals; LR = likelihood ratio.
t
p<.05. **/><.01. <p<.00\. ****p<.0001.
more, the findings indicate that women and men
45 years of age and younger do not differ in
regard to occurrence of EMS (see Figure 1).
Marital satisfaction showed a strong associa-
tion with EMS. Respondents who reported that
their relationships were "pretty happy" and "not
too happy" were two and four times more
likely, respectively, to have reported EMS than
respondents who reported that they were "very
happy" with their relationships. However, there
was also a significant interaction between mar-
ital satisfaction and religious behavior. Consid-
ered by
itself,
religious behavior showed a neg-
ative association with EMS, such that those who
never attended religious services were 2.5 times
more likely to have had EMS than those who
attended religious services more than once a
week. The interaction between religious behav-
ior and marital satisfaction is displayed in Fig-
ure 2, in which separate lines represent the three
categories of marital satisfaction plotted against
attendance at religious services and the proba-
bility of EMS. The different slopes of the mar-
ital satisfaction lines demonstrate that those
who reported "pretty happy" or "not too happy"
marriages showed little or no effect of their
religious behavior and that individuals in "very
happy" marriages showed a strong effect of
their religious behavior.
Both of our opportunity variables showed a
significant relationship with EMS. The categor-
ical variable combining respondent and spouse
work status showed that respondents reported
less EMS when neither partner was employed
than for the other three categories. The combi-
nation in which respondents were working but
spouses were not working was the most indic-
ative of EMS but was significantly different
only from the case in which both spouses were
at home. With respect to respondents' income,
we found that income was not associated with
EMS among survey participants earning up to
$30,000 per year. However, participants earning
more than $30,000 per year showed a positive
UNDERSTANDING INFIDELITY
743
I S-
S
p_
6
Male
Female
20
30 40
70 80
50
60
Age
Figure
1.
Probability
of
extramarital sex by age and gender.
90
relationship between their income
and
reported
EMS.
Survey respondents earning $75,000
or
more
per
year were more than
1.5
times
as
likely
to
have
had
EMS
as
survey respondents
earning
up to
$30,000
per
year.
Participants' education showed
an
increasing
linear association with EMS, such that the more
highly educated
an
individual,
the
greater
the
likelihood
of
her
or
him having
had
EMS.
Par-
ticipants with graduate degrees were 1.75 times
more likely
to
have
had
EMS than participants
with less than
a
high school education. Past
divorce
was
also
a
strong predictor
of
EMS;
respondents
who had
been divorced were
al-
most
two
times more likely
to
report EMS
at
some point during their lives than were respon-
dents
who had
never been divorced.
As
noted
earlier, there
was
also
a
significant interaction
between education and divorce. The association
between increasing education and infidelity
ap-
peared only
for
respondents
who had
been
divorced.
Finally, respondents
who
were first married
at
a
young age showed
a
much higher likelihood
of EMS. Specifically, respondents
who
were
first married
at
16 years
of
age or younger were
the most likely
to
report EMS. Likelihood
of
EMS decreased steadily
as
age
at
which partic-
ipants were first married increased. Survey
re-
spondents who were married
at
16 years
of
age
were almost four times more likely
to
report
EMS than those
who
were first married
at 23
years
of
age (see Figure
3).
Discussion
Infidelity
is a
common problem in marriages.
Although much
has
been written about infidel-
ity
in the
popular press
and
media, research
on
the topic
has
been slow
to
accumulate. Given
the difficulty
of
obtaining information
on
this
sensitive matter, many previous investigations
have encountered
a
variety
of
methodological
difficulties. The current study investigated basic
factors
of
individuals
and
relationships that
might
be
associated with
EMS, and we im-
proved on past research by using
a
multivariable
modeling procedure
and a
nationally represen-
tative sample.
Two variables that
we
considered indexes
of
opportunity
for
EMS, income and employment
status,
were both significantly related with infi-
delity.
The
association between respondent's
income
and
EMS
was in the
hypothesized
di-
rection.
The
positive relationship between
in-
come
and
infidelity
for
participants earning
more than $30,000 per year shows that financial
means are related
to
the likelihood
of
infidelity.
The fact that there
was no
effect
for
those
re-
spondents earning
up to
$30,000
may
reflect
a
744
ATKINS, BAUCOM,
AND
JACOBSON
0.20
in
amantalj
Probability
of
extn
0.05 0.10
000
Not too happy
_
. . _
-^_
_
Pretty happy
_
Very happy
1 8
9
2
3 4 5 6 7
Attendance
at
religious services
Figure
2.
Probability
of
extramarital
sex by
marital satisfaction
and
religious behavior (1
=
never attend
religious services,
9 =
attend religious services more than once
a
week).
"floor effect." That
is, if a
certain level
of fi-
nancial resources facilitates
EMS,
then falling
below $30,000 might
not
provide
the
necessary
means,
and
earning more than $30,000 might
facilitate
EMS. For
example, with increasing
income,
it
might
be
easier
to
hide
the
costs
of
entertainment
or
other expenses incurred
as a
result