Journal of Personality and Social Psychology
2006, Vol. 91, No. 6, 1045–1065
Copyright 2006 by the American Psychological Association
0022-3514/06/$12.00 DOI: 10.1037/0022-3514.91.6.1045
Relationship Perceptions and Persistence: Do Fluctuations in Perceived
Partner Commitment Undermine Dating Relationships?
Ximena B. Arriaga, Jason T. Reed, Wind Goodfriend, and Christopher R. Agnew
Purdue University
The authors propose specific temporal profiles that reflect certainty versus doubt about where a partner
stands with respect to a dating relationship over time. Two multiwave longitudinal studies focused on
within-participant changes in perceived partner commitment. Results from multilevel modeling indicate
that individuals whose perceptions of partner commitment fluctuate over time were more likely to be in
a relationship that eventually ended than were individuals whose perceptions remained relatively steady.
For individuals in recently initiated relationships, the association of fluctuation in perceived partner
commitment with later breakup was significant regardless of the initial level of perceived partner
commitment or the trend, and for all participants, it remained significant when initial level, trend over
time, and fluctuation over time of other meaningful variables were controlled.
Keywords: uncertainty, commitment, perceptions of partner, relationship stability, longitudinal
Juliet:
However, certainty about one’s own sentiments may not allay
concerns about the partner’s sentiments; an individual may feel
confident about the partner’s sentiments one week only to be
surprised the following week by an unexpected partner behavior
that is perceived to reflect less commitment.
We suggest that individuals who vacillate in their perceptions of
their partner’s level of commitment—perceptions of how inclined
the partner is to remain in the relationship—are likely to experience continual doubts over where things stand; they are likely to
wonder whether the partner wants the same level of commitment
or instead wants more or less commitment. We examined this idea
among individuals in relatively newly formed dating relationships
(Study 1) and dating relationships of all durations (Study 2, a
replication study). Partners in novel relationships may wonder
about the uncertain current and future status of their relationship,
given that it is precisely things not yet fully known that elicit a
search for answers (Sanbonmatsu, Posavac, Vanous, & Ho, 2005).
Thus, individuals in newly formed relationships may be particularly interested in, and affected by, their perceptions of the partner’s commitment so as to protect themselves against getting
closer than the partner will allow.
Several studies indirectly suggested that perceiving a partner as
being committed should positively influence the relationship
(Drigotas, Rusbult, & Verette, 1999; Holmes & Rempel, 1989;
Miller & Rempel, 2004; Murray, Bellavia, Rose, & Griffin, 2003;
Murray, Holmes, & Griffin, 2000; Wieselquist, Rusbult, Foster, &
Agnew, 1999). If such positive perceptions were to change—if, for
example, a person loses faith over time that the partner is committed—these studies implied that this would result in negative
outcomes for the relationship. However, not all types of changes
over time in perceived partner commitment are the same. We
identify specific change markers or temporal profiles with specific
implications for a relationship, suggesting that beyond mere increases or decreases over time, temporal patterns marked by fluctuation in perceived partner commitment directly reflect doubts
and are particularly detrimental. We present the results of two
[. . .] O gentle Romeo:
If thou dost love, pronounce it faithfully: [. . .]
Romeo: Lady, by yonder blessed moon I swear
That tips with silver all these fruit-tree tops –
Juliet:
O, swear not by the moon, the inconstant moon,
That monthly changes in her circled orb,
Lest that thy love prove likewise variable.
—William Shakespeare, Romeo and Juliet
In this famous balcony scene from Romeo and Juliet, Romeo
first professes his love to Juliet. Juliet wants to know that Romeo
loves her not just now, but always; that is, she seeks assurance that
his love is constant over time. Just like Juliet, people in lasting
romantic relationships seek assurance that their own relationships
are based on ever present, or temporally stable, factors that keep
the partners together over time (Kelley, 1983). More so than in
marital relationships, in dating relationships such factors are likely
to rely heavily on feelings and perceptions and perhaps less on
tangible investments or structural barriers to leaving (Le & Agnew,
2003). Past research has shown that a dating relationship is disrupted when a person’s own feelings vacillate over time and is
more likely to persist when a person has relatively stable feelings
over time, even when these stable feelings are not particularly
positive toward the relationship (Arriaga, 2001). A person with
relatively unchanging sentiments is afforded a sense of certainty
about where he or she stands with respect to the relationship.
Ximena B. Arriaga, Jason T. Reed, Wind Goodfriend, and Christopher
R. Agnew, Department of Psychological Sciences, Purdue University.
We thank Niall Bolger for his thoughtful comments regarding this
research. We also thank the College of Liberal Arts at Purdue University
for supporting this research through an incentive grant.
Correspondence concerning this article should be addressed to Ximena
B. Arriaga, Department of Psychological Sciences, Purdue University, 703
Third Street, West Lafayette, IN 47907-2081. E-mail: [email protected]
1045
1046
ARRIAGA, REED, GOODFRIEND, AND AGNEW
longitudinal studies of individuals in dating relationships to examine whether those who vacillate over time in their perceptions of
their partner’s commitment are at greater risk for experiencing
relationship dissolution.
Relationship Confidence Versus Doubt
The Role of Perceived Partner Commitment
A growing body of research has suggested that people are
motivated to form a strong sense of conviction about their romantic relationships—a strong sense that “the partner really is the
‘right’ person and can be counted on to be caring and responsive
across time and situations” (Murray, 1999, p. 23; see also Holmes,
2004). This “quest for conviction” account stipulates that conviction is necessary before allowing oneself to feel close to a partner,
lest one become anguished over being tied to a partner with
undesirable qualities or hurt by a partner’s uncaring acts (Murray,
2005). Unfortunately, disappointment is inevitable—no partner is
always perfect or able to act in ways that promote the relationship.
The quest for conviction account suggests that the only way for
relationships to survive such disappointing moments is to adopt a
sense of closure, namely to remove all doubts about the relationship and embrace the idea that even when there are setbacks or
disappointments, these are minor and the relationship remains on
course toward a healthy future (i.e., making a “leap of faith,”
Rempel, Holmes, & Zanna, 1985; see also Holmes, 2004).
In contrast to individuals who follow a quest for conviction,
individuals marred by doubts must continually reevaluate where
the relationship stands and whether it has a viable future. Doubts
about a relationship may arise from any number of sources, such
as from (a) feeling uncertain about one’s own feelings toward the
relationship or partner (Arriaga, 2001; Murray & Holmes, 1999),
(b) feeling that the partner is not concerned with the relationship
(Murray et al., 2000; Holmes & Rempel, 1989), or (c) doubtinducing influences beyond the dyad, such as persistent perceived
disapproval from others (Agnew, Loving, & Drigotas, 2001;
Etcheverry & Agnew, 2004; Lehmiller & Agnew, 2006). In the
current article, we focused on the second source of doubt, namely
doubts about a partner’s commitment.
We suggest that a general inference about a partner’s commitment may be more strongly associated with relationship persistence than perceptions of a partner’s behaviors or even inferences
about those behaviors. Why should perceived partner commitment
have such a strong effect?
There are several reasons to expect that general inferences such
as perceptions of partner commitment are particularly important
for relationship well-being, more so than inferences about a partner’s behaviors, his or her level of satisfaction, quality of alternatives, investments, or other specific partner inferences. First, research has confirmed that inferences made about a partner’s
behavior are just as or even more important than the behaviors
themselves (Bradbury & Fincham, 1990). Second, general inferences about how concerned a partner is with the relationship— his
or her relationship motives— have profound effects on how rewarding a relationship is experienced to be (Kelley, 1979). Third,
general partner inferences form the bases of broad attitudes or
orientations toward a relationship (i.e., macromotives, Holmes,
1981) that temper the perceptions of specific partner actions or
even interpretations of specific actions; perceived partner commitment is precisely such a broadly framed inference (Wieselquist et
al., 1999; see also the discussion of trust by Holmes & Rempel,
1989). Fourth, predicting a broad outcome (relationship stability)
is likely to require a broad-based inference rather than a more
specific inference (Ajzen & Fishbein, 1977). Satisfaction, alternatives, and investments all are specific causes of commitment; as
such, perceptions of these variables are less likely to predict
stability directly than are perceptions of commitment (Rusbult,
Olsen, Davis, & Hannon, 2001).
There is ample extant research suggesting that perceptions of
partner motives have implications for relationship well-being. Perceiving that a partner acts in ways that support the relationship
(i.e., perceiving partner accommodation and partner willingness to
forego his or her own self-interest for the sake of the relationship)
is concurrently associated with trusting the partner (Wieselquist et
al., 1999). Conversely, individuals who have doubts about whether
their partner will be responsive to their needs make negative
inferences about their partner’s behavior and harbor further doubts
(Miller & Rempel, 2004). Moreover, individuals who assume their
partner sees them in more negative ways than the partner actually
sees them (i.e., perceived low regard) report subsequent increases
in their own levels of conflict and ambivalence and have partners
who report subsequent decreases in satisfaction and trust (Murray
et al., 2000). It is important to note that the links between low
partner regard and declines in relationship well-being remain significant even after controlling for a partner’s actual level of regard,
underscoring the importance of perceptions of the partner (see also
Wieselquist et al., 1999). It stands to reason that doubts about a
partner’s commitment also have negative repercussions for the
relationship.
Although these studies suggested positive relationship implications when a partner is perceived to be caring, supportive, and
concerned with maintaining the relationship, there are several
caveats that limit the relevance of these studies to the current
research. First, none of these studies provided direct evidence that
perceptions of the partner predict the ultimate indicator of wellbeing, namely whether a relationship lasts. Second, these studies
examined variables that only indirectly tap subjective perceptions
of partner commitment. For example, Wieselquist et al. (1999)
examined reports that the partner had engaged in prorelationship
behaviors; Miller and Rempel (2004) examined one’s level of
trust. Drigotas et al. (1999) have come closest to examining
absolute levels of perceived partner commitment, namely perceived matches versus mismatches in own and partner commitment. We sought to directly examine perceptions of partner commitment per se given our broader focus on doubt and the strong
link between subjective commitment and resolving doubts (Brickman, 1987).
Most notably, there has been no research that identifies specific
patterns in partner perceptions over extended periods of time,
extended patterns that would meaningfully reflect sustained certainty versus repeated doubts (Kelley, 1979). Two individuals may
exhibit similar levels of perceived partner commitment at two time
periods separated by 6 months, but 1 may have vacillated in his or
her perceptions of the partner’s commitment whereas the other
may have had perceptions that followed a steady path. As we
describe below, these two patterns are distinct in ways that are
theoretically meaningful and are likely to have different outcomes.
Moreover, when examining perceptions, it becomes critically important to capture prospective changes over time; data that capture
prospective changes in perceptions versus retrospective accounts
of changes in perceptions differ descriptively and in their predictive value (Karney & Frye, 2002). In short, much can be gained
theoretically and in predictive value by capturing changes that
unfold over time (Arriaga, 2001; Karney & Bradbury, 1995).
Doubt as Temporal Fluctuations in Perceived Partner
Commitment
Perceived Partner Commitment
FLUCTUATIONS IN PERCEIVED PARTNER COMMITMENT
1047
Profile 1
Profile 2
Profile 3
Profile 4
Time
How do people who do not treat their partner’s commitment as
an open question look different over time from those who do
(Holmes & Rempel, 1989)? Does experiencing sustained certainty
versus repeated doubts make a difference in whether a dating
relationship lasts? To answer these questions, we advance an
analysis of specific within-person temporal patterns, or profiles, in
perceived partner commitment. As elaborated in Shoda’s notion of
a personality signature (Shoda, 1999; Shoda, Mischel, & Wright,
1994), we suggest that distinct patterns of intraindividual variability (in our analysis, intraindividual changes over time) have coherence and are theoretically meaningful (see also Murray,
Holmes, & Collins, 2006).
Our major thesis is that measuring each individual’s idiosyncratic pattern of fluctuation in perceived partner commitment is
critical (Arriaga, 2001; Campbell, Simpson, Boldry, & Kashy,
2005), by which a person’s ongoing state of relationship doubt is
reflected in perceptions that vacillate markedly and sustained certainty is reflected in perceptions that stick to a linear path. This
assertion requires identifying temporal patterns with more precision than is afforded by examining only general changes over time,
which has presented its own challenges but increasingly is captured by a multiwave trend in a variable (i.e., its slope over time,
preferable to averaging between two measurement occasions; Karney & Bradbury, 1995). We identify four specific temporal patterns in perceived partner commitment that capture theoretically
relevant experiences of certainty versus doubt. We are not suggesting that every person exactly fits one of these four profiles or
that data naturally cluster into these four patterns. Rather, these are
four theoretically meaningful profiles, and it is possible to set
cutoffs on continuous data so as to approximate each of these four
profiles. These four patterns are a combination of whether the
general trend is one of increase or decrease over time and whether
the pattern fluctuates along or steadily follows a linear path.1
Figure 1 illustrates a hypothetical example of each profile.
One temporal profile is reflected in steadily increasing levels of
perceived partner commitment (e.g., Profile 1 in Figure 1). These
individuals see their partner as predictably committed and capture
what theoretically has been described as “exaggerating the case for
commitment” (Murray, 1999). Even if the partner acts destructively, the perception of a single destructive act becomes folded
into a broader partner perception that is positive and optimistic.
Thus, these are individuals who have likely developed strong
defenses against potentially negative partner information, defenses
that preempt perceiving declines in a partner’s commitment. By
deepening their positive view of the partner’s commitment over
time, and even enhancing their view in the face of unresolved
problems (see Holmes & Rempel, 1989, high trust individuals),
Figure 1. Temporal profiles in individuals’ perceptions of their partner’s
commitment level, measured repeatedly over time.
these individuals become immune to major relationship threats and
thus are likely to have relatively longer lasting relationships.
A second temporal profile is characterized by individuals who
generally perceive their partner to be committed, or even exhibit a
trend of increasing perceptions of partner commitment, but whose
perceptions fluctuate along that course (e.g., Profile 2). Their
relationship experience is likely to be different, defined at one
moment by optimism and the next moment by relative disappointment. These are individuals who capture what theoretically has
been described as having only moderate trust in the partner
(Holmes & Rempel, 1989); they acknowledge moments when the
partner behaves positively but hesitate in generalizing the positive
moment to future moments in the relationship. Their reactivity to
specific instances becomes apparent in a temporal profile marked
by fluctuation.
The third profile concerns those who have steadily low or
decreasing perceptions of their partner’s commitment, suggesting
a subjective sense of certainty that the partner is decreasingly
committed (e.g., Profile 3). They expect little of their partner, yet
the steady decline affords some level of certainty and predictability
over where their partner stands (Sorrentino, Holmes, Hanna, &
Sharp, 1995). They capture what theoretically has been described
as individuals who sustain their relationships with routines that
may not be satisfying or involve closeness but nonetheless keep
partners intertwined (Berscheid, 1983; see Holmes & Rempel,
1989, low trust individuals). As they avoid having to adjust their
expectations time and again, remaining in their relationship be1
In addition to these two indicators (whether perceived partner commitment follows a trend of increase or decrease and whether it fluctuates),
recent applications of growth curve analysis would suggest also looking at
initial level of perceived partner commitment (the intercept). We were less
concerned with strictly adhering to growth curve variables (a methodological focus) and more concerned with identifying patterns that have clear
theoretical relevance to the experience of doubt. Changes in a variable have
been shown to be more theoretically meaningful than the absolute level of
that variable (e.g., initial level; Karney & Bradbury, 1995), and our
emphasis is on examining increases and decreases that vary in degree of
fluctuation, as fluctuation is theoretically relevant to the experience of
doubt. We later present regression analyses that take into account level (the
intercept), in addition to increase versus decrease (slope) and fluctuation;
absolute level did not predict relationship persistence beyond the effect of
the linear trend over time or fluctuations over time.
1048
ARRIAGA, REED, GOODFRIEND, AND AGNEW
comes a “path of least resistance” albeit not necessarily a positive
path.
A fourth temporal profile is reflected in individuals whose
perceptions of a decreasingly committed partner fluctuate over
time (e.g., Profile 4). In contrast to individuals with steady downward perceptions (Profile 3), these individuals are unable to form
stable expectations of what the future holds given their vacillating
perceptions. Struggling to know where the partner stands, they
remain watchful of their partner’s actions for possible changes thus
readjusting their inferences time and again.
Predicting Breakup Status
The general trend of change, or slope, in perceived partner
commitment is likely to be important in predicting breakup status,
in light of seminal research showing that the linear trend in own
level of commitment (Rusbult, 1983), the trend in own satisfaction
(Karney & Frye, 2002), and the trend in attributions about a
partner (Karney & Bradbury, 2000) each are associated with
potential or actual relationship outcomes. Similarly, perceptions of
partner commitment that decrease over time (e.g., Profiles 3 and 4)
are more likely to result in breakup than those that increase (e.g.,
Profiles 1 and 2).
To establish temporal profiles or signature patterns that reflect
doubt versus certainty in particular, researchers must move beyond
examining increases or decreases only so as to also look at whether
the temporal path fluctuates versus steadily follows a straight line
(Surra, Hughes, & Jacquet, 1999). Arriaga (2001) found that
fluctuations in one’s own level of satisfaction—that is, uncertainty
in one’s own feelings—predicted later dissolution, even controlling for overall increases or decreases over time in one’s level of
satisfaction. However, one’s own feelings are but one source of
doubt and do not address an important additional source of doubt,
whether one thinks the partner is committed. We anticipated that
the extent of fluctuation over time in perceived partner commitment would affect the odds of breakup such that individuals fitting
Profile 2 would be more likely to be in relationships that end
relative to those fitting Profile 1, and the same holds for Profile 4
individuals relative to Profile 3 individuals. In short, fluctuation in
perceived partner commitment should have a main effect on relationship dissolution independent of the trend of increase or decrease in these perceptions or the overall level of these perceptions
(Hypothesis 1). Why should this be so?
The lack of partner predictability embedded in fluctuating perceptions means that these individuals are repeatedly adjusting their
expectations. Research has shown that individuals plagued by
uncertainty (as reflected in fluctuating perceptions of partner commitment) pay special attention to information that violates their
expectations (Driscoll, Hamilton, & Sorrentino, 1991) and thus
adjust their impressions time and again, making them highly
reactive (or sensitive) to each partner interaction (Surra & Hughes,
1997). Relative to low reactivity, such high reactivity is associated
with greater relationship distress, less satisfaction, and less closeness (Campbell et al., 2005; Jacobson, Follette, & McDonald,
1982).
Our logic suggests that certainty over perceived commitment
stands to have an effect that is independent of the general trend in
perceived partner commitment (Sorrentino et al., 1995). We are
not suggesting that steadily declining perceptions of partner com-
mitment (third profile) are good for a relationship; rather, declining
perceptions that fluctuate (fourth profile) are particularly bad for a
relationship, worse than steadily declining perceptions (third profile). They share a pattern of decline in perceptions but differ in
level of fluctuation in perceptions and thus should differ in odds of
breakup; certainty that a partner lacks commitment should predict
more persistence than doubt over a partner’s lack of commitment
because doubts encourage constant monitoring and redefinition of
the status of the relationship whereas certainty retains the status
quo.
An alternate model would suggest that only the first profile—
steadily increasing perceived partner commitment (i.e., a “felt
security” profile, see Murray et al., 2006)—is uniquely associated
with persistence whereas the other profiles are associated with
dissolution. That is, the effect of steady versus fluctuating perceptions may matter more when the perceptions are positive (i.e.,
Profile 1, but not Profile 2, predicts persistence), and less or not at
all when they are negative (i.e., Profiles 3 and 4 predict breakup
equally); we explored the possibility that the trend in perceptions
would moderate the effect of fluctuation but anticipated that fluctuation would exhibit an independent (i.e., main) effect on relationship dissolution.
In keeping with the analysis of the importance of perceived
partner commitment relative to other inferences about the partner
(e.g., behavioral inferences), we also predicted that fluctuations in
perceived partner commitment would be more strongly associated
with relationship dissolution than would fluctuations in perceptions of a partner’s behaviors or fluctuations in inferences about
partner behaviors (Hypothesis 2).
Possible Correlates and Constraints of Fluctuation in
Perceived Partner Commitment
Not all variables that fluctuate over time necessarily reflect
doubt. There are theoretical reasons to expect some variables to
correlate with fluctuation in perceived partner commitment more
than others and to expect some variables to constrain or moderate
the association of fluctuation in perceived partner commitment
with later relationship disruption. We explored possible concomitant, antecedent, and limiting conditions of fluctuation in perceived partner commitment.
Own Level of Satisfaction
Because fluctuations in own satisfaction (Arriaga, 2001) and in
perceived partner commitment both reflect doubt, they are likely to
be positively correlated. Yet, as theoretically distinct and unique
sources of doubt stemming from own feelings versus perceptions
of the partner, they should each provide independent prediction
when examined simultaneously. We predicted that, when examined simultaneously, both should exhibit independent associations
with later relationship dissolution (Hypothesis 3). We did not
examine perceived partner satisfaction, given that perceiving partner changes in overall commitment is likely to have more pervasive effects than perceiving changes in satisfaction only (Rusbult,
1983). Rather than perceived satisfaction, perceiving committed
partner acts more closely captures confidence in a partner’s intentions (Wieselquist et al., 1999).
FLUCTUATIONS IN PERCEIVED PARTNER COMMITMENT
Own Level of Commitment
We examined several possible types of association between own
level of commitment and perceived level of partner commitment.
First, we anticipated that when examined simultaneously, fluctuation in perceived partner commitment and fluctuation in own
commitment both should exhibit independent associations (main
effects) with later relationship dissolution (Hypothesis 4). Both
reflect doubt, and as such, both should influence whether a relationship lasts.
Second, we explored whether own commitment might be an
antecedent of perceived partner commitment. It is possible that a
relatively uncommitted person is inclined to assume a partner is
equally uncommitted—that is, one’s own level of commitment
may color inferences about a partner’s commitment. Yet, it is also
possible that people who perceive a partner to be uncommitted will
adjust their own commitment so as to not be disappointed in the
future (e.g., “My partner doesn’t want this relationship, so I should
not put much into it myself”)—that is, inferences about a partner’s
commitment may cause (and thus precede) adjustments in own
commitment level. We explored whether (a) initial level of perceived partner commitment correlated with subsequent increases
in own commitment and (b) initial level of own commitment
correlated with subsequent increases in perceived partner
commitment.
Third, we explored whether one’s own commitment level at the
outset of the study (initial commitment) constrains the extent of
doubt one experiences. It is possible that people who initially are
relatively uncommitted may not care very much whether the
relationship continues, and so they may not heed their perceptions
of the partner’s commitment. On the other hand, people who very
much want the relationship to continue—those who initially are
highly committed—may seek confirmation of the partner’s commitment; they may be more affected by evidence of a stably
committed partner versus a partner plagued by doubts. That is,
initial commitment level may moderate whether fluctuation in
perceived partner commitment predicts relationship dissolution. It
is equally possible that uncertainty about a partner’s commitment
undermines a relationship regardless of one’s own level of commitment given the pervasive and arguably independent effects of
doubt (Holmes, 2004), suggesting no such moderation.
1049
Overview of Studies
We have suggested that over time, fluctuating perceptions of
partner commitment disrupt a relationship and foretell its ending,
a straightforward proposition with rather complex methodological
implications. To adequately capture specific temporal patterns in
perceptions of the partner, we conducted two longitudinal studies
with multiple measurement occasions relatively close in time. For
each predictor variable, we estimated each individual’s initial level
at the start of the study, the trend of increasing or decreasing
perceptions over time, and the extent of variation or fluctuation
over time. Variation in a variable has been used in previous
research to capture theoretically meaningful patterns (cf. Aronson
& Inzlicht, 2004; Arriaga, 2001; Campbell et al., 2005; Crocker,
Karpinski, Quinn, & Chase, 2003).
Study 1 provided a test of all hypotheses, focusing exclusively
on individuals in relatively newly formed relationships (no more
than 6 months in duration). Participants completed eight measurement occasions on a weekly basis and a follow-up session 2
months after Time 8 to assess breakup. Because this study was
time intensive (it involved weekly sessions), many of the variables
were measured with single items, as has been done in other
time-intensive studies (e.g., Bolger, Zuckerman, & Kessler, 2000)
or studies demonstrating adequate validity of single-item measures
(e.g., Robins, Hendin, & Trzesniewski, 2001, with respect to
self-esteem; Agnew, Van Lange, Rusbult, & Langston, 1998, as
well as Aron, Aron, & Smollan, 1992, with respect to self–partner
cognitive overlap). Study 2 served to validate the single-item
measures used in Study 1 and assess whether findings generalize to
(a) people in relationships of varying durations and (b) profiles
based on multiple occasions separated by greater time spans (i.e.,
nine measurement occasions separated by 4 weeks each rather than
by 1 week).
Method
Study 1 was conducted with undergraduate students who volunteered to
participate as partial fulfillment of a course requirement at a large Midwestern university. Study 2 relied on extant data collected as part of a
major study on substance use in a sample of entering freshman students at
a large Midwestern university.
Own Attachment Style
The temporal profiles described above may have origins in
individual differences that are relevant to close relationships, such
as attachment style (see Simpson & Rholes, 1998, for an overview
of attachment processes). More so than individuals described as
secure or avoidant, those who are anxious gauge their perceptions
of the partner on the basis of their daily partner interactions and
thus may exhibit more volatility in inferring partner motives
(Campbell et al., 2005). We anticipated that fluctuation in perceived partner commitment would be positively correlated with
anxious attachment tendencies (Hypothesis 5) but not necessarily
with avoidant tendencies. We examined the correlations of fluctuation in perceived partner commitment specifically with the
absolute level of each insecure attachment variable, rather than
changes in insecure attachment variables, given that attachment
style is considered to be stable over relatively short periods of time
(Fraley, 2002).
Participants
In Study 1, of the 130 individuals who attended the Time 1 session, 25
participants were eliminated because they did not meet the criteria for
inclusion in this sample. Through a series of checks and assessments, it
became clear that 4 participants made up their responses, 4 were in
relationships longer than 6 months in duration at the start of the study (i.e.,
they ignored inclusion criteria provided during recruitment), 2 were dating
another participant in the study (they were randomly chosen over their
partner to ensure independence of observations), and 15 participants reported at Time 4 that their relationship had ended and thus did not provide
sufficient observations to establish a reliable longitudinal pattern. In the
absence of guidelines on how many observations are necessary to establish
whether a pattern of fluctuation occurs, it seemed reasonable to assume that
three data points would not be sufficient (at most, if they fluctuate, one
would capture a curvilinear pattern). Participants whose relationships
ended after Time 4 but before the follow-up session (n ⫽ 7), or ended by
the follow-up (n ⫽ 21), were coded as a breakup.
1050
ARRIAGA, REED, GOODFRIEND, AND AGNEW
Of the original 130 participants, 23 (18%) were lost because of attrition:
9 participants voluntarily dropped the study after Time 1 whereas 14 could
not be reached for the follow-up session and thus had missing information
regarding their relationship status (persisted vs. ended). These 23 participants did not differ significantly in their Time 1 level of perceived partner
commitment from those who were retained in the Study 1 sample, F(1,
104) ⫽ 1.96, ns. The final Study 1 sample of 82 individuals consisted of
44 women and 38 men in relatively newly formed dating relationships (6
months or less), who had usable data from at least four consecutive
measurement occasions and who subsequently provided data on their
breakup status.
Study 2 examined individuals in relationships of all durations. The larger
study from which the Study 2 sample was derived involved 912 incoming
freshman students who were recruited to provide weekly data on various
topics over the course of their freshman year. Initial recruitment took place
during the summer prior to starting their freshmen year, when students
visited campus as part of an orientation event. Study inclusion criteria
included some past experience with cigarette smoking (i.e., at least one
puff) prior to the study. Participants were paid for completing each weekly
survey, resulting in a subject retention rate of more than 90% over the 35
consecutive weeks of the freshmen year.
Every 4 weeks there were questions assessing whether they were in a
romantic relationship and tapping various characteristics of their relationships; we examined the sample of individuals who reported being in a
relationship (n ⫽ 630) at some point over the course of the study. Of these,
38 (6%) were lost because of attrition (i.e., they left the university or
voluntarily stopped participating in the study before it ended). Given that
these were freshman students, with many involved in fluid, short-term
relationships, many did not meet the criteria for inclusion in this sample:
They did not have a relationship that remained intact for 4 consecutive
times and thus did not provide sufficient observations to establish a reliable
longitudinal pattern (n ⫽ 339). The final Study 2 sample of 253 individuals
consisted of 111 women (44%) and 142 men (56%) who had usable data
from at least four measurement occasions. Unlike Study 1, there was not a
follow-up session to assess breakup; instead, participants who provided
data about an intact relationship for 4 consecutive times and whose relationships subsequently ended over the course of the study were included in
the sample and coded as a breakup.
Study 1 and Study 2 participants were similar, except for the duration of
their relationships. Study 1 participants’ relationships were, on average, 3
months in duration at Time 1; Study 2 participants’ relationships were, on
average, 16 months in duration when they became included in the current
sample. At Time 1, Study 1 participants were 19 years old on average
(SD ⫽ 1.36); Study 2 participants were almost exclusively 18 years old
given that they were college freshmen. The majority were White (in Study
1, 88% White, 7% Asian American, 4% Latino, and 1% African American;
in Study 2, 87% White, 5% Asian American, 2% Latino, 1% African
American, and 3% Other). Sixty-six percent (n ⫽ 54) of Study 1 participants continued to be in their relationships at follow-up whereas 34% (n ⫽
28) were no longer dating their Time 1 partners. Seventy-five percent (n ⫽
189) of Study 2 participants continued in their relationships over the course
of the study whereas 25% (n ⫽ 64) were in relationships that ended. Given
that Study 1 participants were exclusively in relatively newly formed
relationships, it is not surprising that a higher percentage of these relationships ended.
Procedure
Study 1 data were collected over the course of two semesters; the results
did not differ between the two semesters (i.e., there were no main or
interaction effects for semester designation). Data collection sessions for
Time 1 through Time 8 were conducted on a weekly basis in a small
classroom; approximately 10 –20 participants took part in each session. At
Time 1, the experimenter described the study tasks and obtained written
consent from participants. At each time period thereafter, the experimenter
reviewed the activities for the day’s session, assured participants that their
responses would remain confidential, and distributed questionnaires. Each
session lasted approximately 15 minutes. At Time 8, participants completed an additional one-page questionnaire that probed for dishonest
responding. The instructions reiterated that they would receive full credit
regardless of their responses to these final probes. At Time 8, participants
were also debriefed and thanked for their assistance. Follow-up questions
were administered by telephone roughly 2 months after Time 8.
For Study 2, participants logged onto the study website (with a preassigned username and a personally selected password) each week and were
presented with a set of survey questions. The initial weekly survey was
administered at the beginning of the fall semester and the final weekly
survey was administered during the final week of the spring semester.
Surveys were available every week, including winter and spring breaks, for
a total of 35 consecutive weeks. Participants were paid for participation
each week, which varied from week to week, but over 87% of participants
completed all surveys. A set of questions was presented each week to
participants focusing primarily on substance use over the preceding week
(e.g., cigarette use). In addition, participants responded to a differing subset
of questions each week about other aspects of their lives (relationships,
stress, sleep habits, etc.). These additional questions were rotated on a
4-week schedule; thus, relationship variables were administered every 4
weeks for a total of nine times over the course of the 35-week study.
Measures Included in Both Studies
Perceptions of the partner’s commitment was measured by a single item
in Study 1. Participants were asked, “How committed is your partner to the
relationship?” followed by a 9-point response scale ranging from 0 (not at
all committed) to 8 (very committed). In Study 2, this variable was measured with four items (␣ at initial time ⫽ .93), adapted from the Investment
Model Scale (Rusbult, Martz, & Agnew, 1998): “My partner is committed
to maintaining our relationship,” “My partner intends to stay in this
relationship,” “My partner feels very attached to our relationship – very
strongly linked to me,” “My partner is oriented toward the long-term future
of our relationship (for example, imagines being with me several years
from now)”; participants indicated their level of agreement with each item
on a 9-point response scale ranging from 0 (do not agree at all) to 8 (agree
completely).
To validate the single item used to tap perceived partner commitment in
Study 1, we derived two variables, one based on a single Study 2 item
(“My partner is committed to maintaining our relationship”) that is similar
to the single Study 1 item and a second one based on the four Study 2 items
averaged together. We then correlated these two variables at each time. The
average correlation across all times was .95 (range ⫽ .91–.98). Study 2
analyses of perceived partner commitment included the variable based on
an average of the four items.
Own level of satisfaction was measured by a single item in Study 1—“I
feel satisfied with our relationship at the moment”—followed by a 9-point
response scale ranging from 0 (do not agree at all) to 8 (agree completely).
In Study 2, this variable was measured with three items adapted from the
Investment Model Scale (␣ at initial time ⫽ .89): “I feel satisfied with our
relationship,” “My relationship is better than others’ relationships,” and
“Our relationship makes me very happy”; participants indicated their level
of agreement with each item on a 9-point response scale ranging from 0 (do
not agree at all) to 8 (agree completely).
To validate the single item used to tap own satisfaction level in Study 1,
at each time we correlated a variable based on a single Study 2 item (“I feel
satisfied with our relationship”) that was almost identical to the single
Study 1 item with a variable based on the three Study 2 items averaged
together. The average correlation across all times was .95 (range ⫽.91–
.97). Study 2 analyses of own satisfaction level included the variable based
on an average of the three items.
FLUCTUATIONS IN PERCEIVED PARTNER COMMITMENT
Own level of commitment was measured in Study 1 by a single item
from the Investment Model Scale—“I am committed to maintaining my
relationship with my partner”—and in Study 2 by four items from this scale
(␣ at initial time ⫽ .93): “I am committed to maintaining my relationship
with my partner,” “I intend to stay in this relationship,” “I feel very
attached to our relationship – very strongly linked to my partner,” and “I
am oriented toward the long-term future of my relationship (for example,
I imagine being with my partner several years from now).” In all cases,
participants indicated their level of agreement with each item on a 9-point
response scale ranging from 0 (do not agree at all) to 8 (agree completely).
To validate the single item used to tap own commitment level in Study
1, at each time period we correlated a variable based on the Study 2 item
that was identical to the single Study 1 item, with a variable based on the
four Study 2 items averaged together. The average correlation across all
times was .95 (range ⫽ .92–.98). Study 2 analyses of own commitment
level included the variable based on an average of the four items.
Breakup status was measured in Study 1 with the following question:
“Are you still dating the same person that you were when you last
participated in this study?” We derived a two-level breakup status variable
consisting of the group of participants whose relationships persisted (n ⫽
54) versus the group whose relationships ended (n ⫽ 28).
Similarly in Study 2, at each time when participants were asked about
their relationship, they were asked whether they had a romantic partner. If
they indicated that they did have a romantic partner, they were asked to
provide the first name and first letter of the last name of that person.
Subsequently, they were given that person’s name and asked if they were
still involved in a romantic relationship with that person. The persisted
group was based on participants who indicated they had the same partner
throughout the study (n ⫽ 189), and the ended group was based on
participants who indicated they no longer had the same partner from the
previous time (n ⫽ 64).
Measures Included Only in Study 1
Perceptions of positive partner behavior was measured by using two
items (“Based on what has occurred in the past week, how much did your
partner do nice things that matter to you?” “How representative are these
nice behaviors of the type of person your partner is?”), followed by a
9-point response scale ranging from 0 (very little/not at all representative)
to 8 (very much/very representative). These two items were averaged (␣ at
initial time ⫽ .86).
Perceptions of negative partner behavior was measured by using two
items asking participants “Based on what has occurred in the past week,
how much did your partner do negative things that matter to you?” and
“How representative are these negative behaviors of the type of person
your partner is?”, followed by a 9-point response scale ranging from 0
(very little/not at all representative) to 8 (very much/very representative).
These two items were averaged (␣ at initial time ⫽ .74).
Attachment style was measured for other purposes beyond the current
research that demanded using a multiitem scale. We used the 17-item Adult
Attachment Questionnaire (Simpson, Rholes, & Phillips, 1996), which
consists of a series of statements and a 9-point response scale ranging from
0 (do not agree at all) to 8 (agree completely). Nine items captured an
anxious–ambivalent attachment dimension (e.g., “I often worry that my
partner(s) don’t really love me”; ␣ at initial time ⫽ .79); these items were
averaged such that higher numbers reflect anxious–ambivalent tendencies.
Eight items captured an avoidant attachment dimension (e.g., “I don’t like
people getting too close to me”; ␣ at initial time ⫽ .77); these items were
averaged such that higher numbers reflect avoidant tendencies.
For purposes of validating the measure of perceived partner commitment, there was an item in Study 1 only that tapped confidence in a
partner’s dependability. Participants were asked, “How confident are you
that you can depend on your partner?” followed by a 9-point response scale
ranging from 0 (not at all confident) to 8 (very confident). At Time 1,
1051
perceived partner commitment and confidence in a partner’s dependability
were highly significantly correlated, r(82) ⫽ .79, p ⬍ .001, providing
evidence of convergent validity. Furthermore, at Time 1, correlations of
perceived partner commitment with other relationship variables were also
significant but lower in magnitude: own level of commitment, r(82) ⫽ .34,
p ⫽ .002; own level of satisfaction, r(82) ⫽ .62, p ⬍ .001; and the
correlations with individual difference variables were even lower in magnitude, thus providing evidence of divergent validity: anxious–ambivalent
attachment style, r(82) ⫽ ⫺.27, p ⫽ .014; avoidant attachment style,
r(82) ⫽ ⫺.12, ns.
Results
Strategy for Analyzing Change Over Time
Analysis of the data proceeded in two stages. First, we used SAS
software’s PROC MIXED (SAS Institute, 1992; Singer, 1998) and
additional SAS data steps to derive the three within-person change
estimates (i.e., initial level, linear trend over time, fluctuation over
time) for each perception variable. Second, we used these estimates to predict breakup status in subsequent regression models.
We adopted this two-stage strategy to address several challenges
described below.
One challenge concerned the hierarchical (nested) nature of the
data; repeated ratings over time were nested within a participant.
That is, each observation corresponded to a particular time for a
particular participant, so there were multiple observations for each
participant (one for each time). Because observations were clustered by participant, they were likely to violate the assumption of
being independent (Gable & Reis, 1999). By adopting a multilevel
model approach, PROC MIXED accounted for nesting of observations within participant by treating the intercept and slope as
random rather than fixed variables.2
Were repeated ratings clustered (nested) within a person? The
intraclass correlation (␳) provides an indicator of clustering; it
reflects the proportion of total variance explained by betweenperson variation versus within-person variation (i.e., betweenperson variance divided by the sum of between-person variance
and within-person variance; Singer, 1998). In Study 1, the extremely high value (␳ ⫽ .97) indicated little variance within
participants relative to variance between participants (i.e., a particular participant’s responses were homogenous), suggesting a
multilevel model approach over standard (ordinary least squares)
regression. In Study 2, the intraclass correlation (␳ ⫽ .24) indicated less within-participant clustering than in Study 1 but a fair
amount of clustering nonetheless that could inflate the Type 1 error
rate (Kashy & Kenny, 2000).
A second challenge was more difficult to overcome and concerned how to examine whether fluctuation over time in one or
2
By treating the intercept and slope as random variables, we can assume
we sampled ratings from a participant rather than obtained all possible
ratings for each participant (that is, we did not obtain the population of
ratings for each participant). Just as samples vary in the number of
observations included, participants varied in the number of observations
included. It was not essential that all participants have the same number of
observations (i.e., the same number of measurement occasions), and participants who missed an occasion could still be included in the analysis. As
in hierarchical linear and nonlinear modeling (see Karney & Bradbury,
1997), PROC MIXED adjusts participant-specific estimates for their precision.
1052
ARRIAGA, REED, GOODFRIEND, AND AGNEW
more variables predict later breakup status—for example, whether
vacillating versus generally stable levels of (a) perceived partner
commitment and (b) own commitment each predicted later
breakup when examined simultaneously. We first describe our
approach and then explain why we adopted it.
Our strategy was to derive growth curve estimates for each
participant (cf. Karney & Bradbury, 1995). Growth curve analysis
can be thought of as calculating a regression model for each
participant, where his or her repeated perception ratings are regressed onto time. Specifically, we used PROC MIXED to test the
following (Level 1) model, where the time variable was recoded so
that the first time period had a value of zero:
variable measured at all times ⫽
intercept ⫹ slope(time)⫹ residual
This provided output to construct a new data set with several
observations for each participant (one observation at each time).
The new data set included each participant’s actual rating at each
time and the predicted rating based on the estimated linear slope in
his or her ratings from one time to the next.3
For each participant, we derived three change estimates of each
predictor variable. One estimate was a participant’s initial perception at the start of the study, on the basis of the estimated intercept
(i.e., given the way time was coded, each participant’s predicted
level when time equaled zero). A second estimate was a participant’s trend in a given variable over time on the basis of the linear
slope in ratings (i.e., the predicted rating at one time minus the
predicted rating at the previous time), which reflected whether a
participant’s ratings on a variable generally followed a general
pattern of increase versus decrease over time. A third variable was
the amount of fluctuation over time in a variable for a given
participant, on the basis of variation in the participant’s slope. We
used the deviation between an actual (observed) score at a given
time and the predicted score (based on the Level 1 model) to
calculate the standard deviation around the slope line in each
participant’s ratings.4 For each predictor, we derived these three
variables reflecting a given participant’s change estimates— his or
her initial level of a given predictor, the linear trend over time, and
fluctuation over time. We used these participant-specific estimates
in subsequent standard regression analyses predicting breakup
status (persisted vs. ended). We repeated all of the analyses by
using logistic regression; this did not alter the patterns of significance.
Rather than derive change variables in one step (based on the
Level 1 model) and analyze them in a second step (the Level 2
regression model), an alternate strategy might have been an “allin-one” model, in which a single step is used to compare breakup
groups in their intercepts, slopes, and of noted importance, fluctuation. Commonly used multilevel model programs, such as HLM
and SAS software’s PROC MIXED, typically take this approach to
analyze intercepts and slopes as random variables. However, several of our hypotheses required two critical tasks: (a) examining
fluctuation in a variable (based on the within-person residual, rij),
and (b) assessing the association of breakup status with several
Level 1 variables simultaneously—that is, we wanted to see if
breakup groups differed in one fluctuation variable (e.g., fluctuation in perceived partner commitment), controlling for another
fluctuation variable (e.g., fluctuation in own commitment). None
of the commonly used programs accomplish both tasks in a single
step. HLM can be used to model intercepts and slopes for several
variables simultaneously in a single step (in the Level 1 model).
However, it currently cannot be used to compare ended versus
persisted relationships in Level 1 fluctuation in these variables,
which was the crucial variable in this research. On the other hand,
SAS’s PROC MIXED can be used to examine fluctuation in a
single variable, but it is not straightforward how to examine
several fluctuation variables simultaneously (i.e., several variables
in the Level 1 model).5
In short, we could have used SAS’s PROC MIXED to test an
all-in-one model to reflect hypotheses concerning fluctuation in
only one perception variable (e.g., Hypothesis 1, the association of
breakup status with fluctuation in perceived partner commitment,
controlling for the slope and intercept). However, we could not use
this program to examine several fluctuation variables simultaneously, as was specified in several hypotheses. Testing these
hypotheses required adopting the two-stage approach. It is important to note that in cases for which we could test a hypothesis by
using SAS PROC MIXED’s all-in-one approach (i.e., the association of breakup status with fluctuation in perceived partner commitment, controlling for the slope and intercept), we analyzed the
data this way as well as by using the two-step approach (deriving
estimates in one step and analyzing them in a second step); in no
case were the results weakened, and in most cases we obtained
greater significance by using the all-in-one approach. For the sake
of consistency, we present the results of the two-step approach for
all hypothesis tests.
Descriptive Statistics and Correlations
Table 1 presents the estimates and standard deviations for all
within-person change variables for each sample. The standard
deviation for each fluctuation variable was not included because
multilevel model analyses do not yield this information (i.e., the
variance for the residual term). In both studies, participants initially perceived their partner to be relatively committed (7.19 for
Study 1 and 7.43 for Study 2, both on a 0 – 8 scale). In Study 1,
perceptions of partner commitment significantly declined over
time (⫺.06), whereas the decline was a nonsignificant trend in
Study 2 (⫺.01). Overall levels of fluctuation in perceived partner
commitment were comparable in the two studies (.47 and .48).
3
A linear model suggests that, on average, individuals’ repeated ratings
follow a linear pattern more closely than other patterns (e.g., a curvilinear
pattern). Linear patterns have been shown to adequately approximate
change over relatively short periods of time in relationship variables
(Arriaga, 2001), even when the true model follows another pattern (Rogosa, Brant, & Zimowski, 1982).
4
This was the standard error of the estimate, which in this case was the
standard deviation in a linear model with two estimates (an intercept and a
slope). Given that the model had two estimates rather than just one estimate
(such as a model that calculates only the mean level over time rather than
the intercept and slope), we calculated the standard deviation by using n –
2 in the denominator to adjust for two model estimates rather than n – 1.
5
We thank Niall Bolger for showing us a way to use SAS PROC
MIXED to examine fluctuation in a single (Level 1) variable, which cannot
be achieved with hierarchical linear and nonlinear modeling. For more
information on the SAS approach, see the Appendix.
FLUCTUATIONS IN PERCEIVED PARTNER COMMITMENT
1053
Table 1
Estimates and Standard Deviations of Within-Person Change Variables (Study 1 and Study 2)
Study 1 (n ⫽ 82)
Variable
Initial level (intercept)
Perceived partner commitment
Perceived positive partner behavior
Perceived negative partner behavior
Own satisfaction level
Own commitment level
Trend over time (slope)
Perceived partner commitment
Perceived positive partner behavior
Perceived negative partner behavior
Own satisfaction level
Own commitment level
Fluctuation over time (residual)
Perceived partner commitment
Perceived positive partner behavior
Perceived negative partner behavior
Own satisfaction level
Own commitment level
Study 2 (n ⫽ 253)
Estimate
p⬍
SD
p⬍
Estimate
p⬍
SD
p⬍
7.19
6.84
1.80
6.68
7.03
.001
.001
.001
.001
.001
1.01
1.06
1.63
1.25
1.18
.001
.001
.001
.001
.001
7.43
—
—
7.01
7.29
.001
—
—
.001
.001
0.84
—
—
1.04
1.04
.001
—
—
.001
.001
⫺.06
⫺.07
.09
⫺.05
⫺.08
.010
.003
.002
.049
.001
.15
.15
.18
.18
.14
.001
.005
.002
.001
.001
⫺.01
—
—
⫺.01
⫺.03
.195
—
—
.488
.005
.10
—
—
.10
.13
.001
—
—
.001
.001
0.47
0.87
1.27
0.75
0.55
.001
.001
.001
.001
.001
0.48
—
—
0.74
0.58
.001
—
—
.001
.001
Note. Columns labeled p ⬍ indicate whether the corresponding estimate or standard deviation is significantly different from zero. Results of multilevel
models do not yield the standard deviation for fluctuation over time (i.e., the within-person residuals); thus, those estimates are omitted above. Dashes
indicate that perceived positive partner behavior and perceived partner negative behavior were not measured in Study 2.
Table 2 presents these estimates by breakup group for each
study. In both studies, at the univariate level (not controlling for
other variables) individuals in relationships that ended had levels
of perceived partner commitment that decreased and fluctuated
more over time, levels of satisfaction that were initially lower and
fluctuated more over time, and levels of own commitment that
were initially lower, decreased more, and fluctuated more over
time, relative to individuals in relationships that persisted. The
ended group also had perceptions of positive partner behavior that
were initially lower and fluctuated more over time, relative to the
persisted group (examined in Study 1 only). Group differences in
initial level of perceived partner commitment and the trend in own
Table 2
Estimates of Within-Person Change Variables, for Each Breakup Status Group (Study 1 and Study 2)
Study 1 (n ⫽ 82)
Persisted (n ⫽ 54)
Variable
Initial level (intercept)
Perceived partner commitment
Perceived positive partner behavior
Perceived negative partner behavior
Own satisfaction level
Own commitment level
Trend over time (slope)
Perceived partner commitment
Perceived positive partner behavior
Perceived negative partner behavior
Own satisfaction level
Own commitment level
Fluctuation over time (residual)
Perceived partner commitment
Perceived positive partner behavior
Perceived negative partner behavior
Own satisfaction level
Own commitment level
Study 2 (n ⫽ 253)
Ended (n ⫽ 28)
Persisted (n ⫽ 189)
Ended (n ⫽ 64)
Estimate
p⬍
Estimate
p⬍
Estimate
p⬍
Estimate
p⬍
7.42a
7.08a
1.68
6.99a
7.34a
.001
.001
.001
.001
.001
6.77b
6.42b
2.03
6.10b
6.45b
.001
.001
.001
.001
.001
7.54
—
—
7.26a
7.47a
.001
—
—
.001
.001
7.23
—
—
6.35b
6.85b
.001
—
—
.001
.001
.00a
⫺.05
.07
⫺.03
⫺.04a
.753
.036
.020
.117
.007
⫺.17b
⫺.15
.12
⫺.10
⫺.17b
.006
.025
.086
.169
.006
.01a
—
—
.01a
⫺.01a
.123
—
—
.537
.353
⫺.16b
—
—
⫺.09b
⫺.16b
.001
—
—
.038
.002
0.26a
0.67a
1.32
0.60a
0.44a
.001
.001
.001
.001
.001
0.89b
1.27b
1.11
1.08b
0.77b
.001
.001
.001
.001
.001
.35a
—
—
.56a
.42a
.001
—
—
.001
.001
0.97b
—
—
1.48b
1.20b
.001
—
—
.001
.001
Note. Columns labeled p ⬍ indicate whether the corresponding estimate or standard deviation is significantly different from zero. For each study,
estimates within a row with different subscripts indicate a significant difference between persisted and ended, p ⬍ .05. Dashes indicate that perceived
positive partner behavior and perceived partner negative behavior were not measured in Study 2.
ARRIAGA, REED, GOODFRIEND, AND AGNEW
1054
Table 3
Correlations Among Within-Person Change Variables (Study 1)
Variable
Initial level (intercept)
1. Perceived partner commitment
2. Perceived positive behavior
3. Own satisfaction level
4. Own commitment level
Trend over time (slope)
5. Perceived partner commitment
6. Perceived positive behavior
7. Own satisfaction level
8. Own commitment level
Fluctuation over time (residual)
9. Perceived partner commitment
10. Perceived positive behavior
11. Own satisfaction level
12. Own commitment level
* p ⬍ .05.
1
2
3
4
5
6
7
—
.58**
.58**
.49**
—
.76**
.65**
—
.85**
—
.54**
.28*
⫺.11
.38**
.62**
.73**
.22*
.58**
.67**
.65**
.53**
.77**
.59**
.48**
.36**
.82**
—
.64**
.40**
.63**
—
.60**
.57**
⫺.66**
⫺.37**
⫺.34**
⫺.55**
⫺.58**
⫺.74**
⫺.45**
⫺.47**
⫺.54**
⫺.47**
⫺.55**
⫺.65**
⫺.52**
⫺.30**
⫺.44**
⫺.73**
⫺.55**
⫺.49**
⫺.30**
⫺.40**
⫺.28*
⫺.54**
⫺.28*
⫺.30**
—
.56**
.04
⫺.10
⫺.08
⫺.15
8
9
10
11
12
—
.63**
.57**
.52**
—
.49**
.28*
—
.55**
—
—
⫺.45**
⫺.29**
⫺.27**
⫺.49**
** p ⬍ .01.
second sample was composed of the remaining participants (n ⫽
177). The mean relationship duration of participants in the first Study
2 subsample was 3 months, as was the case in Study 1; the mean
relationship duration of participants in the second Study 2 subsample
was 22 months. Thus, the Study 1 sample and first Study 2 subsample
were comparable in the type of participant and sample size, and the
only major difference was the time lag between measurement occasions (1 week in Study 1 vs. 4 weeks in Study 2).
satisfaction level were less robust (i.e., significant for one but not
both studies). None of the results involving perceived negative
partner behavior were significant, so this variable was not examined in further analyses.
Tables 3 and 4 display respectively Study 1 and Study 2 correlations among predictor variables. Correlations among variables
tapping a particular type of change (i.e., a particular change estimate, such as fluctuation) were generally higher in Study 2 than
Study 1. Correlations among the three change estimates for a
single predictor variable tended to be comparable in both studies
and generally fell into a moderate to high range (the average
magnitude was .54).
Testing Hypothesis 1
Do fluctuations in perceived partner commitment uniquely predict breakup? As can be seen in Table 5 (top half for Study 1,
bottom half for Study 2: rows for Perceived Partner Commitment,
Simple Association column), initial level, linear trend, and fluctuation in perceived partner commitment each were significantly
associated with breakup status. In both studies, when examining
initial level of perceived partner commitment, the linear trend, and
fluctuation in a simultaneous regression, fluctuation had a unique
association with breakup status, as did the trend (see the Individual
Comparing Study 1 and Study 2
We assessed support for each hypothesis in Study 1 and Study 2. In
cases where results of the two studies differed in ways that were
relevant to the hypotheses, we divided the Study 2 sample into two
subsamples. The first subsample, comparable to Study 1, included
participants who had been dating 6 months or less (n ⫽ 76); the
Table 4
Correlations Among Within-Person Change Variables (Study 2)
Initial level
Variable
Initial level (intercept)
1. Perceived partner commitment
2. Own satisfaction level
3. Own commitment level
Trend over time (slope)
4. Perceived partner commitment
5. Own satisfaction level
6. Own commitment level
Fluctuation over time (residual)
7. Perceived partner commitment
8. Own satisfaction level
9. Own commitment level
* p ⬍ .05.
** p ⬍ .01.
1
2
Trend over time
3
4
5
—
.75**
.67**
—
.79**
—
.20**
.38**
.17**
.35**
.48**
.31**
.16**
.32**
.16*
—
.67**
.64**
—
.77**
⫺.67**
⫺.47**
⫺.45**
⫺.59**
⫺.68**
⫺.60**
⫺.46**
⫺.41**
⫺.57**
⫺.50**
⫺.42**
⫺.35**
⫺.54**
⫺.55**
⫺.52**
Fluctuation over time
6
7
8
9
—
.67**
.64**
—
.79**
—
—
⫺.36**
⫺.46**
⫺.51**
FLUCTUATIONS IN PERCEIVED PARTNER COMMITMENT
1055
Table 5
Regression Analyses Predicting Breakup Status (Ended vs. Persisted), Study 1 and Study 2
Simultaneous effects
Variable
Simple
association
(Pearson r)
Individual
estimate
(␤)
Overall model
R2
F
df
Study 1 (n ⫽ 82)
Perceived partner commitment
Initial level
Linear trend
Fluctuation
Perceived positive partner behavior
Initial level
Linear trend
Fluctuation
Own satisfaction
Initial level
Linear trend
Fluctuation
Own commitment
Initial level
Linear trend
Fluctuation
.37**
.46**
⫺.48**
.01
.28*
⫺.32*
.28**
10.24**
3, 78
.37**
.31**
⫺.34**
.19
.08
⫺.16
.14**
4.56**
3, 78
.42**
.17
⫺.33**
.36*
⫺.03
⫺.13
.19**
6.04**
3, 78
.44**
.41**
⫺.29**
.35
.13
.03
.20**
6.57**
3, 78
Study 2 (n ⫽ 253)
Perceived partner commitment
Initial level
Linear trend
Fluctuation
Own satisfaction
Initial level
Linear trend
Fluctuation
Own commitment
Initial level
Linear trend
Fluctuation
.28**
.37**
⫺.40**
.09
.24**
⫺.22*
.20**
21.02**
3, 249
.41**
.30**
⫺.41**
.24**
.08
⫺.21*
.21**
21.50**
3, 249
.33**
.31**
⫺.41**
.17*
.16*
⫺.23**
.20**
20.45**
3, 249
Note. Breakup status was coded 0 for ended and 1 for persisted.
* p ⬍ .05. ** p ⬍ .01.
estimate column). We performed a second simultaneous regression
that also included higher order interactions among initial level,
trend, and fluctuation. None of the interactions were significant in
Study 1, but there was a significant two-way interaction between
the linear trend and fluctuation in Study 2, t(253) ⫽ ⫺3.29, p ⫽
.001. To decompose the interaction, we did a median split on the
linear trend. Controlling for initial level, in Study 2 greater fluctuation was highly associated with less persistence among individuals
whose perceptions increased over time (␤ ⫽ ⫺.46, p ⫽ .002) and not
associated with breakup status among those whose perceptions decreased over time (␤ ⫽ ⫺.18, p ⫽ .116). Initial level of perceived
partner commitment did not have a unique association with breakup
status in any of these analyses.
Given the inconsistent finding between Study 1 and Study 2, we
repeated the interaction analysis in the two subsamples of Study 2
described above, one with individuals in more recently initiated
relationships (comparable with Study 1) and another with the
remaining Study 2 participants. When regressing breakup status
onto initial level, the linear trend, fluctuation, and higher order
interactions, the interaction between linear trend and fluctuation
was not significant in the Study 2 subsample of individuals in
newly formed relationships, t(76) ⫽ ⫺1.01, ns, as was the case in
Study 1. However, this interaction was significant in the Study 2
subsample of individuals in relatively established relationships
(longer than 6 months), t(177) ⫽ ⫺2.29, p ⫽ .023, and decomposition of this interaction revealed a pattern similar to the full
Study 2 sample.
This suggests that Hypothesis 1 received robust support among
individuals in newly formed relationships: Controlling for their
initial level and increases or decreases in perceived partner commitment, we found that they were more likely to have relationships
that ended to the extent that their levels of perceived partner
commitment fluctuated over time; the linear trend in perceived
partner commitment also had a unique association with breakup
status, but initial perceived partner commitment did not. Support
for Hypothesis 1 among individuals in more established relationships was conditional: The unique association of fluctuation in
perceived partner commitment with breakup status occurred only
among individuals whose perceptions increased over time and not
among those whose perceptions decreased over time.
1056
ARRIAGA, REED, GOODFRIEND, AND AGNEW
Do specific temporal profiles have different odds of breakup?
The regression results just described can be interpreted in terms of
the four specific temporal profiles outlined in the introduction. The
moderated effect of fluctuation among individuals in longer lasting
relationships suggests that, when one’s overall perception of the
partner’s commitment (initial or absolute level) is controlled for,
the amount of fluctuation in perceptions is linked to breakup status
but only among those individuals whose perceptions are increasing
over time (Profiles 1 and 2) and not among those whose perceptions are decreasing over time (Profiles 3 and 4). In contrast, the
main effect of fluctuation among individuals in newly formed
relationships suggests that, regardless of one’s perception of the
partner at the outset of the study or whether one’s perceptions
increase or decrease over time, greater fluctuation (Profiles 2 and
4) goes with greater odds of breakup.
Tables 6 and 7 provide a concrete illustration of the odds of
breakup for each of the four profiles among individuals in newly
formed relationships, where there was a main effect for fluctuation
(entire sample of Study 1 and subsample of Study 2). We created
two groups to reflect the trend in perceived partner commitment—
those whose levels increased over time (i.e., positively sloped
values; n ⫽ 36 for Study 1 and n ⫽ 47 for the Study 2 subsample)
versus those whose levels decreased (i.e., negatively sloped values;
n ⫽ 46 for Study 1 and n ⫽ 29 for the Study 2 subsample). We
conducted a median split on fluctuation in perceived partner commitment to create two groups—those whose levels fluctuated over
time (n ⫽ 41 for Study 1 and n ⫽ 38 for the Study 2 subsample)
versus those whose levels were relatively steady (n ⫽ 41 for Study
1 and n ⫽ 38 for the Study 2 subsample). Combining these two
categorical variables allowed us to examine frequencies and percentages of ended versus persisted relationships in each of four
groups approximating a temporal profile.6 The results are displayed in the top halves of Table 6 (Study 1) and Table 7 (Study
2 subsample). The four groups were defined by actual levels on the
trend and fluctuation variable and initial levels were not controlled
for; as such, this analysis does not fully correspond to the regression models used to test Hypothesis 1, which controlled for initial
level.
Two findings stand out from this analysis. First, individuals
whose perceptions of their partner’s commitment increased over
time were much more likely to have steady than fluctuating perceptions (28 of 36, or 78%, in Study 1; 29 of 47, or 62%, in the
Study 2 subsample), and individuals whose perceptions of their
partner’s commitment decreased over time were much more likely
to have fluctuating than steady perceptions (33 of 46, or 72%, in
Study 1; 20 of 29, or 69%, in the Study 2 subsample). This is
consistent with the correlations reported in Table 3 (Study 1) and
Table 4 (Study 2): Individuals’ trends over time in perceived
partner commitment and the extent of fluctuation in their trends
were closely aligned: Study 1, r(82) ⫽ ⫺.55, p ⬍ .001; Study 2
full sample, r(253) ⫽ ⫺.50, p ⬍ .001; Study 2 subsample, r(76) ⫽
⫺.24, p ⫽ .038.
Second, when one examines individuals whose perceptions of
partner commitment declined over time (i.e., rows labeled Profiles
3 and 4 in Tables 6 and 7), those whose perceptions followed a
steady decline (Profile 3) were likely to be in relationships that
persisted despite the decline (77% in Study 1 and 78% in the Study
2 subsample). This was not the case for those whose perceptions
followed a fluctuating pattern of decline over time (Profile 4)—
roughly half of these individuals were in relationships that ended
(58% in Study 1 and 60% in the Study 2 subsample). A similar
pattern occurred among individuals whose perceptions followed a
pattern of increase over time (Profiles 1 and 2)—a pattern of
fluctuation increased odds of breakup. We repeated all of the
descriptive analyses by calculating each individual’s mean level of
perceived partner commitment over time and doing a median split
on this variable rather than using his or her slope over time. The
same pattern of results emerged.
Testing Hypothesis 2
Hypothesis 2 suggested that, when examined simultaneously,
perceptions of partner commitment should predict later breakup
status better than perceptions of specific positive and negative
partner behaviors (measured in Study 1 only). We did not examine
perceptions of negative behaviors given that breakup groups did
not differ on any of the three change estimates of this variable (see
Table 2). All three of the change estimates of the perceived
positive behavior variable had a significant correlation with
breakup status (see Table 5, first column) but none had a unique
association (Table 5, second column), possibly because of the high
intercorrelations among these three variables, particularly initial
level of perceived positive behavior with the linear trend and with
fluctuation (see Table 2).
To test Hypothesis 2, we compared the predictive value of each
set of variables—initial level, trend, and fluctuation for perceived
positive partner behavior versus the same three variables for perceived partner commitment. First, when breakup status was regressed onto all six variables, the change in R2 associated with
adding the three perceived partner commitment variables as a
group (i.e., the change in R2 when comparing a model with all six
variables vs. one with the three perceived positive behavior variables only) was significant, F(3, 75) ⫽ 4.82, p ⫽ .004, whereas the
change in R2 associated with adding the three perceived positive
partner behavior variables as a group (i.e., the change in R2 when
comparing a model with all six variables vs. one with the three
6
As stated in the introduction, we do not intend to suggest that
individuals perfectly cluster into these four profiles; individuals vary
continuously on the variables used to approximate the four profiles.
Tables 6 and 7 include the number of participants approximating each
profile for Study 1 and the Study 2 subsample. Not all participants had
levels that increased or decreased over time. In Study 1, 35% of
participants (n ⫽ 29) had slopes that were flat (i.e., they indicated the
same response at each time); in Study 2, 36% of participants (n ⫽ 90)
had flat slopes. Whereas the actual data of someone who provides the
same response over time would suggest a slope of 0, multilevel modeling assigns slope values that are adjusted for the sample (or group)
average slope. Categories in Tables 6 and 7 were based on the slope
values assigned in a multilevel model analysis, some of which were
close to 0 but slightly positive or negative (e.g., ⫺.001). The extent of
increase over time for those whose perceptions increased was less than
the extent of decrease over time for those whose perceptions decreased
(.03 vs. ⫺.12 in Study 1; .02 vs. ⫺.08 in Study 2). When one removes
individuals who started at the highest level and remained stable week
after week at the highest level (i.e., an analysis removing respondents
at the ceiling, as reported in a later section), the extent of increase
versus decrease still differs but not as much (.07 vs. ⫺.12 in Study 1;
.04 vs. ⫺.08 in Study 2).
FLUCTUATIONS IN PERCEIVED PARTNER COMMITMENT
1057
Table 6
Breakup Rates as a Function of Perceived Partner Commitment Over Time for the Entire Study 1 Sample and for a Subsample
Controlling for a Possible Ceiling Effect
Breakup group
Persisted
Perceived partner commitment
%
Ended
n
%
n
86
75
24
6
14
25
4
2
77
42
10
14
23
58
3
19
90
50
9
2
10
50
1
2
75
35
15
9
25
65
5
17
Entire sample (n ⫽ 82)
Perceptions of partner commitment increase over time (n ⫽ 36)
Profile 1: Steady increase (n ⫽ 28)
Profile 2: Fluctuating increase (n ⫽ 8)
Perceptions of partner commitment decrease over time (n ⫽ 46)
Profile 3: Steady decline (n ⫽ 13)
Profile 4: Fluctuating decline (n ⫽ 33)
Subsample controlling for possible ceiling effect (n ⫽ 60)
Perceptions of partner commitment increase over time (n ⫽ 14)
Profile 1: Steady increase (n ⫽ 10)
Profile 2: Fluctuating increase (n ⫽ 4)
Perceptions of partner commitment decrease over time (n ⫽ 46)
Profile 3: Steady decline (n ⫽ 20)
Profile 4: Fluctuating decline (n ⫽ 26)
Note. Profiles 1– 4 correspond to the four temporal profiles depicted in Figure 1 and described in the Introduction. For the entire sample, n for the persisted
group ⫽ 54 and n for the ended group ⫽ 28. For the subsample, n for the persisted group ⫽ 35 and n for the ended group ⫽ 25.
perceived partner commitment variables only) was not significant,
F(3, 75) ⫽ 0.14, ns. Thus, Hypothesis 2 was supported.
We sought to determine whether each perceived partner commitment variable predicted breakup status above and beyond the
effect of the positive behavior variables. We did not attempt to
interpret the effects of specific estimates in a simultaneous regression with six predictors, either in this analysis or subsequent
analyses combining perceived partner commitment with own satisfaction or with own commitment, because of high levels of
multicollinearity among the predictors (i.e., several with variance
inflation factors above 4.0). Instead, we conducted three follow-up
analyses, one for each perceived partner commitment variable,
controlling for the three positive behavior variables. Controlling
for the three perceived positive behavior variables in each of three
separate regressions, breakup status was predicted by initial level
in perceived partner commitment: overall model, F(4, 77) ⫽ 4.87,
p ⫽ .002 (individual estimate for initial level of perceived partner
commitment, ␤ ⫽ .30, p ⫽ .027); by the linear trend in perceived
partner commitment: overall model, F(4, 77) ⫽ 5.78, p ⬍ .001
(␤ ⫽ .39, p ⫽ .005); and by fluctuation in perceived partner
commitment: overall model, F(4, 77) ⫽ 6.85, p ⬍ .001 (␤ ⫽ ⫺.46,
p ⬍ .001). Thus, each of the three perceived partner commitment
variables was a robust predictor of breakup when controlling for
perceived positive behaviors.
Testing Hypotheses 3
We examined whether fluctuation in perceptions of the partner’s
commitment and fluctuation in one’s own level of satisfaction each
had a unique association with subsequent breakup status when
examined simultaneously (Hypothesis 3). In regressing breakup
status onto both variables simultaneously, the overall model was
significant in Study 1, F(2, 79) ⫽ 12.01, p ⬍ .001, and Study 2,
F(2, 250) ⫽ 30.38, p ⬍ .001; fluctuation in perceived partner
commitment provided unique prediction in Study 1 (␤ ⫽ ⫺.42,
p ⬍ .001) and Study 2 (␤ ⫽ ⫺.23, p ⫽ .003), and fluctuation in
own satisfaction provided unique prediction in Study 2 (␤ ⫽ ⫺.26,
p ⬍ .001) but not Study 1 (␤ ⫽ ⫺.08, ns). In the Study 2
subsample that was comparable with the Study 1 sample, each of
the two fluctuation variables uniquely predicted breakup status
(fluctuation in perceived partner commitment, ␤ ⫽ ⫺.28, p ⫽
.025; fluctuation in own satisfaction, ␤ ⫽ ⫺.25, p ⫽ .043). Thus,
Hypothesis 3 was moderately supported, with strong support for
fluctuation in perceived partner commitment and partial support
(in two of three samples) for fluctuation in own level of satisfaction.
We compared the relative predictive value of own satisfaction
variables versus perceived partner commitment variables in analyses parallel to those for Hypothesis 2. First, as seen in Table 5
(rows labeled Own Satisfaction in top half for Study 1 and bottom
half for Study 2), each of the satisfaction variables was correlated
with breakup status except for the linear trend in Study 1 (Simple
Association column). Initial level in own satisfaction had a unique
effect in Study 1 and Study 2 (Individual Estimate column) and
fluctuation had a unique effect in Study 2; the linear trend in own
satisfaction did not have a unique effect. Next, in Study 1, when
entering the three perceived partner commitment variables as a
group, controlling for the three own satisfaction variables, the
change in R2 was significant, F(3, 75) ⫽ 3.74, p ⫽ .015, whereas
the change in R2 associated with adding the three own satisfaction
variables, controlling for the three perceived partner commitment
variables, was not significant, F(3, 75) ⫽ 0.41, ns. In Study 2,
when entering the three perceived partner commitment variables as
ARRIAGA, REED, GOODFRIEND, AND AGNEW
1058
Table 7
Breakup Rates as a Function of Perceived Partner Commitment Over Time for Study 2 in Newly Formed Relationships and for a
Subsample Controlling for a Possible Ceiling Effect
Breakup group
Persisted
Perceived partner commitment
%
Ended
n
%
n
93
83
27
15
7
17
2
3
78
40
7
8
22
60
2
12
95
79
19
11
5
21
1
3
60
40
6
6
40
60
4
9
Entire sample (n ⫽ 76)
Perceptions of partner commitment increase over time (n ⫽ 47)
Profile 1: Steady increase (n ⫽ 29)
Profile 2: Fluctuating increase (n ⫽ 18)
Perceptions of partner commitment decrease over time (n ⫽ 29)
Profile 3: Steady decline (n ⫽ 9)
Profile 4: Fluctuating decline (n ⫽ 20)
Subsample controlling for possible ceiling effect (n ⫽ 59)
Perceptions of partner commitment increase over time (n ⫽ 34)
Profile 1: Steady increase (n ⫽ 20)
Profile 2: Fluctuating increase (n ⫽ 14)
Perceptions of partner commitment decrease over time (n ⫽ 25)
Profile 3: Steady decline (n ⫽ 10)
Profile 4: Fluctuating decline (n ⫽ 15)
Note. Profiles 1– 4 correspond to the four temporal profiles depicted in Figure 1 and described in the Introduction. For the entire sample, n for the persisted
group ⫽ 57 and n for the ended group ⫽ 19. For the subsample, n for the persisted group ⫽ 42 and n for the ended group ⫽ 17.
a group, controlling for the three own satisfaction variables, the
change in R2 was significant, F(3, 246) ⫽ 5.23, p ⫽ .002; the
change in R2 associated with adding the three own satisfaction
variables, controlling for the three perceived partner commitment
variables, was also significant, F(3, 246) ⫽ 5.64, p ⬍ .001. Thus,
perceived partner commitment variables were more predictive of
breakup status than own satisfaction variables in Study 1, and both
sets of variables were predictive in Study 2. We repeated the
analysis in the subsample of Study 2 that was comparable to Study
1. The change in R2 associated with adding perceived partner
commitment variables was significant, F(3, 69) ⫽ 7.62, p ⬍.001,
as was adding own satisfaction variables, F(3, 69) ⫽ 3.45, p ⫽
.021, but the change in R2 was larger from adding the perceived
partner commitment variables (.22 to .41) than from adding satisfaction variables (.32 to .41).
To determine whether each perceived partner commitment variable might have a unique role in predicting breakup status when
controlling for the three own satisfaction variables, we conducted
follow-up analyses, one for each perceived partner commitment
variable. In three analyses of Study 1 data, controlling for the three
own satisfaction variables, breakup status was significantly predicted by the linear trend in perceived partner commitment: overall
model, F(4, 77) ⫽ 6.57, p ⬍ .001 (individual estimate ␤ ⫽ .35,
p ⫽ .011); and by fluctuation in perceived partner commitment:
overall model, F(4, 77) ⫽ 7.19, p ⬍ .001 (␤ ⫽ ⫺.40, p ⫽ .001);
but initial level in perceived partner commitment only approached
significant prediction of breakup status: overall model, F(4, 77) ⫽
5.38, p ⬍ .001 (␤ ⫽ .27, p ⫽ .091). In three analyses of Study 2
data, controlling for the three own satisfaction variables, breakup
status was predicted by the linear trend in perceived partner
commitment: overall model, F(4, 248) ⫽ 20.16, p ⬍ .001 (␤ ⫽
.27, p ⫽ .001); and by fluctuation: overall model, F(4, 248) ⫽
17.33, p ⬍ .001 (␤ ⫽ .16, p ⫽ .046), but not by initial level. Thus,
when controlling for own satisfaction, the linear trend and fluctuation had robust unique associations with breakup status; initial
level did not.
Testing Hypotheses 4
We examined whether fluctuation in perceptions of the partner’s
commitment and fluctuation in one’s own level of commitment
each had a unique association with subsequent breakup status,
when examined simultaneously (Hypothesis 4). In regressing
breakup status onto both variables simultaneously, the overall
model was significant in Study 1, F(2, 79) ⫽ 11.88, p ⬍ .001, and
Study 2, F(2, 250) ⫽ 30.82, p ⬍ .001; fluctuation in perceived
partner commitment provided unique prediction in Study 1 (␤ ⫽
⫺.45, p ⬍ .001) and Study 2 (␤ ⫽ ⫺.23, p ⫽ .002), and fluctuation
in own commitment provided unique prediction in Study 2 (␤ ⫽
–.26, p ⬍ .001) but not Study 1 (␤ ⫽ ⫺.06, ns). The same analysis
in the Study 2 subsample that was comparable with the Study 1
sample replicated Study 1: Fluctuation in perceived partner commitment uniquely predicted breakup status (␤ ⫽ ⫺.29, p ⫽ .023),
but fluctuation in own commitment did not (␤ ⫽ ⫺.21, p ⫽ .103).
Thus, Hypothesis 4 was moderately supported, with strong support
for fluctuation in perceived partner commitment and partial support for fluctuation in own commitment level (only among individuals in established relationships).
We compared the relative predictive value of own commitment
variables versus perceived partner commitment variables in analyses parallel to those for Hypotheses 2 and 3. First, as seen in
Table 5 (rows labeled Own Commitment in top half for Study 1
FLUCTUATIONS IN PERCEIVED PARTNER COMMITMENT
and bottom half for Study 2), each variable tapping own commitment was correlated with breakup status (Simple association column). None of the own commitment variables had unique effects
in Study 1 (the variance inflation factor for initial level of own
commitment was over 5.0), but all had unique effects in Study 2
(Individual Estimate column). In Study 1, when entering the three
perceived partner commitment variables as a group, controlling for
the three own commitment variables, the change in R2 was significant, F(3, 75) ⫽ 4.00, p ⫽ .011, whereas the change in R2
associated with adding the three own commitment variables, controlling for the three perceived partner commitment variables, was
not significant, F(3, 75) ⫽ 1.06, ns. In Study 2, when entering the
three perceived partner commitment variables as a group, controlling for the three own commitment variables, the change in R2 was
significant, F(3, 246) ⫽ 5.21, p ⫽ .002; the change in R2 associated with adding the three own commitment variables, controlling
for the three perceived partner commitment variables, was also
significant, F(3, 246) ⫽ 4.72, p ⬍ .003. Thus, perceived partner
commitment variables were more predictive of breakup status than
own commitment variables in Study 1, but both were predictive in
Study 2. Given this inconsistency, we repeated the analysis in the
Study 2 subsample that was comparable with Study 1; replicating
Study 1, the change in R2 was significant in adding the three
perceived partner commitment variables, F(3, 69) ⫽ 4.06, p ⫽
.010, but not in adding the three own commitment variables, F(3,
69) ⫽ 0.78, ns.
We also conducted follow-up analyses, one for each perceived
partner commitment variable, to determine whether each of these
three variables predicted breakup status beyond the three own
commitment variables. In three analyses of Study 1 data, controlling for the three own commitment variables, breakup status was
significantly predicted by the linear trend in perceived partner
commitment: overall model, F(4, 77) ⫽ 6.60, p ⬍ .001 (individual
estimate ␤ ⫽ .30, p ⫽ .021); and by fluctuation in perceived
partner commitment: overall model, F(4, 77) ⫽ 8.06, p ⬍ .001
(␤ ⫽ ⫺.37, p ⫽ .002); but initial level in perceived partner
commitment only approached significance in prediction of
breakup status: overall model, F(4, 77) ⫽ 6.10, p ⬍ .001 (␤ ⫽ .24,
p ⫽ .051). In three analyses of Study 2 data, controlling for the
three own commitment variables, breakup status was predicted by
the linear trend in perceived partner commitment: overall model,
F(4, 248) ⫽ 19.57, p ⬍ .001 (␤ ⫽ .27, p ⬍ .001); and by
fluctuation in perceived partner commitment: overall model, F(4,
248) ⫽ 17.61, p ⬍ .001 (␤ ⫽ ⫺.21, p ⫽ .007), but not initial level.
Thus, when controlling for own commitment, breakup status is
reliably predicted from the linear trend in perceived partner commitment and fluctuation, but not from initial level.
We conducted several exploratory analyses to further examine
the link between perceived partner commitment over time and own
commitment over time. First, we explored whether own commitment precedes perceptions of partner commitment or vice versa. In
Study 1, the correlation of initial level of own commitment with
the subsequent linear trend in perceived partner commitment,
r(82) ⫽ .59, p ⬍ .001, was higher than the correlation of initial
level of perceived partner commitment with the subsequent trend
in own commitment, r(82) ⫽ .38, p ⬍ .001. In Study 2, these two
correlations were roughly equal: initial own commitment with
trend in perceived partner commitment, r(253) ⫽ .16, p ⬍ .001;
initial perceived partner commitment with trend in own commit-
1059
ment, r(253) ⫽ .17, p ⬍ .001. The Study 2 subsample of individuals in newly formed relationships revealed a pattern similar to
Study 1: initial own commitment with trend in perceived partner
commitment, r(76) ⫽ .39, p ⬍ .001; initial perceived partner
commitment with trend in own commitment, r(76) ⫽ .28, p ⫽
.015.7 With the current correlational design, we attempted to rule
out rather than confirm a possible causal path; neither causal path
could be ruled out (all correlations were significant) rendering both
causal paths possible, but among individuals in newly formed
relationships, there was more support suggesting that own commitment precedes perceptions of the partner’s commitment.
An additional analysis aimed to explore whether one’s own
commitment level at the outset of the study (initial commitment)
constrains the extent of doubt one experiences—that is, whether
one’s initial commitment level moderated the link between fluctuation in perceived partner commitment and breakup. It is possible that people who are not very committed at the outset do not pay
much attention to their partner’s level of commitment and thus are
not affected by the extent that they see their partner as stably
committed. On the other hand, people who are initially highly
committed may seek similar evidence of commitment from their
partner and as such are sensitive to perceived vacillations in their
partner’s commitment; for them, vacillating perceptions may predict later relationship outcomes more than for those who are not
initially committed. We regressed breakup status onto initial level
of own commitment, fluctuation in perceived partner commitment,
and the interaction between these two variables. The interaction
was not significant in either study: Study 1, t(82) ⫽ 0.23, ns; Study
2, t(253) ⫽ ⫺0.17, ns. One’s own initial level of commitment did
not constrain the extent to which fluctuations in perceptions of the
partner’s commitment predicted later breakup status.
Testing Hypothesis 5
Was there evidence that experiencing fluctuations in one’s perceptions of partner commitment has its origins in a dispositional
tendency to be anxiously attached to partner (measured in Study
1)? Consistent with Hypothesis 5, initial level of anxious attachment and subsequent fluctuation in perceived commitment were
positively correlated in Study 1, r(82) ⫽ .39, p ⬍ .001; initial level
of avoidant attachment and subsequent fluctuation in perceived
commitment were also positively correlated, r(82) ⫽ .29, p ⬍
.001. We regressed fluctuations in perceived partner commitment
onto initial anxious attachment style and initial avoidant attachment style simultaneously; initial anxious attachment exhibited a
7
A statistical issue in attempting to interpret these correlations concerns
the correlation between initial level and linear trend within each variable.
The variable (perceived partner commitment vs. own commitment) with a
lower link between initial level and subsequent trend will appear to be a
cause rather than an outcome (Rogosa, 1981). There was little concern in
Study 1, as the correlation between initial level and the linear trend was
higher for own commitment (.82) than for perceived partner commitment
(.54), and yet the cross-correlations (initial level of one variable with linear
trend of the other variable) suggested own commitment as a cause; the
same pattern occurred in the Study 2 subsample. In the full Study 2 sample,
the correlation between initial level and linear trend in own commitment
(.16) was comparable with the correlation between initial level and linear
trend in perceived partner commitment (.20).
1060
ARRIAGA, REED, GOODFRIEND, AND AGNEW
significant unique association with subsequent fluctuation in perceived partner commitment (␤ ⫽ .33, p ⫽ .002), whereas the
association of initial avoidant attachment style only approached
significance (␤ ⫽ .19, p ⫽ .070).
We also conducted exploratory analyses with all of the attachment change estimates. None of the other attachment variables
were correlated with fluctuation in perceived commitment: anxious attachment trend, r(82) ⫽ ⫺.14, p ⫽ .206; anxious attachment
fluctuation, r(82) ⫽ ⫺.11, p ⫽ .343; avoidant attachment trend,
r(82) ⫽ .11, p ⫽ .316; avoidant attachment fluctuation, r(82) ⫽
.00, p ⫽ .999. Also, none of the attachment variables were correlated with later breakup status. Together these findings suggest that
anxiously attached individuals may be predisposed to greater volatility in their perceptions of partner commitment, but this predisposition does not account for the association between fluctuation
in perceptions of partner commitment and later breakup.
Do the Findings Reflect a Measurement Artifact?
A general concern with inferring meaning from temporal profiles generated from self-report responses is that the profiles are
influenced by the response scales. A specific concern is that when
individuals are asked how committed they perceive their partner to
be, those who endorse the highest scale value time after time
would have both high and steady (unchanging) levels of perceived
partner commitment. This would create a ceiling effect in which
initial levels would be high, the slope would be zero or flat (neither
increasing nor decreasing), and there would be no fluctuation. By
virtue of their indicating the highest possible level of perceived
partner commitment, it may be level of perceived partner commitment—not stability per se—that is driving the significant association with breakup status. On the other hand, for those who report
steady levels, but levels that are lower than the absolute highest
level, there would not be a scale ceiling forcing steady levels.
We examined whether support for each hypothesis was robust
when controlling for a possible ceiling effect by eliminating any
individuals who endorsed the highest possible level of perceived
partner commitment at all times. This reduced the sample sizes to
60 participants in Study 1 and 168 participants in Study 2. Despite
the reduction in statistical power, the pattern of significant findings
for each hypothesis test remained the same; there were minor
changes in the specific estimate values (e.g., initial levels of
perceived partner commitment, own satisfaction, and own commitment were slightly lower for this subsample). The results of two
ancillary analyses changed slightly: The simple correlation between fluctuation in own commitment and breakup status became
nonsignificant in Study 1, r(60) ⫽ ⫺.17, ns, and the unique effect
of fluctuation in own satisfaction in predicting breakup, when we
controlled for satisfaction initial level and linear trend, became
nonsignificant in Study 2 (␤ ⫽ ⫺.15, p ⫽ .104).
Descriptively, when categorizing individuals in newly formed
relationships (see bottom half of Tables 6 and 7) into each of the
four temporal profiles, the percentages of ended versus persisted
relationships were similar or slightly more compelling. For example, in comparing the full Study 1 sample with the Study 1
subsample with the ceiling removed, the pattern for Profile 1
remained the same—those whose perceptions of their partner’s
commitment steadily increased over time were overwhelmingly
likely to be in persisting relationships. However, the pattern for
Profile 2 changed; in the full Study 1 sample, the majority of
individuals whose perceptions fluctuated as they increased were
likely to remain in their relationship, but in the subsample with the
ceiling removed, those fitting this profile were equally likely to be
in relationships that ended (50%) versus persisted (50%) despite
perceiving an increasingly committed partner. Other Study 1 comparisons or comparisons for the Study 2 subsample revealed that
removing participants at the ceiling had little effect on the results.
Discussion
Summary of Main Findings and Conclusions
The main findings suggest new avenues for understanding how
certainty versus doubt unfolds over time and how each relates to
later outcomes. Individuals whose perceptions of their partner’s
commitment fluctuated over time were more likely to be in relationships that ended than individuals whose perceptions were
relatively steady. The association of fluctuation in perceived partner commitment with relationship dissolution remained robust
when controlling for fluctuation in one’s own level of commitment, fluctuation in one’s own level of satisfaction, and perceptions of the partner’s positive behavior (consistent with Hypotheses 2, 3, and 4). The link between fluctuation in perceived partner
commitment and relationship dissolution also remained robust
when controlling for initial perceptions of the partner’s commitment among all participants and when controlling for whether
these perceptions increased or decreased over time among individuals in relatively newly formed relationships (consistent with
Hypothesis 1). Among individuals in relationships of longer duration (i.e., longer than 6 months), greater fluctuation in perception
of partner commitment was associated with greater odds of dissolution if those perceptions increased over time but not if they
decreased over time (qualified support for Hypothesis 1).
These findings suggest different conclusions, depending on how
far along a relationship is. Individuals in relatively established
relationships whose perceptions of their partner’s commitment
decrease over time are not affected by the extent of volatility in
their perceptions—things are not good (as inferred from decreasing levels of perceived partner commitment), and perceived ups
and downs bear little on having greater odds of breakup. However,
those whose perceptions of their partner’s commitment increase
over time mirror individuals in relatively novel relationships; they
remain hopeful and thus sensitive to ups and downs in their
perceptions of the partner. The latter findings suggest some counterintuitive (but hypothesis-consistent) conclusions: Even among
individuals who perceived their partner to be increasingly committed, odds of having a relationship that ended increased simply
by virtue of having perceptions that fluctuated over time. Also,
even among individuals who perceived their partner to be decreasingly committed, odds of having a relationship that lasted increased simply by virtue of having perceptions that were steady
over time.
Another difference between individuals in relatively newly
formed relationships versus those in more established relationships
concerned the relative predictive value of perceptions of partner
commitment versus own level of commitment or own level of
satisfaction. In the sample of individuals in relatively established
relationships, all of these variables predicted later relationship
FLUCTUATIONS IN PERCEIVED PARTNER COMMITMENT
persistence versus dissolution. On the other hand, in the two
samples of individuals in more recently initiated relationships,
perceived partner commitment variables accounted for more variance in later relationship status than did own commitment or own
satisfaction. This provides suggestive evidence (albeit not conclusive evidence given the correlational design) that among relatively
novel relationships, perceptions of where the partner stands may
figure more prominently in the ultimate course of the relationship
than a sense of one’s own satisfaction or commitment, whereas in
more established relationships concerns about the partner and
about oneself are similarly implicated.
The current studies also replicated past research focusing on
changes in one’s level of satisfaction or one’s level of commitment. Consistent with Rusbult (1983), participants whose own
level of commitment increased over time were more likely to be in
lasting relationships, and those whose commitment decreased were
more likely to be in relationships that ended. However, when
controlling for level and fluctuation in commitment, this association remained significant only among individuals in established
relationships. Consistent with Arriaga (2001), greater fluctuation
in satisfaction was associated with greater odds of relationship
dissolution, and this association remained robust in two of three
samples (the two Study 2 samples) when controlling for initial
level of satisfaction and the linear trend over time.
We have suggested that the extent of fluctuation in perceptions
of partner commitment reflects varying levels of certainty or doubt
over time about where the partner stands with respect to the
relationship. Not all individuals may experience the same degree
of certainty or doubt over time. Murray et al. (2005) have suggested that each individual may have unique ways of navigating
situations of interdependence with a partner, and they have identified meaningful patterns in individual responses. In a related
vein, we have suggested that each individual exhibits an idiosyncratic temporal pattern in his or her view of a partner’s commitment, and we identified four patterns in individual responses— or
profiles (illustrated in Figure 1)—that reflect distinct and theoretically meaningful experiences of certainty versus doubt.
Origins of Fluctuations in Perceived Partner Commitment
What drives some people to perceive their partner as being
stably committed and others to perceive wavering levels of partner
commitment? There are several possible answers to this question.
One answer that we did not examine in the current research is that
their partners really are stably versus unstably committed. To an
extent, people are accurate in their general perceptions of a partner,
but there is mounting evidence that accurate information is processed in a biased way—for example, individuals may perceive
objective information about partner faults but then integrate it into
a broader impression that downplays the importance of these faults
and sustains a positive perception (cf. Murray, 1999). Because we
did not have partner data, we could not examine this issue directly.
We sought indirect evidence, however, of whether individuals
were relatively accurate in their perceptions of partner commitment. We examined reports of individuals whose relationships
ended and specifically compared own level of commitment and
perceived level of partner commitment among those who initiated
the breakup (i.e., leavers, n ⫽ 19 in Study 1 and n ⫽ 29 in Study
2) versus those whose partners initiated the breakup (i.e., aban-
1061
doned, n ⫽ 9 in Study 1 and n ⫽ 31 in Study 2).8 In general,
leavers made more distinctions between their own level of commitment and their partner’s level in the ways one might expect;
they reported lower initial levels of their own commitment, greater
declines in their own commitment, and more fluctuation over time
in their own commitment, relative to what they reported for the
partner (significantly so for all three variables in Study 2 but only
one variable in Study 1). On the other hand, abandoned individuals
did not make these distinctions; despite ending the relationship, the
partner was not perceived to be less committed initially, less over
time, and less stably. This provides very tentative and indirect
evidence that individuals are cognizant of where their partner
stands relative to them but there is also room for bias, as abandoned individuals were more reluctant or unable to recognize
differences— differences that did not work in their favor—than
were leavers.
We also obtained slightly stronger evidence that one’s own
initial level of commitment predicted subsequent changes in perceptions of the partner’s commitment rather than that initial perceptions of partner commitment predicted subsequent changes in
one’s own commitment. This, too, indirectly suggests that individuals shape their perceptions of the partner’s commitment in light of
their own commitment (a biased process), more than they use what
they believe to be accurate perceptions of the partner’s commitment in adjusting their own commitment.
It is also possible that individuals vary in their propensity to
perceive a partner as stably or not stably committed—that is, the
fluctuation pattern may have some origins in individual differences. Individuals who were predisposed to be anxiously attached
to the partner were more likely to have perceptions of the partner’s
commitment that fluctuated over time (as stated in Hypothesis 5).
This is consistent with research by Campbell et al. (2005), who
showed that anxiously attached individuals (more than securely
attached or avoidant individuals) were highly reactive to daily
interactions with the partner, whereby their impressions of the
partner were closely tied to daily levels of conflict. Similarly,
fluctuation in perceived partner commitment may reflect a predisposition to be highly reactive to partner interactions, stemming
from an anxious attachment style. However, the correlation between fluctuation and anxious attachment was moderate rather
than high, suggesting that fluctuation in perceived partner commitment stems from more than only an anxious attachment style.
It is unlikely that perceiving a partner as being stably or unstably
committed reflects individual differences in response sets. If the
8
The main analyses did not differentiate between leavers and abandoned
for several reasons. First, the pattern of results was the same with or
without the distinction; that is, when compared against people whose
relationships persisted, the differences between leavers and abandoned
were relatively small and not significant. We collapsed these two groups to
keep the results relatively straightforward. Second, the number of individuals in the abandoned group was too small from which to draw firm
conclusions. Third, theoretically, there are competing hypotheses of how
being more committed than the partner might affect one’s breakup status;
it is possible that a less committed partner would be more likely to break
up, but it is also possible that the more committed person would preemptively break up so as to avoid being hurt by the partner. Much remains to
be learned about the implications of disparities between own and partner
commitment—issues that were beyond the current goals.
1062
ARRIAGA, REED, GOODFRIEND, AND AGNEW
extent of fluctuation in perceived partner commitment simply
reflected an individual tendency to respond to survey items in
unstable ways (cf. Gable & Nezlek, 1998), we would have obtained evidence of an unstable response set across several variables—that is, all fluctuation variables would have been highly
correlated. On the contrary, fluctuation in perceived partner commitment was not correlated with all fluctuation variables (e.g.,
fluctuation in attachment variables), but it was correlated with
fluctuation in variables in which the pattern of fluctuation could be
said to reflect doubt (e.g., fluctuation in own satisfaction). Thus,
the theoretically coherent pattern of correlations was unlikely to be
caused by an artifact of using self-report response scales.
In a similar vein, it is unlikely that an individual’s pattern of
fluctuation was caused by an individual’s deliberate effort to
respond a particular way, a self-report bias that can undermine the
validity of self-report measures. The change estimates examined in
the current research were not themselves self-report variables; they
were derived from prospectively observed patterns in each individual’s self-reported data. As such, they are distinct from retrospective recollections of change over time, which are susceptible
to self-report bias (McFarland & Ross, 1987). If prospective reports of change over time (as obtained in this research) are similarly susceptible to self-report bias, they should reveal a data
pattern similar to retrospective reports of change over time; these
two methods when compared directly have yielded distinct data
patterns, suggesting that prospective methods are less susceptible
to self-report biases (Karney & Frye, 2002).
Finally, we explored whether a stable pattern in perceived
partner commitment might have its origins in another response
bias, namely using the top or ceiling of the scale. Individuals who
endorsed the highest possible level of perceived partner commitment time after time would have high and steady levels of perceived partner commitment, making it impossible to determine
whether it is level or stability of perceived commitment that is
driving the association with later breakup status. When we eliminated individuals who time after time consistently endorsed the
highest possible rating of perceived partner commitment, the findings remained the same. Thus, a ceiling effect in ratings of perceived partner commitment did not account for the association
between fluctuation in perceived partner commitment and breakup.
Broader Implications
Are these findings compelling, or obvious and trivial? From a
common sense standpoint, it could seem trivial to report that
people who vacillate in their perception of their partner’s commitment are more likely to be in relationships that end than people
who believe they know where their partner stands and thus vacillate less. The group whose perceptions vacillate experience more
doubt, and all things being equal, it seems self-evident that doubt
could hurt a relationship. This finding is less trivial when one
considers what might have been the case but was not the case.
First, the key variables predicting breakup among novel relationships were not the same as those among more established
relationships. Yet, it is important to advance precise knowledge on
the causal factors that operate at different relationship stages.
When faced with a partner who seems to be decreasingly committed over time (a decreasing trend of perceived commitment),
individuals may react differently depending on the stage of their
relationship. As was shown, unstable perceptions of a decreasingly
committed partner (i.e., high fluctuation) disrupted relatively novel
relationships (negative link between fluctuation in perceptions and
persistence), suggesting it is better to end the relationship than
persist in a state of uncertainty. Fluctuating perceptions were no
different than stable perceptions among individuals in more established dating relationships (no link between fluctuation and persistence)—a decreasing trend of perceived partner commitment
was associated with dissolution regardless of how stable the perceptions were. It is conceivable that among marital relationships or
other relationships that are difficult to end, fluctuating perceptions
of a decreasingly committed partner might be associated with
persistence (a positive link between fluctuation and persistence); if
one feels the relationship must continue no matter what, but the
partner is decreasingly committed, perceiving an occasional commitment increase in an otherwise uncommitted partner may lead to
hope that the partner’s commitment will improve. This remains to
be examined in future research.
Second, and more importantly, common sense would suggest
that what should really matter in predicting a breakup is whether
one perceives the partner to be committed or not—if a partner is
not likely to be there through thick and thin, one would likely leave
the relationship rather than remain vulnerable to being left by the
partner. The common sense view would stop there. However, this
view is overly simplistic given that the amount of fluctuation in
perceptions of a partner’s commitment makes a difference as much
or more than generally seeing a partner as committed or not
committed (Surra et al., 1999). It is not self-evident that people
who are increasingly convinced that their partner wants to leave
(declining partner commitment) would be likely to stay in the
relationship so long as their partner’s lack of commitment is
predictable (i.e., steadily declining). Even less obvious is the
finding that after taking into account changes in perceived partner
commitment (the trend and extent of fluctuation), one’s initial
level of commitment simply did not predict the fate of the relationship. Changes one sees in the partner’s level of commitment
are more important than simply seeing the partner as being committed or not so committed.
Third, common sense might also suggest that perceptions of the
partner should matter only to people who are highly committed
(high initial level of own commitment). Why should those who are
relatively uncommitted care about the partner’s commitment? We
did not find this to be the case, as one’s own initial level of
commitment did not affect (i.e., moderate) the link between steady
perceptions of partner commitment and persistence.
The theoretical message of this research resonates with previous
research on perceived partner regard (e.g., Murray et al., 2003):
When a person has doubts about a partner’s feelings about the
relationship— even when one generally believes that the partner is
generally committed but doubts the consistency of the partner’s
commitment—it becomes difficult to assume that all will turn out
well, particularly when the relationship is not yet well established.
More generally, perceptions of where the partner stands predict the
course of a relationship, particularly in relatively new relationships
(6 months or less). We are not suggesting that a partner’s actual
level of commitment is irrelevant; instead we are underscoring the
importance of perceptions and attributions in directing the course
of relationships (Bradbury & Fincham, 1990; Kelley, 1979). Ultimately, the current research advances research on certainty and
FLUCTUATIONS IN PERCEIVED PARTNER COMMITMENT
doubt in relationships by demonstrating the importance of invariances over time in perceptions of partner motives (see Holmes,
2004).
There are also several methodological implications of this research. One is that multiple measurement occasions afford an
analysis of specific changes over time better than fewer measurement occasions. Designs that attempt to establish a longitudinal
pattern from only two measurement occasions are relegated to
averaging (or correlating) reports over the two time periods (Arriaga, 2001; Karney & Bradbury, 1995). Yet as suggested in
Figure 1, individuals can exhibit the same initial level of perceived
partner commitment, the same mean level over time, and even
similar linear trend over time, and yet have vastly different experiences of certainty versus doubt on the basis of the extent of
fluctuation in their ratings (e.g., Profile 3 vs. Profile 4). As such,
examining fluctuations in meaningful variables not only increases
the pragmatic aim of predicting later relationship outcomes but
also affords more precision in deriving theoretically meaningful
data patterns.
A second methodological implication in studies with multiple
measurement occasions is that checks must be in place to ensure
the results are not attributed to participant response sets. When
participants answer different questions in the same way within a
given measurement occasion, this will artificially inflate the correlations among variables. As Gable and Nezlek (1998) suggested,
it is important to demonstrate divergent validity (as was the case in
this research) by showing that an outcome is associated with
fluctuation in theoretically relevant variables only and not all
variables. Similarly, when participants answer certain questions in
the same way across several measurement occasions (e.g., endorsing the highest possible rating every time), this too creates a
response set that confounds level and variation over time in a
variable. One solution is to eliminate these participants from the
analysis, but this might eliminate a true data pattern. In the current
research, we repeated the analyses excluding participants at the
ceiling of the scale, which eliminated those who truly believed that
their partner was consistently and whole-heartedly committed.
Some studies have attempted to get around this problem by changing the scale endpoints (e.g., “My partner is more committed to
this relationship than anyone in the world will ever be”), but even
this strategy has not been successful in eliminating ceiling or floor
effects (see Arriaga, 2001, for similarly extreme measures). The
best approach might be to use scales with extreme endpoints and
perform analyses on a full sample versus a sample excluding
participants with possible response sets, to test whether the findings possibly reflect a response artifact.
A third methodological issue concerns the lag between measurement occasions. This is less of a concern when examining the
linear slope in that it summarizes changes over time, but more of
a concern when attempting to capture specific patterns of change
(e.g., ups and downs, curvilinear), in which long time lags may
miss the relevant changes. In Study 2 (4-week time lag), the effect
sizes were larger in magnitude for the linear trend in perceived
partner commitment than for fluctuation in perceived partner commitment; in contrast, in Study 1 (1-week time lag) the effect sizes
were similar or slightly larger for fluctuation.9 This suggests that
4-week time lags may gloss over dynamic changes in perceptions
of the partner whereas 1-week time lags may be more appropriate
for this particular type of variable. In the absence of clear guide-
1063
lines, the appropriate time lag to capture specific changes is likely
to depend on the construct of interest; constructs that are more
susceptible to change may require shorter time lags, and those that
are less susceptible may be captured adequately with longer time
lags.
Caveats
The research reported here has some limitations. The most
salient limitation is the focus on individual data patterns. We
demonstrated that one can predict later relationship outcomes from
examining an individual’s ratings over time. Ultimately, however,
developing sound theories of relationship processes will require
understanding partner influences that occur in addition to, and
independently of, individual processes (Kenny, 2006). Another
limitation was that both samples were derived from college students in the Midwestern region of the United States, and one
cannot assume these findings generalize to samples of other college students, other young adults, relatively new dating relationships of older adults with longer relationship histories (e.g., those
who are divorced), or samples of people from different cultures.
Moreover, these are correlational data. Although we have attempted to control for artifacts and discriminate variables that do
versus do not account for the link between perceived partner
commitment and later breakup status, it is still possible that other
variables not measured in this research are causing the observed
changes in both perceived partner commitment and later breakup
status. Finally, we established that different temporal profiles are
predictive of later relationship status but did not provide conclusive evidence on the origins of distinct profiles. We encourage
future research that attempts to overcome these limitations.
9
We examined the effect sizes in the subsample of Study 2 that was
comparable with Study 1 (participants in relationships of similar duration,
similar sample size, differing only in the time lag). When we regressed
breakup onto initial level, linear trend, and fluctuation in perceived partner
commitment simultaneously, the effect size for linear trend (␤ ⫽ .41) was
larger than the effect size for fluctuation (␤ ⫽ ⫺.33); initial level was not
significant. In analyses controlling for the three satisfaction variables, the
effect of adding the linear trend in perceived partner commitment (␤ ⫽ .44)
exceeded the effect of adding fluctuation (␤ ⫽ ⫺.22); adding initial level
was not significant. The same occurred when controlling for the three own
commitment variables (linear trend ␤ ⫽ .43, fluctuation ␤ ⫽ ⫺.25, initial
level was not significant).
References
Agnew, C. R., Loving, T. J., & Drigotas, S. M. (2001). Substituting the
forest for the trees: Social networks and the prediction of romantic
relationship state and fate. Journal of Personality and Social Psychology, 81, 1042–1057.
Agnew, C. R., Van Lange, P. A. M., Rusbult, C. E., & Langston, C. A.
(1998). Cognitive interdependence: Commitment and the mental representation of close relationships. Journal of Personality and Social Psychology, 74, 939 –954.
Ajzen, I., & Fishbein, M. (1977). Attitude– behavior relations: A theoretical analysis and review of empirical research. Psychological Bulletin,
84, 888 –918.
Aron, A., Aron, E. N., & Smollan, D. (1992). Inclusion of other in the self
1064
ARRIAGA, REED, GOODFRIEND, AND AGNEW
scale and the structure of interpersonal closeness. Journal of Personality
and Social Psychology, 63, 596 – 612.
Aronson, J., & Inzlicht, M. (2004). The ups and downs of attributional
ambiguity: Stereotype vulnerability and the academic self-knowledge of
African American college students. Psychological Science, 15, 829 –
836.
Arriaga, X. B. (2001). The ups and downs of dating: Fluctuations in
satisfaction in newly formed romantic relationships. Journal of Personality and Social Psychology, 80, 754 –765.
Berscheid, E. (1983). Emotion. In H. H. Kelley et al. (Eds.), Close
relationships (pp. 110 –168). New York: Freeman.
Bolger, N., Zuckerman, A., & Kessler, R. C. (2000). Invisible support and
adjustment to stress. Journal of Personality and Social Psychology, 79,
953–961.
Bradbury, T. N., & Fincham, F. D. (1990). Attributions in marriage:
Review and critique. Psychological Bulletin, 107, 3–23.
Brickman, P. (1987). Commitment, conflict, and caring. Englewood Cliffs,
NJ: Prentice-Hall.
Campbell, L., Simpson, J. A., Boldry, J., & Kashy, D. A. (2005). Perceptions of conflict and support in romantic relationships: The role of
attachment anxiety. Journal of Personality and Social Psychology, 88,
510 –531.
Crocker, J., Karpinski, A., Quinn, D. M., & Chase, S. K. (2003). When
grades determine self-worth: Consequences of contingent self-worth for
male and female engineering and psychology majors. Journal of Personality and Social Psychology, 85, 507–516.
Drigotas, S. M., Rusbult, C. E., & Verette, J. (1999). Level of commitment,
mutuality of commitment, and couple well-being. Personal Relationships, 6, 389 – 409.
Driscoll, D. M., Hamilton, D. L., & Sorrentino, R. M. (1991). Uncertainty
orientation and recall of person-descriptive information. Personality and
Social Psychology Bulletin, 17, 494 –500.
Etcheverry, P. E., & Agnew, C. R. (2004). Subjective norms and the
prediction of romantic relationship state and fate. Personal Relationships, 11, 409 – 428.
Fraley, R. C. (2002). Attachment stability from infancy to adulthood:
Meta-analysis and dynamic modeling of developmental mechanisms.
Personality and Social Psychology Review, 6, 123–151.
Gable, S. L., & Nezlek, J. B. (1998). Level and instability of day-to-day
psychological well-being and risk for depression. Journal of Personality
and Social Psychology, 74, 129 –138.
Gable, S. L., & Reis, H. T. (1999). Now and then, them and us, this and
that: Studying relationships across time, partner, context, and person.
Personal Relationships, 6, 415– 432.
Holmes, J. G. (1981). The exchange process in close relationships: Microbehavior and macromotives. In M. L. Lerner & S. C. Lerner (Eds.),
The justice motive in social behavior (pp. 261–284). New York: Plenum.
Holmes, J. G. (2004). The benefits of abstract functional analysis in theory
construction: The case of interdependence theory. Personality and Social Psychology Review, 8, 146 –155.
Holmes, J. G., & Rempel, J. K. (1989). Trust in close relationships. In C.
Hendrick (Ed.), Close relationships: Vol. 10. Review of personality and
social psychology (pp. 187–220). Thousand Oaks, CA: Sage.
Jacobson, N. S., Follette, W. C., & McDonald, D. W. (1982). Reactivity to
positive and negative behavior in distressed and nondistressed married
couples. Journal of Consulting and Clinical Psychology, 50, 706 –714.
Karney, B. R., & Bradbury, T. N. (1995). Assessing longitudinal change in
marriage: An introduction to the analysis of growth curves. Journal of
Marriage and the Family, 57, 1091–1108.
Karney, B. R., & Bradbury, T. N. (1997). Neuroticism, marital interaction,
and the trajectory of marital satisfaction. Journal of Personality and
Social Psychology, 72, 1075–1092.
Karney, B. R., & Bradbury, T. N. (2000). Attributions in marriage: State or
trait? A growth curve analysis. Journal of Personality and Social Psychology, 78, 295–309.
Karney, B. R., & Frye, N. E. (2002). “But we’ve been getting better lately”:
Comparing prospective and retrospective views of relationship development. Journal of Personality and Social Psychology, 82, 222–238.
Kashy, D. A., & Kenny, D. A. (2000). The analysis of data from dyads and
groups. In H. T. Reis & C. M. Judd (Eds.), Handbook of research
methods in social and personality psychology (pp. 451– 477). Cambridge, England: Cambridge University Press.
Kelley, H. H. (1979). Personal relationships: Their structures and processes. Hillsdale, NJ: Erlbaum.
Kelley, H. H. (1983). Love and commitment. In H. H. Kelley et al. (Eds.),
Close relationships (pp. 265–314). New York: Freeman.
Kenny, D. A. (2006). The partner (and the participant) in social psychology. Invited address presented at the Seventh Annual Meeting of the
Society for Personality and Social Psychology, Palm Springs, CA.
Le, B., & Agnew, C. R. (2003). Commitment and its theorized determinants: A meta-analysis of the investment model. Personal Relationships,
10, 37–57.
Lehmiller, J. J., & Agnew, C. R. (2006). Marginalized relationships: The
impact of social disapproval on romantic relationship commitment.
Personality and Social Psychology Bulletin, 32, 40 –51.
McFarland, C., & Ross, M. (1987). The relation between current impressions and memories of self and dating partners. Personality and Social
Psychology Bulletin, 13, 228 –238.
Miller, P. J. E., & Rempel, J. K. (2004). Trust and partner-enhancing
attributions in close relationships. Personality and Social Psychology
Bulletin, 30, 695–705.
Murray, S. L. (1999). The quest for conviction: Motivated cognition in
romantic relationships. Psychological Inquiry, 10, 23–34.
Murray, S. L. (2005). Regulating the risks of closeness: A relationshipspecific sense of felt security. Current Directions in Psychological
Science, 14, 74 –78.
Murray, S. L., Bellavia, G., Rose, P., & Griffin, D. W. (2003). Once hurt,
twice hurtful: How perceived regard regulates daily marital interactions.
Journal of Personality and Social Psychology, 84, 126 –147.
Murray, S. L., & Holmes, J. G. (1999). The (mental) ties that bind:
Cognitive structures that predict relationship resilience. Journal of Personality and Social Psychology, 77, 1228 –1244.
Murray, S. L., Holmes, J. G., & Collins, N. L. (2006). Optimizing assurance: The risk regulation system in relationships. Psychological Bulletin,
132, 641– 666.
Murray, S. L., Holmes, J. G., & Griffin, D. W. (2000). Self-esteem and the
quest for felt security: How perceived regard regulates attachment processes. Journal of Personality and Social Psychology, 78, 478 – 498.
Rempel, J. K., Holmes, J. G., & Zanna, M. P. (1985). Trust in close
relationships. Journal of Personality and Social Psychology, 49, 95–112.
Robins, R. W., Hendin, H. M., & Trzesniewski, K. H. (2001). Measuring
global self-esteem: Construct validation of a single-item measure and the
Rosenberg Self-Esteem Scale. Personality and Social Psychology Bulletin, 27, 151–161.
Rogosa, D. (1981). A critique of cross-lagged correlation. Psychological
Bulletin, 88, 245–255.
Rogosa, D., Brant, D., & Zimowksi, M. (1982). A growth curve approach
to the measurement of change. Psychological Bulletin, 92, 726 –748.
Rusbult, C. E. (1983). A longitudinal test of the investment model: The
development (and deterioration) of satisfaction and commitment in
heterosexual involvements. Journal of Personality and Social Psychology, 45, 101–117.
Rusbult, C. E., Martz, J. M., & Agnew, C. R. (1998). The Investment
Model Scale: Measuring commitment level, satisfaction level, quality of
alternatives, and investment size. Personal Relationships, 5, 357–391.
FLUCTUATIONS IN PERCEIVED PARTNER COMMITMENT
Rusbult, C. E., Olsen, N., Davis, J. L., & Hannon, P. A. (2001). Commitment and relationship maintenance mechanisms. In J. Harvey & A.
Wenzel (Eds.), Close romantic relationships: Maintenance and enhancement (pp. 87–113). Mahwah, NJ: Erlbaum.
Sanbonmatsu, D. M., Posavac, S. S., Vanous, S., & Ho, E. A. (2005).
Information search in the testing of quantified hypotheses: How “all,”
“most,” “some,” “few,” and “none” hypotheses are tested. Personality
and Social Psychology Bulletin, 31, 254 –266.
SAS Institute. (1992). SAS Technical Report P-229, SAS/STAT Software:
Changes and Enhancements. Cary, NC: Author.
Shoda, Y. (1999). Behavioral expression of a personality system: Generation
and perception of behavioral signatures. In D. Cervone & Y. Shoda (Eds.),
The coherence of personality: Social-cognitive bases of consistency, variability, and organization (pp. 155–181). New York: Guildford Press.
Shoda, Y., Mischel, W., & Wright, J. C. (1994). Intraindividual stability in
the organization and patterning of behavior: Incorporating psychological
situations into the idiographic analysis of personality. Journal of Personality and Social Psychology, 67, 674 – 687.
Simpson, J. A., & Rholes, W. S. (Eds.). (1998). Attachment theory and
close relationships. New York: Guilford Press.
Simpson, J. A., Rholes, W. S., & Phillips, D. (1996). Conflict in close
1065
relationships: An attachment perspective. Journal of Personality and
Social Psychology, 71, 899 –914.
Singer, J. D. (1998). Using SAS PROC MIXED to fit multilevel models,
hierarchical models, and individual growth models. Journal of Educational and Behavioral Statistics, 24, 323–355.
Sorrentino, R. M., Holmes, J. G., Hanna, S. E., & Sharp, A. (1995).
Uncertainty orientation and trust in close relationships: Individual differences in cognitive styles. Journal of Personality and Social Psychology, 68, 314 –327.
Surra, C. A., & Hughes, D. K. (1997). Commitment processes in accounts
of the development of premarital relationships. Journal of Marriage and
the Family, 59, 5–21.
Surra, C. A., Hughes, D. K., & Jacquet, S. E. (1999). The development of
commitment to marriage: A phenomenological approach. In J. M. Adams & W. H. Jones (Eds.), Handbook of interpersonal commitment and
relationship stability (pp. 125–148). New York: Kluwer Academic/
Plenum Publishers.
Wieselquist, J., Rusbult, C. E., Foster, C. A., & Agnew, C. R. (1999).
Commitment, pro-relationship behavior, and trust in close relationships.
Journal of Personality and Social Psychology, 77, 942–966.
Appendix
Using SAS to Analyze Within-Person Variation in a Single Variable
SAS provides a way of analyzing (Level 2) group differences in the
magnitude of within-person variation (rij) in a Level 1 variable while
controlling for the intercept and slope. We wanted to assess whether the
two breakup groups (ended vs. persisted) differed in the amount of withinperson fluctuation over time. This can be achieved by comparing two
models, one in which the two breakup groups are constrained to have the
same average within-person variation versus one in which the average
within-person variation can vary between the two breakup groups (Niall
Bolger, personal communication, June 4, 2004). As an example, one would
examine an initial model in which the within-person variation in perceived
partner commitment is constrained to be equal for individuals in ended and
persisted relationships by using the following SAS code, where id is a
subject identifier, time is the time period, ppc is perceived partner commitment measured at each time, and breakup is a two-level variable for the
ended versus persisted groups:
proc mixed method ⫽ ml;
class id breakup;
model ppc ⫽ time breakup time*breakup;
random intercept time/subject ⫽ id type ⫽ un grp ⫽ breakup;
run;
The following code would test a second model in which the magnitude of
within-person variation for one breakup group (ended) could vary from that
of the other breakup group (persisted):
proc mixed method ⫽ ml;
class id breakup;
model ppc ⫽ time breakup time*breakup;
random intercept time/subject ⫽ id type ⫽ un grp ⫽ breakup;
repeated/subject ⫽ id grp ⫽ breakup;
run;
The repeated statement elicits an analysis of within-person variance for
each participant (subject ⫽ id) and compares the two breakup groups
(grp ⫽ breakup) in the extent of within-person variance. Each model would
yield a chi-square value; the two chi-square values are compared with a
chi-square difference test. A significant reduction in the chi-square value
would indicate that the two groups differed in the amount of fluctuation
(i.e., within-person variation). Also, in both models, the grp ⫽ breakup
command in the random statement allows Level 2 residuals of the random
effects (␶00 and ␶11 in the Tau or random effects covariance matrix) to vary
by breakup group; this translates into a more conservative test of group
differences in within-person variation, as tested in the second model.
As stated in the text, the limitation of this model comparison approach
is that one can compare within-person variation in only one variable,
without simultaneously controlling for or analyzing within-person variation in other Level 1 variables.
Received November 22, 2005
Revision received May 11, 2006
Accepted May 14, 2006 䡲