Professional Documents
Culture Documents
Collins 201996
Collins 201996
net/publication/14317544
CITATIONS READS
775 4,269
1 author:
Nancy L. Collins
University of California, Santa Barbara
68 PUBLICATIONS 13,429 CITATIONS
SEE PROFILE
All content following this page was uploaded by Nancy L. Collins on 20 May 2014.
Two studies examined attachment style differences in social perception. In Study 1, participants
wrote open-ended explanations for hypothetical relationship events and described how they would
feel and behave in response to each event. Compared with secure participants, preoccupied partici-
pants explained events in more negative ways; they also reported more emotional distress and be-
haviors that were likely to lead to conflict. Avoidant participants also provided negative explanations,
but did not report emotional distress. Path analysis indicated that attachment style differences in
behavior were mediated by explanation patterns and emotional distress. Study 2 was designed to
replicate Study 1 and test the relative importance of attachment style and relationship quality to
predicting each outcome. Results indicated that both variables were significant predictors of expla-
nations, but only attachment style predicted emotional responses. Thesefindingsare consistent with
the idea that adults with different working models of attachment are predisposed to think, feel, and
behave differently in their relationships.
Every situation we meet with in life is constructed in terms of the ment theory, these cognitive, representational models begin to
representational models we have of the world about us and of our- develop in the context of parent-child interactions and are then
selves. Information reaching us through our sense organs is se- carried forward into new relationships, where they guide how
lected and interpreted in terms of those models, its significance for individuals manage their relationships and how they construe
us and for those we care for is evaluated in terms of them, and plans
their social world (Collins & Read, 1994; Shaver, Collins, &
of action conceived and executed with those models in mind. On
how we iaterpret and evaluate each situation, moreover, turns also Clark, 1996). However, the specific processes through which
how we feel. (Bowlby, 1980, p. 229) these models operate remain poorly understood. The purpose
of the current research was to explore these processes by exam-
As individuals enter new relationships, they carry with them ining attachment style differences in the perception of social
a history of personal and interpersonal experiences that shape events. Specifically, two studies tested the hypothesis that adults
how they think and feel about their relationships and how they with different working models of attachment would interpret
behave in those relationships. Recently, attachment theory has and explain relationship events in ways consistent with their ex-
been used as a framework for understanding the specific pro- isting models of self and others. An additional aim was to ex-
cesses through which close relationships in adulthood are in- amine a broader model of social functioning that included cog-
fluenced by each partner's personal and interpersonal history. nition as well as affect and behavior. A model linking attach-
Central to this approach is the notion of working models of self ment style and explanation patterns to emotional responses and
and others (Bartholomew & Horowitz, 1991; Collins & Read, behavioral intentions was proposed and tested.
1990; Main, Kaplan, & Cassidy, 1985). According to attach-
Working Models of Attachment
This article is based on a doctoral dissertation submitted to the Uni- Bowlby(1973,1980, 1982) used the term working models to
versity of Southern California. Special thanks to committee members describe the internal representations that individuals develop of
Steve Read, Norman Miller, Shelley Duval, David Walsh, and Margaret the world and of significant people within it, including the self.
McLaughlin for their valuable insights and suggestions. 1 am deeply in- These representations take root in infancy and early childhood
debted to David Jones, Heidi Lincer, and Michael Purvis for serving as and are presumed to be largely determined by the caregiver's
skillful coders for this project. Finally, I wish to thank Steve Read and emotional availability and responsiveness to the child's needs
Sarah Ullman for helpful comments on previous versions of this article. (Bretherton, 1985; Sroufe& Waters, 1977). According to Main
Funding for this project was provided by a research award from the et al. (1985), a child's early working models are composed of
Science Directorate of the American Psychological Association. The au- schemata that represent his or her attempts to gain comfort and
thor was supported by a fellowship from the John Randolf and Dora security. Over time, these specific experiences become ab-
Haynes Foundation during the conduct of this research, and by Na-
tional Institute of Mental Health postdoctoral fellowship MH15750 stracted into more generalized beliefs and expectations about
during the preparation of this article. the warmth and responsiveness of others and about the worthi-
Correspondence concerning this article should be addressed to ness of the self. Once developed, these beliefs can be used to
Nancy L. Collins, Department of Psychology, State University of New predict and interpret the behavior of others and to act in new
York, Buffalo, New York 14260-4110. Electronic mail may be sent via situations without evaluating each onefromthe beginning.
Internet to [email protected]. Although each child's experiences, and hence working
810
WORKING MODELS OF ATTACHMENT 811
models, are unique, certain regularities have been observed in social situations, and more self-assertive. Secure adults also had
the nature and quality of infant-caretaker relations. On the ba- more positive beliefs about the social world, viewing others as
sis of structured laboratory procedures and home observations, trustworthy, dependable, and altruistic. Those with a more pre-
three distinct behavioral patterns or styles of infant attachment occupied attachment style (which corresponds to the anxious-
have been identified: secure, avoidant, and anxious-ambivalent ambivalent style in infants) measured low in self-worth, social
(Ainsworth, Blehar, Waters, & Wall, 1978). These styles are self-confidence, and assertiveness. Their view of human nature
closely associated with differences in caretaker warmth and re- was also somewhat negative. They believed that people have lit-
sponsiveness, and are thought to reflect differences in the psy- tle control over their lives, and they viewed others as complex
chological organization of the attachment system (for reviews of and difficult to understand. Finally, avoidant adults tended to
this literature, see Bretherton, 1985; Rothbard& Shaver, 1994). have positive views of themselves, indicated by high self-worth
Indeed, Main et al. (1985) suggested that the three attachment and assertiveness, although they viewed themselves as less con-
styles are best understood as "terms referring to particular types fident in social situations and not interpersonally oriented.
of internal working models of relationships, models that direct These adults also had largely negative views of human nature.
not only feelings and behavior but also attention, memory, and Overall, they thought others were not trustworthy or
cognition" (p. 67). dependable.
Of course, representations of self and others continue to Evidence for attachment-style differences in working models
evolve as individuals encounter new relationships throughout is not limited to self-report studies. For example, using a re-
their lives. Nevertheless, attachment theory suggests that cogni- sponse latency paradigm drawn from cognitive social psychol-
tive models that begin their development early in one's personal ogy, Baldwin, Fehr, Keedian, and Seidel (1993) reasoned that
history are likely to remain influential (Collins & Read, 1994). individuals with different attachment styles should be more
On the basis of these ideas, social and clinical psychologists have quick to recognize words that are congruent with their rela-
begun to explore attachment theory as a framework for under- tional schemas. As predicted, when presented with words in a
standing social functioning in adulthood, particularly in the relational context, secure adults were quicker to recognize pos-
context of adult romantic relationships (for reviews, see Bar-
itive words, whereas avoidant adults responded more quickly to
tholomew, 1993; Shaver & Hazan, 1993). Although the re-
negative words. More recently, Mikulincer (1995) found that
search is still in its early stages, three broad conclusions can be
adolescents with different attachment styles differed not only in
drawn from the current literature on adult attachment. First,
the content but also in the structure of their mental representa-
the styles of attachment observed in children appear useful for
tions of their self. For example, compared with insecure adoles-
describing individual differences in adult styles of relating. Al-
cents, secure individuals were found to have more balanced,
though there are a number of unresolved conceptual and mea-
surement issues, there is growing consensus on the prototypic complex, and coherent self-structures.
features of each style (Collins & Read, 1990; Feeney, Noller, & The third general conclusion that can be drawn from the
Hanrahan, 1994; Griffin & Bartholomew, 1994; Hazan & adult attachment literature is that adults with different attach-
Shaver, 1987, 1993). Secure adults describe themselves as com- ment styles differ markedly in the quality of their love relation-
fortable with closeness and intimacy, as willing to rely on others ships (e.g., Brennan & Shaver, 1992, 1995; Collins & Read,
when needed, and as confident that they are loved and valued. 1990; Feeney & Noller, 1990; Hazan & Shaver, 1987; Kirkpat-
In contrast, avoidant adults report being very uncomfortable rick & Davis, 1994; Levy & Davis, 1988; Senchak & Leonard,
getting close to and depending on others, and they tend to be 1992; Simpson, 1990). Studies consistently find that adults
unconcerned about whether others will accept or reject them. with more secure working models report more positive relation-
Finally, preoccupied adults have a strong desire for close rela- ship experiences than those with preoccupied or avoidant
tionships, although they are unsure about whether they can de- models. Secure individuals generally characterize their relation-
pend on others and tend to worry a great deal about being re- ships as intimate, stable, and satisfying. In contrast, avoidant
jected and abandoned. These patterns, measured through self- adults tend to report low levels of intimacy, commitment, and
report and interview methodologies, appear fairly stable over satisfaction, whereas preoccupied individuals report jealousy,
time (Collins & Read, 1990; Kirkpatrick & Davis, 1994; conflict, and high levels of negative emotional experiences.
Scharfe & Bartholomew, 1994) and are associated with memo- Once again, it is noteworthy that these differences are not lim-
ries of relationships with parents in theoretically predictable ited to self-report questionnaire studies. Recent studies have
ways (Collins & Read, 1990; Hazan & Shaver, 1987; Rothbard found differences using behavioral observations (e.g., Feeney,
& Shaver, 1994). Noller, & Callan, 1994; Kobak & Hazan, 1991; Simpson,
The second general conclusion that can be drawn from this Rholes, & Nelligan, 1992), content coding of relationship nar-
literature is that differences in attachment style appear to be ratives (Feeney & Noller, 1991), and even psychophysiological
rooted in cognitive models of self and others (Collins & Read, methods (Feeney & Kirkpatrick, 1996).
1994). Support for this assumption is provided by several stud- Although the correlational nature of these studies does not
ies showing that adults with different attachment styles differ permit us to draw firm conclusions about causality, the underly-
greatly in the way they view themselves and the social world ing assumption throughout this work is that attachment style
(Carnelley & Janoff-Bulman, 1992; Collins & Read, 1990; directly contributes to relationship quality and does not merely
Feeney & Noller, 1990; Hazan & Shaver, 1987). For example, reflect it. Individual differences in working models are assumed
Collins and Read (1990) found that people with a more secure to play an important role by shaping partners' cognitive, emo-
attachment style were higher in self-worth, more confident in tional, and behavioral response patterns. However, the specific
812 COLLINS
Cognitive Response
1990, for a review). Individuals in distressed relationships tend
to make attributions that maximize the impact of their part-
Working model of ner's negative behavior while discounting or minimizing the im-
attachment activated
in memory 1 I Behavioral Response
pact of positive behaviors. Moreover, several studies suggest that
these patterns do not simply reflect relationship quality, but
Emotional Response
may actively contribute to it (Baucom & Lester, 1986; Epstein,
Pretzer,& Fleming, 1987; Fincham & Bradbury, 1987; Fletcher,
Figure 1. Hypothesized model linking working models of attachment Fincham, Cramer, & Heron, 1987; Margolin & Weiss, 1978).
to cognitive, emotional, and behavioral response patterns. Although little is known about the origins of negative attribu-
tional patterns, several researchers have suggested that individ-
ual differences in beliefs and expectations may underlie such
intra- and inter-personal mechanisms are only just beginning to patterns (Baucom, Epstein, Sayers, & Sher, 1989; Fincham,
be documented. 1985; Fincham & Bradbury, 1987, 1989; Fletcher & Fitness,
One way to begin exploring these mechanisms is to consider 1993). Working models associated with attachment may be one
working models of attachment as part of a broader system of important source of these differences. This idea is compatible
cognitive and motivational processes that enable people to with a large body of research on the role of prior knowledge in
make sense of their social experiences and to function in ways social information processing and social judgment. Empirical
that serve their personal needs. On the basic of contemporary work in social psychology clearly demonstrates that many as-
research and theory in personality and social psychology, Col- pects of social perception and inference are heavily influenced
lins and Read (1994) have suggested a general framework for by top-down, theory-driven processes in which existing goals,
understanding how working models may function (see Figure schemas, and expectations shape the way people view new in-
1). According to this framework, working models of attachment formation (for reviews, see Fiske & Taylor, 1991; Markus & Za-
are highly accessible cognitive constructs that will be automati- jonc, 1985). Although most of this research involves thinking
cally activated in memory in response to attachment-relevant about strangers, these processes are increasingly being explored
events. Once activated, they are predicted to have a direct im- in the context of close relationships (e.g., Baldwin, 1992; Bu-
pact on social information processing (including attention, gental, Blue, Cortez, & Reck, 1993; Fletcher & Fincham, 1991;
memory, and inference) and on emotional response patterns. Fletcher & Fitness, 1990; Holmes & Rempel, 1989). Taken to-
Rather than operate in isolation, cognitive and affective re- gether, this work suggests that relationship partners are not sim-
sponses are also predicted to have reciprocal effects on each ply passive recipients of environmental stimuli, but are active
other. The outcome of these processes should then determine participants in the construction of their own reality.
one's choice of behavioral strategies. Furthermore, it is not nec- Thus, like other cognitive representations, working models
essary to assume that people are consciously directing these of attachment are likely to be important knowledge structures
processes, or even that they are aware of them. In fact, much of through which social events arefilteredand understood. When
this system should operate "automatically," that is, spontane- attachment-relevant events occur, attachment models should be
ously, with little effort and outside of awareness (Bargh, 1984). activated in memory automatically. Once activated, these
In short, Collins and Read (1994) argued that the impact of models provide either well-learned and readily available expla-
working models on behavior in any given situation will be nations for many interpersonal events, or the social knowledge
largely mediated by the subjective interpretation of the situation needed for the construction of new explanations "on-line." Fur-
along with one's emotional response. This model is intended thermore, because adults with different attachment styles oper-
only to be a very general framework for exploring a number of ate on the basis of very different models of themselves and the
more specific cognition-emotion-behavior linkages. The pres- social world, they are likely to construe the same events in very
ent study was thefirststep in a program of research designed to different ways. Thus, individuals with different working models
map out these links in greater detail. The primary purpose of of attachment may be predisposed to interpret social events in
these first studies was to explore the link between working ways that are consistent with their existing expectations and
models and cognitive response patterns. Cognitive models asso- beliefs.
ciated with attachment are presumed to guide the appraisal of The primary purpose of the current research was to test the
social situations and to help individuals maintain a coherent idea that working models of attachment bias and shape the ex-
world view and self-image (Bartholomew, 1990; Collins & planation and attribution process. To accomplish this, open-
Read, 1990, 1994). Although there are many cognitive pro- ended explanations were gathered and content coded for the
cesses that should be shaped by working models (see Collins & presence of attachment-relevant themes. It was predicted that
Read, 1994), I began by focusing on one aspect of social cogni- explanations for attachment-relevant events would be shaped in
tion that has already been shown to have important implica- specific ways that reflect underlying differences in models of self
tions for relationship functioning—attribution^ processes. and others. For example, those with a secure attachment style
should filter events through largely optimistic models. As a re-
Attachment Style Differences sult, their explanations should reflect confidence that their part-
in Attributions and Explanations ner is trustworthy and caring, and should minimize the impor-
tance of the event for issues of relationship stability. In contrast,
A large body of research points to the important role of attri- preoccupied and avoidant individuals should be more likely to
butions in relationship functioning (see Bradbury & Fincham, explain events in ways that suggest their partner's lack of re-
WORKING MODELS OF ATTACHMENT 813
sponsiveness and negative motivation and to interpret events as tional responses to an event should be due in part to his or her
having negative implications for the security of their relation- interpretation of the event. To the extent that a partner's behav-
ship. In addition, because of their negative models of them- ior is viewed as selfishly motivated or as a sign that the relation-
selves, preoccupied individuals may be more likely to see them- ship is in jeopardy, emotional distress should result. Thus, as
selves as the cause of their partner's behavior and to provide shown in Figure 1, a path from explanation patterns to emo-
explanations that reflect a lack of self-worth and self-reliance. tional distress was also expected. This link is consistent with a
To demonstrate that these differences are due to the activa- long history of theoretical and empirical work on the role of
tion of attachment-related models, rather than a more general attributions and appraisals in emotional experience (e.g.,
perceptual bias, it was important to include events that were not Berscheid, 1983; Lazarus & Folkman, 1984; Roseman, 1984;
expected to be influenced by attachment style. Although any Weiner, 1982, 1985). (Although the model presented in Figure
relationship event could be relevant to attachment concerns, the 1 also includes a direct link from emotional processes to cogni-
goal was to identify events that were not as likely to activate tive processes, this relation was not investigated in the current
attachment-relevant knowledge structures, or for which these study.)
structures, if activated, would provide less plausible explana- The final step in the model was to examine behavioral re-
tions. For example, explaining why a partner was late for a date sponse tendencies. Negative interpretations of a partner's be-
might be better understood by applying sex-role stereotypes havior and negative emotional responses should lead to corre-
(e.g.,"she took too long getting dressed") or culturally shared spondent behaviors, which are likely to result in conflict and
beliefs (e.g., "he was picking up flowers"). Thus, it was pre- to contribute to relationship distress (Doherty, 1982; Fincham,
dicted that all participants would be more likely to spontane- Beach, & Nelson, 1987; Fincham & Bradbury, 1988; Fincham
ously mention attachment themes (e.g., trust, acceptance, and & O'Leary, 1983). Thus, as shown in Figure 1, behavioral re-
responsiveness) in response to attachment-relevant events than sponses were expected to be directly predicted by the explana-
in response to attachment-irrelevant events. In addition, it was tion given, along with the emotional response to the event. In
predicted that attachment-style differences in the content of ex- summary, it was predicted that the relationship between attach-
planations would be more evident in explanations given for at- ment style and social behavior would be fully mediated through
tachment-relevant events. cognitive and emotional response patterns.
Participants ning by himself/herself, and (d) left you standing alone at a party. The
irrelevant events were chosen as irrelevant only in terms of their likeli-
Participants were 82 female and 53 male undergraduate students hood of activating attachment themes, not in terms of their potential
from the University of Southern California who participated in ex- for negative impact.3 The two events were: Your partner (a) borrowed
change for extra credit in their introductory psychology course. Partici- money from you and didn't pay it back and (b) forgot your birthday.
pants ranged in age from 18 to 25, with a mean age of 18.7. Fifty-five (It is important to note that the difference between the relevant and
percent of participants were currently involved in a romantic relation- irrelevant events was expected to be relative, not absolute—any rela-
ship, 45% were not. tionship event has the potential to be relevant to attachment concerns.)
For all dependent variables, scores were averaged across the four attach-
Materials ment-relevant events and the two attachment-irrelevant events.
Each event was printed on a separate sheet of paper that participants
Participants completed three sets of materials: (a) a background received in counterbalanced order. Directly after each event description,
questionnaire, (b) a relationship events questionnaire, and (c) an attri- participants were asked to provide an open-ended explanation for the
butions questionnaire. partner's behavior. Specifically, they were asked, "Why do you think
Background questionnaire. Participants werefirstasked several de- your partner behaved this way?" They were allowed to write an expla-
mographic questions and several questions about their current and past nation of any length, but if they listed more than one explanation they
relationship experience. Individual differences in attachment style were were asked to circle the most likely one. (Ail analyses are based on the
then measured with a revised version of the Adult Attachment Scale most likely explanation.) Following this task, participants rated the ex-
(AAS) developed by Collins and Read (1990).' This 18-item scale con- tent to which the event would be likely to lead to conflict if it occurred,
tains three subscales (each with 6 items): (a) the close subscale mea- onascalefrom 1 {not at all) to 7{very much).
sures the extent to which a person is comfortable with closeness and Next, to assess affective reactions, participants were asked to take a
intimacy; (b) the depend subscale assesses the extent to which a person moment to think about how they would feel if the event occurred that
is comfortable depending on others and believes that people can be re- day. They then rated the extent to which they believed they would expe-
lied on when needed; and (c) the anxiety subscale measures the extent rience each of 13 emotions on a 7-point scale. These emotional reac-
to which a person is worried about being rejected and abandoned by tions were expected to differentiate people with different attachment
others. Participants were asked to respond to each item in terms of their
styles. A principal-components analysis with oblique rotation was con-
general orientation toward close relationships.
ducted to reduce these items to a smaller set of underlying components.
The intercorrelations among the subscales were similar to those
found in previous work: The close and depend subscales were fairly
strongly correlated (r = .53), and the anxiety subscale was moderately 1
The revised scale includes the following changes. First, one item on
negatively correlated with the close (r = —.34) and depend (r = —.46) the close subscale was replaced to improve reliability. The correlation
subscales. Cronbach's alphas for the close, depend, and anxiety sub- between the revised scale and the original version was r - .98 in a sample
scales were .77, .78, and .85, respectively. Consistent with previ- of undergraduates (N = 295). Second, one item on the anxiety scale
ous work, there were no sex differences on any of the attachment that was poorly worded has been replaced with a similar item. Finally,
dimensions. two items on the anxiety scale concerned with a strong desire to "merge"
The three dimensions measured by the AAS can be used in two with a partner werereplacedwith two new items concerned with "am-
different ways. First, the subscales can be used as continuous measures bivalence" about relationships. Ambivalence in this context refers to an
of the components that underlie differences in adult attachment pat- approach-avoidance conflict between the desire to be close to others
terns (comfort with closeness, comfort with depending on others, and and the simultaneous fear that such closeness leaves one vulnerable to
fear of rejection). Second, through cluster analysis, the subscales can being hurt. The original two merge items were problematic because
be used to categorize people into discrete attachment styles (secure, many respondents were confused about their meaning and because they
avoidant, or preoccupied) according to their profile of scores along all tended to load weakly and inconsistently on factor analyses. As pre-
three dimensions. For example, an individual with a secure attachment dicted, when factor analyzed, the two new items loaded strongly with the
style is comfortable with closeness, believes that people are dependable, remaining four anxiety items. The correlation between the new anxiety
and is not anxious about being rejected or abandoned. Thus, the three subscale and the original version was r = .86 in a sample of undergrad-
dimensions can be used in combination to define discrete styles of at- u a t e s ^ 295).
tachment, but no single dimension corresponds to a single style. As ar- 2
It is important to note that all six behaviors were potentially negative
gued elsewhere (Collins & Read, 1990), an analysis of the continuous in impact. Positive events were not included because they are less likely
attachment dimensions can often provide a more precise understanding in general to elicit active attributional processing (Weiner, 1985). Like-
of attachment processes by specifying which component of one's work- wise, the marital literature finds that spontaneous attributions occur
ing model is most critical to a given relation. This concern is especially more often in response to negative partner behaviors in both distressed
relevant because specific attachment beliefs and concerns may be and nondistressed couples (Camper, Jacobson, Holzworth-Munroe, &
differentially reflected in people's explanations. Therefore, results will Schmaling, 1988; Holtzworth-Munroe & Jacobson, 1985,1988).In ad-
focus primarily on the continuous dimensions, although mean differ- dition, as summarized by Bradbury and Fincham (1987), the occur-
ences for attachment groups will be presented for selected outcomes. rence of negative relationship events better predicts changes in day-to-
Relationship events questionnaire. This booklet contained the pri- day marital satisfaction than do positive events.
mary stimulus materials and dependent measures. Each participant was 3
The relative negativity of each event was determined by pretesting
presented with six potentially negative partner behaviors.2 Four of these on a sample of undergraduate psychology students (N = 295) who rated
were chosen, a priori, as attachment-relevant and two as attachment- the extent to which each of a list of events would cause them distress.
irrelevant. Each attachment-relevant event was designed to tap into one On average, half of the events chosen in each category had been rated as
of four central attachment themes: warmth and responsiveness, emo- moderately distressing (e.g., being left at a party, having a partner want
tional availability when needed (safe haven), separation (proximity to spend an evening alone, and a partner not paying back money that
seeking), and use of one's partner as a secure base. The four events were, was borrowed), and half had been rated as highly distressing (e.g., not
Your partner (a) didn't respond when you tried to cuddle, (b) didn't responding to a cuddle, not being comforted, and a partner forgetting
comfort you when you were feeling down, (c) wanted to spend an eve- one's birthday).
WORKING MODELS OF ATTACHMENT 815
Three factors emerged and were used as the basis for forming three com- consistent with working models associated with attachment. The final
posite indices: (a) Distress (angry, hurt, disappointed, unappreciated, coding dimension was a general evaluation of the extent to which the
sad, jealous, and unloved), (b) Nervousness (nervous, confused, and overall content of the explanation maximized or minimized the nega-
helpless), and (c) Unemotional (unemotional and indifferent). The tive impact of the event. Each explanation was evaluated with respect to
Distress and Nervousness indices were strongly correlated with each the following themes: (a) partner responsiveness, (b) participant's self-
other (r = .75), but were weakly related to the Unemotional index (rs worth, (c) trust in partner, (d) confidence in partner's love, (e) confi-
= —.24 and -.20, respectively). dence that partner is dependable, (f) confidence that relationship is se-
Finally, behavioral intentions were measured by asking participants cure, (g) participant's self-reliance, (h) partner warmth and closeness,
to describe in detail what they would say or do in response to the part- and (i) minimize-maximize negative impact of event.4
ner's behavior. The coding for thefirsteight dimensions was done in two parts. First,
A ttribution questionnaire. Thefinalquestionnaire presented partic- each explanation was coded for whether it explicitly mentioned the at-
ipants with the six events again and asked them to rate the cause of the tachment theme; this was simply coded as either "yes" or "no" for each
event along 10 standard attributional dimensions. These measures were theme. For example, consider the following explanations given by two
presented in a separate booklet to avoid any potential influence on the participants in response to thefirstevent ("y° u r partner didn't respond
content offree-responseexplanations. The 10 dimensions were based to a cuddle"): (a) "My partner is losing interest in me" and (b) "He
on previous research on attribution in the social psychology and marital feels detached, doesn't want to be close to me." The first explanation
literatures (e.g., Epstein etal., 1987;Fincham, 1985; Fincham & Brad- received a "yes" check on the "partner love" theme and a "no" check on
bury, 1987; Holtzworth-Munroe & Jacobson, 1985; Passer, Kelley, & all others. The second received a "yes" check on the "partner warmth-
Michela, 1978). Four items assessed the locus of the cause: Participants closeness" theme and a "no" check on all others. An explanation could
rated the extent to which the behavior was caused by something about receive a check on more than one theme if applicable. High interrater
(a) the partner, (b) the self (i.e., the participant), (c) outside circum- agreement was achieved for this set of ratings. Agreement ranged from
stances, and (d) the relationship. Two items assessed the nature of the 90% to 97%, with an average agreement of 94%. Disagreements between
cause: Participants rated the extent to which the behavior was caused raters were settled by N. Collins.
by something that is (e) stable (not likely to change) versus unstable, Thisfirstcoding simply measured whether an attachment theme was
and (f) global (i.e., affecting many areas of the relationship) versus spe- mentioned, not whether it was positive or negative. The second step was
cific (i.e., affecting only this area). Three items assessed attributions to rate the extent to which the explanation indicated or implied positive
about the partner's motives and intentions: Participants rated the extent or negative information about each attachment theme. All eight attach-
to which their partner's behavior was (g) intended to have an impact on ment themes were rated on a 9-point scale ranging from - 4 to +4, with
them (the participant), (h) voluntary, and (i) intended to be negative. a neutral point in the middle. All scales were anchored, with the nega-
Finally, participants rated the extent to which the behavior was caused tive end reflecting low placement on that dimension (e.g., low self-worth
by (j) the partner's negative attitude toward them (the participant). or lack of trust). Thefinalrating scale, which assessed the overall posi-
Each dimension was rated on a 7-point scale, with higher scores repre- tive or negative implication of the explanation, was rated on a 9-point
senting greater assignment to that dimension. scale ranging from —4 {maximize the negative impact of the event) to
Because several of these dimensions were conceptually related and +4 (minimize the negative impact ofthe event).
were expected to be highly correlated, a principal-components analysis This second set of ratings required that each explanation be consid-
with oblique rotation was conducted and used as a basis for reducing ered with reference to the specific event being explained and the range
these items. This analysis suggested that two sets of ratings should be of possible explanations for that event. An example helps illustrate how
combined. First, three ratings (caused by partner's negative attitude, the explanations were rated. Consider the following explanation that
behavior was intentional, and behavior was intended to be negative) was given for the first event by a male participant: "She just had a bad
were averaged to form an index of the extent to which the behavior was day and wasn't in the mood to cuddle." The following are several al-
negatively motivated. The second index was an average of two ratings: ternatives given by other participants in this sample: (a) "She doesn't
the extent to which the cause of the event was perceived as stable and as like me anymore," (b) "He's losing interest," (c) "He's mad at me and
global. Thefiveremaining attribution dimensions did not form a clear this is his way of punishing me." Given the available alternatives, the
set of factors and were therefore left as separate items. first participant's explanation reflects confidence that his partner cares
about him, that his relationship is secure, and that his partner is not
purposely rejecting closeness. Therefore, this explanation was given pos-
Procedure itive ratings on these dimensions. However, given the event being ex-
plained, this explanation provides no clear basis for making inferences
Experimental procedure. Participants completed materials in small about certain other dimensions, such as trust or self-reliance. Thus the
groups of about 8 to 10. They were given the background questionnaire neutral point on the scale was used when there was no basis for making
first, which contained the attachment scale. Once they completed the inferences about a dimension or when the explanation reflected a neu-
questionnaire, participants were told that the purpose of the study was tral standing on a dimension.
to understand how people explain events that occur in dating relation- Reliability coefficients for this second set of ratings ranged from .77 to
ships. They were told that they would be responding to a series of hypo- .95, indicating fairly strong agreement between raters on all dimensions.
thetical events for afictionaldating partner. Participants who were cur- Reliability coefficients were also computed separately for the attach-
rently in a romantic relationship were instructed to think about a gen- ment-relevant and attachment-irrelevant events and were found to be
eralized partner and not their current partner. Participants then highly comparable. However, for irrelevant events, four of the dimen-
completed the relationship events and attribution questionnaires. sions (self-worth, dependability, self-reliance, and warmth-closeness)
Coding open-ended explanations. Before coding, all open-ended ex- were rated as "neutral" more than 85% of the time by one or both raters.
planations were transcribed onto index cards. Each explanation was This resulted in almost zero variance and very low reliability estimates.
then rated on nine a priori dimensions by two trained coders (one male
and one female) who were graduate students in social psychology. The
purpose of this coding was to assess themes that are likely to differenti-
4
ate the explanations of people with different attachment styles. Eight A more detailed description of the coding dimensions and proce-
of the nine dimensions assessed the extent to which the content was dures is available from N. L. Collins.
816 COLLINS
As such, these four dimensions were not used to test hypotheses for the Table 1
irrelevant events. Correlations Between Attachment-Style Dimensions and
The average of the two coders' ratings was used in all analyses. An Open-Ended Explanation Ratings
examination of the intercorrelations among the eight attachment
themes revealed that most were only moderately or weakly associated, Attachment-style dimensions
indicating that coders were able to successfully discriminate between
the various dimensions (the average intercorrelation was .39, and 20 of Explanation rating Close Depend Anxiety
the 28 correlations were below .50). Nevertheless, three pairs of vari-
ables had substantial correlations: love and secure (.94), responsive and Attachment-relevant events
dependable (. 72), and sel f-worth and self-reliance (. 70). Given the large Love/security .252** .269** -.254**
degree of empirical as well as conceptual overlap, each, pair was averaged Responsive/dependable .288*** .161 -.211*
Self-worth/reliance .15? .140 -.201*
into a single index. Trust .147 .171* -.178*
Coding behavioral intentions. Participants' written descriptions of Partner warmth/closeness .302*** .375*** -.341***
how they would respond to their partner were coded on two dimensions. Global rating
First, drawing from prior research (Fincham, Beach, & Nelson, 1987; Minimize negative impact .261** .280*'* -.271**
Fincham & O'Leary, 1983), behavioral descriptions were coded for the Attachment-irrelevant events
extent to which they would be punishing to their partner on a scale Love/security .003 .024 -.015
ranging from 1 (not at all) to 7 (extremely punishing). A punishing Responsive .026 .103 -.049
behavior was defined as one likely to result in negative consequences Trust .081 .010 -.170*
Global rating
for the partner (e.g., hurt feelings, embarrassment, etc.)- Next, each iii
Minimize negative impact .078 .066
response was rated for the extent to which it would likely lead to conflict
on a scale ranging from 1 (not at all likely) to 7 (extremely likely). Note. N= 135,
Conflict was denned as any negative interaction or exchange between *p < .05. **p< .01. •••/?< .001, two-tailed.
partners. Ratings were conducted by two trained, male coders, each of
whom rated half the questionnaires. Interrater reliability was assessed
by having each coder rate a random selection of the other's question- times at least one attachment theme was mentioned for the two
naires, totaling 30% of the sample (240 responses overall). Interrater
types of events. These scores were then correlated with the at-
reliability was .91 for punishing and ,88 for conflict, indicating strong
interrater agreement on both dimensions. Because the correlation be- tachment dimensions. In response to attachment-relevant
tween the two dimensions was high for both coders (.84 and .79), the events, individuals who were comfortable depending on others
two ratings were averaged to provide an overall measure of the extent to were less likely to spontaneously mention an attachment theme
which a behavior was punishing and likely to result in conflict. (r = -.28, p < .001), and those who were worried about being
rejected were more likely to do so (r - . 19, p < .05). However,
as expected, attachment style was unrelated to the presence of
Results and Discussion
attachment themes for explanations in response to irrelevant
First, correlational results examining the association between events(all rs nonsignificant).
working models of attachment (comfort with closeness, com- Open-ended explanations: Continuous ratings of attachment
fort depending on others, and anxiety about being rejected or themes. To test the hypothesis that attachment style would be
abandoned) and explanations, emotions, and behavioral inten- associated with the positive and negative content of explana-
tions are presented. Next, participants were categorized into at- tions, the attachment scales were correlated with the continuous
tachment styles (secure, avoidant, or preoccupied) on the basis explanation ratings. As shown in the upper panel of Table 1,
of a cluster analysis of the three dimensions, and mean differ- working models of attachment were significantly associated
ences between these groups were examined for each outcome. with the content of explanations given in response to attach-
Finally, the hypothesized multivariate model was tested using ment-relevant events. Individuals who were comfortable with
path analysis. closeness and felt they could depend on others were more likely
to provide explanations that indicated confidence in their rela-
Correlations Between Attachment Dimensions and All tionship and in their partner's love, were less likely to view their
partner's behavior as purposely rejecting closeness and, overall,
Outcome Variables
were more likely to give explanations that minimized the nega-
Open-ended explanations: Explicit attachment themes. To tive impact of the event. In addition, people comfortable with
test the prediction that attachment-relevant events would be closeness were more likely to provide explanations that reflected
more likely to elicit attachment content than irrelevant events, faith that their partner was responsive to their signals and was
the likelihood of having at least one explicit attachment theme dependable when needed. And those who felt they could depend
was compared for the two sets of events. As expected, partici- on others gave explanations that reflected greater trust in their
pants were much more likely to spontaneously mention an at- partner.
tachment theme in response to an attachment event, %2( \,N = As shown in Table 1, participants who were worried about
135) = 47.10, p < .001. On average, 52% of explanations for being rejected and unloved provided explanations that offered a
attachment events contained at least one of the eight attach- much more negative view of their partner and more negative
ment themes, compared with only 22.5% for irrelevant events. interpretation of the event. Consistent with their working
To determine if participants with different working models models, they were more likely to give explanations indicating
were more or less likely to mention explicit attachment themes, that their relationship was in jeopardy and that their partner
composite scores were computed to indicate the number of was unresponsive, not trustworthy, and was purposely rejecting
WORKING MODELS OF ATTACHMENT 817
Table 4
Mean Differences on All Outcome Measures by Attachment Style: Study 1
Outcome variable Secure Avoidant Preoccupied •F(2,132)
= .81, F(6, 260) = 4.93, p < .001. Means for each attachment ple with different working models of attachment interpreted
style and univariate F-tests appear in Table 4. Consistent with and responded to the same events in very different ways. How-
the correlationalfindings,significant univariate effects were ob- ever, these results do not provide information about the interre-
tained for distress and nervousness, but not for unemotional. lations among cognitive, emotional, and behavioral processes.
As predicted, preoccupied individuals were much more likely Thus, the final step was to test the structural model hypothe-
to experience emotional distress. However, the hypothesis that sized to underlie these relations. The hypothesized model spec-
avoidant people would deny emotional responses was only par- ified the following paths: (a) direct paths from working models
tially supported. Although they did report less negative emotion of attachment to explanations and emotions, (b) a direct path
than the preoccupied group (despite providing equally negative from explanations to emotions, and (c) direct paths from ex-
explanations), they were not more likely to report feeling un- planations and emotions to behavioral intentions.7
emotional or indifferent in response to the events. Although it would be ideal to include several indicators for
Behavioral intentions. As shown in Table 4, strong differ- each construct in the model (i.e., to form latent variables of
ences were found on participant's own conflict ratings, with the each construct), the sample size was not large enough to reli-
secure group being much less likely to predict conflict than ei- ably test such a model. Therefore, I tested the model by choos-
ther the avoidant or preoccupied group. Although the F-test for
the behavioral descriptions was only marginally significant, the 6
Because sex differences in explanation were not the focus of this
pattern of means suggests that secure individuals were less likely study, these findings are not presented. However, a summary of these
to behave in ways that would be punishing and likely to result results is available from N.L. Collins.
in conflict. 7
As noted above, the full model presented by Collins and Read
Gender differences. To determine if attachment style (1994) included a reciprocal link from emotional responses to cognitive
differences varied for men and women, a series of 2 (sex) X 3 processing. Although I believe that emotions may have an important
(attachment style) MANOVAs was conducted for all outcome impact on cognitive processing, these processes were not the focus of,
and could not be adequately addressed in, the present studies. Emo-
variables. Although there were several significant main effects
tional responses were made after participants had completed their open-
for gender, there were no significant interactions between at- ended explanations and had responded to additional explanation ques-
tachment style and gender for any of the outcome variables.6 tions that were not analyzed for this article. Because explanations pre-
ceded the emotion ratings temporally in these studies, and because only
Mediational Model Linking Attachment, Explanation, one path could be estimated at a time, it was more reasonable to esti-
Emotion, and Behavior mate the path leading from explanations to emotions. Ideally, both
Specifying the model. Results based on both the continuous paths would be estimated, but such a model could not be properly tested
attachment dimensions and the discrete styles indicate that peo- with these data.
820 COLLINS
ing the single indicator that best represented each theoretical greater emotional distress. Close-depend did not have a direct
construct. First, the continuous attachment dimensions were link to emotions, only an indirect one mediated through expla-
preferred over dummy-coding the three attachment styles be- nations. Finally, both explanations and emotions uniquely con-
cause the dimensions are more sensitive and provided a more tributed to behavioral intentions: Specifically, negative inter-
straightforward means of analysis. However, because the three pretations of the event and emotional distress resulted in in-
subscales were correlated, problems of multicollinearity might tended behavior that was more punishing and more likely to
result if all three were included in the model. Thus, the close lead to conflict.
and depend dimensions were combined into a single scale for An additional model was run in which direct paths were
several reasons. Most important, these two subscales are con- added between the two attachment dimensions and behavioral
ceptually related in that both are thought to be most closely intentions. Neither of these paths were statistically significant.
associated with models of others (vs. models of the self). The Thus, as hypothesized, there were no direct effects between at-
two scales are also fairly strongly correlated (r = .53), and were tachment style and behavioral intentions, only indirect ones
related to most of the outcome variables in a similar way. Sec- mediated through cognitive and emotional response patterns.
ond, to assess explanation patterns I used the judges' ratings Finally, to be sure that the modeling results were not depen-
of the overall negative impact of the open-ended explanations dent on the particular explanation variable used, the same
because it was the best global, summary rating of the explana- model was run with a latent explanation variable that included
tions and because judges' ratings on this global dimension all eight of the attachment theme ratings. The regression co-
seemed to be largely a function of the more specific attachment efficients for this model, and the conclusions drawn from them,
theme ratings. This idea was confirmed by a regression analysis were virtually identical to those for the more simple model. The
that indicated that over 90% of the variance in this global rating only difference in this model was that, in addition to significant
was accounted for by variance in the eight attachment-theme links between the attachment variables and the latent explana-
ratings. Next, the emotional distress index was chosen to repre- tion variable, there were unique relationships between the at-
sent emotional responses because it was the most general mea- tachment variables and the residuals for several explanation
sure of negative affect. Finally, independent ratings of partici- dimensions.
pants' written behavioral responses were used to represent be-
havioral intentions.
Study 2
To test the model, EQS (Bentler, 1989) software for structural
equation modeling was used. Because only one measured vari- Study 1 provides support for the idea that differences in cog-
able was used for each hypothetical variable, this analysis is sim- nitive models associated with attachment styles predispose peo-
ply a path analysis. The goodness offitof the model was assessed ple to think, feel, and behave differently in response to the same
with a joint consideration of the chi-square test and the com- relationship events. Nevertheless, because participants were re-
parative fit index (CFI; Bentler & Bonett, 1980). A nonsignifi- sponding to hypothetical partners on the basis of very restricted
cant chi-square indicates a good-fitting model. The CFI can information, the generalizability of these results to people in
range from 0 to 1.0, a value of .95 or greater indicating a good- ongoing relationships may be somewhat limited. In established
fitting model, although values between .90 and .94 are consid- relationships, couple members should operate on the basis of
ered acceptable (Bentler & Bonett, 1980). several sources of social knowledge, some of which are based on
Testing the mediational model. First, the hypothesized relatively stable models of self and others that are brought into
model was estimated, resulting in a x 2 (2, N= 135) = 4.73,/> = the relationship (i.e., an attachment style), and some of which
.09, CFI = .98. Although these fit statistics were very good, the are based on cognitive representations of their specific partner
path Unking the close-depend composite with emotional dis- and specific relationship (i.e., the history and quality of one's
tress was not significant (p = .06), and the path linking anxiety current relationship). This partner-specific model may include
to explanation patterns was marginally significant (/? - .16, p such things as memories of the partner's past behavior, beliefs
= .07). Thus, the link between close-depend and distress was about the partner's stable dispositions, perceptions of a part-
removed, resulting in a small improvement in model fit, x2(X ner's commitment to the relationship, and so on. Within the
N = 135) = 5.27, p = . 15, CFI « .99. An additional model was presentframework,the representation of one's current relation-
tested in which the marginal path between anxiety and expla- ship is viewed as an additional knowledge structure or working
nation patterns was removed, but this resulted in decreased model that will be activated whenever relationship-relevant
modelfit.In addition, the link between anxiety and explanation events are experienced (Collins & Read, 1994). Of course, con-
patterns resulted in the largest modification index (as indicated crete, relationship-specific models are not expected to be inde-
by the Lagrange multiplier test), which indicated that including pendent of more abstract attachment models, nor are they con-
that path would provide the largest improvement in model fit. sidered objective representations of actual experience. These
Thus, this link was retained. Because model fit was good, no models are best viewed as constructions of social reality that are
additional paths were added. Parameter estimates for the final linked in complex ways with more stable models of self and oth-
model are shown in Figure 2. ers (Collins & Read, 1994).
As hypothesized, working models of attachment predicted The aim of Study 2 was to replicate Study 1 with people in-
explanations, with both close-depend and anxiety providing in- volved in ongoing dating relationships, and to test the relative
dependent contributions. Emotional distress was directly asso- contributions of attachment style and relationship quality to
ciated with anxiety as well as with explanations: Specifically, predicting patterns of explanation, emotion, and behavior. It
fear about being unloved and negative explanations predicted was expected that attachment style would continue to predict
WORKING MODELS OF ATTACHMENT 821
Anxiety
Explanation
pattern ^^43***
Figure 2. Final parameter estimates and goodness-of-fit statistics for the mediational model in Study
1. Numbers on double-headed arrows are correlation coefficients; numbers on single-headed arrows are
standardized regression coefficients. Dashed lines indicate nonsignificant paths that were dropped from the
final model. CFI = comparativefitindex.
outcomes even after accounting for relationship-specific relationship. Relationship length ranged from I to 96 months, with a
models. Once again, attachment-relevant events were com- mean of 16 months and a median of 12 months. Over 75% of the sample
pared to irrelevant events to demonstrate the functional speci- reported that they had been dating their partner longer than 4 months,
and 85% were dating their partner exclusively.
ficity of attachment representations. It was predicted that at-
tachment style differences in explanations would be more evi-
dent in response to attachment-relevant relationship events. Materials
Finally, a mediational model linking attachment style and re- Participants completed three sets of materials: (a) a background
lationship quality to explanation, emotion, and behavior was questionnaire, (b) a relationship events questionnaire, and (c) an attri-
proposed and tested. This model was similar to the model tested bution questionnaire.
in Study 1, but with two additions: (a) Relationship quality was The background questionnaire contained the AAS and measures of
added to the model and was expected to directly predict expla- demographic characteristics that were identical to those in Study 1. Re-
nation patterns, and (b) because attachment style has been liability coefficients for the close, depend, and anxiety subscales in the
shown to predict relationship quality, correlational paths link- present sample were .82, .80, and .83, respectively. The close and depend
subscales were strongly correlated (r = .67), and the anxiety subscale
ing attachment style and relationship quality were also in-
was moderately correlated with the close (r = —.28) and depend (r =
cluded. Consistent with prior research on attributions in mar- —.46) subscales. There were no differences between men and women on
riage, relationship quality was expected to have only an indirect the three attachment dimensions, and there were no significant corre-
relationship with emotional responses, mediated through lations between relationship length and the subscales of the AAS.
explanations. To assess relationship quality, participants completed a 15-item mea-
sure adapted from the one used by Collins and Read (1990) in their
Method work on relationship satisfaction in dating couples. Some of these items
originated from widely used measures of marital satisfaction (Locke &
Study 2 was identical to Study 1 except that all participants were Wallace, 1959; Spanier, 1976), whereas others were developed specifi-
involved in a romantic relationship and were asked to respond to the cally for use with dating couples. Items focused onfiveaspects of rela-
materials in terms of their current partner rather than a hypothetical tionship quality: (a) general satisfaction, (b) communication, (c) feel-
partner. Participants also completed a questionnaire on relationship ing understood and accepted by one's partner, (d) frequency and sever-
quality and responded to one additional attachment-irrelevant event. ity of conflict, and (e) participant's own level of commitment and
The additional irrelevant event was added because, in Study 1, differ- perception of their partner's commitment. The 15 items were standard-
ences between attachment relevant and irrelevant events may have re- ized and averaged to form a single index of relationship quality. Coeffi-
sulted in part from the irrelevant events having not been measured as cient alpha for this index was .84. As expected, relationship satisfaction
reliably. Although it would have been ideal to have had four events in was significantly correlated with the attachment-style dimensions. Indi-
each category, this would have been too taxing for both participants and viduals who reported higher qualityrelationshipswere more comfort-
raters. able with closeness (r = .23, p < .05) and with depending on others (r =
.24, p< .05), and were less worried about being rejected or abandoned
(r=-.29,/><.001).
Participants The relationship events and attribution questionnaires were identical
Participants were 56 male and 73 female undergraduate students to those used in Study 1, except for the addition of one more irrelevant
from the University of Southern California who participated for extra event, making a total of seven partner behaviors (four attachment-rele-
credit in their introductory psychology course. Participants ranged in vant and three irrelevant). This additional event read, "Please imagine
age from 17 to 30 years old, with a mean age of 19.2. To participate, that your partner arrived for your date thirty minutes late." As in Study
it was required that participants be currently involved in a romantic 1, three emotion composites (distress, nervousness, and unemotional)
822 COLLINS
were computed from the emotion ratings, and some of the attribution composite scores were computed to indicate the number of
ratings that were highly correlated and conceptually related were times at least one attachment theme was mentioned for the two
combined. types of events. There were no significant correlations between
these two scores and any of the three attachment dimensions.
Procedure However, participants who were more satisfied with their cur-
rent relationship mentioned fewer attachment themes in re-
The experimental procedure was identical to that in Study 1, except sponse to attachment-relevant events (r = -.21, p < .05), but
that participants were given slightly different instructions directing not to the attachment-irrelevant events (r = -.02).
them to think about their current relationship partner rather than a
fictitious dating partner.
Coding open-ended explanations. Open-ended explanations were Regression Analyses Predicting All Outcomes by
coded on the same nine content dimensions used in Study 1, and the Attachment Dimensions and Relationship Quality
procedures for preparing and coding the explanations were identical to
those followed in Study 1. Ratings were conducted by one male and one Open-ended explanations: Continuous ratings of attachment
female coder. The female rater had also coded Study 1, and the male themes. First, to explore the relative contributions of attach-
rater was new. The reliability coefficients for the overall sample were ment style and current relationship quality to the content of
very comparable to those obtained in Study 1. For the full sample of free-response explanations, the attachment scales and the rela-
events, coefficients ranged from .82 to .95, indicating strong agreement tionship quality index were used as predictor variables in a se-
between raters on all nine dimensions. As in Study 1, reliability esti- ries of simultaneous multiple regression analyses. Because the
mates were also computed separately for the two types of events. Co- close and depend subscales were strongly correlated in this sam-
efficients ranged from .84 to .95 for ratings of attachment-relevant
ple, are conceptually similar, and tend to be related to other
events, and from .74 to .92 for attachment-irrelevant events. However,
for attachment-irrelevant events, two of the dimensions (self-reliance variables in similar ways, the two scales were combined into a
and warmth) were rated as neutral more than 90% of the time by one single index for all regression analyses.
or both raters, which resulted in almost zero variance and very low reli- As shown in the upper panel of Table 5, the regression equa-
ability estimates. As such, these two dimensions were not used to test tion was significant for all variables, with both relationship
relationships between attachment and explanations for irrelevant quality and the anxiety dimension making significant unique
events. contributions. The three variables together accounted for be-
The intercorrelations among the ratings were very comparable with tween 9% and 21% of the variance in the ratings of explanation
those obtained in Study 1. Once again, although many dimensions were content for attachment-relevant events. In every case, partici-
moderately correlated, coders were able to successfully distinguish be- pants who were in relationships that they perceived as satisfying
tween the various dimensions. However, as in Study 1, three pairs of
variables had fairly substantial correlations and were also conceptually
and committed were less likely to interpret their partner's be-
related, so they were averaged as follows; love with secure (r = .89); havior in a way that reflected negative attachment themes. In
responsive with dependable (r= .70); and self-worth with self-reliance contrast, and independent from relationship quality, the extent
(r = .76). to which a person was worried about being rejected and unloved
Coding behavioral intentions. The procedure for coding the written (i.e., anxiety subscale) consistently predicted the presence of
behavioral descriptions was identical to that in Study 1. Coding was negative attachment themes. In contrast to Study 1, a person's
conducted by one male and one female rater. The male rater had also comfort with closeness and ability to depend on others was not
coded Study 1, and the female rater was new. Each rater coded half related to their explanations.
the questionnaires, and interrater reliability was assessed by having the As predicted, individual differences in attachment style did
raters code a random selection of each other's questionnaires, totaling
23% of the sample (210 responses overall). The reliability coefficient
not explain differences in attachment-irrelevant events. As
was .88 for ratings of punishing and .86 for ratings of conflict, indicating shown at the bottom of Table 5, working models of attachment
strong interrater agreement on both dimensions. The two dimensions were not significantly associated with any of the explanation
were strongly correlated for both coders (r = .92 and .94), and were dimensions. Relationship quality was also less consistently re-
therefore averaged. lated to this set of explanations, although several significant
effects did emerge. Specifically, participants with better-quality
relationships provided explanations that reflected high self-
Results and Discussion worth and confidence in their partner's responsiveness. In addi-
Open-Ended Explanations: Explicit Attachment Themes tion, their explanations were rated, generally, as minimizing the
negative impact of their partner's behavior.
To test the prediction that attachment-relevant events would The primary aim of Study 2 was to test the relative contribu-
be more likely to elicit attachment content than irrelevant tions of attachment style and relationship quality to predicting
events, the likelihood of having at least one explicit attachment responses to potentially distressing relationship events. As ex-
theme was compared for the two set of events. As expected, par- pected, individuals in satisfying and committed relationships
ticipants were more likely to spontaneously mention an attach- viewed potentially negative events in a less threatening way,
ment theme in response to attachment events, %2 (1, N = 129) minimizing their impact for broader issues of relationship se-
= 24.60, p < .001. On average, 34% of explanations for attach- curity. However, working models of attachment still remained
ment events contained at least one of the eight attachment an important, independent predictor of explanation patterns.
themes, compared with only 19% for irrelevant events. In contrast to Study 1, however, anxiety was the only attachment
To determine if participants with different working models dimension that consistently predicted. Overall, individuals who
were more or less likely to mention explicit attachment themes, were worried about being rejected and unloved appeared pre-
WORKING MODELS OF ATTACHMENT 823
Table 5
Multiple Regression Analyses Predicting Open-Ended Explanation Ratings From
Relationship Quality and Attachment-Style Dimensions
Predictor variables
Outcome variable RQ Close/depend Anxiety R R2
Attachment-relevant events
Love/security .30*** -.13 -.22* .39 .15***
Responsive/dependable .30*** -.15 -.24** .40 ,16***
Self-worth/reliance .22* -.11 -.18* .30 .09**
Trust .28*** -.08 -.23* .38 .15*"*
Partner warmth/closeness .31*** -.07 -.29*** .46 .21***
Global rating
Minimize negative impact .31*** -.10 -.21* .40 .16***
Attachment-irrelevant events
Love/security .15 .05 -.12 .24 .06*
Responsi ve/dependable .18* .06 -.12 .27 .07**
Self-worth .19* .08 -.14 .30 .09**
Trust .14 .04 -.05 .17 .03
Global rating
Minimize negative impact .30*** .03 -.05 .33 .11**
Note. Unless otherwise labeled, values shown are standardized regression coefficients. RQ = relationship
quality.
*/7<.05. **/?<.01. ***/><. 001.
disposed to interpret events in ways that were consistent with ported lack of emotion. Also, as expected, relationship quality
their expectations, even if their current relationship was going was not associated with emotional responses.
well. As expected, attachment style did not predict the content Behavioral intentions. As shown at the bottom of Table 6,
of explanations for irrelevant events. In addition, for all partici- people who were anxious about being unloved and those who
pants, explanations for attachment-irrelevant events were much were in poor quality relationships were more likely to anticipate
less likely to contain any explicit mention of attachment conflict. However, only anxiety predicted whether the actual be-
themes. These findings replicate those of Study 1 and provide haviors participants said they would engage in were judged to
further evidence that working models of attachment are, to be punishing and likely to result in conflict.
some extent, selectively activated or used. Moreover, these re-
sults suggest that differences found in response to attachment-
Mean Differences for Discrete Attachment Types
relevant events were driven by the activation of attachment
models rather than some more general perceptual bias. Categorizing into styles with cluster analysis. On the basis
Structured attribution measures. Regression analyses for of procedures described in Study 1, a cluster analysis was per-
the structured attribution ratings are presented at the top of formed using scores on close, depend, and anxiety subscales as
Table 6. In contrast to Study 1, these ratings were not strongly the clustering variables. Consistent with Study 1, there was good
predicted by either attachment or relationship quality. Only two evidence for a three-cluster solution. The mean scores on close,
significant equations were obtained. First, individuals who were depend, and anxiety for the three clusters were then examined
in better quality relationships and were not worried about being (see Table 7). The pattern of means was very consistent with
rejected were less likely to attribute their partner's behavior to theoretical accounts of the three attachment styles and with the
something about their relationship. Second, those who were pattern obtained in Study 1. The three clusters were therefore
anxious about being unloved were more likely to believe that retained and labeled accordingly. It is noteworthy that only 9%
the event was caused by something that affected many areas of (« = 12) of the sample was categorized as avoidant, which is
the relationship and was not likely to change. They also had a somewhat lower than typically found in studies of dating
tendency to believe that their partner's behavior was caused by couples.
something about themselves (i.e., about the participant). Open-ended explanation ratings. A MANO\A on the open-
Emotional responses. As shown in the middle of Table 6, ended explanations revealed a significant multivariate effect of
anxiety was strongly associated with increased distress and ner- attachment style, Wilks' lambda = .84, F(12, 242) = 1.86, p
vous feelings, as predicted. Those who were comfortable with < .05. As shown in Table 8, significant univariate effects were
closeness were also more likely to report distress, although not obtained for all dependent variables. For each variable, the se-
nervousness. This finding is consistent with the prediction that cure group was much less likely than the preoccupied group to
secure participants (i.e., those comfortable with closeness and have negative attachment themes. However, in contrast to Study
dependency) should be more willing than avoidant participants 1, the avoidant group gave explanations that were generally as
(who are much lower on these dimensions) to acknowledge dis- favorable as those of the secure group, although the comparison
tress. None of the predictors was significantly related to a re- did not always reach significance.
824 COLLINS
Table 6
Multiple Regression Analyses Predicting Attributions, Emotions, and Behavioral Intentions
Predictor variables
Outcome variable RQ Close/depend Anxiety R R2
Attribution ratings
Partner -.01 .09 .19 .18 .03
Circumstances .12 -.20* -.09 .20 .04
Self -.02 .04 .20* .19 .04
Relationship -.18* .10 .32*** .38 .14***
Negative attitude/intention -.08 .05 .15 .18 .03
Stable/global -.17 .11 .31*** .35 .12***
Voluntary/controllable -.11 -.08 -.13 .15 .02
Emotional responses
Distress -.06 .21* .51*** .49 .24***
Nervousness .04 .17 .47*** .43 .18***
Unemotional .14 -.14 -.11 .20 .04
Behavioral responses
Self-ratings of conflict -.26** .16 .32*** .43 .18***
Ratings of behavioral intentions -.13 .13 .29** .31 .10**
Note. Unless otherwise labeled, values shown are standardized regression coefficients. RQ = relationship
quality.
*p<.05. **p<.01. ***p<.00l.
Structured attribution measures. Consistent with the re- than the avoidant group, but less distressed than the preoccu-
gression analyses, the attribution ratings were not strongly asso- pied group.
ciated with attachment style. A MANOY\ revealed a nonsig- Results for the emotional responses were precisely as pre-
nificant multivariate effect of attachment style, Wilks' lambda dicted. Compared to the secure and avoidant groups, preoccu-
= .85, F( 14, 240) = 1.50, p > .10. Significant univariate effects pied individuals were much more likely to report that they
were obtained only for ratings of relationship as the cause, F(2, would feel emotionally distressed in response to the events. In
126) = 5.93, p < .01, and for ratings of negative motivation, addition, the avoidant group showed the predicted lack of emo-
F(2,126) - 4.32, p < .05. In both cases, the preoccupied group tion. They reported that they would feel less emotionally dis-
gave more pessimistic attributions than either the secure or tressed and more unemotional than individuals in the secure
avoidant group. group. This pattern is consistent with the idea that secure indi-
Emotional responses. A MANO\A on the three emotion viduals are able to acknowledge distress in response to poten-
indices revealed a significant multivariate effect of attachment tially threatening events.
style, Wilks' lambda = .82, F{6, 248) = 4.19, p < .01. Signifi- Behavioral intentions. As shown at the bottom of Table 8,
cant univariate effects were obtained on all three indexes, as secure and avoidant individuals were less likely than preoccu-
shown in the middle of Table 8. As predicted, the avoidant pied individuals to anticipate conflict and to behave in conflict-
group was least likely to report feeling distress and more likely evoking ways, and the avoidant group was less likely than the
to report feeling unemotional in response to their partner's be- secure group to respond negatively to their partner.
havior. Also, as predicted, the secure group was more distressed As predicted, attachment style was related to participants'
own expectations of conflict and to their actual behavioral in-
tentions. People who were anxious about being unloved were
Table 7 more likely to anticipate conflict and to behave in ways that
Mean Differences on Attachment-Style Dimensionsfor were judged as punishing toward their partner and as likely to
Attachment Styles Derived From a Cluster Analysis: Study 2 result in conflict. Relationship satisfaction was also associated
with lower expectations of conflict, but did not predict the na-
Cluster/attachment style ture of participants' written behavioral intentions.
Sex differences. To determine if attachment style differ-
Attachment Secure Avoidant Preoccupied ences varied for men and women, a series of 2 (sex) X 3
dimension {N = 64) (N=12) (TV = 5 3 ) F(2, 126)
(attachment style) MANOVAs were conducted for all outcome
Close 4.41fl 2.36 h 3.29C 115.62*** variables. Although there were several significant main effects
Depend 4.04a 2.29 b 2.89C 92.64*** for gender, there were no significant interactions between at-
Anxiety 1.96. 2.22 a 3.13 b 43.19*** tachment style and gender (see Footnote 4).
Note. Scores can range from 1 to 5. Within rows, means with different Testing the Mediational Model
subscripts differ at p < .05, according to a Newman-Keuls Multiple
Comparison test. The variables used as indicators of each component of the
***/>< .001. model and the procedure for testing the mediational model were
WORKING MODELS OF ATTACHMENT 825
Table 8
Mean Differences on Selected Outcome Measures by Attachment Style: Study 2
Outcome variable Secure Avoidant Preoccupied ' F (2,126)
identical to those in Study 1. The only differences in the model models as well as relationship-specific models will contribute to
were that relationship quality was added as an additional vari- an individual's understanding of events in their relationships.
able that had a direct path to explanation patterns, and correla- However, consistent with the regression analyses, fear of being
tional paths between relationship quality and the attachment- unloved was the only attachment dimension that directly pre-
style dimensions were also added. dicted explanations. Next, consistent with Study I, those anx-
The hypothesized model was estimated first, resulting in ious about being unloved and those who explained their part-
X2(4,Ar= 129)= 1.42,p = .84,CFl = 1.0. Although the model ner's behavior in a negative manner reported greater emotional
fit was very good, the path linking close-depend to explanation distress. As predicted, being comfortable with closeness and de-
patterns was not significant. Therefore, this path was removed pending on others was also associated with greater distress. This
and the model reestimated. All remaining paths were signifi- result is consistent with thefindingthat secure individuals (who
cant, x2 (5, N = 129) = 2.64, p = .76, CFI = 1.0. Parameter are highest on these dimensions) were more willing than avoid-
estimates for thefinalmodel are shown in Figure 3. ant individuals (who are lowest on these dimensions) to ac-
As predicted, relationship quality and attachment style both knowledge emotional distress. As predicted, there was only an
had significant, independent associations with explanation pat- indirect link between relationship quality and emotions, medi-
terns. This finding is consistent with the argument that global ated through explanations. Thisfindingsuggests that the influ-
Relationship
satisfaction
-.29 •*
-+ Anxiety
.26* *
-.42 ***
» Close/depend
Figure 3. Final parameter estimates and goodness-of-fit statistics for the mediational model in Study
2. Numbers on double-headed arrows are correlation coefficients; numbers on single-headed arrows are
standardized regression coefficients. Dashed lines indicate nonsignificant paths that were dropped from the
final model. CFI = comparativefitindex.
826 COLLINS
ence of attachment style on emotional experience will be rela- depending) and anxiety about being unloved were both associ-
tively unaffected by the quality of one's current relationship. ated with explanation patterns; in Study 2, only the anxiety di-
This is especially noteworthy given that emotional responses mension remained strongly predictive.
were such a strong predictor of behavioral intentions. Finally, There are a number of possible explanations for this pattern.
negative explanations and emotional distress resulted in behav- One reason may be that avoidant adults were willing to draw
ioral descriptions that were judged as more punishing and more negative conclusions about their relationship in the context of a
likely to lead to conflict. fictional dating relationship, but were unwilling to do so in the
Overall, the mediational model appeared to be a good repre- context of their current romantic relationship because such
sentation of the underlying relationships among these variables. conclusions may be too threatening to the self. On the one hand,
These results provide further support for the hypothesized un- avoidants' negative view of human nature and of relationships
derlying processes and indicate that the results of Study 1 were might have led them to conclude that their partner was selfish
not unique to hypothetical relationships. However, to provide and rejecting. On the other hand, drawing such a conclusion
further support for the model, two additional models were run. would have implied that they were not valued by others, and
Thefirstmodel included direct paths from the two attachment- that they were concerned about such things—both of which
style dimensions to behavioral intentions. Neither path was sta- would be incompatible with their desire to appear self-confident
tistically significant, indicating that any links between attach- and self-sufficient. Thus, the avoidant group's positive response
ment style and behavior were mediated through explanations pattern in Study 2 may reflect a defensive strategy for dealing
and emotions. The second model included a latent explanation with this potential dilemma and for protecting their self-image.
variable made up of the eight attachment-theme dimensions. This interpretation is consistent with theoretical descriptions of
The results based on this model were virtually identical to those the avoidant style as defensively high in self-esteem and self-
presented for the more simplified model. reliance (Bartholomew, 1990).
Another possible explanation for this difference is that the
General Discussion two samples may have differed in important ways. In a number
of studies (e.g., Collins & Ailard, 1994; Collins & Read, 1990),
Attachment Style Differences in Explanations it was found that the distribution of attachment styles differs in
and Attributions samples of dating and nondating adults. Recall that in Study 1
only half of the participants were currently dating, whereas all
The primary purpose of this research was to test the hypoth- participants in Study 2 were involved in a steady dating rela-
esis that adults with different attachment styles would explain tionship. A comparison of the means on the three attachment
and interpret events in ways consistent with their beliefs and dimensions indicated that, overall, the sample in Study 1 was
expectations about themselves and others. Content coding of less secure; they were less comfortable with closeness, t(260) =
open-ended explanations provided strong support for this hy- -1.88, p < . 10, less willing to depend on others, /(260 ) = -3.01,
pothesis. In both studies, preoccupied individuals were likely to p < .01, and more concerned about being rejected, t(260) =
construct explanations that offered much more negative views 3.O8,p<.01.
of their partner and more negative interpretations of events. However, a more specific and important difference may have
Consistent with participants' expectations, their explanations been in the nature of avoidance in the two samples. Specifically,
were more likely to suggest that their relationship was in jeop- Bartholomew (1990) has argued that attachment researchers
ardy, and that their partner was unresponsive to their needs, not need to distinguish between two types of avoidance. Fearful
trustworthy, and purposely rejecting closeness. Their explana- avoidants avoid getting close to others because they worry about
tions also reflected lower self-worth and self-reliance. In con- being hurt and rejected. These individuals have negative models
trast, secure adults provided much more positive explanations. of both self and others. Dismissing avoidants avoid close rela-
Consistent with their optimistic models of self and others, their tionships because they desire to be independent and self-suffi-
explanations were more likely to communicate confidence in cient. These individuals have negative models of others but pos-
their relationship and in their partner's love, they were less likely itive models of the self. Because fearfuls are least likely to be
to view their partner's behavior as purposely rejecting closeness, involved in steady dating relationships, it is reasonable to as-
and overall they tended to construe events in ways that mini- sume that the avoidant participants in Study I were more likely
mized their negative impact and limited their importance for to have been fearful, and the avoidant participants in Study 2
broader issues of relationship stability. These patterns were fur- were more likely to have been dismissing. Although I did not
ther supported by results on the structured attribution ratings, have a measure of the four styles in the current studies, it was
although the pattern on these measures was stronger in Study 1 possible to examine this issue in an exploratory manner by cre-
than in Study 2. ating the four styles on the basis of scores on the three attach-
Although the findings for secure and preoccupied groups ment dimensions. As noted earlier, the close and depend dimen-
were consistent across the two studies, an interesting pattern sions are more closely associated with models of others, and the
emerged for the avoidant group. In Study 1, avoidant individu- anxiety dimension is more closely associated with models of the
als provided negative explanations similar to those given by pre- self. Therefore, I created four attachment styles by considering
occupied participants. However, in Study 2 they provided posi- each person's profile of scores along the dimensions (close and
tive explanations that were more similar to those given by secure depend werefirstaveraged to form a single score). High scores
participants. Another way of conceptualizing this pattern is were defined as above the midpoint of the scale (i.e., above 3),
that, in Study 1, comfort with relationships (closeness and and low scores were defined as below the midpoint (participants
WORKING MODELS OF ATTACHMENT 827
who fell at the midpoint were given missing values). For exam- are functionally specific. That is, they are selectively activated
ple, a person who had a low score on dose-depend and a low or used in response to relevant environmental stimuli. Second,
score on anxiety was categorized as dismissing, and a person these findings provide evidence that the link between attach-
with a low score on close-depend but a high score on anxiety ment style and explanations was not simply the result of a gen-
was categorized as fearful. A comparison of the breakdown for eral perceptual style or response bias. Indeed, knowledge struc-
the styles in each sample supports the idea that the samples were tures unrelated to attachment seem to have provided a better
somewhat different. In Study 1, the sample included 50% se- basis for understanding the attachment-irrelevant events. It is
cure, 17% preoccupied, 20% fearful, and 12% dismissing. In also likely that attachment-relevant models were activated but
Study 2, the sample included 62% secure, 14% preoccupied, were not used because they simply provided less plausible ex-
11 % fearful, and 14% dismissing. As expected, there was a much planations. Clearly, not all social events should be interpreted
larger percentage of secure participants in the dating sample. on the basis of attachment-related knowledge structures. Many
More important, there were half as many fearful avoidants and events are better understood with reference to commonly
slightly more dismissing avoidants. I then examined mean shared behavioral scripts, person prototypes, sex-role stereo-
differences on all explanation variables across both studies to types, and so on. In the current samples, a review of the expla-
determine if the negative pattern for avoidants in Study 1 was nations suggested that the irrelevant events (e.g., partner was
primarily reflecting the fearful avoidant and not the dismissing late for a date, forgot birthday, or didn't pay back money)
avoidant. As expected, in both studies, the fearful avoidants tended to activate stereotypes of the "typical college student,"
consistently obtained the lowest mean scores on the explanation who was seen as high on stress and very low on financial
ratings. Nevertheless, the dismissing group still tended to be resources.
more negative than the secure group in Study 1 (but not in Another important finding is that, in Study 2, attachment
Study 2). Unfortunately, I could not run the Study 1 analyses style predicted explanations independent of current relation-
separately for dating and nondating participants because only 5 ship quality. This association was important to establish be-
of the dating participants were categorized as dismissing. cause attachment style consistently covaries with relationship
These exploratory analyses provide some support that the quality, and differences in explanations may have simply re-
two samples differed in their degree of insecurity, and that the flected differences in the nature and quality of one's current
negative pattern in Study 1 was being driven primarily by fear- dating relationship. As expected, individuals who had more sat-
ful avoidants. However, there was still some evidence that dis- isfying relationships were more likely to interpret their partner's
missing adults were responding more negatively to the hypo- behavior in positive ways. However, independent of relationship
thetical relationship than to their real relationship. I am cur- quality, individuals who were worried about being rejected in-
rently following up on these findings by measuring the four terpreted their partner's behavior in much more negative ways.
attachment styles more carefully and by manipulating real ver- Comfort with closeness and depending on others was not related
sus fictional relationships in a single study (Collins & Allard, to explanations in this dating sample. Another way to view this
1994). Preliminary results indicate that for both real and hypo- pattern is that, even controlling for relationship quality, secure
thetical relationships, dismissing avoidants provide explana- (and dismissing) adults were more optimistic in their explana-
tions that are more similar to secure respondents', and fearful tions, whereas preoccupied (and fearful) adults were more pes-
avoidants provide negative explanations and attributions that simistic. It may be that, for people in relationships, general
are more similar to preoccupied adults'. Although additional models about the warmth and dependability of others are less
work is needed, thesefindingssuggest that attachment research- salient than perceptions of the current partners caring and re-
ers need to consider the interplay of cognitive and motivational sponsiveness. Thus, concrete models of a specific partner's re-
factors when attempting to understand the explanation process. sponsiveness may provide more plausible explanations and may
For example, our lab is currently planning studies to examine be preferred over more general models about the social world.
the hypothesis that the positive explanations provided by dis- However, if relationship-specific models are preferred over
missing avoidants are the result of a motivated defensive more general models, why does anxiety about relationships con-
strategy. tinue to be an important predictor? One reason may be that
One of the assumptions guiding this research is that working being comfortable with relationships is more closely associated
models of attachment were activated in memory and used to with models of others, whereas anxiety is more closely linked to
shape the social perception process. What evidence do I have negative models of the self (Bartholomew & Horowitz, 1991;
for this assumption? The strongest evidence was provided by Bowlby, 1982;Bretherton, 1985; Collins & Read, 1990;Feeney
comparing the content of participants' free-response explana- & Noller, 1990; Hazan & Shaver, 1987). Individuate are likely
tions for attachment-relevant and attachment-irrelevant events. to develop fairly specific and concrete models or schemas of par-
Over 50% of participants in Study 1 and over 30% of partici- ticular relationships and particular partners; however, models
pants in Study 2 spontaneously mentioned as least one of the of the self are not likely to be completely amended or sup-
eight attachment themes in their explanations. Moreover, across planted in the context of any one relationship (Collins & Read,
all attachment styles, participants were half as likely to mention 1994). As a result, positive or negative models of the self are
an attachment theme in response to attachment-irrelevant likely to remain as important structures through which social
events. Also, as predicted, individual differences in attachment information is filtered. Individuals who have negative self-im-
style were not associated with the content of these explanations. ages and who are concerned that others will abandon them may
These results are important for two reasons. First, they are con- find it difficult to completely set aside their doubts in any rela-
sistent with the argument that working models of attachment tionship. Holmes and Rempel (1989) suggested that people
828 COLLINS
who are unsure about their partner's commitment will con- uals may respond negatively to any situation in which then-
stantly monitor their partner's behavior for signs that indicate needs or expectations are not met by their partner, even if their
lack of caring. They may be waiting for things to go wrong in partner's behavior was not negatively motivated. This reasoning
their relationship and, when faced with a negative partner be- is consistent with the infant literature that indicates that preoc-
havior, will be ready to assume the worst. Preoccupied and fear- cupied children have a low tolerance for distress and tend to
ful adults, who are chronically worried about this issue, may be respond with fear and anxiety whenever there is any disruption
likely to adopt such a strategy and, as a result, negative images in their environment (see Kobak & Sceery, 1988, for further
of the self are likely to be chronically active. One important discussion of emotion regulation). A second factor that may
question is whether attachment style interacts with relationship explain this link is that adults with different attachment styles
quality and relationship length. For example, perhaps negative have different strategies for regulating and expressing emotion,
working models of selfwill lose some of their power in long-term and these strategies may be automatically evoked whenever
relationships that are characterized by high levels of trust and working models are activated (Kobak & Sceery, 1988; Shaver,
faith. A related issue is the extent to which working models are Collins, & Clark, 1996). Finally, working models of attachment
flexible and open to change (see Collins & Read, 1994; Scharfe contain a great deal of emotional content, which may be auto-
& Bartholomew, 1994). These are both important topics for fu- matically triggered whenever working models of attachment are
ture research. activated in memory (Fiske & Pavelchak, 1986). Additional re-
search is needed to explore these and other factors that may
underlie the link between attachment styles and patterns of
Attachment Style Differences in Emotions and Behavior emotions.
The second goal of this research was to provide an initial test The idea that attachment style has a direct influence on emo-
of the multivariate model linking attachment style and explana- tions raises the possibility that emotions have a reciprocal in-
tion processes to emotional response patterns and behavioral fluence on explanation processes. A large body of research
intentions. Drawing from a general model proposed by Collins points to the significant impact that affective experiences have
and Read (1994), it was hypothesized that attachment style on a variety of cognitive phenomena (see Bower & Cohen, 1982;
would be both directly and indirectly related to emotional re- Clark & Isen, 1982; Forgas, 1994). This association has been
sponses, and that the link between attachment style and behav- incorporated into the general model presented by Collins and
ioral descriptions would be mediated by explanations and Read (1994), which was shown in Figure 1. Unfortunately, in
emotions. the present context, a structural model that specified reciprocal
The pattern of results for emotional responses in both studies links between explanations and emotions was attempted, but
was highly consistent with predictions. As expected, preoccu- the model could not be properly estimated (the parameters were
pied adults consistently responded to the events with strong neg- "empirically underidentified"; Hayduk, 1987). Nevertheless,
ative emotion, whereas secure and avoidant adults reported there are good theoretical and empirical reasons to believe that
much lower levels of negative affect. It was also predicted that emotional responses will shape how an event is subsequently
secure adults would be willing to acknowledge moderate levels explained. Thus, the reciprocal influence of emotions and ex-
of distress, but that avoidant adults would actively deny feeling planations remains an important topic for future work.8
distress. This pattern was most clearly supported in Study 2. The importance of explanations and emotions in relation-
Avoidant adults reported less distress than secure adults, and ships lays, in part, on their presumed impact on social behavior.
were more likely to report that they felt unemotional in re- Explanations for events and emotional responses to them were
sponse to the events. both unique predictors of behavioral intentions, despite the fact
There are a number of mechanisms that may explain these that these two variables were highly correlated. It is noteworthy
patterns. First, results from the structural model suggest that that in Study 2, emotional responses were a much stronger pre-
attachment style differences in emotional experience are medi- dictor of behavioral descriptions than were explanation pat-
ated, in part, by variations in the appraisal and interpretation of terns. Thisfindingsuggests that when participants are respond-
events. Thus, in the present context, insecure people felt more ing to a hypothetical relationship, they might overestimate the
emotionally distressed partly because they tended to view their extent to which their behavior would be driven by cognitive
partner's behavior as having negative Implications for them-
selves and for their relationship.
Nonetheless, the path model also supported a direct relation 8
Although it was not possible to test a reciprocal link between expla-
between anxiety and emotions, even after controlling for differ- nation patterns and emotional responses in the current studies, it was
ences in explanation. There are a number of processes that may possible to examine an alternative model in which the path between
account for this direct link between anxiety and emotions. One explanations and emotions was reversed, This model was tested and, in
factor that may be important is differences in personal needs or both studies, was found to be an equally goodfitto the data. Thus, this
alternative model cannot be ruled out on statistical grounds. However;
goals that may be characteristic of people with different attach-
as discussed in Footnote 7, this model is less defensible on conceptual
ment styles (Collins & Read, 1994; Shaver, Coffins, & Clark, grounds given the procedures used in this study. The results of the cur-
1996). For instance, preoccupied individuals are thought to rent studies are consistent with the idea that negative attributions will
have a strong desire for attention and physical affection. When contribute to emotional distress, but the results cannot rule out the re-
this need is not met, or when this goal is blocked (Berscheid, verse, nor were they intended to do so. Future studies are planned to
1983), such individuals are likely to become distressed, regard- focus more specifically on the mutual influence between cognitive and
less of the explanation for the event. Thus, preoccupied individ- emotional processes.
WORKING MODELS OF ATTACHMENT 829
(i.e., rational) factors rather than by emotional factors (which goals) and by proximal factors present in the immediate social
tend to be viewed as more irrational and uncontrollable). How- context (e.g., transitory mood states and social norms). The
ever, replication of this pattern is needed before any conclusions current work suggests that individual differences in working
can be drawn about such processes. Finally, as predicted, the models of attachment may be one important component of the
influence of attachment style on negative behavioral patterns distal context.
was mediated through explanations and emotions. This research differs from related work in this area by focus-
Taken together, thesefindingssupport the hypothesized pro- ing on the content of open-ended explanations. Because of the
cess model that suggests that adults with different attachment considerable time and effort needed to code these materials, one
styles are predisposed to behave differently in relationships must consider the merits of their use. Perhaps the most impor-
largely because they think and feel differently within them- tant reason for gathering open-ended explanations is that the
selves. Moreover, they highlight the importance of incorporat- content of an explanation is vital to understanding the knowl-
ing both cognitive and emotional processes into models of in- edge that was used to construct it (Read & Collins, 1991). For
terpersonal behavior. Biased cognitive processes are likely to set example, the presence of spontaneous attachment themes pro-
into motion a pattern of emotional and behavioral responses vided support for the idea that attachment-relevant knowledge
that shape the interpersonal climate and day-to-day functioning structures were being used to filter and interpret events. This
of one's relationship. These processes may also contribute to idea could not be tested as clearly with only abstract attribution
the maintenance and stability of attachment-relevant knowl- dimensions. In fact, in Study 2, attachment style and relation-
edge structures as people create environments that verify their ship quality both were more strongly associated with the
existing expectations, both positive and negative (Bartholomew, content of the explanations than with the attributional dimen-
1993; Collins& Read, 1994;Swann, 1983). sions. Therefore, if one's goal is to understand how explanations
are constructed, content coding of explanations will be vital.
Conclusions and Caveats Before concluding, several important limitations of the pres-
ent work should be acknowledged. First, participants were
The application of attachment theory to the study of adult asked to respond to hypothetical partner behaviors on the basis
close relationships is still in its earliest stages. Nevertheless, of very restricted information. As a result, one's ability to gen-
many studies have shown that attachment style is an important eralize these findings to explanation processes in the natural
predictor of the nature and quality of adult love relationships. course of relationships may be somewhat limited. Fortunately,
Working models are a central feature of this approach, and are past research on attributions in relationships has found that re-
presumed to guide how people operate in relationships and how sults based on hypothetical events are very consistent with those
they construe their social world, although the mechanisms based on actual relationship events (see Baucom et al., 1989;
through which this occurs have been poorly understood. The Bradbury & Fincham, 1990). A second limitation concerns the
present research is thefirstto directly explore these mechanisms use of written behavioral descriptions to make inferences about
by incorporating attachment style into a model that includes actual interpersonal behavior. Of course, it is difficult to know
cognitive processes as well as patterns of affect and behavior. the extent to which such descriptions accurately reflect behav-
These studies provide strong evidence that attachment styles are ioral tendencies, and future studies will need to address this is-
indeed related to variations in social perception, which has been sue by observing behavior in laboratory orfieldsettings. Addi-
suggested but not yet demonstrated in the attachment litera- tional limitations are apparent in our measurement of emo-
ture. Of course, many other cognitive processes such as person tional response patterns. These measures were limited not only
perception, attention, and memory should also be influenced because I relied on people's predictions about how they would
by working models of attachment (Collins & Read, 1994), and feel, but also because emotional distress was conceptualized in
these deserve attention in future research. a very general way. In future studies I plan to examine the link
This work also has a number of implications for more general between attachment style and emotional responses by using
research on attributions in close relationships. The vast major- self-report as well as behavioral indicators of emotion, and by
ity of research in this area has been concerned with demonstrat- studying a wider range of emotions. Finally, although these data
ing a reliable association between attribution patterns and rela- were consistent with the hypothesized structural models, these
tionship quality based on the assumption that negative patterns data are correlational and cannot be used to draw conclusions
of attribution contribute to, or maintain, marital distress. Un- about causality. Although it is not possible to manipulate at-
fortunately, less attention has been paid to studying the mecha- tachment style, I am currently planning experimental tests of
nisms responsible for variations in attributional patterns. The these relationships by priming certain attachment-related be-
current study contributes to this effort by exploring one factor liefs and experiences in the lab. Longitudinal designs would also
that may underlie such patterns—individual differences in so- be useful in continuing research.
cial knowledge that partners bring into their relationships. This In conclusion, the present work provides insight into the
idea is consistent with recent models of cognition in close rela- mechanisms that may underlie attachment style differences in
tionships, such as Bradbury and Fincham's (1988) contextual relationship quality. More generally, it suggests that to fully un-
model of interpersonal-marital interaction, and Fletcher and derstand interpersonal functioning in adulthood, close relation-
Fincham's (1991) related model of explanation processes in ship scholars will need to explore the representational models
close relationships. These models suggest that cognitive pro- that guide social perception, emotion, and behavior. Attach-
cessing of relationship information will be shaped by relatively ment theory provides one valuable framework for continuing
stable distal factors (e.g., personality, chronic mood states, and this effort.
830 COLLINS
Sroufe, L. A., & Waters, E. (1977). Attachment as an organizational teenth Annual Carnegie Symposium on Cognition. Hillsdale, NJ:
construct Child Development, 48, 1184-1199. Erlbaum.
Swann, W. B. (1983). Self-verification: Bringing social reality into har- Weiner, B. (1985). An attributional theory of achievement motivation
mony with the self. In J. Suls & A. G. Greenwald (Eds.), Psychologi- and emotion. Psychological Review, 92, 548-573.
cal perspectives on the self- Vol. 2 (pp. 33-66). Hillsdale, NJ;
Erlbaum. Received February 13,1995
Weiner, B. (1982). The emotional consequences of causal attributions. Revision received February 23,1996
In M. S. Clark & S. J. Fiske (Eds.), Affect and cognition: The Seven- Accepted February 27,1996 •
The target audience for Review of General Psychology consists of those psychologists who
appreciate both generalism and specialization and who share a vision of psychology as a
unified discipline with common theoretical, methodological, and substantive values.
Authors are encouraged to write their manuscripts from the perspective of more than one
subdiscipline and to review literature that spans at least two subdisciplines in order to attract
a broad readership. In this era of heightened specialization, Review of General Psychology
is committed to the goal of publishing articles from which all psychologists and other social
and behavioral scientists can benefit, regardless of subfield or expertise.
Authors should prepare manuscripts according to the Publication Manual of the American
Psychological Association (4th ed.). All manuscripts must include an abstract containing
a maximum of 960 characters and spaces (approximately 120 words). Manuscripts that
exceed 50 pages in length (not including references, tables, and figures) generally will be
returned withoul review.