Child Development, May/June 2000, Volume 71, Number 3, Pages 684–689
Attachment Security in Infancy and Early Adulthood:
A Twenty-Year Longitudinal Study
Everett Waters, Susan Merrick, Dominique Treboux, Judith Crowell, and Leah Albersheim
Sixty White middle-class infants were seen in the Ainsworth Strange Situation at 12 months of age; 50 of these
participants (21 males, 29 females) were recontacted 20 years later and interviewed by using the Berkeley
Adult Attachment Interview (AAI). The interviewers were blind to the participants’ Strange Situation classifications.
Overall, 72% of the infants received the same secure versus insecure attachment classification in early
adulthood, ␬ ϭ .44, p Ͻ .001. As predicted by attachment theory, negative life events—defined as (1) loss of a
parent, (2) parental divorce, (3) life-threatening illness of parent or child (e.g., diabetes, cancer, heart attack), (4)
parental psychiatric disorder, and (5) physical or sexual abuse by a family member—were an important factor
in change. Forty-four percent (8 of 18) of the infants whose mothers reported negative life events changed attachment
classifications from infancy to early adulthood. Only 22% (7 of 32) of the infants whose mothers reported
no such events changed classification, p Ͻ .05. These results support Bowlby’s hypothesis that individual
differences in attachment security can be stable across significant portions of the lifespan and yet remain
open to revision in light of experience. The task now is to use a variety of research designs, measurement strategies,
and study intervals to clarify the mechanisms underlying stability and change.
INTRODUCTION
One of Bowlby’s primary goals in developing modern
attachment theory was to preserve what he considered
Freud’s genuine insights about close relationships
and development. These included insights
about (1) the complexity of social, cognitive, and emotional
life in infancy, (2) underlying similarities in the
nature of close relationships in infancy and adulthood,
and (3) the importance of early experience.
To preserve these insights, Bowlby recast Freud’s
insights in terms of control systems and ethological
theories. He also placed his own imprint on them by
replacing cathectic bonding with evolved secure base
patterns as the common thread in infant and adult relationships.
He also placed greater emphasis on the
openness of early relationships to change, especially
in light of real-life experiences.
Ainsworth’s observational studies of secure base
behavior at home and in the laboratory (Ainsworth,
Blehar, Waters, & Wall, 1978, Ch. 4, 5, 13) initially focused
on normative trends in infants’ responses to
novelty, separation, and reunion. Her goal was to test
the appropriateness of Bowlby’s control systems
model of infant behavior toward a caregiver. Subsequently,
individual differences designs proved useful
for examining the determinants and developmental
significance of secure base behavior (Ainsworth et al.,
1978, Ch. 7, 8, 14; Colin, 1996).
Working within Mischel’s (1968) critique of the individual
differences paradigm, Masters and Wellman
(1974) examined intercorrelations and stability in several
studies of infant behavior in brief laboratory separations.
They concluded that, consistent with Mischel’s
(1968) situationist critique of the individual
differences paradigm, there was little evidence of
consistency in correlations across discrete “attachment
behaviors” or of stability over intervals of
weeks, days, or minutes. These conclusions carried
considerable weight.
The present study began (Waters, 1978) as an effort
to clarify issues raised by the Masters and Wellman
(1974) review. Strange Situation data were collected
on a middle-class sample at 12 and 18 months of age.
In each episode, we counted the frequency of discrete
“attachment behaviors” and rated key interactive behaviors
(proximity seeking, contact maintaining,
proximity and interaction avoiding, and contact resisting).
In addition, we classified each infant as secure, insecure-avoidant,
and insecure-resistant at each age.
Reliability analysis indicated that most of the discrete
behaviors examined in the Masters and Wellman
(1974) review were far too rare to enable us to obtain
a reliable estimate of an infant’s typical behavior from
brief episodes. That is, measurement failure could explain
much of the negative evidence compiled by
© 2000 by the Society for Research in Child Development, Inc.
All rights reserved. 0009-3920/2000/7103-0013
This is one of three long-term longitudinal studies assessing infant
attachment. See Waters, Hamilton, & Weinfield, “The Stability
of Attachment Security from Infancy to Adolescence and Early
Adulthood: General Introduction,” for an overall view of study
design, measures, and supporting references.
Waters et al. 685
Masters and Wellman (1974). This interpretation was
strengthened by evidence that stability across episodes
and across time was much higher with the
broader (and thus more reliable) rating scales and
classifications. These results addressed the Masters
and Wellman critique in detail and, in doing so, buttressed
an emerging methodological defense of individual
differences research (e.g., Block, 1977; Epstein,
1978).As a result, they too carried considerable weight.
Lacking attachment security measures that could
be applied beyond infancy, few if any researchers in
the mid-1970s planned long-term follow-up assessments.
This obstacle was overcome with the development
and validation of the Berkeley Adult Attachment
Interview (Main, Kaplan, & Cassidy, 1985; see
Crowell & Treboux, 1995, for a review). As Vaughn,
Egeland, Sroufe, and Waters (1979) note, Bowlby’s
theory predicts that secure base use and attachment
representations are significantly stable across time and
yet open to change in light of significant attachmentrelated
experience. The goal of this follow-up study
was to examine the extent of stability and change in
attachment patterns from infancy to early adulthood
and to stimulate research into the mechanisms underlying
these developmental trajectories.
METHOD
Participants and Procedure
Sixty 12-month-olds recruited from newspaper
birth announcements in Minneapolis and St. Paul
were seen in the Ainsworth and Wittig Strange Situation
in 1975 and 1976. Most also participated in a 6month
follow-up at 18 months of age (see Waters,
1978). Fifty of these participants (21 males, 29 females)
were relocated 20 years later and agreed to
participate in the Berkeley Adult Attachment Interview
(George, Kaplan, & Main, 1985). Their ages at
the time of the AAI were from 20–22 years. As was
true for their families in the original study, their socioeconomic
status spanned the lower- to upper-middle
classes. Living arrangements were diverse: 45% lived
at college, 24% at home, 24% independently, 6% in
other arrangements (e.g., military). Seventy-two percent
described their primary occupation as “student”;
18% had completed high-school and were now employed;
4% had completed college and were now
employed; 6% did not mention employment. In most
instances (78%) the participants’ parents had remained
married. Two participants lost a parent before
age 6. Two participants had a child of their own.
Infant attachment assessment. Each participant was
seen in the Ainsworth Strange Situation at 1 year of
age. They were classified as secure, insecure-avoidant,
or insecure-resistant, as described in Ainsworth et al.
(1978). The insecure disorganized classification (Main
& Solomon, 1986) was not yet developed when we
scored these tapes. Independent coders assigned infant
attachment classifications at 12 and 18 months.
Each participant was classified by two independent
coders; eighteen-month data were scored without the
knowledge of 12-month classifications. Raters agreed
on major classifications in 45 out of 50 (90%) of the
cases (see Waters, 1978). Disagreements were resolved
by conference. The distribution of attachment classifications
at 12 months was 29 (58%) secure, 12 (24%)
insecure-avoidant, and 9 (18%) insecure-resistant.
Adult attachment assessment. Adult attachment status
was assessed by using the Berkeley Adult Attachment
Interview (George et al., 1985) when each participant
was from 20 to 21 years of age. Administration and
scoring procedures are summarized in the General
Introduction and detailed in Main and Goldwyn
(1994). The interviews were conducted by three of the
authors. Thirty-seven interviews were conducted in a
private room provided by a community library; three
participants were interviewed in their parents’ homes.
We interviewed 10 participants by telephone, nine
who had moved away from the Minneapolis area and
had no plans to visit and one who was at sea with the
Navy. The interviewers were blind to participants’ infant
attachment classifications.
Before scoring, each interview was typed, compared
with the audiotape, and if necessary corrected.
Two of the authors who had completed AAI training
seminars conducted by Dr. Mary Main served as
coders. Inter-rater agreement was assessed by using
25 of 50 transcripts. Agreement for this sample on the
three major attachment classification was 84%, ␬ ϭ
.72, p Ͻ .001. The distribution of AAI classifications
was 25 (50%) secure, 16 (32%) insecure dismissing,
and 9 (18%) insecure preoccupied. One participant in
each group was classified unresolved.
Negative life events. One of the cornerstones of
Bowlby’s theory is that attachment-related expectations
and working models remain open to revision in
light of changes in the availability and responsiveness
of secure base figures. That is, attachment theory predicts
both stability under ordinary circumstances and
change when negative life events alter caregiver behavior.
To test the hypothesis that changes in attachment
classification would be related to negative life events,
we obtained a score on negative life events from each
participant’s AAI transcript. Negative life events were
defined as (1) loss of a parent, (2) parental divorce, (3)
life-threatening illness of parent or child (e.g., diabetes,
cancer, heart attack), (4) parental psychiatric disorder,
686 Child Development
and (5) physical or sexual abuse by a family member.
The coders who counted negative life events did so
without knowledge of the participants’ Strange Situation
or AAI classification and without training in the
AAI scoring system. To allow time for the impact of
such events to be reflected in the AAI, we limited the
counts to events that had occurred before age 18. To determine
whether results were specific to this method of
ascertaining stressful life events, we examined events
reported by checklist 1 year later. Forty-seven completed
a checklist of life events that included all of the
events identified in theAAIs. This method depends less
on free-recall, the manner in which interview questions
are posed, the participant’s state of mind, and the
amount of material produced in theAAI. These data are
relevant to the present study and to the accompanying
studies that obtained life events from the AAI. Participants
were divided into those reporting none and those
reporting one or more of the target experiences. The one
or greater criterion was set a priori on the basis that all
of the target experiences would be considered major life
events in current research on stress and coping; each
has the potential, on its own, to change expectations
about caregiver availability and responsiveness.
Agreement on life events classification (none versus
one or more) by AAI and checklist was 78.7%, ␬ ϭ
.57, p Ͻ .002). Twenty-two participants were classified
“none” and 15 were classified “one or more” by both
methods. Eight were classified “one or more” by the
checklist but “none” by the AAI. Two were classified
“one or more” by the AAI but “none” by the checklist.
RESULTS
As hypothesized, early attachment security with
mother was significantly related to AAI attachment
security 20 years later (see Table 1). Using three classifications
at each age, 32 out of 50 participants (64%)
were assigned to corresponding classifications in infancy
and early adulthood, ␬ ϭ .40, p Ͻ .005; ␶ ϭ .17,
p ഡ .002 (AAI dependent).1 Thirty-six out of 50 participants
(72%) received the same classification using
the secure-insecure dichotomy, ␬ ϭ .44., p Ͻ .001; ␶ ϭ
.20, p ഡ .002.
Thirty-six percent of the participants changed classification
from infancy to early adulthood. Reliability
and validity problems with the attachment measures
certainly account for some portion of the observed
change. Nonetheless, the results also suggest that experiences
beyond infancy also play a role in adult security.
We examined this by counting the number of
attachment-relevant negative life events mentioned
in each participant’s AAI transcript and relating this
to whether the participant retained or changed attachment
classification across age. These results are
presented in Table 2. When mothers had reported no
stressful life events, attachment stability (three groups
each age) was 72%, ␬ ϭ .465, p Ͻ .009; ␶ (AAI dependent)
ϭ .23, p ഡ .006. For the secure versus insecure
dichotomy, stability was 78%, ␬ ϭ .525, Յ .009; ␶ (AAI
dependent) ϭ .28, p ഡ .003.
1 Cohen’s ␬ is computed from (1) the maximum level of
agreement possible (100%), (2) the proportion of concordant
cases (in the diagonal cells) expected by chance (from cross-multiplying
marginals), and (3) the observed proportion of agreements.
␬ is equal to the proportion of possible agreement over
and above chance that is actually obtained. In addition to the
significance test associated with ␬, the statistic itself can be construed
as an indication of effect size. To determine whether any
of the present results are specific to the statistic used, we also report,
where appropriate, an alternative concordance index
(Goodman & Kruskal’s ␶, by means of SPSS) based on a different
model of chance agreement levels. When computed with AAI
dependent, ␶ reflects the proportional reduction in error when
the Strange Situation classification is used to predict AAI classification.
Complete data from which other indices can be computed
are included in tables.
Table 1 Stability of Attachment Classifications from Infancy to Adulthood
Infant Attachment Classification
(Strange Situation at 12 months)
Adult Attachment
Classification (AAI)
Secure
(B)
Avoidant
(A)
Resistant
(C)
Secure (F) 20 2 3
Dismissing (D) 6 8 2
Preoccupied (E) 3 2 4
Note: S/S ϭ Strange Situation.
Stability:
64% (3 groups each age) ␬ ϭ .40, p Ͻ .005 72% (secure versus insecure) ␬ ϭ .44, p Ͻ .001
␶ (S/S dependent) ϭ .17, p ഡ .002 ␶ (S/S dependent) ϭ .20, p ഡ .002
␶ (AAI dependent) ϭ .17, p ഡ .002 ␶ (AAI dependent) ϭ .20, ഡ .002.
Waters et al. 687
Hierarchical multiple regression analyses were
used to determine whether (1) secure and insecure infants
were equally likely to change attachment classification,
(2) mothers of secure and insecure infants
were equally likely to report stressful life events, (3)
infants whose mothers reported experiencing stressful
life events were more likely to change attachment
classification from the initial to the follow-up assessment,
and (4) secure versus insecure infants whose
mothers report stressful life events were equally likely
to change classification. The analyses used stressful
life events (presence versus absence), infant attachment
classification (secure versus insecure), and their
interaction to predict whether infants’ attachment
classifications (secure versus insecure) changed or remained
the same over the course of the study.2
After first entering stressful life events, R2 change
for infant classification ϭ .01, F(2, 47) ϭ .50, p Ͻ .49.
Thus, there was no difference in the likelihood that
secure infants (31%, 9 of 29) and insecure infants
(28.6%, 6 of 21) would change classification from infancy
to early adulthood. After first entering infant attachment
classification, R2 change for presence or absence
of stressful life events ϭ .09, F(2, 47) ϭ 4.64, p Ͻ
.037. Thus, infants whose mothers reported one or
more stressful life events were more likely to change
attachment classification (44.4%, 8 of 18) than infants
whose mothers reported none (21.9%, 7 of 32). Finally,
after both attachment classification and stressful life
events were included in the analysis, the interaction
2 The results in Table 2 also suggest hypotheses about
changes from insecure to secure attachment in the absence of
stressful life events. These deserve to be pursued with appropriate
statistical power in a larger sample or meta-analysis of data
from several studies. Independent assessment of stressful life
events and caregiver–child interaction at several points between
the initial and follow-up attachment assessments would
also be useful.
term in the analysis was also significant, R2 change ϭ
.14, F(3, 46) ϭ 8.48, p Ͻ .006. Stressful life events were
significantly related to the likelihood of a secure infant
becoming insecure in early adulthood (66.6% if
mother reported one or more events versus 15% if she
reported none, p Ͻ .01) in secure infants. Stressful life
events were not significantly related to classification
changes in insecure infants. Among insecure infants
whose mothers reported one or more such events,
22% became secure as young adults versus 33.3% if
mother reported none (p Ͻ .59).
Although attachment-related stressful life events
were most often associated with changes from secure
to insecure attachment, this was not always the case.
One participant, whose parents responded with consistent
sensitive care to the childhood onset of a lifelong
illness, changed from insecure to secure. The relationship
between life events and attachment patterns
across time was not perfect. Eight participants reported
significant attachment-related stressful life events and
yet retained their infant attachment status in early
adulthood. Similarly, nine participants reported no
such events and yet changed attachment classification.
DISCUSSION
The present data provide strong evidence for the value
of the secure base concept as a conceptualization of attachment
relationships in infancy and adulthood. They
also support Bowlby’s expectation that individual differences
can be stable across significant portions of the
life span. Finally, they confirm the notion that, throughout
childhood, attachment representations remain open
to revision in light of real experience.
The success of the secure base concept as a conceptual
foundation for both the Strange Situation and the
AAI is important support for the notion that early
and late relationships have something in common.
Table 2 Relations of Stressful Life Events to Change in Attachment Classifications
Stability and Change from 12 Months to 21 Years
Number of Stressful
Life Events Reported
Retained Security
Classification on AAI
Changed Security
Classification on AAI
None
Total S/S sample (n ϭ 32) 25 (78%) 7 (22%)
Secure in S/S (n ϭ 20) 17 (85%) 3 (15%)
Insecure in S/S (n ϭ 12) 8 (75%) 4 (25%)
One or more
Total S/S sample (n ϭ 18) 10 (61%) 8 (39%)
Secure in S/S (n ϭ 9) 3 (33%) 6 (66%)
Insecure in S/S (n ϭ 9) 7 (89%) 2 (11%)
Note: S/S ϭ Strange Situation.
688 Child Development
Moreover, the present stability data support the notion
that these relationships are not merely similar in kind
but somehow developmentally related. Processes
that may be contributing to stability include (1) consistency
in caregiver behavior across time, (2) a tendency
toward persistence in early cognitive structures,
(3) the relatively moderate intensity and low
frequency of attachment-related stressful events in
this middle-class sample, (4) the effects of individuals
on their environment, and (5) stabilizing effects of
personality trait variables (Waters, Kondo-Ikemura,
Posada, & Richters, 1991). This study was designed to
stimulate interest and help in the design of research
into the roles that such mechanisms play in the consistency
of attachment stability over time.
A portion of the change noted in this study is attributable
to measurement error. Imperfect scoring
agreement introduces approximately 10% error at
each age. In addition, a similar amount of error is attributable
to the fact that neither the Strange Situation
nor the AAI is perfectly reliable; behavior observed in
a given assessment may not be entirely representative
of the person’s typical behavior (see Ainsworth et al.,
1978, and Crowell & Treboux, 1995, for test–retest data).
Correctly estimating these psychometric factors in
change is important to understanding our results.Accurately
assessing both stability and change is important;
minimizing either would be a mistake.As Vaughn, Egeland,
Sroufe, and Waters (1979) emphasized, Bowlby’s
attachment theory predicts both stability and change.
The portion of change in attachment classifications
that proved correlated with attachment-related stressful
life events provides important support for Bowlby’s
ideas about (1) the openness to change of attachment
representations, and (2) the importance of real-world
experiences in such change. Research on the mechanisms
through which experience leads to change in
attachment representations deserves high priority
in current attachment research. An important conclusion
from this study is that the AAI is sensitive enough
to experience to serve usefully in such work. The types
of events associated with change in attachment security
and the underlying mechanisms of change deserve
careful analysis in shorter-term longitudinal designs.
Middle-class samples offer both advantages and
disadvantages. They represent a large segment of the
population and are ordinarily accessible, cooperative,
and interested in research. This was evident in the fact
that each of the participants we recontacted agreed to
participate in the AAI. The educational level of middleclass
participants is also an asset because the AAI
makes heavy demands on a wide range of conceptual
and verbal abilities. At the same time, stability in
middle-class samples may reflect more than simply
the inherent stability of attachment security. Both a
relatively low rate of negative attachment-relevant experiences
and social support structures that buffer secure
base expectations against such experiences may
also contribute to the stability of secure attachment in
middle-class samples, just as consistent high levels of
stressful events contributes to the stability of insecure
attachment in disadvantaged samples.
Strong social support structures might reduce the
number or impact of negative experiences and thus
increase stability; they could also attenuate links between
negative experiences that occurred and attachment
stability. The best way to address these concerns
is to examine both the stability of attachment in other
populations and the mechanisms of change in close
detail to understand why any participant would stay
the same or change. The accompanying studies provide
important information about stability and change
in populations with very different patterns of caregiving
and life events.
ACKNOWLEDGMENTS
The initial phase of this study was supported by the
Young Scholars in Social and Emotional Development
Program of the Foundation for Child Development.
The Foundation also provided, on very short notice,
critical support for the longitudinal follow-up phase.
This project would not have been possible without
their support, which we gratefully acknowledge.
ADDRESSES AND AFFILIATIONS
Corresponding author: Everett Waters, Department
of Psychology, State University of New York at Stony
Brook, Stony Brook, NY 11794-2500; e-mail: everett.
waters@sunysb.edu. Susan Merrick and Leah Albersheim
are in private practice in Minneapolis, MN; Dominique
Treboux and Judith Crowell are at the State
University of New York at Stony Brook.
REFERENCES
Ainsworth, M., Blehar, M., Waters, E., & Wall, S. (1978). Patterns
of attachment. Hillsdale, NJ: Erlbaum.
Block, J. (1977). Advancing the psychology of personality:
Paradigm shift of improving the quality of research? In
D. Magnusson & N. Endler (Eds.), Personality at the crossroads:
Current issues in interactional psychology. Hillsdale,
NJ: Erlbaum.
Colin, V. (1996). Human attachment. Philadelphia: Temple
University Press.
Crowell, J., & Treboux, D. (1995). A review of adult attachment
measures: Implications for theory and research. Social
Development, 4, 294–327.
Waters et al. 689
Epstein, S. (1978). The stability of behavior: I. On predicting
most of the people much of the time. Journal of Personality
and Social Psychology, 37, 1097–1126.
George, C., Kaplan, N., & Main, M. (1985). The adult attachment
interview. Unpublished manuscript, University of
California at Berkeley.
Main, M., & Goldwyn, R. (1994). Adult attachment rating
and classification systems (version 6.0). Unpublished
manuscript, University of California at Berkeley.
Main, M., Kaplan, N., & Cassidy, J. (1985). Security in infancy,
childhood, and adulthood: A move to the level of
representation. In I. Bretherton & E. Waters (Eds.), Monographs
of the Society for Research in Child Development, 50
(1–2, Serial No. 209, pp. 66–106).
Main, M., & Solomon, J. (1986). Discovery of a new, insecure-disorganized/disoriented
attachment pattern. In
M. Yogman & B. Brazelton (Eds.), Affective development in
infancy. Norwood, NJ: Ablex.
Masters, J., & Wellman, H. (1974). Human infant attachment:
A procedural critique. Psychological Bulletin, 81,
218–237.
Mischel, W. (1968). Personality and assessment. New York:
Wiley.
Vaughn, B., Egeland, B., Sroufe, L., & Waters, E. (1979). Individual
differences in infant–mother attachment: Stability
and change in families under stress. Child Development,
50, 971–975.
Waters, E. (1978). The reliability and stability of individual
differences in infant–mother attachment. Child Development,
49, 483–494.
Waters, E., Kondo-Ikemura, K., Posada, G., & Richters, J.
(1991). Learning to love: Mechanisms and milestones. In
M. Gunnar & L. A. Sroufe (Eds.), Minnesota symposia on
child psychology: Vol. 23. Self processes and development (pp.
217–255). Hillsdale, NJ: Erlbaum.