JOURNAL OF CHILD PSYCHOTHERAPY

VOL. 29 NO. 2 2003 171 – 186

Attachment representations in school-age

children: the development of the child
attachment interview (CAI)

M ARY T ARGET, P ETER F ONAGY &

Y AEL S HMUELI- G OETZ
London

Abstract This paper introduces a new measure of attachment in middle childhood (8 – 13 years), an
interview extensively adapted from the AAI and with a new coding system. We report data from a series of
studies with clinical and normal populations, in which the psychometric properties of reliability, stability
and validity are tested and found to be satisfactory. Although work remains to be done, we see the Child
Attachment Interview as a promising new attachment measure.

Keywords Attachment; latency; middle childhood; child attachment interview.

Introduction
The development of attachment measures began with the assessment of infant
behaviour in the Strange Situation Paradigm. This procedure and the establishment of
its validity led to the ‘move to the level of representations’ (Main et al., 1985), in the
assessment of attachment patterns later in development. The greatest achievement here
was the Adult Attachment Interview (AAI) (George et al., 1985; Main, 1995). The
measure described in this paper has drawn on both the infant and the adult paradigms
and coding strategies, in an effort to produce an assessment of attachment in the middle
school years.
Measures designed to assess attachment organization in infancy and adulthood have
been widely applied and thus well established, but the study of attachment in early
and middle childhood has proven more problematic. The measurement of attachment
in infancy has been rightly restricted to the behavioural level whilst in adulthood it
has been measured through language and representations. As Ainsworth (1990)
argued, the chief concern in using a separation-reunion procedure comparable to the
Strange Situation beyond infancy is that with increasing age, the degree of stress
induced decreases as the child is gradually exposed to everyday separations of greater
length.

Journal of Child Psychotherapy

ISSN 0075–417X print/ISSN 1469-9370 online ª 2003 Association of Child Psychotherapists
http://www.tandf.co.uk/journals
DOI: 10.1080/0075417031000138433
172 M. TARGET ET AL.

In parallel, a plethora of instruments designed to elicit mental representations of

attachment in early and middle childhood have been developed, all sharing the
assumption that inferred mental representations reflect children’s attachment organiza-
tion. Semi-projective measures eliciting mental representations through drawings
(Separation Anxiety Test SAT: Shouldice and Stevenson-Hinde, 1992; Slough and
Greenberg, 1990), family photos and drawings (Main et al., 1985), story stems
(Bretherton et al., 1990), and doll play (Solomon et al., 1995) have also been employed
with mixed results. Whilst these studies demonstrated associations between classifica-
tions derived behaviourally and representationally, the need to replicate such findings
(Main, 1995), low test-retest reliability (Wright et al., 1995), and questions of validity
(Bowers et al., 1994) highlight the need for further work.
Hence we were interested in trying to develop age appropriate measures for assessing
how attachment patterns are manifested in middle childhood. Hitherto, there has been
an assumption that children would not respond meaningfully when asked directly about
attachment experiences. However, Ammaniti and his colleagues have extensive
experience of administering a slightly modified version of the AAI protocol to early
adolescents and pre-adolescents, and the interview material is coded using the usual AAI
coding procedure (Ammaniti et al., 1990, 2003). Similarly, Trowell has used the AAI in
an important London study of sexually abused preadolescent girls, and found it
acceptable (Trowell, personal communication). Adopting a representational approach,
most measures have derived attachment classifications based solely upon an analysis of
children’s verbal responses. However, non-verbal communication, not limited to
separation-reunion behaviour, may be a very useful source of information in identifying
distinct attachment patterns, and would go some way towards integrating representa-
tional and behavioural approaches to the study of attachment.
The Child Attachment Interview (CAI) was thus developed in an attempt to
complement existing attachment measures. Independently of the present authors, Dante
Cicchetti and his colleagues developed a similar protocol, and have been administering it
for a period of 10 years (Cicchetti, personal communication), but without a coding
system. The present paper reports the development of our CAI protocol and coding and
classification system and presents the major psychometric properties of the measure.

Participants
The total sample comprised a number of subgroups: 161 children aged 7 – 12, without
known mental health problems, recruited from urban and rural schools. In addition
there were 65 children referred for psychiatric assessment in specialist child mental
health services. The demographics of these children, divided into referred and not
referred, are shown in Table 1.
The mean age at interview was somewhat higher in the non-referred group, as was the
proportion of girls, of middle class families, non-white children and those living in two-
parent households. With the exception of age, which was significantly higher among the
non-referred children because of the inclusion of a group of early adolescents, none of
these differences approached statistical significance, that is, the referred children were
comparable to those in the ‘normal’ group.
 ATTACHMENT REPRESENTATIONS IN SCHOOL-AGE CHILDREN 173

Table 1 Demographic characteristics of the two sample groups (not referred and referred)

Psychiatric referrals vs. non-referred

Not referred Referred

(n = 161) (n = 65) Statistic

Mean age (years) (SD) 11.1 (1.6) 10.2 (1.3) t = 4.1, df = 224, p 5 0.001
Mean verbal IQ (SD) 99.2 (18.8) 102.9 (18.3) t = 1.1, df = 156, n.s.
% boys 50.3% 58.5% w 2 = 1.2, df = 1, n.s.
% middle class 40.2% 33.9% w2 = 0.65, df = 1, n.s.
% white 70.0% 82.0% w2 = 4.4, df = 2, n.s.
% living with two parents 47.1% 44.4% w2 = 0.50, df = 1, n.s.

Procedure

Administration
Two interviewers with experience in the administration of the interview conducted the
assessments. The CAI formed part of a larger battery of measures including, amongst
others, measures of expressive language and IQ, administered over one to two sessions.
The CAI was completed first and conducted in a private room with interviewer and
child sitting face to face. Before the beginning of each assessment, the interviewer
explained the nature of the study and ensured that the child felt at ease and consented to
take part. The duration of the interview ranged from 20 min to 1 h, and the sessions
were videotaped. For 28 children, all those recruited in the second phase of data
collection, the AAI was also administered to the mother. All parents were asked to
complete the Child Behaviour Checklist (CBCL) (Achenbach and Edelbrock, 1983).
Following an interval of approximately 2 months, children recruited in the second phase
were assessed again by the same interviewer, when only the CAI was completed, for
evaluation of test-retest reliability.

Coding
There were three independent judges familiar with current attachment assessment
methodologies, and involved in developing the current coding system. The first (YSG),
then a doctoral student, coded the total sample, the second and third judges, final year
Clinical Psychology trainees, each coded one half of the sample. Coding was based on
the video-recorded interviews, to allow a behavioural as well as a linguistic analysis.

Measures

The CAI Protocol

The development of the interview protocol was conceptually based on the Adult
Attachment Interview (George et al., 1985) with several criteria in mind. First, akin to
174 M. TARGET ET AL.

the AAI, the CAI needed to activate the attachment system so as to elicit attachment-
related information. Second, whilst the interview needed to be constructed so as to reveal
structural variations in presentation, it also needed to be flexible enough to help children
with the demands placed upon them, but without compromising validity. Third, in
contrast to the AAI, we decided that the CAI should focus on recent attachment-related
events and how the current relationships with each parent were represented.
Guided by the above criteria, the questions comprising the interview were initially
taken from the Berkeley Autobiographical Interview (Main et al., 1985) and the AAI,
and adapted for use with children in the 7 – 12 years age range. The version of the
interview reported on below (a second version, extensively modified following piloting)
comprised 14 questions plus probes:

(1) Who is in your family? (lives with you in your house).

(2) Tell me three words that describe yourself (examples).
(3) Can you tell me three words that describe what it’s like to be with your mum
(examples)?
(4) What happens when mum gets upset with you?
(5) Can you tell me three words that describe what it’s like to be with your dad
(examples)?
(6) What happens when dad gets upset with you?
(7) Can you tell me about a time when you were upset and wanted help?
(8) What happens when you’re ill?
(9) What happens when you hurt yourself?
(10) Has anyone close to you ever died?
(11) Is there anyone that you cared about who isn’t around anymore?
(12) Have you ever been away from your parents for the night or for longer than a
day?
(13) Do your parents sometimes argue? Can you tell me about a time when that
happened?
(14) In what ways do you want/not want to be like your mum/dad?

The CAI coding and classification system

We adopted several principles in developing the CAI coding and classification system.
First, we would not assume that the existing AAI coding system would be appropriate to
the CAI. Second, we should assign attachment classifications separately for mother and
father, and assess whether there were singular or multiple internal working models
within this age range. Third, we segmented the interview into descriptions of
interactions with parents, termed Relationship Episodes (REs). The concept of REs was
informed by Luborsky’s Core Conflictual Relationship Theme method, in which REs
identified from psychotherapeutic sessions were studied (Luborsky and Crits-Christoph,
1990). Identifying REs revealed the richness of the information elicited and highlighted
the importance of not only the linguistic content and form of the narrative, but also
non-verbal communication as a key source of information.
 ATTACHMENT REPRESENTATIONS IN SCHOOL-AGE CHILDREN 175

The majority of the scales given below aimed to assess the child’s overall current state
of mind with respect to attachment, a state of mind which is assumed to be reflected in
the narrative as a whole. However three of the scales, namely, Preoccupied Anger,
Idealization and Dismissal were rated separately for mother and father and all ranged
from 1, denoting a low score to 9, denoting a high score.

Emotional Openness
The Emotional Openness scale was developed in order to assess the child’s ability to
express and label emotions, and to ground them in descriptions of interactions with
attachment figures. We were influenced by Sroufe’s (Sroufe and Fleeson, 1986) affect-
regulation model, and studies which have identified emotional openness as an important
aspect of children’s attachment-related narratives and a marker of security of attachment
(Oppenheim, 1997; Slough and Greenberg, 1990; Wright et al., 1995).

Preoccupied Anger
The CAI Preoccupied Anger scale was developed as an age-appropriate modified version
of the Involving Anger scale of the AAI (Main and Goldwyn, 1994). We found that it
was vital to underline the involving nature of the anger, and (in contrast to the AAI) to
include involving denigration or contempt, as well as anger itself.

Idealization
The CAI Idealization scale was also conceptually based upon the AAI Idealization scale
but was modified to reflect the responses given by children. It aimed to measure the
extent to which the child attempted to present an unsupported picture of an ‘ideal’
parent.

Dismissal
This scale was used to assess active denial of attachment and the presentation of parents
and attachment experiences as unimportant.

Self-organization
This scale attempted to capture the child’s internal representation of self-efficacy, based
on the presence of self-initiated and constructive conflict resolutions (Cassidy, 1988;
Oppenheim, 1997; Sroufe et al., 1983).

Balance of positive/negative references to attachment figures

This scale was based on the assumption that secure children would more readily
recognize and integrate positive and negative aspects of parental figures, thus presenting
a better-integrated and more balanced description of attachment figures.
176 M. TARGET ET AL.

Use of examples
Children’s ability to provide relevant and elaborated examples was also considered a
possible marker of security of attachment, as in the AAI where this is a key aspect of
coherence.

Resolution of conflicts
Children’s ability to describe constructive resolutions to conflicts has been closely linked
to attachment security (Oppenheim, 1997) and was thus included in the CAI.

Overall coherence
Whilst no a priori assumptions were established concerning the centrality of the
coherence of transcript in determining the child’s attachment classification, it was
considered an important dimension. The scale was rated on the basis of scores for
‘Idealisation’, ‘Preoccupied Anger’, ‘Dismissal’, and the ‘Use of Examples’, together with
a consideration of the overall qualities of consistency, development and reflection.
Alongside the linguistic analysis, a simple behavioural analysis of children’s responses
to the interview situation and questions was included. Maintenance of eye contact,
changes in tone of voice, marked anxiety, changes of posture in relation to the
interviewer and contradictions between verbal and nonverbal expressions were
considered when assessing emotional openness, coherence, idealization, preoccupied
anger, etc.
Attachment classifications with respect to mother and to father, independently, were
arrived at using an algorithm for combining the scale ratings. For instance, to obtain a
Secure classification, the child must have been assigned a rating of approximately five or
above on all CAI scales with the exception of the Idealization, Dismissal and
Preoccupied Anger Scales where a score of three or less was expected. In our first coding
scheme, the results of which are reported in the present paper, we further assigned a
level of security: Secure/Very Secure/Insecure/Very Insecure, with respect to mother and
father. Again, we specified algorithms for making this judgement. (Since the analyses
reported below, and others, we have considerably developed the classification scheme,
for example by incorporating qualitative differences within this basic framework).
A copy of the complete CAI Protocol, and Coding and Classification Manual, can be
obtained from the first author.

Results
The results are presented in four main sections. The first question we looked at was
whether coders could agree on their ratings, so first we report inter-rater reliabilities for
scales and main and sub-classifications. Our next question was whether the secure-
insecure types emerged from the scale scores. We address this by reporting the internal
consistency of the coding system. Thirdly, we examined whether the child’s interview
behaviour would be consistent over time (the interview would be of little use as a
 ATTACHMENT REPRESENTATIONS IN SCHOOL-AGE CHILDREN 177

measure if the child said different things in different ways on each occasion): test-retest
reliability over a 2-month period is presented, followed by an examination of AAI-CAI
concordance for a sub-sample. Fourthly, we examined discriminant validity, to assess
whether the child’s attachment classification was significantly related to variables such as
age, gender and IQ. Finally, we examined aspects of predictive validity: does the CAI
relate in expected ways to mother’s AAI? We report the results of these analyses below.

Internal consistency of CAI scales

Internal consistencies between three sets of scales were calculated: the five ‘state of mind’
scales (use of examples, balance, emotional openness, conflict resolution, and coherence)
were highly intercorrelated, with a standardized item alpha of 0.92. This indicates
statistically that the scales were tapping into a single construct. We also calculated the
association between the four scales rated separately for mother and father (anger,
dismissal, idealization and level of security). The standardized item alphas for these
scales were moderate: in relation to mother 0.65, and father 0.55; this indicates that the
three types of ‘insecure’ narrative, together with the overall level of security, when
examined for representations of mother and father separately, did not cohere such that
these scales seemed to be measuring a single entity. However, these three areas of our
measurement of child attachment (state of mind, representations of mother and father)
were very highly correlated: the standardized item alpha for these scales together was
0.94. Thus, a single variable summarizing security of attachment was created from these
three sets of variables, and the measure was taken to be measuring a single construct
(attachment).

Inter-rater reliability for CAI scales

Inter-rater reliability was computed twice, initially for three coders, and then with an
improved coding system for two coders. Table 2 presents the intraclass correlations
(ICCs) across three coders, and Pearson’s correlations between pairs of coders. These
statistics measure the extent to which a group of raters or a pair of raters, respectively,
agree in their independent judgements. In the first attempt at establishing inter-rater
reliability, only one scale (idealization of father) yielded unacceptable ICCs, with the
confidence interval including a negative correlation. The median ICC was 0.88 which
indicates very strong agreement between the three coders. The second assessment of
inter-rater reliability, across 50 cases, also showed a high correlation between two raters,
the median r being 0.87 with no unacceptably low agreements.

Inter-rater reliability for CAI main classifications

Main classifications, namely Secure or Insecure with respect to mother and father, were
assigned. Agreement was assessed using the kappa statistic and Kendall’s tau-b, which
are standard measures of agreement between independent coders on a categorical
judgement, e.g. secure vs. insecure. The relationships between classifications given by
two and three coders are shown in Table 3; they are consistently high. For the three
178 M. TARGET ET AL.

Table 2 Inter-rater reliability of scale scores

ICC for 3 coders (cases = 30) Pearson r for 2 coders (cases = 50)

Emotional Openness 0.92 (0.85 – 0.96) 0.91 (0.85 – 0.95)

Balance 0.80 (0.63 – 0.90) 0.83 (0.72 – 0.90)
Use of Examples 0.87 (0.76 – 0.93) 0.87 (0.78 – 0.92)
Anger with Mother 0.82 (0.66 – 0.91) 0.94 (0.90 – 0.97)
Anger with Father 0.75 (0.52 – 0.88) 0.66 (0.47 – 0.79)
Idealization of Mother 0.71 (0.46 – 0.85) 0.89 (0.81 – 0.94)
Idealization of Father 0.38 ( 7 0.15 – 0.69) 0.74 (0.58 – 0.84)
Dismissing of Mother 0.94 (0.89 – 0.97) 0.79 (0.66 – 0.88)
Dismissing of Father 0.94 (0.89 – 0.97) 0.79 (0.66 – 0.88)
Conflict Resolution 0.88 (0.79 – 0.94) 0.84 (0.73 – 0.91)
Coherence 0.90 (0.82 – 0.95) 0.90 (0.83 – 0.94)
Level of security: Mother 0.91 (0.83 – 0.95) 0.89 (0.81 – 0.94)
Level of security: Father 0.90 (0.81 – 0.95) 0.89 (0.81 – 0.94)

Table 3 Agreement on attachment classifications across 3 coders, and 2 coders

3 coders (cases = 30) 2 coders (cases = 50)

median kappa (range) Kappa

Mother Father Mother Father

Secure/insecure 0.92 (0.84 – 0.92) 0.92 (0.85 – 0.92) 0.79 0.84

3-way 0.84 (0.84 – 0.85) 0.86 (0.78 – 0.93) 0.58 0.66
4-way 0.83 (0.74 – 0.89) 0.86 (0.77 – 0.89) 0.60 0.54
Disorganization n too small to estimate 0.79 0.88

Median Kendall’s tau-b (range) Kendall’s tau-b

Level of security 0.58 (0.52 – 0.59) 0.67 (0.56 – 0.74) 0.82 0.84

coders, the number of disorganized classifications was too small to estimate agreement,
but with a larger group of cases yielded acceptable kappas, although they were somewhat
low for three- and four-way categorization. The level of security rating was substantially
improved.

Test-retest reliability (stability)

Forty-six children were retested 3 months after the first CAI (Tables 4 and 5) to find out
whether their attachment representations were similar on the two occasions.
On the whole, stability coefficients were quite high, and the median is 0.63. There
was considerable variability in the stability of the scales, for example while Anger with
Mother appeared to be highly stable across 3 months, Anger with Father was far less so.
Also, Idealization of both parents was somewhat unstable, but by contrast, Emotional
Openness, the Use of Examples, and Coherence seemed highly consistent between
 ATTACHMENT REPRESENTATIONS IN SCHOOL-AGE CHILDREN 179

Table 4 Test-retest reliability (stability) of scale scores

Test-retest: Test-retest:
3 months 1 year
(n = 46) (n = 33)

Pearson r
Openness 0.70 0.63
Balance 0.55 0.35
Examples 0.66 0.57
Anger M 0.90 0.54
Anger F 0.29 0.25
Idealize M 0.52 0.25
Idealize F 0.42 0.08
Dismiss M 0.71 0.44
Dismiss F 0.63 0.39
Conflict Res 0.58 0.34
Coherence 0.68 0.75

Table 5 Test-retest reliability (stability), at 3 months and 1 year, of classifications

of security with mother and father

3 months (n = 46) 1 year (n = 33)

Mother Father Mother Father

Kappa
Secure/insecure 0.74 0.68 0.73 0.68
3-way 0.77 0.64 0.79 0.71
4-way 0.78 0.67 0.78 0.66
Disorganization 1.00 1.00 0.72 0.52
Kendall’s tau-b
Level of security 0.79 0.75 0.65 0.60

testings. Of the parent-specific scales, only Dismissing was highly stable for both mother
and father. The classification across 3 months was quite stable across 46 cases. The
security classification for representation of mother was 0.75 or above, and for father was
0.65 or above. Interestingly, all the children who were coded disorganized on one
occasion were coded the same 3 months later.
Thirty-three children were retested 1 year after their initial assessment (Tables 4 and
5). Not surprisingly, the stability of the scale scores was moderate across this longer
interval, median correlation 0.40. Again, there was considerable spread: Emotional
Openness, Use of Examples, and particularly Coherence, were quite stable, whereas the
parent-specific scales particularly Idealization and Anger with Father, were quite
variable. Nevertheless, the classification arrived at by coders was relatively stable, and
only slightly below the coefficients obtained with a gap of 3 months between testings.
 180
M. TARGET ET AL.
Table 6 Relationship between attachment classification with mother and father, and demographic variables and verbal IQ

Mother Father

Secure Insecure Statistic Secure Insecure Statistic

Age: mean (SD) 11.3 (1.5) 10.8 (1.7) F = 3.2, df = 1,184, 11.2 (1.5) 10.9 (1.7) F = 2.1, df = 1,174,
n.s. n.s.
Verbal IQ: mean 99.5 (18.9) 101.0 (18.6) F = .20, df = 1,117, 99.6 (19.6) 99.9 (18.6) F = 0.01, df = 1,109,
(SD) n.s. n.s.
No. (%) boys 54 (46.2%) 38 (55.1%) w2 = 1.1, df = 1, n.s. 51 (48.1%) 36 (51.4%) w251, df = 1, n.s.
No. (%) middle class 35 (48.6%) 17 (35.4%) w2 = 1.5, df = 1, n.s. 32 (49.2%) 17 (34%) w2 = 2.1, df = 1, n.s.
No. (%) white black 86 (73.5%) 23 47 (70.1%) 13 w2 51, df = 2, n.s. 77 (72.6%) 22 50 (72.5%) 12 w251, df = 2, n.s.
asian (19.7%) 8 (6.8%) (19.4%) 7 (10.4%) (20.8%) 7 (6.6%) (17.4%) 7 (10.1%)
No. (%) living with 34 (46.6%) 31 (45.6%) w251, df = 1, n.s. 33 (53.2%) 32 (47.1%) w251, df = 1, n.s.
both parents
 ATTACHMENT REPRESENTATIONS IN SCHOOL-AGE CHILDREN 181

Overall, these test-retest reliability figures are encouraging, and suggest that generally
speaking children’s security classifications can be expected to be stable.

Discriminant validity
The vital question arises as to whether stability in manner of attachment representation
is observed because of actual consistency in attachment status, or whether it reflects
other stable aspects of the child, such as IQ, gender, socio-economic status, expressive
language capacity, or ethnicity. This issue of discriminant validity is addressed, for the
non-referred sample, in Table 6.
There was no statistically significant difference between the mean age of children
classified as secure or insecure with each parent. There was a slight and non-significant
tendency for children insecure with their mothers to be younger. Neither gender nor
social class predicted security of attachment, and the prevalence of black or Asian
children was comparable in the secure and insecure groups. In this non-clinical sample,
the percentage of children living with both parents was no higher in the secure than in
the insecure group. Importantly, Verbal IQ was almost identical among children with
secure vs. insecure representations of attachment security with each parent. This is
crucial given the weight attached to linguistic coherence in the coding of attachment
representations. On a subsample of 88 children, expressive language scores were also
collected. There was a slight but statistically non-significant superiority for children
secure with their mothers in their expressive language at the time of the CAI
administration (F = 2.9, df = 1,86, n.s.). Security with respect to father was associated
with expressive language to a similar degree (F = 2.7, df = 1,82, n.s.).

Predictive validity
The relationship between mothers’ current state of mind with respect to attachment as
assessed by the AAI, and their children’s attachment status as assessed by the CAI, was
examined for 75 children. The correspondence between main attachment classifications
for mother-child dyads was highly significant (64% agreement; k = 0.29, p 5 0.01).
Twenty-one of the 39 children rated as Secure as assessed by the CAI had Secure
mothers as assessed by the AAI (54%). Twenty-seven of the 36 children classified as
Insecure by the CAI had Insecure mothers as classified by the AAI (75%). Interestingly,
none of the seven children classified as Preoccupied, and only one of the six children
classified as Disorganized, had mothers with AAIs classified as Secure (w2 = 10.2, df = 3,
p 5 0.02). The association was just as strong between mothers’ AAIs and the child’s
attachment security with respect to father (65% agreement; k = 0.29, p 5 0.01). Again,
none of the eight children classified as Preoccupied with respect to Father, and only one
of the six children classified as Disorganized, had mothers with AAIs classified as Secure
(w2 = 11.4, df = 3, p 5 0.01). Exploring these associations from the point of view of
mothers’ attachment classification, it once again seemed that Preoccupied and Secure
classifications were more predictive (in opposite directions) of the child’s security: 20 of
29 (69%) of Secure/Autonomous mothers had children whose CAIs were secure with
respect to mother, and 19 of 29 (66%) had children who were secure with respect to
182 M. TARGET ET AL.

father; 18 of 25 Preoccupied mothers had children with Insecure classifications with

respect to mother (72%), and 82% were insecure with respect to father. Unresolved (U/
d) classification on the AAI (n = 36) did not predict child insecurity in relation to either
father or mother. Whilst U/d did not predict insecurity, Atypical (disorganized) CAI
classifications with respect to both mother and father were only found in cases where the
mother’s AAI had been classified as Unresolved.

Discussion
In addressing the existing ‘measurement gap’ in measures of attachment for middle
childhood, the series of studies reported here had three aims. First, to construct a
developmentally sensitive interview protocol for the assessment of attachment in middle
childhood, and to develop a coding and classification system. The second was to
establish the psychometric properties of the newly developed system. The third was to
suggest where these studies need to go next. The summary and discussion of findings
will follow this logic used in developing the measure.

The CAI protocol and coding system

Underpinning the development of the CAI protocol was the assumption that children
would be able to comprehend and thus respond to direct questions concerning
attachment experiences and relationships, and that variations in the presentation of
these experiences would reflect their internal attachment organization. Whilst piloting
Version I of the CAI interview protocol clearly demonstrated that children could
understand and respond coherently to direct questions concerning attachment-related
themes, it also highlighted the need for refinements. Version II of the CAI protocol was
subsequently devised and included more focused prompts used as ‘scaffolding’ to assist
children in producing attachment related narratives.
We are still refining and experimenting with means of coding and classifying these
interviews. The version reported here involved a set of nine-point scales adapted from
the AAI for this age range, a binary secure-insecure distinction, four levels
(subclassifications) of security-insecurity, and a four category qualitative coding which
largely maps onto the AAI coding system. Our current priority is to improve the
psychometric properties of the system as a whole by further defining the Preoccupied
and Disorganized categories, by studying tapes of interviews with children from clinical
samples, and developing a simpler coding scheme (using a procedure more often used in
personality research, called a Q-Sort), requiring less inference and knowledge of
attachment theory, for reliable use by less experienced coders. Our early results on this
are promising.

Psychometric properties
Internal consistency of the scale scores was high, and supported the assumption that the
measure was indeed tapping a coherent construct and quality in the children’s responses.
However, these results might reflect the measurement of a construct associated with
 ATTACHMENT REPRESENTATIONS IN SCHOOL-AGE CHILDREN 183

security of attachment such as intelligence, hence the need for further investigations as
reported.
The next step was to establish the degree of agreement between raters. Agreement
between judges for all CAI scales was high, and was improved to an excellent level by
refining the coding manual. Inter-rater agreement for the secure-insecure distinction
and for the overall level of security was in addition shown to be high for both mother
and father, showing that the current CAI coding and classification system allows
experienced raters consistently to distinguish between Secure and Insecure interview
responses. Agreement for the four main classifications with respect to both mother and
father was acceptable, with the Preoccupied category the least frequent and least reliable.
This mirrors experience with the AAI, in which Preoccupied attachment status is
relatively unusual and probably harder to recognize confidently than Secure and
Dismissing patterns. We were encouraged that disorganization of attachment strategy,
reflected at either a behavioural or a representational level, was easy to agree on for this
sample.
Test-retest reliability (stability) across a 3-month or 1 year period, using a different
coder for the two time points, produced somewhat mixed results. Consistency in
scale scores was adequate at 3 months, but only found for certain scales after a gap
of a year. However, the stable scales included the most important ones: Coherence,
Use of Examples and Emotional Openness were all very stable, while Dismissing
ratings were also reasonably consistent. Test-retest reliability for the Secure-Insecure
distinction and for the four main attachment classifications and with respect to both
mother and father were all remarkably high, even after an interval of 1 year. This
stability of the Secure-Insecure distinction was comparable with reported infant and
adult data (in adulthood, Bakermans-Kranenburg and van IJzendoorn (1993),
reported stability across 1 – 15 months ranging from 77 – 90%; Lamb et al. (1985),
found 77.1%; Waters (1978), found 96% stability in infants across a 6-months
interval). Stability of scales has been scarcely reported with respect to other
attachment measures. Waters (1978) reported that reliability of discrete behaviour
variables in the Strange Situation was very low across a 6-month period. Wright et
al. (1995) reported that test-retest reliability for the Separation Anxiety Test
following a 4-week interval did not reach statistical significance. AAI stability data
has not been presented, to our knowledge.
We had been apprehensive about whether we would find that the CAI showed
discriminant validity. Would relatively ‘secure’ attachment narratives turn out to be
simply a product of higher IQ or expressive language skills, of being older, middle class,
a girl, or of any other extraneous variable? The evidence thus far is that, at least in
normal samples, these concerns have not been borne out. It seems that the CAI is
measuring something which is weakly related to but not accounted for by major
demographic and cognitive variables.
Predictive validity for attachment within mother-child dyads (AAI-CAI) was shown to
be high, and comparable to that reported for infant-mother concordance rates based on
the Strange Situation. The prediction seemed to be strongest in cases where the child
was insecure, whereas a relatively high proportion of secure children had insecure
mothers. Preoccupation and Disorganization of attachment behaviour or narratives were
184 M. TARGET ET AL.

particularly likely to be associated with insecure attachment classification of the mother’s

AAI.
The findings presented here suggest that it is unnecessary to adopt a projective
approach in assessing attachment status in middle childhood. Children can respond to
direct questioning concerning attachment-related experiences and their responses appear
to reflect their internal attachment organization.
There are many further refinements which we are working on, particularly to the
coding procedure. We have space only to give one example. One of our aims in the
development of the CAI coding and classification system was to integrate both linguistic
and behavioural information. However, the attachment coding reported here was based
primarily upon a linguistic analysis of the content and form of attachment-related
narratives, and the integration of nonverbal information was, although guided by theory
and by the behavioural coding of attachment behaviour in preschool children, still
largely intuitive. The need to develop a coding system that incorporates detailed
behavioural information is important, not only because such information will
potentially illuminate differences in attachment organization that may otherwise not
be detected, but also because such an approach would go some way to bridge the gap
between the study of attachment in infancy and adulthood. The CAI could be a unique
tool in that the child’s behaviour during the interview forms the background against
which the child’s linguistic representation of attachment figures and relationships can be
assessed. Since the presentation summarized here, we have begun to explore the use of
some principles from facial-action coding and from mother-infant interaction studies
(the work of Anstadt et al. (1997) and Beebe et al. (1997) respectively).
Although beyond the scope of the present report, we have undertaken qualitative
analyses of the relationship between self-representation and attachment relationships.
The self-concept question that opens the interview suggests that – as one might expect
from the developmental and clinical literatures – secure children differ from their
insecure counterparts in describing the self. Their self-descriptions are likely to be more
mixed and better rooted in specific examples. It is also noticeable that secure children
often introduce humorous, self-deprecating descriptions and examples (e.g. ‘I think I’m
funny, but my family don’t laugh at my jokes . . . ’ ‘Why do you think that is?’ ‘Probably
because I’ve told them a thousand times before!’), which suggest more acceptance and
liking of the self than do the idealized self-descriptions (‘great at football’, ‘normal’)
sometimes offered by dismissing children, or the often troubled and troubling ones
offered by preoccupied children (‘always there to help my Mum’, ‘hidden’). There is a
wealth of material to explore in these interviews, both quantitative and qualitative.
We believe that, taking the above findings and our continuing work together, we are
on the track of a useful new assessment tool for research and theory in both the
developmental and the clinical realms, and we look forward to offering further reports of
the work as it proceeds.
Subdepartment of Clinical Health Psychology
University College
Gower Street
London WC1 UK
e-mail: m.target@ucl.ac.uk
 ATTACHMENT REPRESENTATIONS IN SCHOOL-AGE CHILDREN 185

References
ACHENBACH, T.M. and EDELBROCK, C.S. (1983) Manual for the Child Behaviour Checklist and
Revised Child Behavior Profile. Burlington, VT: University of Vermont, Department of
Psychiatry.
AINSWORTH, M.D.S. (1990) ‘Epilogue: Some considerations regarding theory and assessment
relevant to attachment beyond infancy’. In GREENBERG, M.T. CICCHETTI, D. and
CUMMINGS, E.M. (eds) Attachment in the Pre-School Years: Theory, Research and
Intervention. Chicago: University of Chicago Press.
AMMANITI, M., CANDELORI, C., DAZZI, N., DE CORO, A., MUSCETTA, S., ORTU, F., POLA,
M., SPERANZA, A.M., TAMBELLI, R. and ZAMPINO, F. (1990) Intervista sull’attaccamento
nella latenza. Unpublished.
AMMANITI, M., SPERANZA, A.M. and TAMBELLI, R. (2003). ‘Intervista sull’attaccamento nella
latenza’. Attachment and Human Development. In press.
ANSTADT, T., MERTEN, J., ULLRICH, B. and KRAUSE, R. (1997) ‘Affective dyadic behavior,
core conflictual relationship themes, and success of treatment’. Psychotherapy Research, 7:
397–419.
BAKERMANS-KRANENBURG, M.J. and VAN IJZENDOORN, M.H. (1993) ‘A psychometric study
of the Adult Attachment Interview: reliability and discriminant validity’. Developmental
Psychology, 29: 870–9.
BEEBE, B., LACHMANN, F. and JAFFE, J. (1997). ‘Mother – infant interaction structures and
presymbolic self and object representations’. Psychoanalytic Dialogues, 7: 113–82.
BOWERS, L., SMITH, P.K. and BINNEY, V. (1994) ‘Perceived family relationships of bullies,
victims and bully/victims in middle childhood’. Journal of Social and Personal Relationships,
11: 215–32.
BRETHERTON, I., RIDGEWAY, D. and CASSIDY, J. (1990) ‘Assessing internal working models of
the attachment relationship: An attachment story completion task’. In GREENBERG, M.T.
CICCHETTI, D. and CUMMINGS, E.M. (eds) Attachment in the Preschool Years: Theory,
Research and Intervention. Chicago: University of Chicago Press.
CASSIDY, J. (1988) ‘Child–mother attachment and the self in six-year-olds’. Child Development,
59: 121–34.
GEORGE, C., KAPLAN, N. and MAIN, M. (1985) The Adult Attachment Interview. Unpublished
manuscript, Department of Psychology, University of California at Berkeley.
LAMB, M.E., THOMPSON, R.A., GARDNER, W. and CHARNOV, E. (1985) Infant–Mother
Attachment: The Origins and Developmental Significance of Individual Differences in Strange
Situation Behaviour. Hillsdale, NJ: Erlbaum.
LUBORSKY, L. and CRITS-CHRISTOPH, P. (1990) Understanding transference: The CCRT
method. New York: Basic Books.
MAIN, M. (1995) ‘Recent studies in attachment: Overview, with selected implications for
clinical work’. In GOLDBERG, S., MUIR, R. and KERR, J. (eds) Attachment Theory: Social,
Developmental, and Clinical Perspectives. Hillsdale, NJ: Analytic Press, Inc.
MAIN, M. and GOLDWYN, R. (1994) Adult Attachment Rating and Classification System, Manual
in Draft, Version 6.0 . Unpublished manuscript: University of California at Berkeley.
MAIN, M., KAPLAN, N. and CASSIDY, J. (1985) ‘Security in infancy, childhood and adulthood:
A move to the level of representation’. In BRETHERTON, I. and WATERS, E. (eds) Growing
Points of Attachment Theory and Research. Monographs of the Society for Research in Child
Development (Vol. 50, pp. 66 – 104). Chicago: Chicago University Press.
186 M. TARGET ET AL.

OPPENHEIM, D. (1997) ‘The Attachment Doll-play Interview for preschoolers’. International

Journal of Behavioral Development, 20: 681–97.
SHOULDICE, A. and STEVENSON-HINDE, J. (1992) ‘Coping with security distress: The
separation anxiety test and attachment classification at 4.5 years’. Journal of Child Psychology
and Psychiatry, 33: 331–48.
SLOUGH, N.M. and GREENBERG, M.T. (1990) ‘5-year-olds’ representations of separations from
parents: Responses from the perspective of self and other’. New Directions for Child
Development, 48: 67–84.
SOLOMON, J., GEORGE, C. and DEJONG, A. (1995) ‘Children classified as controlling at age six:
Evidence of disorganized representational strategies and aggression at home and at school’.
Development and Psychopathology, 7: 447–63.
SROUFE, L.A. and FLEESON, J. (1986) ‘Attachment and the construction of relationships’. In
HARTUP, W. and RUBIN, Z. (eds) Relationships and Development. Hillsdale, NJ: Erlbaum.
SROUFE, L.A., FOX, N. and PANCAKE, V. (1983) ‘Attachment and dependency in
developmental perspective’. Child Development, 54: 1615–27.
WATERS, E. (1978) ‘The reliability and stability of individual differences in infant–mother
attachment’. Child Development, 49: 483–94.
WRIGHT, J.C., BINNEY, V. and SMITH, P.K. (1995). ‘Security of attachment in 8 – 22 year olds:
A revised version of the Separation Anxiety Test, its psychometric properties and clinical
interpretation’. Journal of Child Psychology and Psychiatry, 36: 757–74.