Abstract
The quality of early parent-child relationships has been linked to a variety of clinical and developmental outcomes [1],[2]. Different psychometric instruments have been used to measure similar aspects of these relationships, based on different theoretical positions. Two such instruments, the Adult Attachment Interview (AAI) [3] and the Parental Bonding Instrument (PBI) [4] both purport to measure the nature of early, affectionate ties between parents and children as recalled by adults. The first is a semistructured interview technique; the second a brief self-report questionnaire.
The AAI has been found to correspond closely to observational measures of attachment [5] and is, therefore, often considered the ‘gold standard’ in the field. Unfortunately, the instrument is complex, and the coding of interview transcripts requires substantial training, time and resources.
Self-report measures of attachment, such as the PBI can be scored more easily and quickly, but it is unclear whether they measure the same construct. Despite this lack of attention to construct validity, some studies have used the PBI to obtain attachment information, or have used the terms ‘bonding’ and ‘attachment’ interchangeably, resulting in confusion. Only one study has used both the PBI and the AAI, comparing the responses of patients with borderline personality disorder with those of dysthymic patients [6]. However, comparisons of the instruments within participants were not reported. The lack of attention paid to the issue of construct validity among developers of self-report questionnaires pertaining to attachment has been cited as an important gap in the literature [7].
In this study, the AAI and the PBI were both administered to non-psychotic adolescents referred for psychiatric treatment. Attachment-related information obtained from the two instruments was compared within participants. We hypothesised that the AAI information and the PBI information would be similar but not identical, as described below.
Adult attachment interview
Based on Bowlby's [8] attachment theory, this 1-h long, semistructured interview asks participants to describe their childhood attachment experiences and the impact of those experiences on their development and personality [9]. Responses are transcribed verbatim, and coding of the transcripts allows participants to be classified as having one of four particular ‘states of mind’ with respect to attachment, corresponding to distinct styles of discourse. These styles of discourse were discovered by studying parents of infants with known attachment classifications, and are thought to relate to adults' early attachment experiences [10].
Autonomous discourse is highly coherent and is thought to reflect either optimal attachment experiences or the participant's psychological resolution of suboptimal attachment experiences. Unresolved discourse shows specific lapses in the monitoring of reasoning when the individual is discussing traumatic events such the loss of an attachment figure or abuse by an attachment figure and is considered least optimal. Dismissing and preoccupied discourse are considered intermediate, and are thought to reflect rejecting and inconsistent attachment experiences, respectively. A small percentage of participants do not fit any of these categories, and are labelled ‘cannot classify’.
Coding the AAI requires a 2-week specialised training course, followed by extensive reliability testing. Most coders require about half a day per transcript to code the AAI. High concordance between parental AAI and infant attachment classification (done by coders blind to AAI results) [11] and concordance between parental AAIs and those of adult offspring [12] have established the AAI's validity. Test-retest reliability was found to be 78% (K = 0.63) [13] and 90% (K = 0.79) [12] in two studies, and discriminant reliability with respect to verbal intelligence, non-attachment related memory, and social desirability has also been established [14].
Because AAI classification is based on styles of discourse, the attachment history described by the participant is not necessarily indicative of classification. To determine participant classification, several nine-point scales are used to code transcripts (see Table 1) [3]. The five Experience Scales represent the participant's probable attachment experiences as judged by the coder, who uses both participant report and an estimate of participant believability to assign a rating. The State of Mind Scales represent aspects of the participant's discourse that may be affected by attachment-related feelings. Final classification on the AAI is derived from the State of Mind Scales, the most important of which is the Coherency scale.
The Scales of the Adult Attachment Interview (AAI)
Parental bonding instrument
The PBI is a 25-item self-report questionnaire probing adult recollections of parental behaviours and attitudes towards the subject in childhood. It is completed once for each parent. No specialised training is required to score it, and participant responses are taken at face value.
The PBI was developed using factor analyses from self-reports of experiences with parents in childhood [4], yielding two factors: care and overprotection. High care and low overprotection is considered optimal, while low care and high overprotection (‘affectionless control’ [15]) is considered least optimal, and the two scales are inversely correlated. The care factor superficially resembles the AAI ‘loving/unloving’ Experience Scale (representing caring, nurturing parental behaviour); the overprotection factor the AAI ‘involving/reversing’ Experience Scale (representing mutually protective behaviour between parent and child; see Table 1). No PBI factors correspond to the other three AAI Experience Scales, nor to the AAI State of Mind Scales.
The authors suggest that the PBI corresponds, to some degree, to objective reality [16] because of high correlations between participants' ratings of their parents and ratings by others [16],[17]. Studies showing higher rates of reported affectionless control in clinical than in non-clinical participants [18], and higher rates in more seriously disturbed than in less seriously disturbed participants [6] have demonstrated its discriminant validity. High test-retest reliability, and concurrent and predictive validity have been established [15],[19].
Relationship between the Adult Attachment Interview and the Parental Bonding Index
By using the information above, three hypotheses about the relationship between the PBI and the AAI were made. Given the similarity of the concepts of bonding and attachment, it was hypothesised that care and overprotection scores on the PBI would differ significantly in relation to attachment classification (hypothesis 1), with participants autonomous on AAI reporting more optimal early experiences with parents on PBI than participants with other AAI classifications.
The experience scales of the AAI that appear to correspond most closely to the PBI Care and Overprotection Scales are (respectively) ‘loving/unloving’ and ‘involvement/role reversal’ (see Table 1). Other AAI Experience Scales reflect different aspects of suboptimal parenting, suggesting that they might be negatively associated with care on the PBI. Therefore, it was hypothesised that the Care and Overprotection Scales of the PBI would relate to the Experience Scales of the AAI as follows: positive association between PBI care and AAI loving/unloving; positive association between PBI overprotection and AAI involvement/role reversal; possible negative associations between PBI care and AAI involvement/role reversal, AAI rejection, AAI neglect, and AAI pressuring to achieve (since all of these reflect suboptimal parenting; hypothesis 2).
The final hypothesis is based on scoring differences between the AAI and the PBI. Parental Bonding Instrument responses are taken at face value, but AAI Experience Scales are based partly on participant report and partly on the coder's judgement of how believable that report is. If the participant appears to be either minimising parental flaws (rated on the AAI Idealisation Scale) or exaggerating parental flaws (possibly related to the AAI Involving Anger Scale), the coder adjusts experience scores accordingly. Therefore, it is likely that the prediction of PBI care and PBI overprotection scores using the corresponding AAI scales (loving/unloving and involvement/role reversal, respectively) can be improved by taking into account participant idealisation and participant anger as judged by AAI coders (hypothesis 3).
Method
Participants
The participants were 145 adolescents (13–19 years old) consecutively recruited into a study of suicidality in adolescents [20] from one of five participating treatment centres in three Canadian cities. Inpatients and outpatients were included, but participants with active psychosis, organic brain disease or central nervous system disorder were excluded. Inclusion criteria were age between 13 and 19 years, and informed written consent from the adolescent (or from the parent or guardian if the participant was under the age of 18). Approximately 65% of adolescents presenting to the treatment centres during the 24-month study period met these criteria.
Fifteen of 145 participants were eliminated from the study for failure to appear for the AAI interview, incomplete interviews or a failure of the recording system. The final sample was composed of 57 females (44%) and 73 males (56%) with a mean age of 15.3 ± 1.47 years. Participants were 84.5% white, and 53% were in residential treatment. There were no significant differences between the male and female adolescents in age, (t128 = 0.12, p = 0.90) or residential treatment status, (X 2 1,128 = 1.05, p = 0.31).
Procedure
Two interview sessions were held with each participant, within 2 weeks of each other. During the first session, descriptive, demographic and psychosocial data were collected and the PBI for each parent was administered. In the second session, the AAI was conducted. Each AAI was audiotaped and transcribed verbatim. Three trained coders scored the transcripts independently. None of the coders conducted interviews. Each transcript was rated by at least two coders and final scores for each scale and final classifications were reached by consensus after discussion. The mean concordance rate for all AAI classifications for all pairs of coders was 78.6%. The mean kappa for primary classifications (excluding U) was 0.71.
Analyses and results
Table 2 shows the means and standard deviations on the maternal care and overprotection scales of the PBI for each of the four largest attachment groups (‘cannot classify’ cases were excluded, as this group contained only four cases). To test the first hypothesis, a multivariate analysis of variance (MANOVA) for AAI attachment classification was done with PBI maternal care and PBI maternal overprotection as the dependent variables. Significant differences by AAI attachment classification were found (F3,122 = 2.79, p = 0.012). A MANOVA for AAI attachment classification was also done with PBI paternal care and PBI paternal overprotection as the dependent variables, but results were not significant. Therefore, all subsequent analyses were done using maternal data only. Univariate F-tests on these data revealed significant differences for both PBI care (F3,122 = 3.37, p = 0.021) and PBI overprotection (F3,122 = 3.94, p = 0.010) by AAI attachment classification. Significant results from post hoc comparisons (two-tailed) included differences between the AAI Unresolved and AAI Autonomous classifications on both PBI care and PBI overprotection (t = −3.38, p = 0.001; t = 3.06, p = 0.003, respectively), and between the AAI Unresolved and AAI Dismissing classifications on PBI overprotection (t = 2.95, p = 0.004). Unresolved participants reported lower PBI maternal care and higher PBI maternal overprotection than AAI Autonomous participants, and they reported higher PBI maternal overprotection than AAI dismissing participants. These differences remained significant after applying the Bonferroni correction for multiple t-tests.
Means and standard deviations for maternal care and maternal overprotection by attachment category
To test the second hypothesis, Pearson product-moment correlations between all AAI experience scales and the care and overprotection scales of the PBI were determined (see Table 3). As predicted, the PBI care and AAI loving/unloving scales were correlated (R = 0.39, p <0.001), as were the PBI overprotection and AAI involving/role reversing scales (R = 0.40, p< 0.001). Negative correlations were found between the PBI care scale and the AAI involvement/role reversal, rejection and neglect scales (R=−0.25, p = 0.006; R=−0.41, p< 0.001; R=−0.46, p< 0.001, respectively), but not between the PBI care scale and the AAI pressuring to achieve scale.
Correlations between Parental Bonding Index scales and Adult Attachment Index experience scales (n = 126)
Beyond hypothesis 2, the matrix also reveals: a negative correlation between PBI care and PBI over-protection, consistent with previous findings on the PBI [4]; negative correlations between AAI loving/unloving and AAI rejection, and AAI loving/unloving and AAI neglect, and a positive correlation between AAI rejection and AAI neglect (indicating internal consistency among these AAI experience scales); and a positive correlation between the PBI overprotection scale and the AAI neglect scale (R and P-values as shown in the table). When this correlation matrix was run separately for each attachment classification, the highest scale correlations across instruments occurred in the autonomous category (for example, R = 0.48 for PBI care and AAI loving/unloving).
Stepwise regressions for maternal care and maternal overprotection (n = 126)
Before testing hypothesis 3, correlations between the care and overprotection scales on the PBI and the ‘idealisation’ and ‘anger’ state of mind scales on the AAI were determined. Adult Attachment Interview idealisation and AAI anger were both significantly correlated with PBI care (R = 0.34, p< 0.001; R = −0.44, p < 0.001, respectively) and significantly correlated with PBI overprotection (R = −0.34, p< 0.001; R=0.36, P< 0.001, respectively). The correlations were in the expected directions. Thus, highly idealising participants on the AAI rated mothers as more optimal on the PBI. Similarly, highly angry participants on the AAI rated mothers as less optimal on the PBI.
Having established these correlations, hypothesis 3 was tested. The proportion of the variance accounted for by a given correlate is R2. Thus, the proportion of the variance for PBI care accounted for by the AAI loving/unloving scale is (0.39)2 = 0.15, and the proportion of the variance for PBI overprotection accounted for by the AAI involving/role reversing scale is (0.40)2 = 0.16. To determine whether or not these predictions could be improved by accounting for participant idealisation and participant anger, two stepwise regressions were done with PBI care and PBI overprotection as the dependent variables, respectively (see Table 4). In the first regression, the variables AAI loving/unloving, AAI idealisation, and AAI anger were entered to predict PBI care. The statistical program used [21] automatically enters the best predictor variable first, then the second best predictor variable, and so on. The adjusted R2 was 0.35, indicating that these variables together account for 35% of the variance of the PBI care scale (large effect size) [22] and supporting hypothesis 3. In the second regression, the variables AAI involving/role reversing, AAI idealisation, and AAI anger were entered to predict PBI overprotection. The variable AAI anger was no longer a significant predictor after AAI involving/role reversing and AAI idealisation had been entered. The adjusted R2 was 0.20 (moderate effect size) [22], indicating that the AAI involving/role reversing and AAI idealisation variables together account for 20% of the variance of the PBI overprotection scale, also supporting hypothesis 3. To test whether the inclusion of other State of Mind Scales from the AAI in these regressions might further improve the prediction of PBI care and overprotection, each of the other State of Mind Scales was entered in turn with the previously established predictors. Only the AAI unresolved trauma scale further improved the prediction of PBI care (adjusted R2 = 0.39 for all variables together, see Table 4), and no AAI scales further improved the prediction of PBI overprotection.
Discussion
The results illustrate a significant association between maternal PBI scale scores and AAI attachment classification. The lack of association between paternal PBI scale scores and AAI attachment classification may reflect the predominance of mother-child attachment relationships in most cases in determining the adult state of mind with respect to attachment.
Autonomous participants on AAI reported high maternal care and low maternal overprotection on PBI, the pattern considered most optimal. Unresolved participants on AAI obtained the least optimal PBI scores. These results are consistent with findings that unresolved attachment is frequently associated with psychopathology [23],[24], while autonomous attachment is considered optimal [3], and that the PBI distinguishes abused participants (many of whom would fit ‘unresolved trauma’ on AAI) from non-abused participants (many of whom would fit other AAI categories) [25].
Adult Attachment Interview dismissing participants reported less PBI maternal overprotection than AAI unresolved participants. Low reported maternal overprotection in AAI dismissing participants is consistent with attachment theory, as these participants are thought to be insecure in response to caregiver rejection, rather than overinvolvement [3]. Unresolved participants on AAI may rate their mothers as highly overprotective as part of their overall perception of having received suboptimal parenting. Some studies have also found a relationship between the unresolved and preoccupied classifications [26–28]. When unresolved participants are assigned their ‘best alternative’ classifications, many are found to be preoccupied, explaining their high reported caregiver overprotection on PBI.
The PBI did not distinguish AAI preoccupied participants from any of the other groups, but this may have related to the small number of these participants. The PBI also failed to distinguish AAI autonomous from AAI dismissing participants, likely due to the tendency for AAI dismissing participants to minimise difficulties with parents [3], resulting in optimal PBI ratings that resemble those of AAI autonomous participants. A further reason for discrepancies between instruments is that AAI classifications use self-report data about mothers and fathers, while the PBI scores are specific to one parent. Thus, it is possible that some AAI classifications were influenced by paternal data, which would not be reflected in the maternal PBI scores.
Within the AAI scoring system, two factors appeared relevant to the PBI/AAI association. First, certain experience scales of the AAI closely match the PBI subscales in content. The PBI care scale is correlated with the AAI loving/unloving scale, and the PBI overprotection scale is correlated with the AAI involvement/role reversal scale. AAI experience scales representing various types of parenting difficulty showed negative correlations with the PBI care scale and positive correlations with the PBI overprotection scale, consistent with high care and low over-protection being most optimal on the PBI.
The second factor is the tendency for participants to idealise their parents or show anger towards them, as reflected in the AAI idealisation and involving anger subscales. Including these subscales in regression analyses improved the prediction of PBI scale scores from the corresponding AAI experience scales. Including the AAI unresolved trauma scale also improved the prediction of PBI maternal care, possibly due to the fact that this AAI scale reflects frightening or abusive parental behaviour, suggesting extremely suboptimal parenting.
Autonomous participants on AAI showed the best correspondence between PBI care and AAI loving/unloving scale scores. This finding is consistent with the contention of AAI coders that autonomous participants are most believable in reporting their attachment experiences [3], resulting in similar PBI self-ratings and AAI coder ratings.
In summary, PBI scale scores differed in relation to attachment classification, and were correlated with the corresponding AAI experience scales. The PBI also distinguished the most optimal from the least optimal attachment types, but failed to make finer distinctions among the four attachment categories. Minimisation of attachment-related difficulties (due to idealisation of parents) or exaggeration of them (due to anger towards parents) accounted for some of the discrepancy between the two instruments. Including these factors in regression analyses improved the AAI ratings' prediction of PBI scale scores. The AAI unresolved trauma scale was also predictive of PBI maternal care, suggesting that this factor must also be considered in comparing the two instruments. Finally, participants reporting the least attachment-related difficulty showed the greatest agreement across instruments.
The generalisability of these findings may be limited by sample characteristics. Other age groups and non-clinical populations warrant further study. The sample characteristics on each of the two instruments, however, are consistent with the literature. A care/overprotection correlation of −0.54 was found on the PBI scales for mothers (−0.47 in developers' original sample [4]). The mean values of the PBI scale scores were consistent with other clinical adolescent samples [29],[30] as was the distribution of major attachment categories on the AAI [31],[32].
The above findings imply that the PBI provides a reasonable approximation of AAI attachment ratings only for samples containing many participants reporting optimal attachment-related experiences, few participants likely to regard their childhood attachment experiences with ongoing idealisation or anger, and few participants that have experienced childhood abuse. Therefore, results from the two instruments would be expected to match more closely in non-clinical samples. Scales pertaining to participant idealisation, participant anger, and unresolved trauma contribute to the complexity of the AAI, but also appear to strengthen its validity as a measure of attachment in clinical populations.
Relatively large sample sizes (130 participants, in this case) appear to be required to obtain good correspondence across instruments. Therefore, use of the PBI to assess attachment in small, clinical samples would not be advisable, unless estimates of participant idealisation or anger towards parents were also available.
Nevertheless, the PBI may provide a useful screen for attachment-related difficulties in large, non-clinical populations. It could, therefore, be used to select high risk subgroups for subsequent detailed examination using the more complex AAI. Targeting those subgroups for early parenting intervention would be a useful clinical application. Furthermore, our results do not negate the utility of the PBI as a measure of perceived parenting style.